2015 International Conference on

Computing, Communication and Security

(ICCCS)
4 - 6 December, 2015

Le Meridien

Pointe aux Piments, Pamplemousses, Mauritius

ISBN 978-1-4673-9353-9

IEEE Catalog Number CFP15CSG-DVD

MAURITIUS SUBSECTION

www.icccs.in
Preface

2015 International Conference on Computing, Communication and Security (ICCCS)

Copyright and Reprint Permission: Abstracting is permitted with credit to the source. Libraries
are permitted to photocopy beyond the limit of U.S. copyright law for private use of patrons
those articles in this volume that carry a code at the bottom of the first page, provided the per-
copy fee indicated in the code is paid through Copyright Clearance Center, 222 Rosewood Drive,
Danvers, MA 01923. For reprint or republication permission, email to IEEE Copyrights Manager
at pubs-permissions@ieee.org. All rights reserved. Copyright ©2015 by IEEE.

IEEE Catalog Number CFP15CSG-DVD

ISBN 978-1-4673-9353-9

All technical inquiries regarding your DVD must be directed to:

The Society of Information Technologists and Entrepreneurs
Port Louis, Mauritius, Republic of Mauritius
Tel. (+230) 213 0032
Email: sitemauritius@gmail.com

The Society accepts no responsibility as to the content of the articles published inside this DVD
which is solely the responsibility of the authors.

Only registered and presented papers will be sent for inclusion into the IEEE Xplore Digital
Library.

IEEE Notice:
Please be advised that conference proceedings must meet IEEE's quality standards and IEEE reserves the
right not to publish any proceedings that do not meet these standards.
 i

Message from the General Chairs

2015 International Conference on Computing, Communication and Security (ICCCS) is aimed to be an annual
conference conceptualized to foster students, researchers, academicians and industry persons in the field of
Computer Science and Engineering, Communication and Security. 2015 International Conference on
Computing, Communication and Security (ICCCS) is organized by the Society of Information Technologists and
Entrepreneurs and the IEEE Mauritius Subsection on 4-6 Dec, 2015 at Le Meridien, Mauritius.

This conference will include invited keynotes and oral presentations. 2015 International Conference on
Computing, Communication and Security (ICCCS) will provide a forum for researchers and engineers in both
academia and industry to exchange the latest innovations and research advancements in Innovative
Computing, Communication and Security. 2015 International Conference on Computing, Communication and
Security (ICCCS) also provides the attendees the chance to identify the emerging research topics, as well as the
future development directions in all field of Computer Science.

The primary goal of the conference is to promote research and developmental activities in advanced
Computing, Communication and Security challenges. Another goal is to promote scientific information
interchange between researchers, developers, engineers, students, and practitioners working in and around
the world.

Authors have contributed original research papers which have been accepted after a double blind review
process. All registered papers presented in the conference will be submitted for inclusion into the IEEE Xplore
Digital Library provided they meet quality standards.

We cordially invite you to join us at the conference and we look forward to welcoming you in Mauritius.

With best wishes,

Prof. Subramaniam Ganesan, Prof. Sri Niwas Singh, Prof. Vijay Kumar,
Oakland University, Rochester, Indian Institute of Technology Kanpur, University of Missouri-Kansas City,
USA India USA
 ii

Keynote Speech

Prof. Subramaniam Ganesan

Oakland University, Rochester, USA

Automotive Embedded System Security and

Challenges

KEYNOTE ABSTRACT
In this presentation, automotive embedded system security, software and hardware solutions, Vehicle to
Vehicle communication (V2V), security threat due to V 2 V, CAN bus simulations for Security, MiL (model in the
loop) and SiL (software in the loop) for modeling security of the Vehicle environment are covered. Solutions for
security threats due to advanced wireless sensors, handheld mobile devices, over the air software updating,
data routers with cloud based computing, and connected vehicles are presented.

Keynote Speech

Prof. Sri Niwas Singh

Indian Institute of Technology Kanpur, India

Intelligent System Applications in Solving Engineering

Problems
KEYNOTE ABSTRACT
Due to increased complexity of the systems, cost reduction and detail modeling of the systems, the
requirements of optimization have been increased. The conventional methods, which guarantee to provide the
optimal solution, fail to solve many practical problems due to several requirements of these methods. The most
familiar conventional optimization techniques fall in two categories viz. calculus based method and
enumerative schemes. Though well developed, these techniques possess significant drawbacks. Calculus based
optimization generally relies on continuity assumptions and existence of derivatives. Enumerative techniques
rely on special convergence properties and auxiliary function evaluation. Moreover, these optimizations are
generally single path search and stuck with the local optima.

Intelligent based optimization methods such as genetic algorithm (GA), particle warm optimization (PSO),
bacteria foraging, ant colony, neural networks, etc. are multi-path search and provide solution near to the
global optima. They do not require derivatives of objective function and constraints. This presentation briefly
covers some of the important techniques of optimization along with scope and future challenges.
 iii

Keynote Speech

Prof. Sanjay Misra

Covenant University, OTA, Nigeria

Measurement Theory Principles and Software Complexity

Metrics
KEYNOTE ABSTRACT
The scientific base of software complexity metrics should be evaluated through the principals of measurement
theory. However, it is common observations that the majority of developers of new metrics do not consider for
applying the principals of measurement theory on their new metrics. This talk evaluates the different
measurement theories criteria developed by different scientist (which are applicable to software metrics) and
have shown that how to evaluate a new metric against the principles of measurement theory. This work will
guide and help to the software professionals and developer of new metrics: how to use the measurement
theory in their software metric program.

Keynote Speech

Prof. Andrew Ware

University of South Wales, United Kingdom

Friend or Foe? What is the potential for AI systems?

KEYNOTE ABSTRACT
Recent advance in intelligent system paradigms have been rapid and significant. Moreover, the pace of
development shows no sign of abating. We now live in a world where robots help in such activities as running
hotels, take a lead role in the manufacturing of an almost unmanageable range of products, and even provide
care for the sick and ageing. Autonomous vehicles assisted by intelligent control algorithms are now being
tested on roads used by the general public and unmanned drones are capable of flying more safely than
manned planes. The benefits that the advancements in AI afford humanity seem almost endless.

Set against the many advances and achievements of such intelligent systems and machines there have been
stark warnings from such notable scientists as Professor Stephen Hawking who highlight the very real danger
of some of the possible intended and unintended misuses of such technology. Moreover, there are still myriad
unanswered questions about the legal and ethical ramifications of the use of such systems in everyday life.

The talk will highlight some of the advances made in the area of intelligent systems and investigate the pros
and cons of such systems. In short, the talk will seek to inform discussion on whether such systems are friend
or foe.
 iv

Keynote Speech

Prof. Zoran Bojkovic

University of Belgrade, Serbia

Recent Advances in Mobile Cloud Computing for Fifth

Prof. Zoran Bojkovic
Generation (5G) Networks

KEYNOTE ABSTRACT
Mobile computing and cloud computing are two technologies converging into a rapidly growing field of mobile
cloud computing (MCC). By the definition of National Institute of Standards and Technology (NIST), cloud
computing is a model for enabling on demand network access to a shared pool of configurable computing
resources (e.g. networks, servers, storage, applications, and services) that can be rapidly provisioned and
realized with minimal management effort or service provider interaction. Future applications enabled by 5G
will have impacts on almost every aspect of digital lives. On the other hand, the 5G system deployed initially
2020, is expected to provide approximately 1000 times higher wireless area capacity and save up to 90
percent of energy consumption per service when compared to the current fourth generation (4G) system. Also,
10 times higher battery life of connected devices and 5 times reduced end-to-end latency are anticipated.

The objective of this presentation is to address recent advances in mobile cloud computing for 5G networks. In
this work, the topics of interest for both academia and industry are classified in the following categories
related to many aspects of MCC in 5G:

 5G Heterogeneous cloud radio access networks (HCRANs).

 Emotion-aware mobile cloud computing (EMC) in 5G designed to provide services by recognizing user’s
emotional changes.
 Mobile cloud sensing
 Energy efficient cloud computing.
 Mobile ad-hoc networks (MANET) in 5G
The presentation starts with emphasizing that the motivation of heterogeneous radio access networks is to
embed the cloud computing technology into heterogeneous networks (HetNets) to realize the large- scale
cooperative signal processing and networking functions. Thus, spectral efficiency and energy efficiency
performance is improved beyond existing HetNets and cloud radio access networks (CRANs). Cloud computing
is used to fulfill the centralized large-scale cooperative processing for suppressing co-channel interferences.
The goal of EMC is to provide personalized, human-centric, intelligent emotion-aware services in 5G through
maximizing user’s quality of experience (QoE), while optimizing resource allocation among the mobile
terminals and the remote cloud, under dynamic network environments. The presentation continuous with the
problem how to process the sensor data before sending it to the cloud. The architecture of mobile cloud
sensing is analyzed, too. In order to save energy, many techniques are adopted by the cloud servers, such as
 v

dynamic voltage and frequency scaling, as well as power gating. Cloud centers need to support on-demand
dynamic resource provisioning, where clients can at any time submit virtual machine requests with various
amounts of resources. Cloud-assisted mobile ad-hoc networks are expected to be popular in 5G mobile
networks because the significantly faster performance of 5G communications enables clouds to provide
realistic services. On the other hand, the energy consumption problem will become serious due to the highly
increased cloud computing speed. In closing, many challenges remain to be addressed ranging from MCC
architecture/5G network design, resource/mobility management, security enhancement and protection, to
networking protocol development and new services. This will stimulate further research of MCC in 5G
networks.
 vi

General Chairs
Prof. Subramaniam Ganesan, Oakland University, Rochester, USA

Prof. Sri Niwas Singh, Indian Institute of Technology, Kanpur, India

Prof. Vijay Kumar, University of Missouri-Kansas City, USA

Organising Chairs
Vishal Kumar, BTKIT, India

Sameerchand Pudaruth, Mauritius

International Advisory Committee

Prof. Zoran Bojkovic, University of Belgrade, Serbia

Prof. (Dr.) Fushuan Wen, Institut Teknologi Brunei, Brunei

Prof. Rajat Moona, Director General, C-DAC, India

Prof. (Dr.) Lalit Awasthi, ABV-GIET, India

Prof. (Dr.) M. P. Gupta, IIT Delhi

Prof. (Dr.) Subarna Shakya, Tribhuvan University, Nepal

Dr. Sanjay Madria, Missouri University of Science and Technology, USA

Dr. Jai Govind Singh, Asian Institute of Technology, Bangkok

Dr. Anurag K. Srivastava, Washington State University, USA

Dr. Al-Sakib Khan Pathan, International Islamic University Malaysia, Kuala Lampur

Dr. Sanjay Kumar, Birla Institute of Technology Mesra, India

Dr. Sunil Kumar, San Diego State University

Anupam Srivastava, Associate Dean, Middle East College, Knowledge Oasis Muscat, Oman

Sanjeev K. Singh, Siemens, USA

 vii

Technical Program Committee

Vinaye Armoogum, University of Technology, Mauritius

Parma Nand, School of Computing Science & Engineering, Galgotias University, India

Dilip Sharma, GLA University, India

Balasubramani R, NMAM Institute of Technology, India

Carlos Caldeira, University of Evora, Portugal

Vedpal Singh, Universiti Teknologi Petronas, Malaysia

Hima Gupta, Jaypee Business School, Noida, India

Anamika Ahirwar, Maharana Pratap College of Technology, Gwalior, India

Hlaing Htake Khaung Tin, Computer University, Loikaw, Kayah State, Myanmar

Yiannis Koumpouros, Technological Educational Institute of Athens, Department of Informatics, Greece

G. Shobha, R V College of Engineering, Bangalore

Cheng-Hsiang Liu, National Pingtung University of Science and Technology, Taiwan

Gaurav Bajpai, College of Science and Technology, University of Rwanda

Pramod Kumar Tula's Institue, India

Shekhar Verma, Associate Professor, Indian Institute of Information Technology, Allahabad, India

Ruchir Gupta, Department of Computer Science, IIIT, Jablpur, India

K V Arya, Associate Professor, ABV-Indian Institute of Information Technology, Gwalior

Bharti Suri, GGS Indraprastha University, Delhi, India

Ranjan Dasgupta, NITTTR, Kolkata, India

Vijay Bhaskar Semwal, IIIT Allahabad, India

Publicity Committee
Shilpi Saxena, Graphic Era Hill University, India

Sudhakar Chauhan, NIT Kurushetra, India

Manoj Satyarthi, GGSIU, New Delhi, India

 viii

Sampson D. Asare, University of Botswana, Botswana

Naganna Chetty, MITE, Mangalore, India

Paul Macharia, Data Manager, Kenya

Vijay Saini, WIT, India

Upasana Gitanjali Singh, University of KwaZulu-Natal, South Africa

Deepika Saxena, DMM, Kurukshetra, India

Local Organizing Committee

Sameerchand Pudaruth, University of Mauritius

Dr. Nitish Chooramun, University of Mauritius

Financial Committee
Hoshiladevi Ramnial, Mauritius

Nadeem Nazurally, Mauritius

Publications Committee
The Society of Information Technologists and Entrepreneurs
ICCCS-2015 Table of Contents

Table of Contents
Hardware Implementation of Ultralightweight Cryptographic Protocols . . . . . . . . . . . . . . . . . . . 1
Umar Mujahid, Najam Ul Islam Muhammad and Qurat Ul-Ain
Physical Layer Secrecy Solution for Passive Wireless Sensor Networks . . . . . . . . . . . . . . . . . . . . 2
Avinash Thombre and Aditya Trivedi
Visualising and Analysing Online Social Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Manoj Maharaj and Kambale Vanty Muhongya
An Efficient Key Management Scheme in Hierarchical Wireless Sensor Networks . . . . . . . . . . 4
Xinyang Zhang and Jidong Wang
Experiential Analysis of the Effectiveness of Buck and Boost DC-DC Converters in a
Basic Off Grid PV system . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Osamede Asowata, Ruaan Schoeman and Pienaar Christo
Sixsoid: A New Paradigm for k-Coverage in 3D Wireless Sensor Networks . . . . . . . . . . . . . . . . 6
Manjish Pal and Nabajyoti Medhi
Identifying Ideal Values of Parameters for Software Performance Testing . . . . . . . . . . . . . . . . . . 7
Charmy Patel and Ravi Gulati
Reducing Structured Big Data Benchmark Cycle Time using Query Performance
Prediction Model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Rekha Singhal
Robust Blind Watermarking Technique for Color Images using Online Sequential
Extreme Learning Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Ankit Rajpal, Anurag Mishra and Rajni Bala
A Detection Mechanism of DoS Attack using Adaptive NSA Algorithm in Cloud
Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Sumana Maiti, Chandan Garai and Ranjan Dasgupta
Vehicle Security and Forensics in Mauritius and Abroad . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
Pravin Selukoto Paupiah
Performance Evaluation and Analysis of Layer 3 Tunneling between OpenSSH and
OpenVPN in a Wide Area Network Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
Irfaan Coonjah, Clarel Catherine and K. M. S. Soyjaudah
Overview of Data Quality Challenges in the Context of Big Data . . . . . . . . . . . . . . . . . . . . . . . . . 13
Suraj Juddoo
A Classification Method to Classify High Dimensional Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Amit Gupta
Experimental Performance Comparison between TCP vs UDP tunnel using OpenVPN. . . . 15
Irfaan Coonjah, Pierre Catherine and K.M.Sunjiv Soyjaudah
6to4 Tunneling Framework using OpenSSH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Irfaan Coonjah, Pierre Catherine and K.M.Sunjiv Soyjaudah
ICCCS-2015 Table of Contents

Comparative Study of Wireless Sensor Network Standards for application in Electrical

Substations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Fabrice Labeau, Akash Agarwal and Basile Agba
Classification Performance Analysis of MNIST Dataset using a Multi-resolution Technique 18
Ramesh Kumar Mohapatra, Banshidhar Majhi and Sanjay Kumar Jena
Context-Aware Mechanisms for Improving the Cognitive Processes of Users in Mobile
Learning Platforms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Brita Curum and Kavi Kumar Khedo
Outputs Analysis and Dual Immersion Method for Chaotic Systems . . . . . . . . . . . . . . . . . . . . . . 20
Kokou Langueh, Octaviana Datcu, Jean-Pierre Barbot, Gang Zheng and Krishna
Busawon
Comparison of Different Antenna Arrays for the Channel Capacity . . . . . . . . . . . . . . . . . . . . . . 21
Shu-Han Liao, Chien-Hung Chen, Chien-Ching Chiu and Min-Hui Ho
Least Squares SVM Approach for Abnormal Brain Detection in MRI using
Multiresolution Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Deepak Ranjan Nayak, Ratnakar Dash and Banshidhar Majhi
Introducing FERPS: A Framework for Enterprise Resource Planning Selection . . . . . . . . . . . . 23
Upasana Singh and Omesh Singh
Performance Evaluation of Matrix- and EXIF-Based Video Rotation Methods in
MJPEG-Based Live Video Encoder. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Dejan Skvorc, Mirela Ostrek and Matija Horvat
Integrating Learning Analytics for Higher-Order Hhinking e-Assessments . . . . . . . . . . . . . . . . . 25
Sandhya Gunness and Upasana Singh
A Brief Perspective of Context-Aware Recommender System and Services . . . . . . . . . . . . . . . . 26
Ebunoluwa Ashley-Dejo
A Standard Compliant Blood Bank Management System with Enforcing Mechanism. . . . . . 27
Amarjeet Cheema, Siddharth Srivastava, Praveen Srivastava and Dr. B K Murthy
A Survey of Recommender System Feedback Techniques, Comparison and Evaluation
Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Omobayo Esan, Ebunoluwa Ashley-Dejo, Masupha Lerato, Ngwira Sm and Tranos Zuva
Review of Channel Allocation Methods for TVWS Devices Using PAWS Protocol. . . . . . . . . 29
Elesa Ntuli, Seleman Ngwira and Tranos Zuva
Personal Preferences Analysis of User Interaction based on Social Networks . . . . . . . . . . . . . . . 30
Cheng-Hung Tsai, Han-Wen Liu, Tsun Ku and Wu-Fan Chien
Multimodal Biometric Scheme for E-assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Samuel Ojo, Tranos Zuva and Selelman Ngwira
Survey of Biometric Authentication for E-assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Samuel Ojo, Tranos Zuva and Selelman Ngwira
A Framework for the Support of Mobile Knowledge Workers in South Africa. . . . . . . . . . . . . . 33
Michelle Robberts, Klarissa Engelbrecht and John Andrew Van Der Poll
ICCCS-2015 Table of Contents

Characteristics of Transfer Function of Power Lines having Tapped Branches in the

Propagation Path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Banty Tiru
IDPS based Framework for Security in Green Cloud Computing and Comprehensive
Review on Existing Frameworks and Security Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Komal Gill and Anju Sharma
Face Recognition Techniques, their Advantages, Disadvantages and Performance
Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
Lerato Masupha, Tranos Zuva, Seleman Ngwira and Omobayo Esan
Designing Budget Forecasting and Revision System Using Optimization Methods . . . . . . . . . 37
Mohammad Jafar Tarokh and Kaban Koochakpour
Ameliorate Pursuance of Transport Control Protocol in Cognitive Radio Mobile Ad Hoc
Network Utilize Accessible Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Senthamarai Selvan Annamalai and Selvaradjou Ka
An Improved Energy Efficient Chain Base Routing in Wireless Sensor Networks . . . . . . . . . . 39
Matimu Ngobeni and Zanele Mahlobogwane
An Integrated Model for ICT Governance and Management Applied to the Council for
Evaluation, Accreditation and Quality Assurance of Higher Education Institutions in
Ecuador (CEAACES) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Carlos W. Montenegro and Denys A. Flores
Experiencing Game-Based Learning through Mobile Devices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
Roopesh Sungkur and Upasana Singh
Privacy Challenges in Proximity Based Social Networking: Techniques & Solutions . . . . . . 42
Asslinah Mocktoolah and Kavi Kumar Khedo
Cloud-based Technologies for Addressing Long Vehicle Turnaround Times at Recycling
Mills . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Joubert van Tonder and John van der Poll
How to Detect Unknown Malicious Code Efficiently? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Jaehee Lee, Hyoungjun Kim, Hyunsik Yoon and Kyungho Lee
Green ICT Maturity Models - Towards a Generic Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Ravi Foogooa, Chandradeo Bokhoree and Kumar Dookhitram
QoS-Aware Single Service Selection Mechanism for Ad-Hoc Mobile Cloud Computing . . . . . 46
Ayotuyi Akinola and Alaba Akingbesote
First Principle Leakage Current Reduction Technique for CMOS Devices . . . . . . . . . . . . . . . . . 47
Hippolyte Djonon Tsague and Twala Bhekisipho
Customer Engagement Manifestations on Facebook Pages of Tesco and Walmart . . . . . . . . . 48
Swaleha Peeroo, Martin Samy and Brian Jones
The Impact and Opportunities of e-Tutoring in a Challenged Socio-Economic
Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Petra Le Roux and Marianne Loock
ICCCS-2015 Table of Contents

Weighting Technique on Multi-timeline for Machine Learning-based Anomaly Detection

System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Kriangkrai Limthong
Algorithm Performance Indexing Through Parallelism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
Peter Manyere and Andre Nel
Designing of Licensed Number Plate Recognition System Using Hybrid Technique from
Neural Network & Template Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Mohit Bansal and Binay Binod Kumar
A Classification Method to Detect if a Tweet Will be Popular in a Very Early Stage . . . . . . 53
Xianghui Zhao, Xiaoyi Wang, Zheng Zhan, Yong Peng and Yuangang Yao
Load Frequency Control of a Nonlinear Two-Area Power System . . . . . . . . . . . . . . . . . . . . . . . . . 54
Rajeshree Ramjug-Ballgobin, Sayed Zahiruddeen Sayed Hassen and Sonia Veerapen
Observer-based Control for Biomass Regulation in Wastewater Treatment Plants . . . . . . . . . 55
Rajeshree Ramjug-Ballgobin, Krishna Busawon, H C S Rughooputh and Richard Binns
Rough Set Theory Based Method for Building Knowledge for the Rate of Heat Transfer
on Free Convection over a Vertical Flat Plate Embedded in a Porous Medium . . . . . . . . . . . . 56
Hossam Mohamed and Mohamed M. Abdou
Full-Half Duplex Mixed Strategy for Capacity Maximization in MIMO Interference
Channels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Emmanuel Migabo, Thomas Olwal and Thato Phate
Design and development of Interval arithmetic library using Q format data
representation : A FPGA approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Vidya Mansur and Rajashekar Shettar
Computation Beyond Moore’s Law: Adaptive Field-Effect Devices for Reconfigurable
Logic and Hardware-Based Neural Networks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Udo Schwalke
Initial Proposal of a Framework in the Context of Cyberdefense to Articulate Controls
and Metrics Associated. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Pablo Sack and Jorge Ierache
Teaching Global Warming Through Blended Learning: A Case Study in a Private
Secondary School in Mauritius . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Teena Gomes and Shireen Panchoo
Performance Analysis of Parallel CBAR in MapReduce Environment . . . . . . . . . . . . . . . . . . . . . 62
Sayantan Singha Roy, Chandan Garai and Ranjan Dasgupta
Using a Location-Based Mobile Application to Locate a Traditional Health Practitioner
in South Africa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Johannes Kekana and Robert Hans
A New Efficient Algorithm for Executing Queries over Encrypted Data . . . . . . . . . . . . . . . . . . . 64
Rasha Refaie, Abd El-Aziz Ahmed, Nermin Hamza, Mahmood Ahmed and Hesham Hefny
Memristor Model For Massively-Parallel Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Dalibor Biolek, Viera Biolkova and Zdenek Kolka
ICCCS-2015 Table of Contents

Stability of Digitally Emulated Mem-Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

Zdenek Kolka, Viera Biolkova and Dalibor Biolek
Depiction of Experimental findings for a Cloud Enterprise Architecture . . . . . . . . . . . . . . . . . . . 67
Promise Mvelase
Agent Development Platforms for Bioinformatics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Ajit Kumar Gopee, Geerish Suddul, Yasmina Jaufeerally-Fakim and Admiral Chikudo
Confused which Educational Video to Choose? Appropriateness of Educational
Programmes on YouTube Education for use by educators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
Perianen Appavoo, Meera Gungea, Teenah Jutton and Preetamsing Dookhun
Analysis of IEEE 802.11n Network Access Categories in EDCA Non-Saturated Networks . 70
Prosper Sotenga, Peter Chuku and Thomas Olwal
Security in the Internet of Things through Obfuscation and Diversification . . . . . . . . . . . . . . . 71
Shohreh Hosseinzadeh, Sampsa Rauti, Sami Hyrynsalmi and Ville Leppänen
Proactive Digital Forensics in the Cloud using Virtual Machines . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Dirk Ras and H.S. Venter
Fusion of Local and Global Features for Face Recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Bibi Nazmeen Boodoo-Jahangeer and Sunilduth Baichoo
Rapid Prototyping with a Local Geolocation API . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Geerish Suddul, Kevin Nundran, Jonathan L. K. Cheung and Morgan Richomme
Role of Attributes Selection in Classification of Chronic Kidney Disease Patients. . . . . . . . . . 75
Naganna Chetty, Kunwer Singh Vaisla and Sithu D Sudarshan
Do You Know Big Brother is Watching You on Facebook? A Study of the Level of
Awareness of Privacy and Security Issues among a Selected Sample of Facebook Users in
Mauritius . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Mrinal Sohoraye, Vandanah Gooria, Suniti Nundoo-Ghoorah and Premanand Koonjal
UOM Multimodal Face and Ear Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Bibi Nazmeen Boodoo-Jahangeer and Sunilduth Baichoo
An Optimized CRM Adjusted to Big Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Daif Abderrahmane and Fatima El Jamiy
Requirements Determination of Object-Oriented Programming Software Tools for
Novice Learners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Yeeshtdevisingh Hosanee and Dr. Shireen Panchoo
Design and Implementation of an Automatic Staff Availability Tracking System . . . . . . . . . . 80
Kyle Stone and Jan Spies
Proposed Model for Evaluation of mHealth Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Chel-Mari Spies
An Improved Image Steganography Scheme with high Visual Image Quality . . . . . . . . . . . . . . 82
Sumit Laha and Rinita Roy
A Quantum-Inspired Cuckoo Search Algorithm for the Travelling Salesman Problem . . . . . . 83
Sumit Laha
ICCCS-2015 Table of Contents

EWSA: An Enriched Workflow Scheduling Algorithm in Cloud Computing . . . . . . . . . . . . . . . 84

Shilpi Saxena and Deepika Saxena
An Analysis of the Suitability of Cloud Computing Services in the Nigerian Education
Landscape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Sanjay Misra and Adewole Adewumi
Developing a Mobile Application for Taxi Service Company in Nigeria . . . . . . . . . . . . . . . . . . . . 86
Sanjay Misra and Adewole Adewumi
Enhancing Data Security in Open Data Kit as an mHealth Application . . . . . . . . . . . . . . . . . . . 87
Paul Macharia, Matthew Dunbar, Betsy Sambai, Felix Abuna, Bourke Betz, Anne
Njorege, David Bukusi, Peter Cherutich and Carey Farquhar
Performance Evaluation of Classification Techniques for Devanagari Script . . . . . . . . . . . . . . . . 88
Sanjay Misra, Chitra Dhawale and M.S. Sonawane
Enhanced Rule Extraction by Augmenting Internal Representation of Feed Forward
Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Vinita Srivastav, Chitra Dhawale and Sanjay Misra
An Investigation on Residential Exposure to Electromagnetic Field from Cellular Mobile
Base Station antennas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Amar Renke and Mahesh Chavan
Parallelization of Association Rule Mining : A Survey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Shivani Sharma and Durga Toshniwal
Effectiveness of LSP Features for Text Independent Speaker Identification . . . . . . . . . . . . . . . . 92
Sharada Chougule and Mahesh Chavan
Analysing Road Accident Data Using Association Rule Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Sachin Kumar and Durga Toshniwal
Distribution Methods of Web Pages and Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Neelaksh Sheel and Kunwar Singh Vaisla
 Hardware Implementation of Ultralightweight
Cryptographic Protocols
Qurat Ul Ain
Department of Electrical Department of Electrical
Engineering, Bahria University Engineering, Bahria University
Islamabad, Pakistan Islamabad, Pakistan
aine.khalid@gmail.com umarkhokhar1@hotmail.com
Abstract—RFID (Radio Frequency Identification) is among the
most emerging and tremendously growing technologies in the
field of automatic identification. The technology is far better than
its other contending systems such as Barcodes and magnetic
tapes as it provides optimal communication link with non line of
sight capability. Several researchers have proposed various
ultralightweight RFID authentication protocols to provide cost
effective security solutions. However, the proper hardware
implementation of these ultralightweight authentication protocols
has been neglected long which develops a chaos of their practical
feasibility. In this paper, we have addressed this problem and
proposed generic hardware architecture for EPC-C1G2 tags. We
have also simulated four ultralightweight authentication
protocols (SASI, LMAP, David-Prasad and RAPP) using
ModelSim PE and synthesized using Xilinx Design Suite 14.5.
Algorithm has been successfully implemented in hardware using
Spartan 6 (Low Power) and Vertex 7 Field Programmable Gate
Array (FPGA). We believe that the proposed architecture and
simulated results will act as the stepping stone for the future
implementation of these low cost RFID tags for ASIC.
Keywords—RFID; Ultralightweight; Mutual Authentication;
Tags; FPGA; SASI
I. INTRODUCTION
RFID systems are the rapidly developing technology across
the world because of its enriched features and functional haste.
RFID system is basically composed of the three main modules
which include: Reader, Tag and the Backend Server
(commonly known as database).
Tag Reader Database
Fig 1: Architecture of RFID based Systems
Tag uses radio signals (wireless) for communication with
the reader and then reader further communicates with the
backend server. It is the assumed that the communication
channel between the reader and database is secure as here we
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Umar Mujahid M.Najam-ul-islam
Department of Electrical
Engineering, Bahria University
Islamabad, Pakistan
najam@bui.edu.pk
can use traditional cryptographic algorithms (such as AES,
3DES, Public or Private key cryptography). However, the
channel of communication between tag and the reader is
considered to be insecure because of limited computational
power at reader side. As this link is wireless, which an
adversary can easily access to launch passive or active attacks.
In order to enhance the security of this wireless channel, a large
number of ultralightweight mutual authentication protocols
(UMAP) have been proposed. However, almost all of these
algorithms apparently failed to provide optimal security. The
generic solution for the implementation of security protocols is
to use encryption of data that is to be communicated. Main
focus of this paper will be on the efficient hardware
implementation of the four protocols; SASI [1], LMAP [2],
David Prasad [6] and RAPP [14] of UMAP family. UMAP
family mainly involves simple bitwise logical operators such as
AND, OR, XOR and some ultralightweight primitives
(Permutation, Cyclic Left Rotation). For a low cost RFID tag,
only 2500 to 3000 gates can be allocated for the
implementation of security related tasks.
After these protocols many active and passive attacks came
into existence against these protocols. N.J.Hopper et.al
proposed the first lightweight authentication protocol (HB
protocol [11]) for passive RFID tags. A. juels et.al extended the
protocol (HB+ [10]). However in 2005, H. Gilbert presented an
active attack against HB+ protocol in [12] and challenged its
security claim. In 2006, Pedro Peris Lopez et.al proposed a
new security frameworks and introduced two new
ultralightweight authentication protocols LMAP [2] and EMAP
[3] using simple triangular functions (XOR, AND, OR).
In 2007, H-Chein proposed a new ultralightweight
primitive "Rotation" and proposed SASI protocol [1]. Protocol
seemed to be more sophisticated and robust than its contending
protocols. However, in 2010, Avoine et.al highlights the
vulnerabilities in SASI protocol [5]. Similarly, in [9] Hung –
Min et.al has provided the strong attacks on SASI. In [4],
meanwhile in 2009, Neeli R. Prasad et.al in [6] came up with a
new ultralightweight mutual authentication protocol that
provides strong security and high privacy for Low-Cost RFID
systems. In 2012, a new protocol RFID Authentication
Protocol using Permutation (RAPP [16]) has been proposed.
However, Zahra Ahmadein et.al [17] presented termed as SASI [1], LMAP [2], RAPP [14] [15] and David-
Desynchronization attack on RAPP. Recently, two new Prasad [6]. The description of these protocols involves the
protocols have appeared in literature that belongs to UMAP different logical operations such as Exclusive – OR, Logical
family. These two protocols, RCIA and ASASI, were presented OR, Logical AND etc. [Table 1] provides the logical results of
in 2014 by Umar Mujahid et.al and Qurat Ul Ain et.al in [7] these operations.
and [8] respectively. In [14] [15] the RAPP protocol based on TABLE 1: TRUTH TABLE FOR LOGICAL OPERATORS
permutation is being discussed and the cryptanalysis on the
protocol has also been provided. A B A⊕ 𝐁 A˄B A˅B
0 0 0 0 0
The organization of paper is as follows. Section II provides 0 1 1 0 1
the brief overview of several ultralightweight mutual 1 0 1 0 1
authentication protocols. Section III highlights the proposed 1 1 0 1 1
hardware schematics of these ultralightweight protocols.
Whereas, Section IV presents the results and analysis. Finally, Figure 2 below, represents the generic hardware schematic
section V concludes the overall discussion. of the ultralightweight mutual authentication protocols.
Generally, inside the tag, there exists a cryptographic
II. ULTRALIGHTWEIGHT MUTUAL AUTHENTICATION processor. The function of this processor is to implement the
PROTOCOLS security protocols in tag. An ALU (Arithmetic Logical Unit) is
Mutual Authentication protocols provide corroboration to used in all the cryptographic processors for the execution of the
both tag and reader that they are communicating with either a mathematical operations. Similarly registers and memory are
valid reader or tag. Whenever tag and reader communicates used for storage purposes. Multiple selection pins are used
with one another, they follow some mechanism which might be where required for the selection of specific task and operations.
called as two way handshake method. Using this mechanism
tag and reader authenticates the validity and authenticity of
Bit stream Bit stream
each of the tag or reader.
MUX

Sequencer
H-Chein in [1] has presented the classifications of
authentication protocols based on cryptographic functions that Register
Memory ALU
can be used at tag's end: Output

i. Full-fledged: This is the most powerful class of

mutual authentication in which we can incorporate
traditional cryptographic solutions such as symmetric
encryption, one way hash functions and even public
key cryptography.
ii. Simple: This class is weaker as compared to full-
fledged class because we can only use pseudorandom
number generator and one-way hash functions.
iii. Lightweight: This class is even weaker than simple
authentication protocols; in this class we can use
lightweight pseudorandom number generators and
some simple functions such as Cyclic Redundancy
Check (CRC) but no hash functions at Tag side.
iv. Ultralightweight: This is the weakest class; we
cannot incorporate even pseudorandom number
generator at Tag end. We can only use simple bitwise
logical functions such as XOR, OR, AND etc. Hence,
randomness can only be generated from Readers. Rest
of the paper will be focused on the application and
working of this category.
In this section of paper, four main security protocols have
been discussed that belongs to the Ultralightweight Mutual
Authentication Protocol (UMAP) family. These protocols are
Fig 2: Generic Hardware Schematic of UMAP Family
2.1 LMAP
"Lightweight Mutual Authentication Protocol" (LMAP)
was the very first proposal in the UMAP family that was
proposed in 2006 by Pedro Peris – Lopez [2]. Just like SASI,
LMAP is divided into four main stages that include
identification, Mutual authentication and the key updating.
Operations used for the explanation of LMAP are:
 Logical Exclusive – OR operator (⊕)
 Logical OR operator (˅)
 Logical AND operator (˄)
 Logical SUM (+)
Tag has total of six specific entities including IDS, ID and
the key (K). The key is further divided into four parts
represented as 𝐾1 , 𝐾2 , 𝐾3 , 𝐾4 . Where,
𝐾 = 𝐾1 ||𝐾2 | 𝐾3 |𝐾4
𝑝𝑟𝑖𝑜𝑟
Each key is further categorized as 𝐾𝑛𝑛𝑒𝑥𝑡 and 𝐾𝑛 .
Similarly, these entities are also specified for Reader. IDS, ID
and the K are an n-bits string according to the architecture of
 𝑝𝑟𝑖𝑜𝑟 prior
UMAP family set by EPC class 1 generation II, where n can be 𝐾3𝑛𝑒𝑥𝑡 = 𝐾3 ⊕ n1 + K1 ⊕ ID 8
32, 48, 64 or 96 bit string. 𝑝𝑟𝑖𝑜𝑟 prior
𝐾4𝑛𝑒𝑥𝑡 = 𝐾4 ⊕ n1 + K2 ⊕ ID (9)
Figure 3 explains the mathematics of LMAP protocol.
Whenever tag and reader communicates, the authentication 2.2 SASI
steps take place. Reader transmits a HELLO message towards
tag. On receiving the message, tag replies with its IDS. Reader SASI "Strong Authentication and Strong Integrity" was
searches for the same IDS in backend server. Afterwards proposed by H-Chein [1] in 2007. He proposed this protocol
calculates the messages A, B and C. for Low-Cost passive RFID tags. Basically the description of
protocol involves three operations:
𝐴 = 𝐼𝐷𝑆 ⊕ K1 ⊕ n1 1  Logical Exclusive-OR operator (⊕)
𝐵 = 𝐼𝐷𝑆 ˅ 𝐾2 + 𝑛1 (2)  Logical OR operator (˅)
𝐶 = 𝐼𝐷𝑆 + 𝐾3 + 𝑛2 (3)  Left Rotation operator [ROT(x, y)]

*Where, 𝑛1 and 𝑛2 are the two random numbers with a *Where x & y are the strings of n- bit length (According to
length of n-bits. EPC-C1G2, n can be 32 bit, 48 bit, 64 bit or 96 bit)

The protocol follows various authentication steps to

Hello
confirm the validity of reader or tag. The SASI protocol is
Reader Tag being presented in Figure 4.
IDS
{𝐼𝐷, 𝐼𝐷𝑆, 𝐾1 , 𝐾2 , A||B||C
{𝐼𝐷, 𝐼𝐷𝑆, 𝐾1 ,
𝐾3 , 𝐾4 } D 𝐾2 , 𝐾3 , 𝐾4 } The protocol is divided into four different steps:
1) Initialization of authentication process.
𝐴 = 𝐼𝐷𝑆 ⊕ 𝐾1 ⨁ 𝑛1 2) Verifying reader
𝐵 = 𝐼𝐷𝑆 ∨ 𝐾2 + 𝑛1 3) Verifying tag
𝐶 = 𝐼𝐷𝑆 + 𝐾3 + 𝑛2 4) Pseudonyms and key updating
𝐷 = (𝐼𝐷𝑆 + 𝐼𝐷)⨁ 𝑛1 ⨁ 𝑛2
Reader Hello Tag
𝑝𝑟𝑖𝑜𝑟
𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 = (𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟 +(𝑛2 ⨁ 𝐾4 )) ⊕ 𝐼𝐷 IDS
𝑝𝑟𝑖𝑜𝑟 𝑝𝑟𝑖𝑜𝑟
{𝐼𝐷, 𝐼𝐷𝑆 𝑛 , 𝐾1𝑛 , 𝐾2𝑛 } {𝐼𝐷, 𝐼𝐷𝑆 𝑛 , 𝐾1𝑛 , 𝐾𝑛2 ,
𝐾1𝑛𝑒𝑥𝑡 = 𝐾1 ⨁𝑛2 ⨁(𝐾3 + 𝐼𝐷) A||B||C
𝐼𝐷𝑆 𝑛 −1 , 𝐾1𝑛−1 , 𝐾2𝑛−1 }
𝑝𝑟𝑖𝑜𝑟 𝑝𝑟𝑖𝑜𝑟 D
𝐾2𝑛𝑒𝑥𝑡 = 𝐾2 ⨁𝑛2 ⨁(𝐾4 + 𝐼𝐷)
𝑝𝑟𝑖𝑜𝑟 𝑝𝑟𝑖𝑜𝑟
𝐾3𝑛𝑒𝑥𝑡 = 𝐾3 ⨁𝑛1 + (𝐾1 ⨁𝐼𝐷) 𝐴 = 𝐼𝐷𝑆 ⊕ 𝐾1 ⨁ 𝑛1
𝑝𝑟𝑖𝑜𝑟 𝑝𝑟𝑖𝑜𝑟
𝑛𝑒𝑥𝑡
𝐾4 = 𝐾4 ⊕ 𝑛1 + (𝐾2 ⊕ 𝐼𝐷) 𝐵 = 𝐼𝐷𝑆 ∨ 𝐾2 + 𝑛2
𝐶 = 𝐾1 ⊕ 𝐾2 + (𝐾1 ⨁𝐾2 )
𝐾1 = 𝑅𝑂𝑇((𝐾1 ⨁ 𝑛2 ), 𝐾1 )
Fig 3: LMAP Protocol 𝐾2 = 𝑅𝑂𝑇((𝐾2 ⨁ 𝑛2 ), 𝐾2 )
𝐷 = 𝐾2 + 𝐼𝐷 ⊕ ( 𝐾1 ⨁𝐾2 ⋁ 𝐾1 )
On accepting the message sequence A||B||C from the
reader, tag extorts random number 1 (𝑛1 ) from message A, 𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 = 𝐼𝐷𝑆 + 𝐼𝐷 ⨁ (𝑛2 ⨁ 𝐾1 )
random number 2 (𝑛2 ) from B and calculates C using its own 𝐾1𝑛𝑒𝑥𝑡 = 𝐾1
keys and the extracted random numbers. If calculated C' = C, 𝐾2𝑛𝑒𝑥𝑡 = 𝐾2
only then the reader is a valid reader.

𝐷 = 𝐼𝐷𝑆 + 𝐼𝐷 ⊕ n1 ⊕ n2 4 Fig 4: SASI Protocol

On obtaining D transmitted by the tag, for verifying the tag, Tag has some specific entities which play vital role in the
reader calculates D' and verifies tag by checking D' = D. description of the protocol. These entities involve Pseudonym
Important element of tag was revealed in message D which is Identity (IDS), tag ID (ID) and the two keys. IDS is further
generated after the verification reader. Key updating follows divided into new IDS (𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 ) and old IDS (𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟 ). Two
the following pattern: set of secret keys have been divided in similar manner
𝑝𝑟𝑖𝑜𝑟 𝑝𝑟𝑖𝑜𝑟
(𝐾1𝑛𝑒𝑥𝑡 ,𝐾1 ,𝐾2𝑛𝑒𝑥𝑡 ,𝐾2 ). Tag shares its IDS and keys with
𝑝𝑟𝑖𝑜𝑟
𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 = 𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟 + 𝑛2 + 𝐾4 ⊕ ID 5 reader for mutual authentication. The most significant entity
𝑝𝑟𝑖𝑜𝑟 prior for any tag is its ID which is revealed in the last authentication
𝐾1𝑛𝑒𝑥𝑡 = 𝐾1 ⊕ n2 ⊕ K3 + ID 6 step. Keeping in mind that the IDS, ID and keys are a string of
𝑝𝑟𝑖𝑜𝑟 prior
𝐾2𝑛𝑒𝑥𝑡 = 𝐾2 ⊕ 𝑛2 ⊕ K4 + ID 7 n-bit value. Reader initially sends a HELLO message to tag. In
reply to the hello message, tag transmits its IDS towards the
reader. Reader now checks if the IDS is same and if it is 𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 = 𝐼𝐷𝑆 + 𝐼𝐷 ⊕ n2 ⊕ K1 (16)
present in the database. If so, reader continues with the 𝐾1𝑛𝑒𝑥𝑡 = 𝐾1 17
authentication process. Reader calculates the messages A, B 𝐾2𝑛𝑒𝑥𝑡 = 𝐾2 18
and C using the ALU through the equations provided next:
2. 3 David - Prasad Protocol
𝐴 = 𝐼𝐷𝑆 ⊕ K1 ⊕ n1 (10)
𝐵 = 𝐼𝐷𝑆 ˅ 𝐾2 + 𝑛2 11 In September 2009, David and Prasad [6] proposed a new
𝐶 = 𝐾1 ⊕ K 2 + K1 ⊕ K 2 12 protocol from UMAP family for the passive Low – Cost RFID
Where, Tags. The main aim of the protocol was to provide the security
𝐾1 = 𝑅𝑂𝑇 𝐾1 ⊕ n2 , K1 13 within limited resources (Hardware and Power consumption).
𝐾2 = 𝑅𝑂𝑇 𝐾2 ⊕ n1 , K 2 (14) In this protocol, before inquisitive tag, reader needs to get a
one day certificate from Certificate Authority (CA). The
*Where 𝑛1 and 𝑛2 are the two randomly generated numbers logical operations involved in the protocol account are:
of n-bits length.
 Logical Exclusive – OR operator (⊕)
Rotation operation is being explained in Figure 5. As reader  Logical AND operator (˄)
uses its own set of keys and IDS in order to generate the  Logical NOT operator (~)
messages A, B and C for the encryption purposes, not of 𝐾1 and
𝐾2 has been used. To create a string of data, reader transmits Figure 6 represents the David-Prasad Protocol. In this
message A, B and C as a single message towards tag by protocol, IDS and keys shared by both the tag and reader are
concatenating A||B||C. again of string length n bits. After the initialization of the
conversation by the reader through HELLO message and on
"LEFT ROTATION OPERATION" receiving the IDS from tag, it generates A, B and D. In these
messages reader uses two randomly generated numbers that are
Rotation operation ROT (x, y), left rotates the bits of x 𝑛1 and 𝑛2 . The messages A, B and D are formed in the
with respect to the hamming weight of y. Where, following manner:
hamming weight of y is the number of 1's bits in string y.
Hello
Reader IDS Tag
Let's assume x and y to be 5 bits value: 𝐼𝐷, 𝐼𝐷𝑆, 𝐾1 , {𝐼𝐷, 𝐼𝐷𝑆, 𝐾1 ,
A||B||D
To calculate ROT (10110, 01101) 𝐾2 𝐾2 }
E||F
x = 10110
y = 01101 𝐴 = 𝐼𝐷𝑆 ∧ 𝐾1 ∧ 𝐾2 ⊕ 𝑛1
𝐵 = 𝐼𝐷𝑆 ∧ 𝐾2 ∧ 𝐾1 ⊕ 𝑛2
Step 1: Count Number of one's in y. According to above 𝐷 = 𝐾1 ∧ 𝐾2 ⊕ (𝐾2 ∧ 𝑛1 )
example, number of one's in y is three. 𝐸 = 𝐾1 ⊕ 𝑛1 ⊕ 𝐼𝐷 ⊕ (𝐾2 ∧ 𝑛2 )
𝐹 = 𝐾1 ∧ 𝑛1 ⊕ (𝐾2 ∧ 𝑛2 )
Step 2: Rotate left x by three bits since the number of
one's in y was three. 𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 = 𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟 ⊕ 𝑛1 ⊕ 𝑛2

ROT (10110, 01101) = 10101 Fig 6: David Prasad Protocol

Fig 5: Rotation Operation 𝐴 = 𝐼𝐷𝑆 ˄ 𝐾1 ˄ 𝐾2 ⊕ n1 19

𝐵 = 𝐼𝐷𝑆 ˄ 𝐾2 ˄ 𝐾1 ⊕ n2 20
On receiving A||B||C, to verify reader, tag takes out 𝑛1 from
message A and 𝑛2 from message B and then calculates its own Tag on obtaining A||B||D, pull out 𝑛1 from A and 𝑛2 from
message C'. If C = C', then the reader has been verified by the B. Tag calculated D' using the same equation as of reader.
tag. As said earlier, the most significant entity of tag is its ID;
hence, after verifying the authenticity of reader, tag reveals its 𝐷 = 𝐾1 ˄ 𝑛2 ⊕ K 2 ˄ 𝑛1 21
ID in message D and transmits D to the reader. Message D is
been calculated using the following mathematical computation: After testing the validity of reader by calculating D' = D, on
𝐷 = 𝐾2 + 𝐼𝐷 ⊕ ( K1 ⊕ K 2 ˅ K1 (15) a successful match, tag updates its pseudonym and reveals its
Reader verifies tag with the help of message D. After the ID in E and transmit the string E||F towards the reader.
verification of these three authentication steps, pseudonym and
key updating takes place using the following equations: 𝐸 = 𝐾1 ⊕ n1 ID ⊕ K 2 ˄ n2 22
𝐹 = (𝐾1 ˄ 𝑛1 ) ⊕ K 2 ˄ n2 (23)

Reader on the other hand locally calculates E and F
messages to verify tag. After authentication, reader updates its
pseudonym.
𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 = 𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟 ⊕ 𝑛1 ⊕ n2
3.4 RAPP Protocol
"RFID Authentication Protocol Based on Permutation"
(RAPP) was proposed by Nasour Bagheri [14]. He in his
protocol introduced a new primitive named as Permutation.
The justification of the protocol involves the following
operations:
 Logical Exclusive – OR operator (⊕)
 Left Rotation operation [Rot(x,y)]
 Permutation operation [Per(x,y)]
Tag extracts 𝑛1 from message A and calculates the local
version of B. If local value of B and received value of B is
same then tag transmits message C towards reader.
𝐶 = 𝑃𝑒𝑟 𝑛1 ⊕ 𝐾1 , 𝑛1 ⊕ K 3 ⊕ ID 27
(24)
On receiving message C, it will compare it with the local
pseudonym again if a match occurs then it will generate
another pseudonym 𝑛2 . Reader calculates D and E and
transmits them towards tag.
𝐷 = 𝑃𝑒𝑟 𝐾3 , 𝐾2 ⊕ n2 28
𝐸 = 𝑃𝑒𝑟 𝐾3 , 𝑅𝑜𝑡 𝑛2 , 𝑛2 ⊕ Per n1 , K 3 ⊕ K 2 29
Tag extracts 𝑛2 from D and computes E'. If the two values of E
are same tag and reader update their respective pseudonym in a
manner as shown in Figure 8. The proposed logical hardware
scheme for the RAPP protocol will be shown in next section.

The usage of permutation operation avoids the usage of

unbalanced AND & OR operations. Figure 7 explains the Reader Hello
permutation operation. {𝐼𝐷𝑆, 𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟 , IDS Tag
𝐼𝐷𝑆 𝑛𝑒𝑥𝑡 , 𝐾1
𝑝𝑟𝑖𝑜𝑟
, 𝐾1𝑛𝑒𝑤 , A, B {𝐼𝐷, 𝐼𝐷𝑆, 𝐾1 ,
𝑝𝑟𝑖𝑜𝑟 𝑛𝑒𝑤 𝑝𝑟𝑖𝑜𝑟 𝐾2 , 𝐾3 }
"PERMUTATION OPERATION" 𝐾2 , 𝐾2 , 𝐾3 C
𝑛𝑒𝑤
,𝐾3 } D, E
Permutation operation Per(x,y) shuffles the
data of string x with respect to the string y. 𝐴 = 𝑃𝑒𝑟 𝐾2 , 𝐾1 ⊕ 𝑛1
𝐵 = 𝑃𝑒𝑟( 𝐾1 ⊕ 𝐾2 , 𝑅𝑜𝑡(𝑛1 , 𝑛1 )) ⊕ 𝑃𝑒𝑟(𝑛1 , 𝐾1 )
Let's assume two 6-bit strings, x and y. 𝐶 = 𝑃𝑒𝑟( 𝑛1 ⊕ 𝐾1 , (𝑛1 ⊕ 𝐾3 )) ⊕ 𝐼𝐷
𝐷 = 𝑃𝑒𝑟(𝐾3 , 𝐾2 ) ⊕ 𝑛2
x=1011101 𝐸 = 𝑃𝑒𝑟(𝐾3 , 𝑅𝑜𝑡(𝑛2 , 𝑛2 )) ⊕ 𝑃𝑒𝑟(𝑛1 , 𝐾3 ⊕ 𝐾2 )
y=0111010
𝑝𝑟𝑖𝑜𝑟
Step 1: Check if the first bit of y is 0 or 1? 𝐼𝐷𝑆 𝑛𝑒𝑤 = 𝑃𝑒𝑟 𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟 , 𝑛1 ⊕ 𝑛2 ⊕ 𝐾1
𝑝𝑟𝑖𝑜𝑟 𝑝𝑟𝑖𝑜𝑟
Step 2: If the first bit was 0, shift the first bit ⊕ 𝐾2 ⊕ 𝐾3
of x to the right most position. Else shift it to 𝑝𝑟𝑖𝑜𝑟 𝑝𝑟 𝑖𝑜𝑟
𝐾1𝑛𝑒𝑤= 𝑃𝑒𝑟(𝐾1 , 𝑛1 ) ⊕ 𝐾2
the left most position. 𝑝𝑟𝑖𝑜𝑟 𝑝𝑟𝑖𝑜𝑟
𝐾2𝑛𝑒𝑤= 𝑃𝑒𝑟(𝐾2 , 𝑛2 ) ⊕ 𝐾1
Step 3: Follow the same pattern and shuffle all 𝑝𝑟𝑖𝑜𝑟
the bits of x 𝐾3𝑛𝑒𝑤 = 𝑃𝑒𝑟(𝐾3 , (𝑛1 ⊕ 𝑛2 )) ⊕ 𝐼𝐷𝑆 𝑝𝑟𝑖𝑜𝑟

Per (1011101, 0111010) = 0110111 Fig 8: RAPP Protocol

Fig 7: Permutation Operation III. PROPOSED HARDWARE MODELS

Figure 8 explains the RAPP protocol. In RAPP protocol,
tag stores four values IDS, K1, K2 and K3 (all are n bit long).
To avoid Desynchronization attack in addition to current
pseudonyms reader also stores the old values of these
pseudonyms. Reader in order to start communication with tag
sends a HELLO message to the tag which replies with its IDS.
Reader then calculates message A and B using the permutation,
rotation and exclusive OR operations present in ALU. The
messages are calculated using the equations defined below:
Dealing with passive RFID tags, tags need to have less
number of logic gates in order to reduce the cost of a RFID tag
or chip. Typically a system can have maximum of 5000 to
10,000 (5K - 10K) total of the logic gates out of which only
2,500 to 3,000 (2.5K – 3K) can be devoted for the
implementation of security algorithms [13]. In order to reach
and fulfill such sturdy demands, the complex mathematical
operations need to be performed by the reader itself. Hence the
focus will be on passive tags.

𝐴 = 𝑃𝑒𝑟 𝐾2 , 𝐾1 ⊕ n1 (25) The standard for Ultralightweight Mutual Authentication

𝐵 = 𝑃𝑒𝑟((𝐾1 ⊕ K 2 ), 𝑅𝑜𝑡(𝑛1 , 𝑛1 ) ⊕ Per n1 , K1 (26) Protocols was set in 2004 by the EPC global UHF. The
standard was implemented in 2006. According to Electronic
Product Code Class 1 Generation 2 (EPC-C1G2), 96 bits were O_D Output Data
set as the standard for these algorithms and 32 bits for the I_D Input Data
ultralightweight mutual authentication protocols (for UMAP
1, 2 Reader Input & Output
family). In 2007, SASI [1] was set as a standard for
ultralightweight protocols. E_Flag End Flag
S_Flag Start Flag
Generic Architecture for EPC-C1G2 Protocols: The R,W,P Read, Write, Position
section represents the generic architecture of the protocols that
belongs to the UMAP family. The generic hardware
architecture is being shown in Figure 9. and the number of LUT's (Look Up Table) required. This
section will focus on the efficient hardware implementation of
SASI [1], LMAP [2], EMAP [3] and David-Prasad [6]
according to the 32 Bit architecture. All of the designs are
Timer composed according to the hardware architecture for EPC-
R
C C1G2 protocols derived from the generic architecture
Activation
presented in Figure 9.
T – Flag

3.1 Hardware Implementation of LMAP

Q 1, 2
O_D
32 Bit Architecture: The 32 Bit Architecture for the
I/P proposed model of LMAP is being shown in Figure 10. The
R &
REG W ACK
proposed model is derived from the basic hardware architecture
R O/P
P Next
of EPC-C1G2. The proposed schematic basically consists of
C I_D Actual
three major components namely; the memory representing the
R stored strings such as IDS, the ALU which is even the basic
/
C module as it performs all the mathematical and logical
T
E- Flag operations and finally their present are the registers used to
CR/P store values. In coalesce; the module is basically a
R FSM
cryptographic processor. ALU unit in case of LMAP protocol
S-Flag basically consists of the four basic mathematical and logical
C
function named as XOR, AND, OR and SUM operation.
Fig 9: Hardware architecture for a generic EPC-C1G2 Protocol
Bit Stream
Bit Stream
The module consists of input and output devices, registers, MUX
timers and a cryptographic processor. There also exists a Selection Pin 1
module of Finite State Machine (FSM). The module further
composes Input Output module and a clock. Another module XOR ALU
IDS
named as cryptographic processor is of great importance. This Sequencer
ID
specific processor is used for the implementation of security K1 AND Register
Output
algorithms. Cryptographic processor has an ALU, registers, K2
K3 OR
memory, sequencer and much other minor equipment. The K4
UMAP family algorithms will be discussed in this section with Memory
SUM

respect to their efficient hardware implementation. Table 2

defines the important variables being used in Figure 9.
In this section, we will be discussing the efficient proposed
hardware high level modules of the protocols discussed above
along with the number of logic gates, flip flop, occupied slices
TABLE 2: DESCRIPTION OF EPC-C1G2 PARAMETERS
Parameters Description
R Reset
C Clk
REG Registers
CR/P Cryptographic Processor
Q Query
Selection Pin 2
Fig 10: Logic Scheme of LMAP
3.2 Hardware Implementation of SASI
32 Bit Architecture: Figure 11 shows the proposed generic
hardware architecture for the implementation of SASI protocol.
The proposed results are according to the 32 Bit architecture.
The implemented hardware consists of a memory that stores
the values such as IDS, ID and the secret keys K1 and K2. The
ALU unit of SASI hardware composed of the four main
functions that are Rotation, XOR, or and Sum operation.
Furthermore registers used are shown by a register block. The
inputs to the passive tag, after going through the process of
complex calculations generates the output which is further The output from the tag is further used for communication with
transmitted towards reader. the Reader.

3.3 Hardware Implementation of David-Prasad

Bit stream Bit stream
32 Bit Architecture: Similarly, another efficient hardware MUX
O
model is proposed for the implementation of David-Prasad U
Protocol. The proposed protocol is represented in Figure 12. Selection Pin 1 T
The model is in accordance to the 32 bit architecture that P
follows the EPC-C1G2 standard. The proposed schematic U
IDS Permutation
T
basically consists of three major components namely; the ID
memory representing the stored strings such as IDS, the ALU K1 Rotation Register
which is even the basic module as it performs all the K2
mathematical and logical operations and finally their present K3 XOR
are the registers used to store values. Selection Pin 2
Sequencer
Bit Stream Bit Stream

MUX Fig 13: Proposed Hardware scheme of RAPP

Selection Pin 1
IV. RESULTS & ANALYSIS
ID Rotation ALU
Sequencer
Table 3 presents the specific design parameters according
IDS XOR
Register to the hardware architecture of the protocols presented in
Output
K1 OR previous sections. The table compares the implemented results
K2
on the two FPGA kits that are Spartan 6 (Low Power) and the
SUM
Vertex 7. Total number of LUT's available in vertex 7 kit is
Memory Selection Pin 2 204,088 whereas in case of FPGA kit Spartan 6, the available
LUT's are 2400. Similarly, the total occupied slices are 51000
Fig 11: Logic Scheme of SASI Protocol and 1600 in case of vertex 7 and Spartan 6 kits respectively.

Occupied slices are 78 for Spartan 6 kit for the

Bit stream Bit stream
ultralightweight mutual authentication protocols i-e LMAP and
MUX
for SASI the value increases up to 324 occupied slices. For the
Selection Pin 1 David Prasad protocol, the occupied slices are 187 for Spartan
6 kit whereas RAPP occupies 198 slices in Spartan 6 kit.
ID AND Looking at the occupied slices from the Vertex 7 kit, LMAP is
PID2 using least of the occupied slices that are 66. SASI uses 299
OR Register
K1 Output occupied slices each whereas David Prasad and RAPP use 158
K2 NOT
and 168 respectively.

Figure 13: Logic Scheme for DavidSelection

Sequencer Prasad Pin 2
Fig 12: Proposed Hardware scheme of David-Prasad Protocol
4.4 Hardware Implementation of RAPP
32 Bit Architecture: Finally, the suggested hardware model
according to the EPC-C1G2 is proposed in this section. The
schematic architecture is presented in Figure 13. In this case,
another new function is used in the ALU that is permutation
and which has been explained in detail in the previous section.
RAPP also utilizes the Rotation function and XOR operation.
Number of Flip Flop used keep on increasing with the
increasing the occupied slices both in case of Spartan 6 and
Vertex 7 kits. In case of SASI and LMAP the FF are 46 & 22
respectively for Spartan 6, 43 and 32 for Vertex 7. David
Prasad protocol utilizes 35 & 41 FF in case of Spartan 6 and
vertex 7 respectively. RAPP, however, utilizes 188 LUT's and
29 FF for Spartan 6 kit and for Vertex 7 kit, it uses, 190 LUT's
and 34 FF. However, there always exists a tradeoff between the
cost, performance and the security of the RFID systems [17].
Virtex 7 family is based on 28 nm design technology whereas
Spartan 6 family is based on 45nm technology. Spartan 6
FPGA kit is the best solution for the low cost wireless
communication systems.
 TABLE 3: SIMULATION RESULTS
FPGA Kits Parameters LMAP SASI David Prasad RAPP
Occupied Slices 78 324 187 198
Spartan 6 LUT's 64 388 158 188
LUT's FF 22 46 35 29
Occupied Slices 66 299 158 168
Vertex 7 LUT's 52 345 144 190
LUT's FF 32 43 41 34

V. CONCLUSION [8] Qurat Ul Ain, ; Yusra Mahmood,.; Umar Mujahid.; M. Najam-ul-islam,

In this paper we have presented an optimal architecture and
implemented four ultralightweight authentication protocols:
LMAP, SASI, David-Prasad and RAPP using two different
FPGA kits. We believe that the proposed architecture and
simulated results will act as the stepping stone for the future
implementation of the low cost RFID tags.
REFERENCES
[1] Hung-Yu Chein, "SASI: A New Ultralightweight RFID Authentication
Protocol Providing Strong Authentication and Strong Integrity", IEEE
Transaction on Dependable and Secure Computing, Vol. 4, No. 4, pp. 337
– 340, 2007.
[2] Pedro Peris – Lopez, Julio Hernandez – Castro et.al. "LMAP: A Real
Lightweight Mutual Authentication Protocol for low-cost RFID Tags.",
Proceedings of 2nd workshop on RFID Security, Austria, pp. 100 – 112,
2006.
[3] Peris-Lopez, Pedro, Julio Cesar Hernandez-Castro, Juan M. Estevez-
Tapiador, and Arturo Ribagorda. "EMAP: An efficient mutual-
authentication protocol for low-cost RFID tags." In On the Move to
Meaningful Internet Systems 2006: OTM 2006 Workshops, pp. 352-361.
Springer Berlin Heidelberg, 2006.
[4] Paolo D'Arco, Alfredo De Santis, IEEE Transactions on dependable and
secure computing, "On Ultralightweight RFID Authentication Protocols",
Vol. 8, No. 4, July/August 2011.
[5] Avoine, Gildas, Xavier Carpent and Benjamin Martin, Radio Frequency
Identification: Security and Privacy Issues, "Strong Authentication and
Strong Integrity (SASI) is not that strong", pp. 50-64, Springer Berlin
Heidelberg, 2010.
[6] David, Mathieu, and Neeli R Prasad, providing strong secure and high
privacy in low cost RFID networks.", International Conference on
Security and Privacy in mobile information and communication systems,
itlay, pp. 172 – 179, 2009.
[7] Umar Mujahid, M. Najam-ul-Islam, and M. Ali Shami, “RCIA: A New
Ultralightweight RFID Authentication Protocol Using Recursive
Hash,” International Journal of Distributed Sensor Networks, vol. 2015,
Article ID 642180, 8 pages, 2015. doi:10.1155/2015/642180
"Cryptanalysis of mutual ultralightweight authentication protocols: SASI
& RAPP," Open Source Systems and Technologies (ICOSST), 2014
International Conference on , vol., no., pp.136,145, 18-20 Dec. 2014
doi: 10.1109/ICOSST.2014.7029334
[9] Hung – Min Sun, Wei – Chih Ting et.al, "On the Security of Chien's
Ultralightweight RFID Authentication Protocol", IEEE Transaction on
Dependable and Secure Computing, Vol. 8, No. 2, pp. 315 – 317, 2011.
[10] A. Juels and S. Weis, "Authenticating Pervasive Devices with Human
Protocols", Proc. CRYPTO '05, pp. 293-308, 2005.
[11] N.J.Hopper and M. Blum, "Secure Human Identification Protocols",
Proc Int'l Conf. Theory and Application of Cryptology and information
security: Advances in Cryptology (ASIACRYPT '01), pp. 52-66, 2001.
[12] H.Gibert, M.J.B. Robshaw, and H. Sibert, Electronics Letters, vol. 41,
no. 21, pp. 1169-1170, "An active attack against HB+ a Provably Secure
Lightweight Authentication Protocol", 2005.
[13] Honorio Martin, Enrique San et.al, "Efficient ASIC implementation and
Analysis of two EPC-C1G2 RFID Authentication Protocols", IEEE
Sensors journal, Vol. 13, No. 10, October 2013.
[14] Bagheri Nasour, Masoumeh Safkhani et.al, "Cryptanalysis of RAPP, an
RFID Authentication Protocol", IACR Cryptology ePrint Archive, 2012.
[15] Wang Shao-hui, Han Zhijie et.al, "Security Analysis of RAPP: An
RFID Authentication Protocol based on Permutation", College of
computer, Nanjing University of posts and telecommunications Nanjing
210046, China, Jiangsu High Technology Research Key Laboratory for
Wireless sensor Networks, Nanjing Jiangsu, 210003, China.
[16] Z. Ahmadian, M. Salmasizadeh, and M. R. Aref, “Desynchronization
attack on RAPP ultralightweight authentication protocol,” Information
Processing Letters, vol. 113, no. 7, pp. 205–209, 2013.
[17] Maarof, A.; Labbi, Z.; Senhadji, M.; Belkasmi, M., "Security analysis
of low cost RFID systems," in Codes, Cryptography and
Communication Systems (WCCCS), 2014 5th Workshop on , vol., no.,
pp.11-16, 27-28 Nov. 2014.
[18] U. Mujahid, M. Najam-ul-islam, J. Ahmed, and U. Mujahid,
“Cryptanalysis of ultralightweight RFID authentication protocol,”
Report 2013/385, 2013,https://eprint.iacr.org/2013/385.
[19] Zilong Liu; Dongsheng Liu; Lun Li; Hui Lin; Zhenqiang Yong,
"Implementation of a New RFID Authentication Protocol for EPC Gen2
Standard," in Sensors Journal, IEEE , vol.15, no.2, pp.1003-1011, Feb.
2015.
 Physical Layer Secrecy Solution for Passive
Wireless Sensor Networks
Thombre Avinash Sarjerao∗ , Aditya Trivedi
ABV-Indian Institute of Information Technology and Management, Gwalior, India 474015
Email: ∗ avinash.thombre14@gmail.com, atrivedi@iiitm.ac.in †
Abstract—The backscatter communication system has tremen-
dous potential in commercial applications, still very less work has
been done to study the benefits of it. Backscatter communication
system is the backbone of many low cost and low power dis-
tributed wireless systems. The data transmission between various
nodes in wireless communication system always comes with the
risk of third party interception. This leads to privacy and security
breaches of the information. In this paper, physical layer security
of backscatter wireless system for multiple eavesdropper, single
tag, and single reader case is studied. Unique characteristics of
the channel are used to provide security to signal transmission.
In order to degrade the reception of the signal by eavesdropper,
a noise injection scheme is proposed. The advantages of this
approach are discussed for various cases while evaluating the
impact of the key factors like antenna gain and location of
the eavesdropper on the secrecy of the transmission. Analytical
results indicate that if properly employed, the noise injection
scheme improves the performance of backscatter wireless system.
Index Terms—Backscatter communication system, physical
layer secrecy, wireless sensor networks, artificial noise injection.
I. INTRODUCTION
Physical layer security focuses on exploiting the physical
layer properties of channels to secure signal transmission
against eavesdropping. It can also complement current crypto-
graphic techniques as the two approaches operate in different
domains, one protects the communication phase while the
other protects the data processing after the communication
phase. In addition, physical layer security techniques can also
be adopted for secret key generation and distribution, by
exploiting the rich randomness and dynamics available in the
wireless medium.
Radio Frequency Identification (RFID) is wireless technol-
ogy used to identify and track objects using tag and reader.
Tags, also called transponders or labels, which are attached
to the object being detected or transplanted inside the body
of the animal to be tracked. Every tag consist of a chipset
and a antenna used to store information about the object and
transmit information to the reader as well as harvest power,
respectively. Most common types of tags do not have their own
energy source; instead they harvest energy from continuous
carrier wave. Reader, also called interrogator, interrogates tags.
The interrogation takes place over wireless media and there is
no requirement of the line of sight between the two.
Backscatter wireless communication systems is an emerging
wireless technology that has been very extensively used in
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
†
distributed passive wireless sensor networks and RFID systems
[1]-[10]. RFID systems allows interconnection of the various
objects using tags [1]. Various technical problems must be
dealt right form tag circuit design to advanced backscatter
signal processing, to rip the benefits of backscatter based
wireless systems [1]-[11]. The standardization and practical
issues related to RFID and backscatter communication security
are discussed in [5]. Different issues addressed to the authenti-
cation of RFID tags and their associated protocols are studied
in [6], [8], and [9].
Backscatter wireless communication system’s secrecy con-
stitutes a main design problem. Malicious attacks in communi-
cation, such as eavesdropping, may result into data interception
as well as serious security or privacy issues like owner tracking
or identity or data modification [11]. The pioneering work of
Wyner in [12] is the first to discuss the usage of the wireless
communication channel characteristics as a tool to secure
wireless communication. The problems of securing backscatter
wireless communication systems face basic problems from the
practical limitations, in terms of size, cost, and computation,
which encourage novel approaches to wireless communication
security [13], [14]. To overcome some of the limitations of
cryptography in backscatter wireless communication systems,
physical layer (PHY) secrecy techniques are developed which
does not depend on the secret key generation and distribution.
Physical layer secrecy schemes can be used as either for
supplementing to the cryptography or as a complement that
can strengthen available cryptographic schemes by providing
a secure communication channel for key distribution and
exchange.
In subsequent section II, system model is explained. Multi-
ple eavesdropper single reader and single case is illustrated
with the help of received signal and SNR expressions. In
section III, noise injection scheme is discussed. The effects
of it on signal reception and corresponding changes in SNR
are studied. In section IV, conditions for positive secrecy
are evaluated. Section V, contains the results of the paper.
Conclusion and future work is discussed in section VI.
II. SYSTEM MODEL
Consider a backscatter communication system consisting of
a single reader single tag and multiple eavesdroppers. The
tag does not initiate communication on its own. To wake up
the tag, reader transmits the continuous wave (CW) signal i.e.
carrier. An RF voltage is induced over the antenna which in

Fig. 1. Illustration of multiple eavesdropper backscatter system.
turn powers the tag. The activated tag sends back the stored
information by controlling the attenuation factor after impact
of the CW signal. The reader continuously broadcasts the CW
signal to wake up the tag, as the tag does not have its power
source, at the same time receives the backscattered signal
which contains the stored information inside the tag, i.e., the
reader should operate in full duplex mode.
A. Backscatter System Model
Considering the discrete time signal model in the baseband,
the received signal at the reader is
yR = hT R hRT xs + ηR + hT R ηT
where x is the transmitted signal by the reader, s is the
tag’s information signal, hT R and hRT are the tag-reader and
reader-tag channel gains, ηR is Additive White Gaussian Noise
2
(AWGN) at the reader with power σR . ηT is AWGN at the
tag which is backscattered to the reader with power σT2 . The
power of the signal transmitted by the reader is denoted by
Px . The fraction of the received power reflected back in the
tag’s useful information signal is Γ. Note that Γ < 1 is due to
the passive nature of the tag [1].
In this paper, Friis equation model is used to model the
power loss during signal propagation. For the reader with
transmit power Px , received power after backscattering is
PRrx = Px ΓG2RT K 2 d−4
RT
where GRT represents the combined transmitter receiver an-
tenna gain of the reader tag link. dRT is the reader-tag
distance, K = (λ/4π)2 is the constant depending on the
carrier wavelength. The SNR at the reader is
Px ΓG2RT K 2 d−4
RT
γR = 2 + σ2 G −2
σR T RT Kd RT
where γR is the SNR at the reader. The aim of this paper
is to secure the transmission of the data from eavesdroppers
while transmission using physical layer security techniques.
The received signal at the reader is
yE = hT E hRT xs + ηE + hT E ηT E
where hT E is the tag-eavesdropper channel gain, ηE is the
2
AWGN at the eavesdropper with power σE , ηT E is the AWGN
backscattered from tag to the eavesdropper with power σT2 E .
Note that, when eavesdropper is listening to tag, reader is also
transmitting its CW signal as well. Therefore eavesdropper
receives the superposition of CW signal and tag’s information
signal and assume that eavesdropper knows CW signal and
removes it from the received signal.
Using the Friis equation to model the power loss due to
propagation, the SNR at eavesdropper is
Px ΓK 2 GRT GT E d−2 −2
RT dT E
γE = 2 + σ2 G −2 (5)
σE T E T E KdT E
where GT E is combined transmitter-receiver gain of tag-
eavesdropper link and dT E is the tag-eavesdropper distance.
The condition for positive secrecy is γR > γE . This
condition can be easily violated if eavesdropper has a sensitive
2 2
receiver than reader, i.e. σE  σR and tag’s backscattered
2
noise power (σT E ) is small. This is a critical because the
receiver noise at the eavesdropper is not controllable. A noise
injection scheme is proposed that will improve the security
in backscatter communication and allow secure transmission
even when the eavesdropper is present.
III. NOISE INJECTION SCHEME
Unfortunately, the conventional CW signal is known to the
(1) eavesdropper and hence, does not interfere with the reception
of backscatter signal by eavesdropper. The authors in [16]
have proposed a noise injection scheme, which superimposes
a noise like random signal generated privately on the con-
ventional CW signal which is much similar to the AWGN so
that eavesdropper should not distinguish it from conventional
receiver noise. Therefore, x + z is transmitted instead of x
, where z is injected noise signal with noise power Pz . The
received signal at the reader after backscattering is
yR = hT R hRT xs + hT R hRT zs + ηR + hT R ηT (6)
where first term is useful signal and the remaining terms
are combined noise. The reader will be able to attenuate the
(2) received noise effectively as it is aware of injected noise.
But the eavesdropper may find some trouble in attenuating
the same due to a) the eavesdropper is not aware of the
random noise signal superimposed over the signal b) an
eavesdropper is often considered as a small device which
can perform limited computations and do not have advanced
receiver structure.
(3) For above noise injection scheme SNR at the reader is,
Px ΓG2RT K 2 d−4
RT
γR = (7)
kPz ΓGRT K 2 d−4
2
RT + σ 2 + σ2 G
R
−2
T RT KdRT
where k represents noise attenuation factor. The additional
interference caused at the eavesdropper while receiving the
(4) backscattered signal from tag is,
 yEn = hT En hRT xs + hT En hRT zs + hREn z 
+ ηEn + hT En ηT En (8)
where z is the received backscattered noise signal from tag
while z  is the injected noise signal received at eavesdropper
directly from the reader, i.e., z and z  represent the transmitted
noise signals that are received by the eavesdropper at different
time instants. Note that the power of the directly received noise
z  is typically much larger than the power of the backscattered
noise z. Neither z nor z  is known to the eavesdropper. But
the eavesdropper, if equipped with a directional antenna, can
minimize or potentially zero its antenna gain towards the
reader, effectively removing the third term in (8) from its
received signal. However, in this case, the noise injection
would still be beneficial as the impact of the backscattered
noise seen in the second term of (9).
The generalised expression for SNR of the received signal
at the eavesdropper is given by,
Px ΓGRT GT En K 2 d−2 −2
RT dT En
γEn =
Pz GREn Kd−2 2 −2 −2
REn + Pz ΓGRT GT En K dRT dT En
2
+ σEn + σT2 En GT En Kd−2
T En (9)
where GREn represents the combined transmitter receiver
antenna gain of the reader and nth eavesdropper link, and
dREn is the distance between reader and nth eavesdropper.
Comparing with eavesdropper SNR without noise injection,
given in (4), the benefit of noise injection is clear as reader can
now limit the eavesdropper’s SNR by controlling the injected
noise power.
So, among n eavesdroppers scattered around tag and reader,
the one with highest value of SNR is selected as it is having
highest chances of posing security threat. Because with such
high a value of SNR, it can easily detect and decode the signal
compared to other eavesdroppers. This SNR value is set as
benchmark for further evaluations as it represents worst case
of eavesdropper SNR values. If the SNR at eavesdropper is
greater than the SNR at reader then eavesdropper will easily
detect and decode the signal. Hence the purpose of noise
injection is not be fulfilled.
Figure 2 shows that, as the distance between tag and eaves-
dropper increases SNR at eavesdropper falls rapidly. This in-
dicates that with increase in the tag-eavesdropper distance the
interception of backscattered signal by eavesdropper decreases.
Hence, more security to the backscatter communication. But
as the distance between tag and eavesdropper decreases i.e.
eavesdropper gets closer to the tag, the SNR at eavesdropper
rises drastically towards infinity. Suppose if eavesdropper is
just at the tag, then it becomes very difficult to attain secrecy.
In such situation SNR of the eavesdropper will be infinity,
theoretically, and it can decode any signal backscattered by
tag and attaining secrecy becomes a tough job.
Fig. 2. Variation in the SNR at eavesdropper as per tag-eavesdropper distance
IV. CONDITION FOR POSITIVE SECRECY
The performance limits of physical layer security are often
characterized by maximum secrecy rate achievable for given
secure transmission scheme. This parameter gives the maxi-
mum rate at which the transmission of confidential information
can be decoded by the legitimate receiver with arbitrarily
small error while perfect secrecy against the eavesdropper is
maintained. Performance of backscatter system can also be
characterised by deriving the achievable secrecy rate C0S of
the proposed noise injection technique. Due to non-Gaussian
distribution of the combined noise at the eavesdropper, it is
difficult to derive exact secrecy rate expression. Even though,
the derived SNR expressions cab be used to get an approximate
secrecy rate expression that can quantify the overall secrecy
of the transmission. Therefore, secrecy rate is given by
 +
C0S = C0R − C0E
+
C0S = (log (1 + γR ) − log (1 + γE )) (10)
where, C0R is the capacity of the receiver channel and C0E
is the capacity of the eavesdropper channel.γE and γR in
(10) would represent the capacity or SNR of the eavesdropper
and relay with the strongest signal reception respectively. As
discussed previously, the receiver noise at the eavesdropper is
uncontrollable and unknown, a robust design approach should
aim to provide secrecy in the worst-case scenario by assuming
2
σE = σT2 E = 0 (with such a worst-case assumption, secrecy
is not achievable without noise injection). Therefore, in the
subsequent sections of the paper, the performance analysis and
design for the worst-case situation is provided by assuming no
noise at eavesdropper.
For securing communication against eavesdroppers, SNR at
reader should be greater than maximum of SNRs at eaves-
droppers, i.e.,
γR > γEmax (11)
where γEmax = max{γEn }. Only then it is possible to
transmit the data securely over wireless transmission channel.
Fig. 3. The minimum required power for noise injection Pz for a range of
Fig. 4. The minimum required power for noise injection Pz for a range of
noise attenuation factors k.
receiver-eavesdropper distances dT E .

V. RESULTS
Using the SNR expressions in (7) and (9), with worst case
2
condition of eavesdropper’s noise( i.e. σE = σT2 E = 0), this
condition is given by

 2
dT E GT E
>
dRE GRE
 
d2RT σR
2
+ KGRT σT2 −2
− (1 − k)ΓGRT KdRT (12)
KGRT Pz
It is clear that noise injection, Pz > 0, is important for
positive secrecy rate, which depends on the relatively how far
is the eavesdropper. When eavesdropper is very close tag then
maintaining secrecy becomes very difficult and strong noise
signal has to be introduced.
In order to fulfil the condition in (12) without caring about Fig. 5. Optimal value of α versus receiver-eavesdropper distance dT E .
the eavesdropper’s parameters, the right side of the above
equation should be positive. Therefore
d4RT σR
2
+ d2RT KGRT σT2 d2RE d2RT σR
2
+ KGRT σT2
Pz > (13) Pz > (14)
(1 − k)ΓG2RT K 2 d 2T E K
The right side of the equation is finite as long as k < 1. Figure 4 shows minimum required power for noise injection
Therefore, if reader is able to adjust the power of noise depends on location of eavesdropper. As the eavesdropper
injection according to (13) secure communication is possible gets closer to the tag, the power required for noise injection
regardless of the location and gains of the eavesdropper increases towards infinity and achieving positive secrecy be-
antennae. comes more challenging. Therefore, it is interesting to study
Figure 3 shows the minimum required power for noise how close the eavesdropper can get to the receiver for practical
injection for a range of noise attenuation factors. As the atten- values of the noise power generated by reader.
uation capability of the reader gets weaker, i.e. as k increases, Secure communication at a positive secrecy rate can usually
there is need for a stronger noise signal to maintain positive be achieved by inserting a small amount of noise at the reader.
secrecy. Also lower curve with reader to tag distance of 2 Now, the problem of optimally allocating the total transmission
meter indicates minimum required power for noise injection power at the reader between the conventional CW signal and
is lesser than that for reader to tag distance of 4 meter. Which the injected noise is discussed, in order to maximize the
means as the reader gets closer to the tag required power for achievable secrecy rate given in (10). To perform this power
noise injection decreases. allocation, the reader must be able to estimate the reader-tag
Consider, reader does not go for noise processing (i.e. k = channel. To do so, the reader can use the backscatter signal to
0) and all antennae are omnidirectional. Then condition for estimate the channel. This can be done either jointly during
achieving positive secrecy becomes the signal detection or the initial training phase with the tag.
Fig. 6. Optimal value of α versus eavesdropper’s antenna gain ratio
GT E /GRE .
To study the optimal power allocation problem, we define
the ratio of power allocated between conventional CW signal
and noise as α ∈ (0, 1].
Px = αP and Pz = (1 − α)P
The optimization of the power gives optimized equation of the
α as
 proved SNR at eavesdropper and thus affects the transmission.
a(a + k)[a(b − 1) + (b − k)] − a(1 − k)
α=1− (15)
a(b − 1) + k(b − k)
2 4
σR dRT GRE d2RT d2T E
where a = and b = 1 +
P ΓG2RT K 2 ΓGRT GT E Kd2RE
Figure 5 shows the optimal value of α versus the receiver-
eavesdropper distance dT E ranging from 0 to 1 m. The optimal
value of α is very close to 1 for nearly all possible range of the
tag-eavesdropper distance (including the values of dT E > 1
not shown in the figure), which implies that only small fraction
of power is required for noise injection in order to achieve the
optimal physical layer security performance. Only when the
receiver-eavesdropper distance approaches 0, does the optimal
α starts to drop significantly.
The impact of eavesdropper’s antenna gains, when it is
equipped with a directional antenna with potentially high
directivity, can be studied by either minimizing the antenna
gain towards the reader GRE or maximizing the antenna gain
towards the tag GT E , the eavesdropper is able to improve its
SNR. Expression for γE in (9) shows (with the assumption
of σE = 0) the ratio of the eavesdropper antenna gain, i.e.
GT E /GRE , is an important factor and range of values of it are
considered. For simplicity, consider dRT = dRE = dT E = 2m
or dRT = dRE = dT E = 4m.
Figure 6 shows the optimal value of α versus the eavesdrop-
per’s antenna gain ratio GT E /GRE ranging from 1 to 105 .
The optimal value of α is very close to 1 for a wide range of
practical antenna gains. Only when the gain ratio goes beyond
103 , does the optimal α start to drop, but still remains at a
Fig. 7. Variation in the SNR at eavesdropper as per optimal power allocation
ratio
large value even if the gain ratio reaches 105 . This implies
that, for most practical antenna gains, a small portion of the
power is needed for noise injection.
Figure 7 shows that, as the optimal power allocation ratio
increases SNR at the eavesdropper also increases. The increase
in power allocation ratio indicates that there is decrease in the
noise power that is being transmitted. This results into im-
This increases vulnerability of the signal being detected by the
eavesdropper and hence it reduces the secrecy of backscatter
communication. Thus for low SNR at eavesdropper at more
noise power has to be injected and in turn less power allocation
ratio should be maintained.
VI. CONCLUSION
In this paper, we analyzed the characteristics of physical
layer security in multiple-eavesdropper single-reader single-
tag case and proposed conditions for positive secrecy. A noise
injection scheme is proposed using the characteristics of the
channel. Conditions for positive secrecy are derived for various
cases. Optimal power allocation is discussed and analyzed the
performance under different situations.
In future, investigations may be done for multi-reader/tag
case, designing more efficient and backscatter oriented secrecy
achieving codes with low complexity under various backscatter
radio propagation environments.
R EFERENCES
[1] D. M. Dobkin, The RF in RFID: Passive UHF RFID in Practice. Newnes,
2007.
[2] T. Philips, T. Karygiannis, and R. Kuhn, ”Security standards for the RFID
market,” IEEE Security Privacy, vol. 3, no. 6, pp. 8589, Nov. 2005.
[3] A. O. Bicen and O. B. Akan, ”Energy-efficient RF source power control
for opportunistic distributed sensing in wireless passive sensor networks,”
in Proc. 2012 IEEE Symp. Comput. Commun.
[4] D. Arnitz, U. Muehlmann, and K. Witrisal, ”Wideband characterization
of backscatter channels: derivations and theoretical background,” IEEE
Trans. Antennas Propag., vol. 60, no. 1, pp. 257-266, Jan. 2012.
[5] A. Juels, ”RFID security and privacy: a research survey,” IEEE J. Sel.
Areas Commun., vol. 24, no. 2, pp. 381-394, Feb. 2006.
[6] A. Juels, ”Minimalist cryptography for low-cost RFID tags,” in Proc.
2004 Int. Conf. Security Commun. Netw.
[7] E. Vahedi, R. K. Ward, and I. Blake, ”Security analysis and complexity
comparison of some recent lightweight RFID protocols,” in Proc. 2011
Int. Conf. Computational Intelligence Security Inf. Syst.
[8] S. Piramuthu, ”SASI: a new ultralightweight RFID authentication pro-
tocol providing strong authentication and strong integrity,” IEEE Trans.
Dependable Secure Comput., vol. 4, no. 4, pp. 337-340, Dec. 2007.
[9] P. Peris-Lopez, J. C. Hernandez-Castro, J. Estevez-Tapiador, and A.
Ribagorda, ”M2AP: a minimalist mutual-authentication protocol for low
cost RFID tags,” in Proc. 2006 Int. Conf. Ubiquitous Intelligence Comput.
[10] T. Li and G. Wang, ”Security analysis of two ultra-lightweight RFID
authentication protocols,” in Proc. 2007 IFIP SEC.
[11] H.-J. Chae, D. J. Yeager, J. R. Smith, and K. Fu, ”Maximalist cryp-
tography and computation on the WISP UHF RFID tag,” in Proc. 2007
RFID Security.
[12] A. D. Wyner, ”The wire-tap channel,” Bell System Tech. J., vol. 54, no.
8, pp. 1355-1387, 1975.
[13] P. K. Gopala, L. Lai, and H. El Gamal, ”On the secrecy capacity of
fading channels, IEEE Trans. Inf. Theory, vol. 54, no. 10, pp. 4687-4698,
Sept. 2008.
[14] Z. Li, W. Trappe, and R. Yates, ”Secret communication via multi antenna
transmission,” in Proc. 2007 Conf. Inf. Sciences Syst.
[15] M. Bloch and J. Barros, Physical-Layer Security: From Information
Theory to Security Engineering. Cambridge University Press, 2011.
[16] Walid Saad, Xiangyun Zhou, Zhu Han, and H. Vincent Poor. ”On
the Physical Layer Security of Backscatter Wireless Systems.” Wireless
Communications, IEEE Transactions on 13, no. 6 (2014): 3442-3451.
 Visualising and Analysing Online Social Networks
Kambale Vanty Muhongya and Manoj Sewak Maharaj
School of Management, Information Technology and Governance
University of KwaZulu-Natal
Durban, South Africa
vantykyb@gmail.com, Maharajms@ukzn.ac.za
Abstract—The immense popularity of online social networks
generates sufficient data, that when carefully analysed, can
reveal unexpected realities. People are using them to establish
relationships in the form of friendships. Based on data collected,
students’ networks were extracted, visualized and analysed to
reflect the connections among South African communities using
Gephi. The analysis revealed a slow progress in terms of
connections among communities from different ethnic groups in
South Africa. This was facilitated through analysis of data
collected through Netvizz as well as by using Gephi to visualize
social media network structures.
Keywords—Online social network; betweeness centrality;
closeness centrality; race; visualization; graph; Gephi
I. INTRODUCTION
Data visualization is becoming an increasingly important
component of analytics in the age of big data. Businesses
trying to understand global markets, governments concerned
about transformation and well being of their citizens and many
other challenges. A variety of tools like Gephi, Vizster,
Leximancer, NodeXL etc. are used around the world to
visualize, analyses large scale data.
Online social networks are formed by individuals and/or
content and the relationship between them. Facebook remains
the leading social network today. Facebook is known for
connecting individuals in forms of friendships claiming to be
transforming distant communities. In this paper, the
researchers present an analysis of student Facebook networks
using Gephi, an open source software that works like a
database and assists to visualize and analyse obscured large
scale online.
II. BACKGROUND
A primary use of social networking sites (SNSs) is
communicating and sharing information with friends. When it
comes to social network sites statistics, African nations rank
low. However, South Africa currently ranks 29th on
Facebook’s international customer record, and show many
resemblances with larger nations [1]. According to
Socialbakers [2], with a large 82% membership, Facebook is
the prominent public media site used in Southern Africa. Over
half of South African Facebook users access the site via their
mobile phones. Eighty percent of the respondents indicated
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
that they have a MySpace profile, with MXit at 29%, and
Twitter at a close 28%. Additional findings from the same
survey indicated that 74% surf the Internet to visit SNSs, 74%
access Facebook at least once a day, 25% have met more
friends on SNSs than they have in real life, 24% have gone on
a real-world date with someone they met on social media, and
16% use SNSs to advertise their businesses.
In general, Facebook contains a wall, a friend page, a news
feed and an email page. A wall is an area where the user or
friends can post notes or add multimedia. A friend page shows
the number and a list of the friends a user is connected to. A
news feed informs the user about some Facebook events and
about the activities of Facebook friends. Facebook has an
embedded email service available to users to send private
messages to other Facebook users. To see and view profiles of
individuals on Facebook, a user needs to subscribe. A valid
existing email is required for someone to subscribe and be
able to use the network. Facebook allows searches and
discloses personal information. Users are not obliged to
disclose information. Users can decide to restrict access to
their profiles by changing their privacy settings in the system
[3]. Nevertheless, by default, anyone can search and read other
people’s profiles on the network.
Facebook is a social network used by university students,
high school pupils, and others. Facebook profiles show contact
details, including physical addresses and telephone numbers,
and additional information not often found on other social
networks thus enhancing its capacity to form both online and
real-world friendships. Facebook users can add friends to their
profiles asking for friendship. This is done by sending a friend
request to another user. When the other party accepts the
request, the connection is shown in the network of friends.
Considering the descriptions above, it might be difficult to
understand the link among events, peoples, contents that
happens on social networks like Facebook. However with tool
like Gephi, this has become possible. Initiated by Mathieu
Jacomy in 2006, Gephi visualizes and analyses all kind of
networks and complex datasets, dynamic and hierarchical
graphs [4]. It runs on Windows, Linux and Max OS and is free
of charge. We have used Gephi to visualize and analyse large
Facebook networks of students in South African universities.
Gephi work with other compatible software such as Netvizz a
.gdf file (a simple text format that specifies an undirected
graph) [5].
Netvizz assits in the collection of Facebook networks in a
form of a file that is then imported into Gephi for further
analysis. Other compatible files include .gexf, .gdf, .gml,
GraphML, Pajek Net, GraphViz DOT, .csv, UCINET DL,
Tulp TPL, Netdwaw VNA and Spreadsheets. The data once
imported is presented in a form of a dataset with rows and
columns. More columns can as well be added if more data
need to be added to the data source. Gephi is easier to use than
other network analysis tools where the user might need to
write scripts in order to visualise networks.
Gephi has been used to visually map @reply conversation
networks on twitter [6], to visualise and analyse business
intelligence data [7] but has not been used to map connections
among different ethnic groups. In this paper we use Gephi to
map and analyse students’ relationship based on race.
III. NETWORK VISUALIZATION
Visualizing a network means to represent a network in a
graphical form. This graphical form is made of nodes
(individuals, contents etc) and edges (links). For instance if A
and B are members of a network then A and B are nodes and
if there is something that connect A and B then it is called
edge. For example A (node 1) and B (node 2) are friends
(edge) or for instance node A is following node B. These
relationships can be directed or undirected.
A. Navigation
Gephi is easy to navigate. Just like any other software, there
is a menu bar on top of the window which includes standard
commands. Below the menu bar, there are three most
important buttons (this can be called a tool bar). These three
buttons assist users to perform different task on a specific
network. An overview button where the partition, ranking,
layout and statistics on nodes and edges are performed; the
data laboratory button where all the data of a network can be
imported, adjusted and modified; and a preview button which
assists in the viewing of the network and exportation of the
network after all necessary modifications have been
performed.
B. Manipulation of graphs
Graphs can be visualized using different colours,
background colours of the working place, shapes of nodes
(e.g. 3 Dimensions), labels of nodes, font, size, labels
attributes (from the data source). Colours, labels, sizes of
nodes tell different stories. They can determine communities
in a network, influential nodes in the network. Node colours
and sizes can be used to determine communities in the
network. The separation of community can be achieved by
different layouts.
A network can be partitioned. This partition uses labels
from the data source. For instance if there is a label called
gender, race etc. in the data source, they can be used to
distribute groups in the network. The partition present nodes
in form of frequencies with percentage associated to it. The
percentage determines how many nodes belong to a particular
group. A network filter is used to perform statistics link to ego
networks, degree range etc. Ego networks can be studied
separately from the main network. This is made possible by
taking the number identifying the node from the database and
searching for it in the filter settings of ego network. Analysis
in terms of path length, coefficient, modularity (betweeness
and closeness centrality), network density can be performed.
The usability of the application is of great importance as it
manipulates data presenting the data in forms of graphics and
perform statistical analysis in form of frequency that are
performed by application such as SPSS or Excel.
C. Data source
Before the analysis was performed, students were consulted
to allow the researcher to use their Facebook networks without
disclosing their identity. After permission was given from
students, a sample of 25 Facebook networks was randomly
extracted for analysis. The data source is accessible by
clicking on the data laboratory button. The data source is in a
form of rows and columns and contains two most important
sheets (nodes and edges). The data source includes peoples
and/or content and how they are linked. As previously
mentioned, edges are used to link individuals in a network and
nodes are the individuals. These connections and individuals
can be manually adjusted or changed if necessary. New
columns, new nodes and new edges can be added to the data
source. Excel files can be imported here as well. Application
like NodeXL which allow the extraction of data from different
website into an Excel file might be used to collect the data and
then import the file into Gephi database. Also Gephi allow
users to export the data source into Excel. This compatibility
is of great importance as it assists researchers to use different
applications (different functions and outputs) to manipulate
networks. In this window, the data is verified and cleaned for
correctness of outputs.
D. Network Preview
There are embedded setting in the Preview window. Users
can change and personalize the look of the graph output.
Properties like border, colours, width are available for both
edges and nodes. Once the user is satisfied with the look of the
graph, he can decide to export the network using .svg, .pdf and
.png type of files.
 IV. NETWORK ANALYSIS
As it was mentioned before, networks are formed
online by individuals with different characteristics such
as race, gender, language etc. to create a profile. The
analysis presented in this section is based upon race
factor and how it affecting South African university
students to connect on Facebook. A cursory look at the
student networks does not reveal underlying trends that
are exposed by a Gephi analysis.
Fig. 1 is an Indian network extracted from Facebook, where
according to the results presented by Ghephi, red nodes
represent 79.75% of his friendship with Indians, the light blue
colour represents 10.25% of his friendships with Blacks, and
purple nodes represent 9% of his friendships with Whites. The
analysis was performed on a Black network and the result
revealed a 93.48% Black friendships, 4.82% Indian
friendships, 0.71% White friendships, and 0.71% Coloured
friendships.
The same analysis was done on several networks of
individuals from different ethnic groups. As mentioned in the
experiment above, the graphs keep on telling the same story;
individuals are friends with people who are mostly from the
same ethnic group. That was just on the surface network.
Getting into ego-networks or looking at how nodes are
connected in a network the result did not change. The
outcome of this exploration revealed that people befriend
mostly those who come from the same ethnic group. However,
this does not mean they are not connecting with people from
other ethnic groups. Gephi include different metrics for data
analysis [8, 9, 10]. These include: the average shortest path,
closeness centrality, and betweenness centrality between
nodes in the selected networks.
A. Average shortest path
The average shortest path is defined as the minimum
number of edge-hops required to traverse the network, starting
from node u and ending at node y [10, 13]. This disstance can
be explored to understand the distance between different
individuals from different ethnic groups in a network. In this
case, the researcher evaluated Indian, Black, White, and
Coloured networks where it was seen that if the network
belongs to a specific ethnic group, the average distance of the
network will be high compared with other friends in the same
network who are not from the same ethnic group.
With the 25 networks collected, using the Gephi average
shortest path test in an Indian network, the distance for
Indians, Whites, and Blacks to become friends with Indians
was evaluated. Indians were found to have a 1.69 average
shortest path to become friends with other Indians, whereas
the Black average shortest path was 1.2 and the White average
shortest path was 1.2, representing the lower average shortest
paths. The researcher did the same in a Black network, White
network, and Coloured network and reached the conclusion
that if, for example, a network belongs to an Indian, there will
be more chances that his or her friends in the network will be
Indians and they will have more chances to be connected to
each other than to other ethnic groups.
 Fig. 1: Visualization of race of students’ network
B. Betweenness centrality test
This metric indicates how frequently a node is found on a
shortest path between two nodes in the network in order for it
to connect with another node [9, 13]. Figure 2 refers: if the
number is high, it means in between two nodes there are more
chances that they become connected. The lower the number,
the lower the probability for a node to connects to another
node.
From Fig. 2, it is observed that in network one Black
individual have a very high betweenness centrality number.
This means that Blacks will have more chances to befriend
Blacks. Network two represent Indians in network one and
reveal that the chance for Blacks to befriend Indians is lower.
It has been mentioned that the majority of friends are black in
network one, even though there are other friends from other
ethnic groups, Blacks friend will have a greater tendency to
become friends with Black than with Indian students.
When applying the same logic and method in an Indian
network represented by network three and network four in
Figure 2, researchers observed that Indians have a very high
betweenness centrality number. This reveals that the chance
for Indians to befriend Blacks is lower. The true meaning of
network three and four is that the majority of friends are
Indians, even if there are other friends from other ethnic
groups, Indian will have a higher tendency to become friends
with Indians than with Blacks.
High centrality scores show that the node can reach others on
relatively short paths or that the node lies on significant zones
of shortest paths linking others. By performing more
betweenness centrality tests with different ego-networks, the
researcher was able to confirm the findings and come to the
same conclusion, namely an individual from a particular
ethnic group will have more chance of having friends from his
own ethnic group than from another ethnic group in South
African universities. It also shows that there are networks
where individuals have friends from their ethnic group only.
The result presented above confirms that people are
more likely to become friends with people who are like
themselves, thus reflecting the ‘birds of a feather flock
together’ [12].
Network 1- Blacks betweenness centrality in a Black
Facebook network
Network 2 - Indians betweenness centrality in a Black
Facebook network
Network 3 - Indians betweenness centrality in an Indian
Facebook network
Network 4- Blacks Betweenness centrality in an Indian
Facebook network
Fig. 2: Betweeness centrality Gephi visualisation
C. Closeness centrality test
This test indicates how long it will take someone to become
a friend in the network [11, 13]. For instance, if a person in
the network is from a Black ethnic group and while
subscribing to a social network site chose white as a profile
race, his/her chance of obtaining friendships from another
ethnic group that are white will be the same (the same applied
to variables such as picture, language etc). When a person
choses Black and the user on the other side is white, the
chances to become friend to a white person become lower.
The same logical exploration is true with users profile
language. Without considering other variables such as profile
picture, a White Facebook user with an IsiZulu profile can
become friend to a Black Facebook user who speaks IsiZulu.
If the White person keeps his profile language as English,
his/her chance of obtaining friendship from Black community
will be lower. Genetically no one can change the colour of his
skin but anyone can change or learn another person’s
language. This is to say that language is an equal parameter
that can allow an individual to become friends with anyone
from any ethnic group. In this case, it takes a long time for a
person who is from a different ethnic group to be accepted as a
friend on Facebook.
For as long as different ethnic groups have no racial
conflicts, their chances of becoming friends are approximately
the same. Some of them may have more friends than others,
depending on their individual relationships in the community.
It will take longer for a node that is not connected to any other
node in a network to connect with another.
V. CONCLUSION
Apparently, despite 20 years of post-Apartheid democracy
in South Africa, the impact and effects of enforced segregation
are still manifest, at least insofar as friendship relationships
between university students. Our analysis here shows that
students mostly befriend those on Facebook who are of the
same ethnic background as themselves. While we
acknowledge that there are many factors that contribute to
whether students become online friends, our analysis reveals
an unmistakable pattern. Parallel research on this aspect,
relating to the conversion of online friendships to real-world
friendships reveals a similar pattern. In this paper, Gephi was
used to visualize and analyse relationships amount students
belonging to different ethnic groups. Gephi is a useful visual
environment for the manipulation and analysis of both online
and offline social networks with quantifiable data. There are
opportunities for designers and developers to initiate
integrative graph software that manipulate both qualitative and
quantitative data in forms of graphs and other statistical
measures. This will revolutionize how research is currently
being performed. Obscured networks are easily observed and
understood. The understanding of individual and community
behavior in relation to demographic factors can easily be
analysed. It should be highlighted that the application is easy
to use.
 REFERENCES

[1] D. Taylor, Social Networking Usage in South Africa (2010). Available

at:
https://recruitingunblog.wordpress.com/2011/09/14/social-south-africa-
trusa/
[2] Socialbakers, South Africa Facebook Statistics (2013). Available at:
http://www.socialbakers.com/facebook-statistics/south-africa
[3] D. Boyd and E. Hargittai, Facebook privacy settings: Who cares? Peer-
Reviewed Journal on the Internet, 15(8) (2010). Available at:
http://firstmonday.org/article/view/3086/2589
[4] M. Bastian, S. Heymann and M. Jacomy,. Gephi: An open Source
Software for Exploring and Manipulating Networks. Paris, Association
for the Advancement of Artificial Intelligence (2009), pp. 361-362.
[5] Netvizz, Netvizz. (2012).
Available at:
https://tools.digitalmethods.net/netvizz/facebook/netvizz/?fb_source=sea
rch&ref=br_tf&code=AQAA0gJiMT70OKNhQdRjQBner1AlcAfu56C
H4jkWfdAzJ-
wXe6LzffrBbAZ7906zmGnZaS58dUnsA_y4qAuhHUOiDIF7ojveoP_s
wf53WsQ9RvuDrnDpyWy5TqIKFQns9EiFDosr83exczFLx-
U_j856JuR7jiPJiJp-
[6] A. Bruns, How Long is a Tweet? Mapping Dynamic Conversation
Netwoks on Twitter Using Gawk and Gephi. Information,
Communication & Society 15(9) (2012), pp. 1323-1351.
[7] S. Heymann, and B, Le Grand, Visual Analysis of Complex Networks
for Business Intelligence with Gephi. 17th International Conference on
Information Visualisation 1(4) (2013), pp. 307-312.
[8] Y. Shang, Distinct Clusterings and Characteristic Path Lengths in
Dynamic Small-World Networks with Identical Limit Degree
Distribution. Journal of Statistical Physics 149(3) (2012), pp. 505-518.
[9] A. Spitz and E.-Á. Horvát, Measuring Long-Term Impact Based on
Network Centrality: Unraveling Cinematic Citations. PLOS ONE 9
(2014), e108857.
[10] E. Otte, and R. Rousseau, Social network analysis: a powerful strategy,
also for the information sciences. Journal of Information Science, 28(6)
(2002), pp. 441-453.
[11] U. Brandes, A Faster Algorith for Betweenness Centrality. Journal of
Mathematicl Sociology, 25(2) (2001), pp. 163-177.
[12] D. Boyd, White Flight in Networked Publics? How Race and Class
Shaped American Teen Engagement with MySpace and Facebook. In: L.
Nakamura & P. Chow-White, eds. Race After the Internet. New-York:
Routledge Press (2007), pp. 203-222.
[13] S.P. Borgatti, Centrality and Network Flow. Social Networks 27 (2005),
pp. 55-71.
 An Efficient Key Management Scheme in
Hierarchical Wireless Sensor Networks
Xinyang Zhang
School of Electrical and Computer Engineering
RMIT
Melbourne VIC , Australia
Abstract—In this paper, a secure efficient hierarchical key
management scheme (SEHKM) for wireless sensor networks is
proposed. It satisfies several requirements of WSN. The design of
the scheme is motivated by the constraints of sensor networks
and high resources cost in traditional key management
techniques. The key management scheme supports the
establishment of three types of keys to encrypt messages sent
among sensor nodes. They are network key, group key and
pairwise key. Network key encrypts broadcast messages and
authenticates new nodes. Group key is shared by all nodes in the
same cluster. Pairwise key is shared with a specific pair of nodes.
In hierarchical WSN, cluster heads are very important due to the
structure of the network. In order to improve security strength
and reduce resource cost and risk after cluster head compromise,
an assistant node is introduced in the scheme. SHEKM includes
key establishment, key transportation, mechanisms after node
operation and dynamical freshness of keys. The evaluation of
SEHKM is presented as well as the comparison with some
existing key management schemes. The performance analysis and
simulation show that SEHKM is very efficient in computation,
communication, and storage while the security strength is high.
Keywords—hierarchical wireless sensor networks, security, key
management, assistant node, key update
I. INTRODUCTION
Wireless sensor network (WSN) is a multi-hop ad-hoc
network formed by a large number of low-cost micro-sensor
nodes which communicate through radio channel [1]. It is
widely used in many areas in modern society, such as military,
agriculture, environment and health care. Because information
protection is essential, security for WSN is important and it
gets a lot of attentions from the public. As one of the important
issues, key management includes key establishment, key
transportation, key update and mechanism after node operation.
Based on current researches in this area, some existing key
management schemes are assessed under different
requirements and constraints of WSN.
In this paper, a key management scheme for hierarchical
WSN (HWSN) using different keys to achieve security at each
level is proposed. There are three types of keys: network key,
group key and pairwise key. Network key is shared by all
nodes in network and it is utilized for securing broadcast
message and authentication. Group key is shared by the nodes
in the same cluster. Pairwise key is shared by two nodes. The
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Dr. Jidong Wang
School of Electrical and Computer Engineering
RMIT
Melbourne VIC, Australia
scheme also includes key updating and mechanism after node
operation.
II. RELATED WORK
A. Constraints of WSN
In a WSN, sensor nodes are independent tiny devices. Each
of them has individual battery and hardware. This leads to
constraints [2, 3]. The first one is energy. Once a node is
deployed in a network, the battery is not rechargeable in many
applications. The battery has to serve the lifetime of a node for
all functions. As a result, reasonable energy consumption is
very important for a sensor node and it affects the overall
performance of the whole network. Then the physical size and
prize of the node device decide it could not have large and
expensive chips. So the hardware of memory and processor in
a node is limited. However, a sensor node built on an
embedded system has three functions: sensor interface, data
processing and network interface [4]. A sensor node has to
carry out all these functions with limited hardware. The
processor also requires energy too.
In the other hand, all nodes in a WSN communicate with
each other by radio channel which is open and can be accessed
by anyone in the same range. This makes a great challenge for
security. In addition, WSN can be deployed in different
environment depend on different applications. The entire
network is affected by this environment condition. Moreover,
due to unreliable channel and severe environment, there are
much more packet loss and fault in WSN than traditional
networks. Table 1 summarizes constraints of WSNs.
TABLE I. CONSTRAINTS OF WSN
Constraints Details
Severe environment, limited resources of memory,
Physical
energy and computation
Unreliable channels and limited bandwidth, collision
Communication
and latency.
B. Requirements of WSN Key Management [5]
Due to the constraints in WSNs, the key management
scheme employed in a WSN should provide the required
security strength and works efficiently under the limitation of
power, memory, computation and communication in WSN.
According to Carman [2] and Sastry [6], key management
scheme should satisfy several requirements based on security
and function of the WSN. Table 2 shows the requirements.
TABLE II. REQUIREMENTS OF WSN KEY MANAGEMENT
Requirements Details
Confidentiality, authentication, data integrity,
Security
freshness and robustness
Reasonable communication, computation, storage
Efficiency
and power consumption
Operation Flexibility, scalability and accessibility
C. Network Structure
WSNs can be classified as hierarchical networks and flat
networks by the way that sensor nodes formed network [1]. In
an HWSN, all nodes are divided into several clusters and each
cluster has a head [5]. A cluster head leads a group of general
sensor nodes and all cluster heads are led by the head of cluster
head. Every node belongs to a cluster and communicates with a
base station (BS) through cluster head. In a flat network, all
nodes have the same capability. They communicate with
neighbors and transmit data to BS one by one. Comparing with
flat WSN, HWSN has many advantages [7]. First, in a
hierarchical network, cluster heads and BS manage the network.
Normal nodes can only wake up when they are needed for data
transmission or collection. Hence, the energy consumption
would be reduced for this. Then, cluster head and cluster
members exchange information in a cluster, it helps cluster
head concludes the local information. Lastly, there is less
competition between nodes for communication channel as
cluster head transmits most data, more nodes can be deployed
in the network. The hierarchical WSN has better scalability.
Compare with key management schemes for flat WSN [8, 9],
schemes for HWSN [10, 11] are more efficient. However, a
cluster head manages the keys of all nodes in the same cluster.
Hence, the cluster will face to serious problems if the head is
compromised or damaged [5].
D. Existing Key Management Schemes [5]
Based on TDGH [12], Panja [13] proposed a key
management mechanism which is suitable for HWSN. The
structure is a cluster head leads a group of general sensor nodes
and all cluster heads are led by the head of cluster head. All
sensor nodes pre-install a symmetric key for key transport
encryption. Every node has a partial key to generate group
keys. The leaf nodes generate random numbers to calculate
their partial keys. The partial keys of parent nodes are
calculated by the partial keys of the leaf nodes. There are two
types of group keys: intra-cluster key and inter-cluster key. The
intra-cluster group key is used for encryption/decryption of
messages inside a sensor network group, whereas the inter-
cluster group key is used for groups of cluster heads. A cluster
head computes the intra-cluster group key by using all the
partial keys in the group as arguments. The inter-cluster group
key is generated in the same approach by using the partial keys
of cluster heads. This scheme is simple and easy to implement.
The energy consumption is reasonable as Panja analyzed. It
also satisfies the requirement scalability, accessibility,
authentication and flexibility. But the scheme has weaknesses
too. In Panja scheme, the group key generation with blind
factor will cost more communication resource. This scheme
implements one key in a cluster, so if a node is compromised,
the group key may lost and the whole cluster will be in danger.
The security strength is not high enough for some applications.
LEAP (Localized Encryption and Authentication Protocol)
[10]is a key management protocol with the operation of four
types of keys: individual key, pair wise key, cluster key and
group key. The individual key is for each node shared with BS.
The pair wise key shared with another node. The cluster key
shared with all nodes within the cluster and the group key is for
BS communicates with all other sensor nodes in the network.
These four keys are generated by a pre-distributed key called
initial key. Firstly, the individual key is calculated by a
function with the ID of the node. Secondly, nodes broadcast
their IDs in the neighbor discovering and the receiver uses a
function with initial key to establish the pair wise key shared
with the neighbor. Then all nodes delete the initial key after the
pair wise key generation phase. Next is cluster key distribution.
Cluster head secures the key with the pair wise key and
broadcasts it in the cluster. Lastly, the BS broadcasts the group
key. LEAP uses μTimed Efficient Streaming Loss-tolerant
Authentication Protocol (μTESLA) [14] to authenticate the
broadcast of BS which make sure that packet with group key is
just sent by BS. LEAP has many advantages. The security
strength of it is high and it meets the requirement of
accessibility. However LEAP failed to satisfy the flexibility
requirement because no new node can be added into the
network after deletion of initial key. In addition, if initial key is
disclosed, the adversary will be able to establish pairwise key
with any node in the network.
III. SEHKM: A SECURITY EFFICIENT HIERARCHICAL
KEY MANAGEMENT SCHEME FOR WSN
SEHKM provides three types of keying mechanisms to
secure WSNs. An overview of keys is introduced at first. Then
the assistant node which increases security strength of the
scheme is presented. Next are the establishments of each type
of keys, node operation and key update. At last, node
compromise is described.
A. Overview
With the consideration of advantages of hierarchical
structure and existing schemes, SEHKM is proposed. It
assumes the adversary only can eavesdrop during in
initialization phase and BS is always secure.
SEHKM is a key management scheme for HWSNs and
provides three types of keys to encrypt message. A network
key shared by all nodes in the network, a group key shared by a
group of nodes and a pairwise key shared by a specific pair of
nodes. Each node pre-distributes a network key and a number
IN.
 Network key this is a key shared by all nodes in the
network. It is pre-installed and helps with the initialization
of network. After initialization phase, the key will be
updated by BS. It works for BS to encrypt broadcast
messages and authentication.
 Group key a group key is a key shared by all nodes in the
same group. In lower level of network, a cluster can be a
 group. Cluster head is group leader. In higher level, BS and
all cluster head is a group and BS is the group leader.
Figure 1 shows the groups in different levels. Group key
secures messages broadcast in a group.
 Pairwise key A pairwise key is a shared by two nodes. In
this scheme, not any pair of nodes has a pairwise key. BS
has pairwise keys with all cluster heads, BS shares a key
with assistant node in each cluster, a cluster head has
pairwise keys with all other nodes in the cluster and
assistant node has pairwise keys that shared with all
normal nodes in the cluster. During the key establishment,
a disposable pairwise key is used to secure the phase and it
will be erased later. Pairwise key secures unicast message
and can be used for authentication.
In the following sections, assistant node, key establishment,
transportation and updating will be presented. The keying
mechanisms during node operation such as node addition,
revocation and replacement are described in the next.
B. Assistant Node
Hierarchical structure has many advantages compare with
flat structure. In a HWSN, cluster heads are the most important
nodes in their cluster. A cluster head organizes other nodes in
the cluster, processes data aggregation and manages security
keys. As a cluster head plays such important roles, the
compromise of a cluster head is fatal in the cluster and it
affects all the nodes in the cluster. In order to combat the
security problem caused by cluster head compromise, an
assistant node is proposed.
In a cluster, besides the cluster head, another management
node, named assistant node is arranged. An assistant node is
randomly chosen by the cluster head at cluster setup phase. It
shares pairwise keys with each member nodes and the cluster
head. The pairwise key establishment will be covered in next
section. All member nodes in the cluster are candidates to be an
assistant node. Whenever a cluster head cluster head is
compromised, replaced or removed, the assistant node will kick
in and will take the management role in the cluster. Figure 1
shows assistant nodes in a two-layer wireless sensor network.
Fig. 1. Assistant nodes in two-layer wireless sensor network.
An assistant node has IDs of all nodes in the same cluster as
well as the information of gateway nodes. The gateway nodes
enable the assistant node to find routes to the BS above its own
cluster head in the hierarchical structure. It will find routes to
the BS to get command from it when the cluster head is
compromised or damaged.
The assistant node finds a route to the BS in 6 steps:
1) Initialization
 Set Searching Depth to 0;
 Create a route request packet with following parameters
i. Cluster head ID ( source CH ID)
ii. Routing Path : Assistant Node ID
iii. Searching Depth
2) Assistant node passes the request packet to a gateway node
never encountered before.
3) The gateway node adds its own ID to the routing path and
sends the request to a gateway node in a neighboring
cluster.
4) Then this neighboring gateway node adds its own ID to
the Routing Path and passes the request packet to its
cluster head.
5) Each time a cluster head receives a request, it will
increase Searching Depth by 1. Then it will check if the
source CH ID is its member. If it is, then the route is
found. Otherwise it will check if Search Depth is larger
than 3. If it is then go back to step 1. Otherwise it adds its
ID to the Routing Path and passes the request to its own
head, then goes back to step5.
6) End.
C. Establishment of Disposable Pairwise Keys
In the network clustering, each member node u joins the
cluster of the head node i and generates the disposable pairwise
key 𝐷𝑃𝐾𝑢 ,𝑖 shared by node u and cluster head i using IN and
node u’s ID by function f.
𝐷𝑃𝐾𝑢 ,𝑖 = 𝑓𝐼𝑁 𝐼𝐷𝑢  
As a cluster head knows all IDs of its member nodes, it is
able to calculate the disposable pairwise key with each member
node. After network clustering and generation of disposable
pairwise key with head node, each node erases IN and only
remains disposable pairwise key 𝐷𝑃𝐾𝑢 ,𝑖 . Cluster head may
transport this key to another node if it is needs. So this key may
be shared by more than two nodes.
D. Establishment of Group Keys
Group key is shared by a group of nodes. The first group
key is a random number generated by the group leader and it is
sent to each member node encrypted by disposable pairwise
key.
After initialization of all keys, group key needs to be update
to keep secure. The generation of following group keys is
different from the first group key. The algorithm will be
described in the key update section.
E. Establishment of Network Key
All nodes in the network shared one Network key. This key
is pre-installed and secures messages in initialization phase.
After group key establishment, BS generates a random number
as the new network key. BS encrypts the key by group key and assistant node. After establishments of all pairwise keys,
broadcasts the key to all cluster heads, then each cluster head disposable pairwise keys should be erased.
broadcasts the key encrypted by its group key to its own
member nodes. Network key is always a random number G. Node Operation
generated by BS In a WSN, the operations of node addition, revocation and
replacement are needed. A list of notations is defined below:
F. Establishment of Pairwise Keys
In SEHKM scheme, there are three types of pairwise keys  BS : base station.
in the scheme which distinguished by their generation and  CH : cluster head.
storage. The first type is group associated pairwise keys which  MN : member node.
include the key shared by group leaders with their member  NK : network key
nodes and the pairwise keys shared by BS with cluster heads.  GK: group key.
The second type of pairwise keys is assistant node associated.  PK : pairwise key.
This type includes the key shared by assistant node with  DPK: disposable pairwise key.
member nodes in the same cluster and the key shared by  IK : temporary key.
assistant node with BS. The last type is disposable pairwise key 1) Node addition: The new node pre-installs the current
and it will be erased after establishment of all keys. The network key and IN to authenticate and join a cluster. Then the
establishment of disposable pairwise key is introduced above. cluster head generates a random number as pairwise key and
1) Group associated pairwise key establishment: The sends it, network key and group key to new node. The process
generation of group associated pairwise key uses Diffie-
Hellman algorithm. For example, in a group, the leader i
′
generates a random seed 𝑔1𝑖 , then calculates 𝑔1𝑖 as:
′
 𝑔1𝑖 = 𝛼 𝑔 1𝑖 𝑚𝑜𝑑 𝜌 
′
The group leader broadcasts 𝑔1𝑖 encrypted by group key in
′
the group. Then a member node u generates g 2𝑢 and sends 𝑔2𝑢
′
to group leader i. 𝑔2𝑢 is encrypted by the disposable pairwise
′
key. Therefore group leader gets 𝑔2𝑢 . Then the pairwise key
between member node u and group leader i is:

 𝑃𝐾𝑢 ,𝑖 = 𝑓(𝑔1𝑖 , 𝑔2𝑢 )  of node addition is presented in figure 2.

Fig. 2. Node addition.
Where the function f is:
2) Node revocation: The keys that the removed node has
𝑓 𝑘1 , 𝑘2 = 𝛼 𝑘 1 𝑘 2 𝑚𝑜𝑑 𝜌 should all refresh. Node revocation includes two steps: first is
to remove the node from network, second is to update all keys
In this function, 𝜌 is a prime number and 𝛼 is the primitive that the removed node has. During the time period after node
root of 𝜌 , 𝑘1 , 𝑘2 < 𝜌.
remove and before key update, node addition is not allowed.
2) Assistant node associated pairwise key establishment:
An assistant node is selected beside the cluster head in each
cluster. The assistant node associated pairwise key is
generated by a random function. 𝑃𝐾𝑢 ,𝑣 is pairwise key shared
by node u and v. m is a number only known by generator node
v. The function is:

𝑃𝐾𝑢 ,𝑣 = 𝑓𝑚 (𝐼𝐷𝑢 )

Then generator node v sends 𝑃𝐾𝑢 ,𝑣 to node u. Node u

stores 𝑃𝐾𝑢 ,𝑣 while node v only stores node u’s ID. It will
generate 𝑃𝐾𝑢 ,𝑣 if it is needed. The node revocation process is shown in figure 3.
Fig. 3. Node addition.

For a key shared by the assistant node and a member node,

the assistant node plays the role of generator. However BS is 3) Node replacement: This step acts as first node
the generator of the pairwise key shared between BS and an revocation and then node addition. For replacement of cluster
head, the new head joins the network first, then BS generates
and sends the intermediate key to the new head and all normal
nodes join the cluster of new head.
H. Key Update
For the purpose of efficiency, in SEHKM, only the network
key and group keys will be updated periodically. Network key
is generated by the BS while group key is established by
collaborating of all nodes in the cluster. The network key
should be sent to each cluster head with the command of group
key update. The mechanism of the group key generation in the
upper layer is same as in the cluster level. Therefore the
following discussion on the proposed key update algorithm
focuses on the group key update on the cluster level. The main
feature of the key update is that member nodes send small
random numbers (SRNs) along with normal traffic data
together to reduce transmission energy. In current cryptography
algorithm, some of them using block cipher technique which is
treats a block of plaintext as a whole to be encrypted [15]. If
the message is not long enough as the block, 0 is used to fill the
empty parts. So these empty parts can be used to carry the
SRNs.
The group key update includes four steps: (1) member
nodes send SRNs to the cluster head; (2) the cluster head stores
and transforms SRNs; (3) the group key generation; (4) the
group key distribution.
1) In the normal network operation, a member node
attaches an r-byte SRN with its normal application data
regularly. Not all packets have a SRN attached. The frequency
of SRN attachment at each node is fixed.
2) A cluster head stores these SRNs collected and
assembles a number 𝑁1 of K bytes. K is the key size where
K>r. If only r-i bytes of last SRN are used, then the remaining
i bytes will be used for assembling 𝑁2 . When 𝑁2 is assembled,
the function f is used to generate a new number and 𝑁1 is
updated with this new number. Only the very first 𝑁1 is
assembled from the collected SRNs. The following 𝑁1 is from
the recursive function f (4):
𝑁1 = 𝑓 𝑁1 , 𝑁2 
The following SRNs are used to assemble 𝑁2 only. SRNs
are from different nodes. The order of them is not important as
long as they are not from one node only. If a node sends SRNs
more than [K/n] times continually, the cluster head should
erase [K/2] bytes and generates [K/2] bytes of random number
to re-assemble 𝑁1 or 𝑁2 . The process keeps going till the
cluster head receives a command to update the group key. Then
all the following SRNs are discarded.
3) The third step is group key generation. When the cluster
head needs to update the group key, if the received SRNs are
not enough to form a new 𝑁2 , then the cluster head generates a
random number to pad the last 𝑁2 . The group key is generated
by the function (4) using the last 𝑁1 and 𝑁2 :
𝐺𝐾 = 𝑓 𝑁1 , 𝑁2 
If a node compromise or damage is detected, then the 𝑁1 ,
𝑁2 generated from SRNs should be discarded as it is hard to
figure out which parts of them are from the compromised
nodes. Then the cluster head will uses the new network key as
𝑁1 and generates 𝑁2 by itself to calculate a new group key.
4) The cluster head distributes the new group key and new
network key together. It is mentioned before that the new
network key has been distributed to the cluster heads with
group key update command. The key distribution can be
encrypted by previous group keys. However, if node
compromise is detected, the distribution should be carried out
securely by pairwise keys.
I. Node Compromise
In SEHKM, each node keeps several keys. Once a node is
compromised, the keys that the node has should all be updated.
First, the compromised node is marked and removed from
network. Then the cluster with compromised node updates
group key. The cluster head encrypts the new group key with
pairwise key and sends it to each member node besides
compromised node. The next step is network key distribution.
BS encrypts the key with group key and broadcast it to cluster
heads. In the cluster which existing compromised node, cluster
head secures network key with new group key and broadcast it.
Other cluster heads broadcast new network key encrypted by
original group key. At last, the pairwise keys shared by
compromised node and other nodes are deleted.
IV. EVALUATION OF SEH
The evaluation of proposed scheme includes performance
evaluation and security analysis
A. Performance Evaluation
In SEHKM, network key and group keys are updated
regularly, therefore only the cost for update the two types of
keys are considered. The update of network key and group key
are same: first the BS generates and broadcasts the network key
and the group key of its cluster, then each cluster head
calculates its group key and broadcasts the group key with
network key in the cluster. Thus the computation and
communication costs are analyzed in the group key update
process.
It assumes the network size is N, m is number of node in a
cluster, d is network degree and all clusters have same number
of nodes in HWSN. It assumes one encryption or a decryption
of the same length of key consumes 1 computation energy unit
E; A transmission of the key consumes 1 transmission energy
unit T; One key needs 1 storage unit S. In the following
sections, the cost of group key update in a two-layer network is
analyzed.
1) Computation cost: Computation cost includes
encryption and decryption of keys. First, BS encrypts the key
and all cluster heads decrypt the key. Then each cluster head
encrypts the key and member nodes decrypt the key. The cost
on cryptography of group key distribution is [(N-1)/m +N] (E).
 2) Communication cost: The communication cost includes higher than 32bytes. However in AES the cost is constant if
the energy consumption on key transmission. First BS data length is less than 128bytes.
broadcasts the key to all cluster heads, then each cluster head Figure 4 shows the energy cost in Panja and SEH when
broadcasts the key in the cluster. The group key broadcast network size is constant and network degree is increasing.
costs [1+ (N-1)/m] (T) AES always costs more energy than DES. SEHKM consumes
3) Storage cost: BS, cluster heads, assistant nodes and less energy than Panja. If network degree increases, the both
normal nodes have different number of keys. The storage cost schemes cost more energy.
of each node is shown in table 3.
a) Base station: It stores 1 network key, 1 group key
and pairwise keys shared with cluster heads. The cost is [2+
(N-1)/m](S).
b) Cluster head: Each cluster head stores 1 network key,
2 group keys shared in cluster of BS and its own. There are m-
1 pairwise key shared with member nodes and 1 key shared
with BS. The cost is [m+3](S).
c) Assistant node: It stores 1 network key, 1 group key
and 2 pairwise keys shared with cluster head and BS. The cost
is 4(S).
d) Normal node: Each normal node stores 1 network
key, 1 group key and 2 pairwise keys shared with assistant
node and cluster head. The cost is 4(S). Fig. 4. Encryption energy cost in Panja and SHEKM

TABLE III. STORAGE COST (S) IN DIFFERENT NODES Table 5 presents the time cost of group key update in Panja
and SEHKM. It shows Panja cost much more time than
Node Network Group Pairwise Key Total(S)
Key (S) Key (S) (S) SEHKM. When network degree increases, time consume for
BS 1 1 (N-1)/m 2+ (N-1)/m group key update in Panja increases while in SEHKM it is
Cluster 1 2 m m+3 constant. Time for network key and group key update together
head in SEHKM grows while network degree rises.
Assistant 1 1 2 4
node
TABLE V. TIME COST (SECOND) OF PANJA AND SEH
Member 1 1 2 4
node Networ Netwo- Number of Group Key NK and
4) Simulation: According to above analysis result, the k Size rk Nodes in A Panja SEHKM GK in
simulation result of energy cost using DES for encryption in Degree Cluster SEHKM
two-level HWSN on a popular sensor node, the MICA from 22765 3 28 34.55 0.04 0.16
CrossBow is presented. The key size is 32bytes and unit is 22621 4 12 75.13 0.04 0.20
joule. The results are shown in table 4. 19608 5 7 94.01 0.04 0.24
19531 6 5 105.53 0.04 0.28
TABLE IV. ENERGY COST (JOULE) OF GROUP KEY UPDATE IN 21845 7 4 119.64 0.04 0.31
SEHKM, LEAP AND PANJA
Network Size LEAP Panja SEH
B. Security and Operation Analysis
100 0.33 0.27 0.03
200 0.6 0.54 0.05 In WSN, compromise detection is one of most critical
300 1.00 0.81 0.07 security requirements and almost all key management schemes
400 1.34 1.08 0.10 rely on it. Under detected compromise, SEHKM provides
500 1.68 1.36 0.12 nodes revocation mechanism to delete compromised node. This
600 2.01 1.63 0.14
mechanism avoids new key disclosure to the compromised
700 2.35 1.90 0.17
800 2.68 2.17 0.20
node and improves security strength of sensor network.
900 3.02 2.45 0.22 Assistant node improves security level too. It is able to
1000 3.69 2.99 0.27
communicate with BS safely if cluster head is compromised
The result shows SEHKM always costs less energy than because it has pairwise key with BS. As it describes in section
Panja and LEAP. B and G of chapter 3, assistant node gets command from BS
But in HWSN, network degree can be more than two and let other nodes which are not compromised join other
levels. It is simulated that the energy cost in SEHKM and clusters. This mechanism saves many normal nodes while in
Panja under different network degrees. other hierarchical schemes they are not able to be saved. The
DES and AES are different encryption algorithms. AES is method reduces resource cost and improves security.
stronger than DES but cost more energy and time. In DES,
The proposed scheme satisfies security requirement of
energy and time cost of encryption for 24bytes of data both are
WSN. First, data confidentiality can be achieved by encryption.
Three types of keys provide protection on different
communication types. Then, data authentication can be
provided by pairwise key or group key using a symmetric
message authentication code. Data integrity is achieved
through data authentication.
In SEHKM, each node in the network is able to exchange
information with any other nodes. This is accessibility. On the
other hand, this scheme supports node addition, node
revocation and node replacement services. Thus flexibility is
provided. Beside, scalability of network is provided due to the
hierarchical structure. The network is able to have large scale
of nodes. Therefore, the operation requirement of WSNs is
satisfied in SEHKM.
V. CONCLUSION
In this paper, a secure and efficient scheme for hierarchical
wireless sensor network (SEHKM) is proposed. It has the
following features:
 SEHKM supports establishment, transportation and
updating of three types of keys. A network key shared by
all nodes in the network. A group key shared by a group of
nodes and a pairwise key shared by a specific pair of nodes.
 SEHKM includes assistant node to help to organize
member nodes in the same cluster when cluster head is
removed, compromised or replaced. The group key update
algorithm is safe and efficient.
 The resource cost of SEHKM is reasonable. The key update
procedure is efficient and the memory cost for keys storage
is small.
REFERENCES
[1] J. Zheng and A. Jamalipour, Wireless sensor networks: a networking
perspective: John Wiley & Sons, 2009.
[2] D. W.Carman, P. S. Kruus, and B. J. Matt, Constraints and Approaches
for Distributed Sensor Security: NAI Labs Tech. Rep.#00-010, 2000.
[3] S. Gajjar, S. Pradhan, and K. Dasgupta, "Wireless sensor network:
Application led research perspective," in Recent Advances in Intelligent
Computational Systems (RAICS), 2011 IEEE, 2011, pp. 025-030.
[4] I. F. Akyildiz and M. C. Vuran, Wireless Sensor Networks. Singapore:
Markono Print Media Pte Ltd, 2010.
[5] X. Zhang and J. Wang, "Key Management in Wireless Sensor Networks:
Development and Challenges," in Applied Mechanics and Materials,
2014, pp. 654-660.
[6] N. Sastry and D. Wagner, "Security considerations for IEEE 802.15. 4
networks," in Proceedings of the 3rd ACM workshop on Wireless
security, 2004, pp. 32-42.
[7] R. Rajagopalan and P. K. Varshney, "Data aggregation techniques in
sensor networks: A survey," 2006.
[8] L. Eschenauer and V. D. Gligor, "A key-management scheme for
distributed sensor networks," presented at the Proceedings of the 9th
ACM conference on Computer and communications security,
Washington, DC, USA, 2002.
[9] W. Du, J. Deng, Y. S. Han, P. K. Varshney, J. Katz, and A. Khalili, "A
pairwise key predistribution scheme for wireless sensor networks," ACM
Trans. Inf. Syst. Secur., vol. 8, pp. 228-258, 2005.
[10] S. Zhu, S. Setia, and S. Jajodia, "LEAP+: Efficient security mechanisms
for large-scale distributed sensor networks," ACM Trans. Sen. Netw.,
vol. 2, pp. 500-528, 2006.
[11] B. Maala, H. Bettahar, and A. Bouabdallah, "TLA: A Tow Level
Architecture for Key Management in Wireless Sensor Networks," in
Sensor Technologies and Applications, 2008. SENSORCOMM '08.
Second International Conference on, 2008, pp. 639-644.
[12] M. Steiner, G. Tsudik, and M. Waidner, "Key agreement in dynamic
peer groups," Parallel and Distributed Systems, IEEE Transactions on,
vol. 11, pp. 769-780, 2000.
[13] B. Panja, S. K. Madria, and B. Bhargava, "Energy and communication
efficient group key management protocol for hierarchical sensor
networks," in Sensor Networks, Ubiquitous, and Trustworthy
Computing, 2006. IEEE International Conference on, 2006, p. 8 pp.
[14] A. Perrig, R. Szewczyk, J. D. Tygar, V. Wen, and D. E. Culler, "SPINS:
Security protocols for sensor networks," Wireless networks, vol. 8, pp.
521-534, 2002.
[15] S. William and W. Stallings, Cryptography and Network Security, 4/E:
Pearson Education India, 2006.
[16] G.Yang, W. Chen and X. Cao. “Wireless sensor network security,”
Beijing Science Press. 2010
[17] R. Watro, D. Kong, S. Cuti, C. Gardiner, C. Lynn, and P. Kruus,
“ TinyPK: Securing sensor networks with public key technology” , in
Proceedings of the 2nd ACM Workshop on Security of Ad Hoc and
Sensor Networks (SASN ’ 04) , Washington, DC, Oct. 2004 , pp. 59 –
64 .
[18] J. Zheng and A. Jamalipour, Wireless Sensor Networks: A Networking
Perspective: A John & Sons, Inc, and IEEEE, 2009
[19] C. Chih-Chun, S. Muftic, and D. J. Nagel, "Measurement of Energy
Costs of Security in Wireless Sensor Nodes," in Computer
Communications and Networks, 2007. ICCCN 2007. Proceedings of
16th International Conference on, 2007, pp. 95-102.
[20] N. A. Pantazis and D. D. Vergados, "A survey on power control
issues in wireless sensor networks," Communications Surveys &
Tutorials, IEEE, vol. 9, pp. 86-107, 2007.
[21] G. de Meulenaer, F. Gosset, O. X. Standaert, and O. Pereira, "On the
Energy Cost of Communication and Cryptography in Wireless Sensor
Networks," in Networking and Communications, 2008. WIMOB '08.
IEEE International Conference on Wireless and Mobile Computing,
2008, pp. 580-585.
[22] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci,
"Wireless sensor networks: a survey," Computer Networks, vol. 38, pp.
393-422, 3/15/ 2002.
[23] A. S. Wander, N. Gura, H. Eberle, V. Gupta, and S. C. Shantz,
"Energy analysis of public-key cryptography for wireless sensor
networks," in Pervasive Computing and Communications, 2005.
PerCom 2005. Third IEEE International Conference on, 2005, pp. 324-
328.
 Experiential Analysis of the Effectiveness of Buck
and Boost DC-DC Converters in a Basic off Grid PV
System
Asowata Osamede, Ruaan Schoeman and HCvZ Pienaar
Dept of Electronic,
Vaal University of Technology,
Vanderbijlpark, SOUTH AFRICA-1900

Abstract— The major renewable energy potential on earth is DC converters [7]. Fig. 1 reveals a basic off-grid PV system
provided from solar radiation and solar photovoltaics (PV) are connected to a DC-DC converter via a Data logging interface
considered a promising technological solution to support the circuit (DLIC). The geographical position of a country
global transformation to a low-carbon economy and reduce strongly affects the operational application of solar energy
dependence on fossil fuels. The aim of this paper is to carry out
[8].This is attributed to the fact that the intensity of solar
an experiential analysis of the effectiveness of power conditioning
devices with a focus on the Buck and Boost DC-DC converters energy is unevenly distributed on the surface of the earth.
(12 V, 24 V and 48 V) in a basic off grid PV system with a fixed
load profile. Improper choice of power conditioning device in a
basic off-grid PV system can attribute to power loss, hence the Solar Panel Power DC-DC Power Power
Converter
need for a right choice of power conditioning device to be coupled
with the system of the essence. This would assist in harnessing
more of the available solar energy. The practical setup consists of
Current Sensor
Voltage Sensor

a PV panel that is set to an orientation angle of 0o N, with a

corresponding tilt angle of 36o, 26o and 16o. Preliminary results,

Control via DLIC

which include data analysis showing the power loss in the system
and efficiency, indicate that the 12V DC-DC converter coupled
with the load profile had the highest efficiency for a latitude of
26o S throughout the year.
MPPT Controller Power
Keywords— Poly-crystalline PV panels, DC-DC converters, tilt
and orientation angles, direct solar radiation, load profile.

I. INTRODUCTION Figure 1: Basic off-grid PV system [9]

In recent years, the world economic growth and population
increase has necessitated the need for more energy [1]. Energy
is essentially important for the socio-economic development
of developing countries [2]. The increase in this demand is the
major increase in carbon monoxide (CO) emission which is
attributed to the combustion of fossil fuel [3] as predicted by
the United States Department of Energy. The need to reduce
this CO emission is of essence due to its hazardous effect.
Renewable energy sources are inexhaustible, intrinsically
clean, replenishing themselves and allowing energy harvesting
at a rate decided by nature [4]. This makes renewable energy
sources a key interest in present times. Solar energy which is a
type of renewable energy is generated from the sun and can be
converted into electrical energy and heat energy through its
solar cells [4]. Solar energy can be converted to electrical
energy by means of two methods: the first one is a direct
method with PV systems and the second is an indirect method
called solar thermal power generation [5]. PV modules
provide direct conversion of solar radiation into electrical
energy and play an important role in the field of renewable
energies [6] at standard connections of 12 V, 24V or 48 V DC-
The output power from a PV panel is determined by
numerous limiting factors, which thus result to a loss in the
power disseminated to the load profile. These factors include
the temperature [10] incident on the PV panel, choice of power
conditioning device coupled with the PV system as well as tilt
and orientation angles. The purpose of this paper is therefore
aimed at evaluating the effectiveness of Buck and Boost DC-
DC converters which is a type of power conditioning device
hence ascertain the most effective with certain load profiles.
This would further assist research in PV systems to reduce
cost, decrease power loss and increase the overall efficiency of
the system. In the Vaal Triangle, South Africa where this
research is carried out, it’s perceived that the 36o tilt angle is
recommended for optimum output power, hence preference is
given to the 36o tilt angle. [11]. In this paper the maximum and
minimum temperatures of the geographical region were this
research was carried out are presented in graph form. A brief
overview and vivid comparison of the Buck and Boost DC-DC
converters is done. This is followed closely with the practical
set up. The average on-time per week is used to ascertain the

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

efficiency of different DC-DC converters. This contrast is done
to evaluate the most effective of the DC-DC converters. The
results are presented in a number of graphs and are used to A V ER A G E T EM P ER A T U R E ( A P R I L - J U LY )

formulate the conclusions. Minimum temperature °C Maximum temperature °C

II. LITERATURE REVIEW

38

38
Among the factors that unfavourably influence the output

28
25
power from a PV system, temperature effect is a major factor.
[10] as well as the appropriate choice of power conditioning
devices in PV systems. A rise in surface temperature of a PV Jan 4 2
module negatively affects its output power. It is important to Feb 5 3

4
3
note that the direct beam radiation received by a PV panel is

1
Mar 6 4
directly related to its power conversion [11]. In line with the AAp
PR 3Y
MA 7 JUN JUL

-3
above the average minimum and maximum day temperature in May 6 2
degrees for the months of April to July 2012 is represented in
Table 1 and Fig. 2. This is done to give a concise description of Figure 2: Average minimum and maximum temperature (April – July,
the temperature that was recorded during the sampling period. 2012).
The average maximum and minimum day temperatures are
derived using MS EXCEL from the day temperature data
downloaded from the DAQPRO data logger. Further analysis
which involves a bar chart proves that the month of April and
May had the highest day temperature with the Months of July
having the lowest recorded temperature. The temperature on
the surface of a PV panel basically affects the PV output power
which is supplied to the DC-DC converter (Buck or Boost).
Furtherance to the graph presented in Fig. 2 a Pearson
correlation is done between the PV voltages against the
average temperature per week and is presented in Fig. 3.
Correlation is a statistical technique that can show whether and
how pairs of variables or data are related [12]. Fig. 3 clearly
shows the correlation of the PV voltage with an increase in day
temperature. It is advised that PV system should be operated at
the maximum power point in order to reduce the overall cost of
the system [13]. So power regulations circuits such as DC-DC Figure 3: Graph of the correlation of the PV voltage against the day
converters should be incorporated in the set-up of a PV system. temperature for the month of June, 2012.
Power regulation circuits used in PV systems include
Maximum Power Point trackers, DC-DC converters and solar Diffused radiation is basically the radiation from the
chargers, but this article is limited to the empirical test of the surroundings of the PV panel which is not sufficient to drive
use of DC-DC converters. The correlation is further used to the load successfully. The third portion of the graph (82% to
determine the effect that different radiation conditions exert on 100%) indicates the effect of direct beam radiation on the DC-
the output of a DC-DC converter. Fig. 4 shows an example of DC converter voltage. The point at which power is delivered
the effect that different radiation conditions exert on the output
to the load is indicated by a triangle and represents the start of
of a DC-DC converter The first portion of the graph (0% to
the conversion-time as a result of the direct beam radiation
56%) indicates no output of the DC-DC converter as no solar
radiation exists (attributed mainly to night time). The second (100% - 81% equaling 19% for this given example). The
portion of the graph (56% to 81%) is where the DC-DC triangle also indicates a voltage of about 12 V which coincides
converter produces an output voltage which is not its stated with the output voltage specification of the DC-DC converter.
operating voltage (12 V). This is due to the fact that diffused The effect of the minimum and maximum temperature of on
radiation exists and not direct beam radiation. the output voltage from the DC-DC converter during the
sampling period has been presented in Fig. 3. Table 2.
TABLE I. AVERAGE MINIMUM AND MAXIMUM
TEMPERATURES FOR THE MONTHS OF (APRIL-JULY, 2012) represents a succinct comparison of the types of DC-DC
converters that exist and their basic application. This is done
Month Average Minimum Average Maximum
temperature (oC) temperature (oC) to have an overview of the characteristics of the Buck and
April 3 38 Boost DC-DC converters, hence assist to make a right choice
May 4 38 to incorporate in PV systems with specific load profile.
June 1 25
July -3 28
 measuring instrument which must be controlled on at least two
Radiation conditions for the output
occasions [14]. The PV panel was connected to a 12 V, 24 V
of the DC-DC converter
14
and a 48 V MEANWELL DC-DC converter, which are Buck
and Boost converters and can be used for power conditioning
DC-DC converter voltage

12
No radiation Diffuse radiation Direct radiation [15]. A 12 V, 24 V and 48 V DC-DC were tested in this
10 research because they maintain a low power loss in the
8 conversion process. [A DLIC using hall-effect current sensors,
6
LTS 6-NP (LEM product) connected the PV system to a
R² Value DAQPRO 5300 data logger for recording input and output
4
voltages and currents]. The DAQPRO 5300 data logger has 8
2 analogue input channels and a 16 bit sampling resolution that
0 makes it ideal for logging purposes in this type of empirical
0 10 20 30 40 50 60 70 80 90 100
Percentage (Time)
research [16]. The DC-DC converter was connected to a
varying load resistance.
Figure 4: Radiation conditions for the output of a DC-DC converter.
However, it is possible that the PV voltages could be as high
as 29 V. Bearing in mind the 29 V limitation reveals that the
TABLE II. TYPES OF DC-DC CONVERTERS THAT EXIST AND THEIR BASIC
APPLICATION
data loggers stated above will not be able to accommodate
these high voltages, as most of them have an extreme input
DC-DC converters Step-up switching Step-down switching
converters regulators
voltage of 2.5 V. A data logging interface (DLI) circuit was
Known names Boost switching Buck switching incorporated in the system. It is afterwards necessary to
regulators regulators condition the voltage to make it less than the maximum input
Working Principle Provide a higher Provide a regulated
voltage output than voltage level that is voltage required by the data loggers. The DLI must also be
the input voltage less than the input able to provide DC current monitoring using hall-effect
voltage
Advantages Easy to analyse and High efficiency, high
current sensors. The electrical design of the test setup is
has a simple circuit switching frequencies presented in Fig. 5.
built-in circuit
protections at low
cost
DAQPRO 5300 DAQPRO 5300
Disadvantages Not suitable for high High voltage ripples
Data Logger Data Logger
power conversion at the output
Input power Output power
Application Designed for driving They are available
Vi x Ii Vo x Io
strings of LEDs. with internal
architectures that are DLIC DLIC
optimized for
applications that have
specific goals, such as
power saving and high
efficient PV systems. SOLARWORLD
SW220 MEANWELL
PV Panel DC-DC Converter Load resistor
Tilt angles = 12 V/24 V 6Ω
16°, 26° and 36° power regulation
III. PRACTICAL SET UP Orientation angle =
0°

In the practical set up, A SOLARWORLD (SW220 poly-

Figure 5: Electrical design of the practical set-up
crystalline PV panel) was used owing to its lower cost and
better performance in areas of direct solar radiation. The PV IV. TEST RESULTS
panel was placed at a fixed orientation angle of 0 o with the tilt
angle varied to 16o, 26o and 36o on a weekly basis during the
The efficiency of the different DC-DC converters was also
sampling period of four months. These changes in strictures
calculated using the four months of collected data (see Tables
are used to evaluate the effectiveness of each of the three
3, 4 and 5 for the three different DC-DC converters). The
different DC-DC converter with all wide-open to the same
different tilt angles, PV voltage influenced by Temperature,
design of PV panel, enabling an effective relationship as only
current and power inputs, DC-DC converter voltage, current
one variable is altered. Vanderbijlpark, South Africa (Latitude:
and power outputs, efficiency and on-time are shown. The
26°42, 649’ S and Longitude: 27°51, 809’ E). Cost efficiency
input voltage (PV voltage) data was analysed below 27 V,
is achieved by the use of a single PV panel in this research. A
while the DC-DC voltage was reduced 2 V below its
sampling period of four months was done which was from
specification (e.g. using a 24 V DC-DC converter resulted in a
April 2012 through July 2012
filtered voltage of 22 V). The contrast of the efficiency of the
12V, 24V and 48V DC-DC converters are used in obtaining
Samples were gathered for each tilt angle for this particular
the plot presented in Fig. 3. This was done in order to make
period. The samples ensured test-retest reliability of the
allowance for the response time of the DAQPRO 5300 data logger responsible for recording the data.

TABLE III: DATA FOR A 48 V DC-DC CONVERTER (BUCK DC-DC CONVERTER).

Samples Tilt angle PV voltage (Vin) PV current (Iin) Power in DC voltage DC current (Iout) Power out Efficiency On - time/wk in
(°) (V (A) (W) (Vout) (V) (A) (W) (%) hrs
1 16° 29.72 2.81 83.5 48.26 1.404 67.75 81.14 81.9
2 16° 30.74 2.81 86.3 48.05 1.404 67.46 78.09 33.11
3 26° 30.67 2.81 86.1 48.81 1.404 68.52 79.51 77.73
4 26° 31.01 2.81 87.1 48.31 1.404 67.82 77.83 30.05
5 36° 30.52 2.81 85.8 48.75 1.404 68.44 79.80 37.19
6 36° 29.63 2.81 83.3 48.54 1.404 68.15 81.85 14.07
Average 30.38 2.81 85.35 48.45 1.404 68.02 79.70 NA

TABLE IV: Data for a 24 V DC-DC converter (Buck DC-DC converter).

Samples Tilt angle PV voltage (Vin) PV current (Iin) Power in DC voltage DC current (Iout) Power out Efficiency On- time/wk in
(°) (V (A) (W) (Vout) (V) (A) (W) (%) hrs
1 16° 27.86 4.115 114.4 22.95 4.078 93.59 81.63 4.23
2 16° 29.43 4.115 121.0 24.21 4.078 98.72 81.52 4.77
3 26° 31.8 4.115 130.9 24.27 4.078 98.97 75.63 33.06
4 26° 30.37 4.115 124.9 25.17 4.078 102.64 82.13 28.39
5 36° 31.26 4.115 128.3 24.01 4.078 97.91 76.11 14.71
6 36° 30.26 4.115 124.1 24.16 4.078 98.52 79.12 16.91
Average 30.16 4.115 123.93 24.12 4.078 98.39 79.35 NA

TABLE V: DATA FOR A 12 V DC-DC CONVERTER (BUCK DC-DC CONVERTER).

Samples Tilt angle (°) PV voltage (Vin) PV current Power DC voltage DC current Power out Efficiency (%) On- time/wk
(V (Iin) (A) in (W) (Vout) (V) (Iout) (A) (W) in hrs
1 16° 31.35 2.7 84.7 12.44 6.0 74.64 88.18 30.76
2 16° 29.62 2.7 79.9 11.95 6.0 71.7 89.65 15.72
3 26° 31.15 2.7 84.2 12.51 6.0 75.06 89.24 33.93
4 26° 32.03 2.7 86.4 12.68 6.0 76.08 87.97 33.28
5 26° 31.96 2.7 86.2 12.71 6.0 76.26 88.37 29.71
6 36° 31.17 2.7 84.2 12.61 6.0 75.66 89.90 25.02
7 36° 31.47 2.7 84.9 12.6 6.0 75.6 88.97 33.16
8 36° 29.55 2.7 79.6 12.22 6.0 73.32 91.89 19.37
Average 31.03 2.7 83.76 12.46 6.0 74.79 89.27 NA

These individual PV and DC-DC voltages were averaged and P

n(%) = out x100 (3)
the input and output powers were calculated as follows: Pin
67.75
Input power (PV power) = PV voltage x PV current (1) n(%) = x100
83.5
Pin = Vin x Iin 81.84 %
Pin = 29.72 x 2.81
The significance of Tables 3, 4 and 5 is to contrast the
Pin = 83.5 W efficiency of the different types of DC-DC converters applied
in this research, hence ascertaining which DC-DC converter
was most efficient when used with a 220 W poly-crystalline
Output power (DC - DC power) = DC - DC voltage x DC - DC current) (2)
PV panel. From the results gathered in Tables 3 - 5, the 12
Pout = Vout x Iout DC-DC converter was most efficient as it shows an overall
efficiency of 89.27%.
Pout = 48.26 x 1.414
Pout = 67.75 W The similarity of these converters is that they provide an
output voltage that rightly matches their specification at points
The efficiency of the PV system for each week for the data when the PV voltage was optimum (29 V). In DC–DC buck
obtained of this research was calculated using the efficiency (step-down) converters, the output voltage magnitude is
formula which is presented in equation 3. always lower than the input voltage magnitude [17] and is
directly proportional to the duty cycle [18]. This indicates that
the 12 V and the 24 V DC-DC converters are buck (step
down) converters. In like manner, in a boost (step-up)
converter the output voltage magnitude is higher than the input to formulate results in some previous article by the authors.
voltage magnitude, making the 48 V a boost (step-up) Furtherance to the effectiveness of the Buck Converter it’s
converter. The inductance in DC-DC buck converters also applicable in use for powering high performance micro-
basically reduces with the increased current corresponding to processors due to their optimized load transient response.
increased incident solar irradiation. This contributes to a 75% Hence proves that its conversion efficiency outperformed the
smaller inductor as the stable step response changes in respect 24 V and 48 V DC-DC converters. It is recommended that
to the input solar power [17]. The disadvantage of the boost stand-alone PV systems make use of a Buck converter owing
DC-DC converter over the buck DC-DC converter is that to its high conversion efficiency.
under low irradiation condition, a boost converter cannot track
the MPP as this point is in the non-operating region. All three
types of DC-DC converters start delivering power to the load VI. ACKNOWLEDGEMENT
(start of the on-time) when their output voltage reaches their
rated specification which is attributed to effect of direct beam I hereby wish to express my gratitude to the following persons
radiation incident on the PV panels. The contrast of the and organisation who facilitated this research and enabled this
efficiency of the 12V, 24V and 48V DC-DC converters are paper to be completed successfully on time.
presented in Table 6 and the values are used in obtaining the •Prof H C Vz. Pienaar
plot presented in Fig. 6. •Research Department of Vaal University of Technology
•Telkom South Africa Ltd, TFMC Pty LTD, M-TEC and
TABLE VI: DATA FOR A THE BUCK AND BOOST DC-DC CONVERTER THRIP
DC-DC Average Power in Power out Power loss REFERENCES
converter efficiency (W) (W) (%)
12 V 89 84 75 11
24 V 79 124 98 21 [1] D. E. Furkan, “The analysis on photovoltaic electricity generation status,
48 V 80 85 68 20 potential and policies of the leading countries in solar energy,”
Renewable and Sustainable Energy Reviews, vol. 15, pp. 713-720, 2010.
[2] Energy in the World Economy [Online]. Available at
http://www.fao.org/docrep/003/x8054e/x8054e04.htm. Accessed: 2015-
2-2.
Efficiency curve for the 12V, 24V and 48V DC-
DC conveters [3] USDOE. US. Energy Information Administration. (2008). [Online].
Available at: http://www.eia.doe.gov/emeu/iea/popgdp.html. Accessed:
140 2015-02-23.
120
[4] C, A, Grimes, K. V. Ooman, R. Sudhir. Light, water and hydrogen: The
100 solar generation of hydrogen by water photoelectrolysis. New York,
80 USA: Springer science+Business media LLC. 2008.
60 [5] Y. F. Nassar and A. A. Salem. “The reliability of the photovoltaic
40 utilization in southern cities of Libya. Desalination,” vol. 209, pp. 86-90,
20
2007.
0 [6] K. Stefan. Solar electric power generation - photovoltaic energy
Efficiency (%) Power in (W) Power out (W) Power loss (%) systems, modeling of optical and thermal performance, electrical yield,
12 V 24 V 48 V
energy balance, effect on reduction of greenhouse gas emissions. Berlin:
Springer. 2006.
[7] F. A. Farret and M. G. Simoes. Integration of alternative sources of
Figure 6: Efficiency curve of the Buck and Boost DC-DC converters (12 V, energy. New Jersey: John Wiley Press. 2006
24V, 48 V).
[8] C. Luis and S. Santiago. Modelling photovoltaic systems using PSpice.
Chichester: John Wiley. 2002.
[9] Maximum power point tracking photovoltaic system homepage on bryan
buckley. [Online]. Available:
V. CONCLUSION http://bryanwbuckley.com/projects/mppt.html.
[10] Ozemoya, J. Swart, H.C. Pienaar, R.M. Schoeman. Factors impacting on
In review, the purpose of this paper was to carry out an the surface temperature of a PV panel controlling the ambient
experiential analysis of the effectiveness of Buck and Boost temperature of a PV panel to maintain high conversion efficiency. South
African Telecommunications and Network Application Conference
DC-DC converters with a certain load profile by observing an SATNAC 2013, Stellenbosch, South Africa.
empirical testing of the difference DC-DC Converters. The [11] O. Asowata, A. J. Swart and C. Pienaar, 2012. Optimum tilt angles for
optimization and evaluation was carried out using the photovoltaic panels during winter months in the Vaal Triangle, South
efficiency rate of the different DC-DC converters to drive the Africa. Smart Grid and Renewable Energy (SGRE), Scientific Research
load. Data used in this study was taken from April through Publishing, USA, vol.3 No.2, May 2012.
July 2012. The results presented in Tables 3, 4 and 5 and [12] Creative Research Systems. [Online]. Available at
Figure 6 clearly show that the 12 V DC-DC converter (buck http://www.surveysystem.com/correlation.htm. Accessed: 2015-05-3.
converter) was most efficient with this specific load profile.
Some information, data presented in this paper has been used
[13] K, Y., Hyunmin, J. & Deokjung, K. August. A new peak power tracker
for cost-effective photovoltaic power system. Energy Conversion
Engineering Conference. IECEC 96, vol.3: pp. 1673-1678. 1996.
[14] Welman, F. Kruger and B. Mitchell. Research Methodology. South-
Africa: Oxford University Press. 2005.
[15] W. Chen, H. Shen, B. Shu., H. Qin. And T. Deng. Evaluation of
performance of MPPT devices in PV systems with storage batteries.
Renewable Energy, vol. 32: pp. 1611-1622. 2007.
[16] Zamir & Berman. (2004). Fourier Systems Data Acquisition Solutions
for Industry. [Online]. Available at:
http://www.fouriersystems.com/pdf/fourier_catalogue.pdf. Accessed
2012-06-17.
[17] M. H. Taghvaee, M. A. M. Radzi, S. M. Moosavain, H. Hizam and, H.
M. Marhaban. A current and future study on non-isolated DC-DC
converters for photovoltaic applications. Renewable and Sustainable
Energy Reviews, vol. 17, pp. 216-227. 2013.
[18] Optimizing the Load Transient Response of the Buck Converter (R.
Richard, P. Brian and Z. Zansky): Applied Power Electronics
Conference and Exposition, 1998. APEC '98. Conference Proceedings
1998. Thirteenth Annual, Volume: 1

Asowata Osamede received his MTech in 2013. His research interest

includes renewable energy sources with a focus on photovoltaics. He is
currently studying towards a DTech in Electrical Engineering at the Telkom
Centre of Excellence, Vaal University of Technology, Vanderbijlpark, South
Africa
 Sixsoid: A new paradigm for k-coverage in 3D
Wireless Sensor Networks
Manjish Pal
Department of Computer Science and Engineering
National Institute of Technology Meghalaya, Shillong
India - 793003
Email: manjishster@gmail.com
Abstract—Coverage in 3D wireless sensor network (WSN)
is always a very critical issue to deal with. Coming up with
good coverage models implies more energy efficient networks.
K-coverage is one model that ensures that every point in a given
3D Field of Interest (FoI) is guaranteed to be covered by k sensors.
When it comes to 3D, coming up with a deployment of sensors
that gurantees k-coverage becomes much more complicated than
in 2D. The basic idea is to come up with a convex body that is
guaranteed to be k-covered by taking a specific arrangement
of sensors, and then fill the FoI will non-overlapping copies
of this body. In this work, we propose a new shape for the
3D scenario which we call a Sixsoid. Prior to this work, the
convex body which was proposed for coverage in 3D was the
so called Reuleaux Tetrahedron. Our construction is motivated
from a construction that can be applied to the 2D version of the
problem in which it imples better guarantees over the Reuleaux
Triangle. Our contribution in this paper is twofold, firstly we
show how Sixsoid gurantees more coverage volume over Reuleaux
Tetrahedron, secondly we show how Sixsoid also guarantees a
simpler and more pragmatic deployment strategy for 3D wireless
sensor networks. In this paper, we show the constuction of Sixsoid,
calculate its volume and discuss its implications on the k-coverage
in WSN.
I. I NTRODUCTION
Wireless Sensor Networks (WSN) consists of energy con-
strained sensor nodes having limited sensing range, commu-
nication range, processing power and battery. The sensors
generally follow different hop-by-hop ad-hoc data gathering
protocols to gather data and communicate. The sensors can
sense the information which-ever lies in its sensing range
using RF-ID (Radio Frequency Identification). The sensed
data can be communicated to another sensor node which lies
within communication range of the sender. The gathered data
finally reaches the base station which may be hops apart from
any sensor. Since sensor transceivers are omni-directional, we
assume the sensing and communication ranges as spheres of
certain radii. The network is called homogeneous when all
the sensors have the same radii and heterogeneous otherwise.
Coverage of a certain FoI and deployment of sensors are an
issue of research where the aim is to make energy efficient
networks. Sensor deployment and coverage in 2D requires
simpler strategies and protocols as compared to 3D. 3D sensor
network is used generally for underwater sensor surveillance
[4], floating lightweight sensors in air and space, air and water
pollution monitoring, forest monitoring, any other possible 3D
deployments etc. Real life applications of WSNs are mostly
confined to 3D environments. The term k-coverage in 3D is
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Nabajyoti Medhi
Department of Computer Science and Engineering
National Institute of Technology Meghalaya, Shillong
India - 793003
Email: nabajyoti.medhi@nitm.ac.in
used to describe a scenario in which k sensors cover a common
region. More precisely, a point in 3D is said to be k-covered
if it lies in the region that is common to the sensing spheres of
k-sensors. k is termed as degree of coverage. The ultimate aim
of this project is to come up with a deployment strategy for
sensors that guarantees k-coverage of a given 3-dimensional
Field of Interest (FoI) for large values of k. A first step to
address this issue is to come up with a 3-dimensional convex
body (tile) that is guaranteed to be k-covered by a certain
arrangement of k (or more) sensors, and then fill the FoI with
non overlapping copies of that shape by repeating the same
arrangement.
The term Sixsoid has been coined in this paper to signify
a geometrical shape that resembles a super-ellipsoid [14].
Sixsoid is created by the intersection of six sensors, each
having the same sensing radius, which are placed on the
six face centers of a cube of side length r where r is the
radius of the sensing spheres. We compare the implications
of this convex body with the previously proposed model on
3D k-covergae based on Reuleaux Tetrahedron. Recall that
the Reuleaux Tetrahedron, is created by the intersection of
four spheres placed on the vertices of a regular tetrahedron of
side length r. In an attempt to guarantee 4-coverage of the
given field, [2] considers a scenario in which four sensors
are placed on the vertices of a regular tetrahedron of side
length equal to the sensing radius r. Their intersection is
the so called Reuleaux Tetrahedron. Unfortunately it is not
possible to obtain a tiling of the 3D space with non-overlapping
copies of Reuleaux Tetrahedron [1]. In fact, such a tiling is
not possible even with a tetrahedron [15]. In [1] a plausible
deployment strategy is hinted that exploits this construction
by overlapping two Reuleaux Tetrahedrons, gluing them at a
common tetrahedron’s face, but this deployment doesn’t seem
to be pragmatic. In this paper, we propose another 3D solid
(the Sixsoid) for this purpose and an extremely pragmatic
deployment strategy. We show that, the volume guaranteed to
be 6-covered (by Sixsoid) is more than the volume guranteed
to be 4-covered (by Reuleax Tetrahedron). Indeed, 6-coverage
is more desirable than 4-coverage if it takes less number of
sensors to k-cover the same (FoI). We also discuss how the
fraction of the volume of Sixsoid inside a cube changes as a
function of sensing radius.
Another key ingredient of the work of [1], [2] is a result
that states that any convex object with breadth (the maximum
distance between any two points) at most r (where r is the
sensing radius) can be k-covered, by placing all the sensors
inside that convex object. Thus to apply this result to Reuleaux
Tetrahedron they need to take the side of the tetrahedron as
r
1.066 because the breadth of Reuleaux Tetrahedron is 1.066r
if we take r as the side of the tetrahedron. It is well known
that the volume of of a Reuleaux Tetrahedron constructed out
of a regular tetrahedron of side length r0 is approximately
0.422r03 . Thus, the volume of of the Reuleaux Tetrahedron with
0.422 3
a breadth equal to the sensing radius r is roughly 1.066 3 r . In
our case, the Sixsoid obtained by placing the sensors on the
faces of a cube of side length R = r has a breadth of r and
an approximate volume of 0.685r3 . Thus, according to our
construction a Sixsoid is guaranteed to be k-covered.
A. Prior Work on 3D k-coverage
There are relateively fewer works done to address the
problem of 3D k-coverage as compared to the 2D version of
the problem. In [1], [2] the authors make significant progress
on this problem. They discuss the relevance of Reuleaux
Tetrahedron in various issues dealing with connectivity and k-
coverage in 3D. Prior to that the following works disscuss the
3D version; in [13], authors propose a coverage optimization
algorithm based on sampling for 3D underwater WSNs. [9]
proposes an optimal polynomial time algorithm based on
voronoi diagram and graph search algorithms, in [6] authors
suggest algorithms to ensure k-coverage of every point in
a field where sensors may have same or different sensing
radii, [8] defines the minimum number of sensors for k-
coverage with a probability value, [12] studies the effect of
sensing radius on the probability of k-coverage, [5] proposes
an optimal deployment strategy to deal with full coverage and
2-connectivity, [3] brought forward a sensor placement model
based on Voronoi structure to cover a 3D region, in [11],
authors provide a study on connectivity and coverage issues in
a randomly deployed 3D WSN. In this paper, we compare our
proposed Sixsoid based k-coverage model with the existing
Reuleaux Tetrahedron based model [1]. Comparison has been
done in terms of volume of k-coverage, sensor spatial density
in 3D and placement and packing in 3D.
II. O UR C ONTRIBUTION
In a previous work [10], the authors used a geometrical
shape that resembles a super-ellipse for k-coverage in 2D
homogeneous WSN. This model [10] showed much better
efficiency in terms of area of k-coverage, energy consumption
and requirement of less number of sensors as compared to
the k-coverage model with Reuleaux Triangle [1]. This work
done in [1] is further extended to 3D by taking Reuleaux
Tetrahedron. Motivated by the fact that considering a “super-
elliptical” tile has proven to be much better than the Reuleaux
Triangle based model [10], in this work we extend that idea
to 3D. The main hurdle was to compute the volume of the
solid (Sixsoid) generated out of our construction. Unlike the
Reuleaux Tetrahedron its volume was not already known.
Another reason for opting our construction is its resemblence
to a superellipsoid because of which it fits much better in
3D rather than a Reuleaux Tetrahedron. Moreover, Sixsoid
provides a practical and easier packing in 3D. Once we fill
the FoI with cubes of side length R, a Sixsoid is formed by
the overlapping of six sensors lying on the six face centers of
a cube. So, every cube contains a Sixsoid inside it. When we
consider packing in 3D, cubes are space filling. We propose
a result in a later section that states that our deployment
strategy ensures 4-coverage of entire FoI along with fact that
approximately 68% of it is 6-covered. Thus, the Sixsoid based
model ensures at least 4-coverage of the entire FoI. Packing
of Reuleaux Tetrahedron in 3D as proposed in [1] is harder to
achieve and it is not feasible for practical deployment. Cubical
structure is easy to manage in space and sensor placement
on the surfaces of a cube also becomes an easier and faster
process. In the subsequent sections, we calculate the volume
of the Sixsoid with the supporting theorems and compare
our results with the Reuleaux Tetrahedron based model. In
this work we don’t address the issue of connectivity. In the
remaining part of the paper, we discuss the computation of
volume of a Sixsoid in Section 3. In Section 4 we discuss
the deployment and packing strategy in 3D that exploits the
structure of Sixsoid. Section 5 is devoted to the comparison of
our model with the results proposed in [2]. We end the paper
with conclusion and potential research directions for future.
III. C OMPUTING THE VOLUME OF S IXSOID
The Sixsoid which we denote by S(R) is defined as the 3-
D object obtained due to the intersection of six spheres placed
on the centers of the faces of a cube of side-length R. The
radius of each sphere is R. In this section, we compute the
volume V of S(R). Our basic strategy is to take a horizontal
cross section (via a sliding plane) of the cube at a distance of x
from the top of the cube and perform the following integration,
 R  R/2
V = A(x) dx = 2 A(x) dx
0 0
where A(x) is the area of the cross section and the last equality
follows due to the symmetry of D(R).
Henceforth, we will use the following notation re-
garding the geometry of the objects involved in this
section. Let Fx1 (top), Fx2 (bottom), Fy1 (left), Fy2 (right),
Fz1 (front), Fz2 (back) be the faces of the cube and we de-
note the respective spheres (centered at face centers) by
Sx1 , Sx2 , Sy1 , Sy2 , Sz1 , Sz2 . We will denote the sliding plane
by Px and let rαi be its radius. Let Cαi be the circle
which is obtained by the intersection of Px with Sαi where
α ∈ {x, y, z} and i ∈ {1, 2}. In the remainder of this section
we show the behavior of the cross-section as the plane Px
slides from x = 0 to x = R2 .
Observation 1: For all 0 ≤ x ≤ R/2, Cx1 and Cx2 are
concentric circles and rx2 ≤ rx1 . Also as x varies from
0 to R/2, rx1 monotonically decreases, rx2 monotonically
increases and Cx1 = Cx2 at x = R/2.
Observation
 2: For 0 ≤ x ≤ R/2, ry1 = ry2 = rz1 =
3R2 2
rz 2 = 4 + Rx − x .
Lemma 1: A(x) 2
√ = π(2Rx − x ) from x = 0 to x = l1 ,
3− 7
where l1 = R( 4 ) ≈ 0.0886R
Proof: We observe that, A(x) is exactly Cx2 as Px
slides below x = 0. So to compute A(x) we only need to
compute
 the radius rx2 of Cx2 . This value can be computed
as R2 − (R − x)2 , which proves the claim regarding A(x).


to see that α = 2 sin−1 GH 2r where r = rx2 and GH is the

length of the line segment joining G and H. We first show
how to compute GH in terms of R and x.
Computing GH: Notice that G and H are points of intersec-
tion of the circles Cy1 and Cx2 . Let us write the equation of
these circles assuming O to be the origin and lines parallel
to y and z axis passing though it as x and y  axes. The
equations of Cy1 and Cx2 in this coordinate system would
be x2 + (y  − R2 )2 = ry21 and (x − R2 )2 + (y  − R2 )2 = rx22
respectively.
 Subtracting
 the later from the former gives x =
2
1 2 2 R
R ry1 − rx2 + 4 . Now if we conside the 
equation of Cy1

2
as a quadratic equation in y  with x = R1 ry21 − rx22 + R4
then the modulus of the difference of the roots of this equation
will exactly be the length of GH. Simple algebraic manipula-
tion shows that

GH = 2 3Rx − R2 /4 − 2x2

The desired area can be decomposed into two parts. (i) 4

times the area of sector EOG (ii) 4 times the area of region
GOH. We compute each of these as follows:
(a) Area of sector EOG is
Fig. 1. Cross-Section of A(x) is the enclosed region EF GHIJKLE, for   
l2 ≤ x ≤ l 3 1   π GH
r2 · θ/2 = · 2Rx − x2 · − 2 sin−1 √
2 2 2 2Rx − x2
Now inorder to know l1 (the event when the cross section (b) Area of the region GOH = Area of (ΔGOH) + Area of
changes), we need to find the value of x at which Cx2 is the Cap, which
 is equal to

tangential to Cz1 , this is also the instance at which Cx2 is 1

· GH · r 2 − GH 2 + 1 · r 2 · 2 sin−1 GH
2 4 2 2ry1
tangential to each Cy1 , Cy2 and Cz1 . To find this x we need   
2
to solve the following equation, rx2 = R/2 + rz1 − 12 · GH · ry1 2 − GH 4

3R2  R
+ Rx − x2 = 2Rx − x2 + Adding the aforementioned two values, multiplying by 4, and
4 2 replacing the values of GH, r, ry1 , gives us the expression of
3R2 R 2
+ Rx − x2 = 2Rx − x2 + + R 2Rx − x2 A(x) in the lemma.
4 4 Computing l2 : To find the value of x at which the cross-
R  section changes from the above mentioned 8-sided region, we
−x = 2Rx − x2
2 need to look at the instant when the circle Cx2 circumscribes
R2 the region formed due the intersection of Cy1 , Cy2 , Cz1 , Cz2 .
2x2 − 3Rx + = 0 At this instant the following equality holds;
4
√
The root√of the above quadratic equation which is less than R/2 is r x2 = 2 (z − R/2)
 √ 
x = (3− 4 7)R . This proves the lemma. √ R + 5R2 + 8Rx − 8x2 R
   = 2 −
4 2
Lemma 2: A(x) = 12 · r2 · π2 − 2 sin−1 2√2Rx−x
Y
2 √
√ 2 
    5R + 8Rx − 8x2 R
= 2 −
+ 21 · Y · r2 − Y4 + 12 · r2 · 2 sin−1 2rYy
2
4 4
   1
 √ 2 
√ 5R + 8Rx − 8x 2 R
1
− 2 · Y · ry1 − 42 Y2
where Y = 2Rx − x2 = 2 −
4 4
 √ 
2 7R
2
+ 3Rx − 2x2 , r = 2Rx − x2 and 5R2 − 8x2 + 8Rx = 16Rx − 8x2 + R (rearranging and squaring)
4 
2
6x2 − 8Rx + R2 = 0 (after a few algebraic manipulations)
ry 1 = Rx − x2 + 3R 4 from x = l1 to x = l2 where √ √
√ √ which solves to x = ( 23 + 610 )R and x = ( 23 − 610 )R, since
l1 = R( 3−4 7 ) and l2 = ( 23 − 10
6 )R 0 ≤ x√ ≤ R/2, the only root which is of concern for us is x =
Proof: Recall that above x = l2 , the cross-section is a ( 23 − 610 )R.
circle and at x = l2 , Cx2 is tangential to Cz1 . As P (x)  
Lemma 3: A(x) = 4 12 · s2 · θ − 12 · z · (2z − R) +
slides slightly below x = l2 , the cross-section is a 8-sided  2 z− R
closed region drawn in Fig.1. Let O be the center of the 4 (y − R2 ) . where θ = ∠AOB = 2 sin−1 z 2 and s =
 √
square and ∠EOF = ∠GOH = ∠IOJ = ∠KOL = α, Rx − x2 + 3R2
z= R+ 5R2 −8x2 +8Rx
4 , 4 .
∠F OG = ∠HOI = ∠JOK = ∠LOE. Our aim is to find
out the area of the region EF GHIJKLE. From the notation Proof: We notice that for l2 ≤ x ≤ l3 , A(x) is the
used in the Fig. 1 we deduce that α + θ = π2 . Also it is easy area which is common to Cy1 , Cy2 , Cz1 , Cz2 . Fig.2 shows the
 IV. D EPLOYMENT S TRATEGY IN 3D
Consider a tessellation of a 3D Field of Interest (FoI) using
cubes of side length R. We place the sensors each of sensing
radius r = R on the centers of the faces of every cube. Clearly,
every sensor will be shared by 2 cubes in the tiling (except
the boundary ones).
Theorem 2: In the aforementioned packing every point is
either 6-covered or 4-covered.
Proof: By construction, every point inside a Sixsoid
formed will be covered by 6 sensors. We just have to prove
that every other point will be 4-covered. According to the
notations used in the construction of Sixsoid in Section 3,
which we borrow here, this is equivalent to proving that every
point in the cross-section (which is a sqaure of side length R)
obtained by slicing the cube with the plane Px is 4-covered
for all 0 ≤ x ≤ R. Again by symmetry of S(R), we only
have to prove this for all values between 0 ≤ x ≤ R/2. As
we have noticed in Section 3 that due to the nature of Sixsoid
Fig. 2. Cross-Section of A(x) is the enclosed region ABCD for l3 ≤ x ≤ l4 the topology of the cross section changes at two values of x,
namely x = l1 and x = l2 . Consider the arrangement of circles
formed by the circles on the square of side length R. It can be
geometry of A(x). Again let O be the center of the square We verified that for all the ranges when the topology of the cross
divide it into two parts, (i) The square ABCD (ii) Four caps section remains fixed (i.e. from 0 to l1 , l1 to l2 and l2 to R/2)
surrounding the square each of equal area. We compute each every point inside the square is 4-covered.
of these areas as follows:
From the above result we can infer that using the afore-
(a) The area of the square is l2 where l is the side mentioned packing we are guaranteed to have 4-coverage of
length of the square,

which according to the fig- 100% of the FoI and 6-coverage of approximately 68% of it.
ure is {2 z − R2 }2 . It can be verified
 that z =
√
R+ 5R2 +8Rx−8x2
and s = rz1 = 3R 2 2 V. C OMPARISON OF T WO M ODELS
4 4 + Rx − x .
Thus area of the square is In this section we present the comparison of our model with
√ 2 the Releaux Tetrahedron Model based on our Matlab programs.
5R2 − 8x2 + 8Rx − R We perform the numerical integration of the expression in
2· Theorem 2 and compare it with the volume of Reuleaux
4
Tetrahedron with varying sensing radius. This comparison is
(b) The area of a cap (say the cap AB) can be obtained presented in Fig.3. Next, we evaluate the sensor spatial density
by subtracting the area of ΔOAB from the area of for k-coverage in Sixsoid and compare it with the Reuleaux
sector OAB which is equal to Tetrahedron. Recall that the minimum sensor spatial density
1 2 1 R per unit volume needed for full k-coverage of a 3D field is
k
· s · θ − · z · (z − ) defined as a function of sensing radius ras λ = vol(S(R))
2 2 2
where R = r. We take the similar parameter for Reuleaux
z− R k r
3 where r0 = 1.066 .
−1
θ = ∠AOB = 2 sin
where 2 Tetrahedron from [1], [2], which is 0.422r
z√ and 0
2 R+ 5R2 −8x2 +8Rx Fig.4 shows the comparison of the minimum sensor spatial
s = Rx − x2 + 3R
4 , z = 4 densities and we find that Sixsoid requires much less sensors
Putting the values of z and s and multiplying the area of the per unit volume to guarantee k-coverage.
cap by 4 gives us the desired result. We use observation 1, to
conclude that the cross section doesn’t change for all values VI. C ONCLUSION AND F UTURE W ORK
of x between l3 and R/2. Thus l4 = R/2. In this paper, we discussed a new geometric model for
As a consequence of the previous results, we can conclude addressing the problem of k-coverage in 3D homogeneous
the following: Wireless Sensor Networks which we have named as the
Sixsoid. We show in a number of ways how this model
Theorem 1: outperforms the existing model of Reuleaux Tetrahedron [1],
    [2], namely the volume of the convex body involved, which in
l1 l2 R/2
V =2 A(x) dx + A(x) dx + A(x) dx turn implies improved spatial density and a better (in terms of
0 l1 l2 coverage) and more pragmatic deployment strategy for sensors
in a given Field of Interest. From the point of view of geometry,
In order to get a good estimate on the value of the above our construction of Sixsoid and its volume computation might
definite integral, we use Matlab to perform the numerical be of independent interest. We suspect our construction might
integration. The value of V turns out to be approximately have interesting consequences for non-homogenoeus networks
0.685R3 . as well which we leave as a direction of future research.
 Wireless Sensors to Achieve Both Coverage and Connectivity. In:
Proc. ACM MobiHoc, pp. 131-142, 2006
[6] Huang, C.F., Tseng, Y.C.: The Coverage Problem in a Wireless
Sensor Network. In: Proc. ACM Intl Workshop Wireless Sensor
Networks and Applications (WSNA), pp. 115-121 (2003)
[7] Jaklic, A., Leonardis, A., Solina, F.: Superquadratics and their geo-
metrical properties, computational imaging and vision, chap. 2, vol.
20, Kluwer Academic Publishers, Dordrecht. In: http://lrv.fri.uni-
lj.si/ franc/SRSbook/geometry.pdf (2000)
[8] Kumar, S., Lai, T.H., Balogh, J.: On k-Coverage in a Mostly
Sleeping Network. In: Proc. ACM MobiCom, pp. 144-158 (2004)
[9] Megerian, S., Koushanfar, F., Potkonjak, M., Srivastava, M.: Worst
and Best-Case Coverage in Wireless Sensor Networks. In: IEEE
Transactions in Mobile Computing, vol. 4, no. 1, pp. 84-92, Jan/Feb
(2005)
[10] Medhi, N., Sarma, N.: On Demand k-coverage with Purposeful Mo-
bility in Wireless Sensor Networks. In: IEEE Indicon, Hyderabad,
India (2011)
[11] Ravelomanana, V.: Extremal properties of three-dimensional sensor
networks with applications. In: IEEE TMC, 3(3), pp. 246-247, Jul.-
Sep. (2004)
[12] Wan, P.J., Yi, C.W.: Coverage by Randomly Deployed Wireless
Sensor Networks. In: IEEE Trans. Information Theory, vol. 52, no.
Fig. 3. k-covered volume comparison between Sixsoid and Reuleaux 6, pp. 2658-2669, June 2006
Tetrahedron [13] Xiaoyu, D., Lijuan, S., Linfeng, L.: Coverage Optimization Algo-
rithm Based on Sampling for 3D Underwater Sensor Networks. In:
International Journal of Distributed Sensor Networks (2013)
[14] http://www.wolframalpha.com/input/?i=superellipse
[15] mathworld.wolfram.com/Space-FillingPolyhedron.html

Fig. 4. Minimum spatial Density comparison between Sixsoid and Reuleaux

Tetrahedron

R EFERENCES
[1] Ammari, H. M.: On the problem of k-coverage in 3D wireless
sensor networks: A Reuleaux Tetrahedron-based approach. In: 7th
international conference on Inteligent sensors, Sensor networks and
Information processing (2011)
[2] Ammari, H.M., Das, S.K.: A Study of k-coverage and Measures of
Connectivity in 3D Wireless Sensor Networks. In: IEEE Transac-
tions on Computers , vol. 59, no. 2, February (2010)
[3] Alam, S., Haas, Z.: Coverage and connectivity in three dimensional
networks. In: Proc. ACM MobiCom, pp. 346-357 (2006)
[4] Akylidiz, I.F., Pompili, D., Melodia, T.: Underwater Acoustic
Sensor Networks: Research Challenges. In: Ad Hoc Networks, vol.
3, pp. 257 -279, Mar (2005)
[5] Bai, X., Kumar, S., Xuan, D., Yun, Z., Lai, T.H.: Deploying
 Identifying Ideal Values of Parameters for Software
Performance Testing
Charmy Patel
Shree Ramkrishana Intritute of Computer
Education & Applied Sciences,
Surat, Gujarat
charmyspatel@gmail.com
Abstract—Performance is the most imperative feature
concerned with the quality of software. Performance testing
provides the proof of performance of the product and set a
baseline for further enhancement in the application. Software
performance is dependent on various factors like response, speed
and underling resources. In this paper we have identified critical
factors affecting software performance and derived the measures
and matrices related to performance. We have performed
statistical mining to find the confidence interval of critical
parameters which plays important role in software performance.
Our research for Software Performance analysis is a systematic,
quantitative and qualitative way in the direction of supporting
performance sensitive software systems development that meet
performance objectives and prevents performance problems.
Keywords—Performance testing; performance
measurements; linear regression; correlation; confidence interval
I. INTRODUCTION
Performance is a marker which refers to a software system
that accumulates its requirements for relevance. To gain
Software system’s main goal- response time and throughput,
Responsiveness is a main skill [1]. Software performance
testing validates the stability, scalability or speed features of
the application/system under test. Performance is
apprehensive to achieve resource-utilization, response times
and throughput that meet the performance objectives for the
product/project.
In most of the software projects, testing is done at the last
phase. Lead the industry to performance driven approach and
stress diminution, Critical problem relevant to performance
and quality are to be solved at early phase of implementation.
So, to achieve the software quality, at the time of application
development, developer requires an indication about the
performance of a page/module created by him/her. This
indication can be given in terms of critical factors which
affects software performance. Rigorous analysis and statistics
is required for the identification of such parameters.
This paper is structured as follows; section 2 covers the
concept of performance measurements. Section 3 about
methodology to find performance measures by applying
statistical mining in which relationship between various
parameters is established using the correlation coefficient and
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Ravi Gulati
Department Of Computer Sceince
Veer Narmad South Gujarat University,
Surat, Gujarat
rmgulati@gmail.com
simple linear regression technique. Afterwards section 4
describes our experimental setup, process and results. We
finally conclude by demonstrating our derived results in the
form of critical performance affecting factors and performance
measurements matrix.
II. PERFORMANCE MEASUREMENTS
In our previous work we analyzed that software quality and
performance related problems are commenced at early stage of
implementation but not fixed till they become very much
expensive and difficult. Different performance testing tools in
market are available with their own static boundaries like
platform, browser compatibility etc. No consolidated
framework is available which analyze any web application to
find its performance affected critical factors. To make
application efficient, performance management and
improvements are done manually based on suggestions of the
team leader and/or the project manager. Above discussion
state that to maintain software quality and performance at the
time of software development and deployment is a research
challenge which leads us to the development of integrated
performance analysis framework in the support of software
quality assurance. So, to address this challenge we had
proposed a framework which
 Extracts the value of performance parameters.
 Compare these values with ideal performance
measurements.
 And provides feedback for performance
improvement.
By the pilot survey and analysis among industry people we
have identified key factors responsible for software
performance.
Performance testing tools helps to test system performance in
practically virtual environment. However Performance of
Software and Web application depends on various critical
performance factors [3]. For website performance testing is
dependent on Speed, Scalability and Stability of the page. So,
number of http requests, various objects’ size and their loading
times are very important parameters which effects on webpage
performance.
Our work is divided into two key process areas 1)
Identification of critical performance factors and 2) Derivation
of ideal performance measurements matrix. In the first phase
the key elements responsible for making software
performance down is derived by applying regression analysis
and in the next phase a performance measurement matrix is
created to define ideal ranges for parameters like size, number
of requests and response time.
III. Exertion of Statistical Mining
To achieve performance related factors and measurements the
analysis of historical data with help of statistical mining and
regression analysis is performed.
Regression analysis rivets discovering the association among
a dependent variable and one or more independent variables.
In parametric statistics, simple linear regression method is
widely used to analyze variable’s mean response. Various key
forms of inferential statistical analysis has base of this
concept. Correlation and regression analysis are related with
each other because both work with relationships amongst
variables. To show how or to what level two or more variables
are associated with each other, correlation coefficient which is
a measure of linear association is used. It measures only the
extent of linear association among two variables.
The Confidence Interval represents the accuracy of an
estimate. It is an experimental interval that normally includes
the parameter of interest if the test is repeated. Confidence
level or confidence coefficient is used to check how frequently
the observed interval contains the parameter.
IV. EXPERIMENTS AND EVALUATION
Our experiment is designed in following phases for the
derivation of performance parameters.
A. Data Collection and Big Query Processing
To find the ideal values for all performance sensitive
parameters we have taken HttpArchive data set. Steve Souders
created the HttpArchive dataset and Pat Meenan's
WebPagetest system provides platform to built it [4].
HttpArchive is a treasure trove of web performance data. It
crawls millions popular sites twice a month and records the
performance relevant data of each webpage [5]. Approx 4 GB
data is available in the dataset. All the data are retrieved using
IE9 browser with the default internet connection speed of 5.0
mbps.
HTTPArchive data set is available on Google Big Query by
Ilya Grigorik who is a web performance developer advocate
and engineer at Google, where he works on building the web
fast and performance best practices.
We have explored HTTPArchive dataset for our research
using GitHub as a service. First we developed a new project in
which we imported HTTPArchive- a web performance
dataset. Dealing out on massive datasets can be time
consuming and high-priced but we easily retrieved the dataset
with the help of Google Big Query using fast processing SQL
queries [8]. This dataset has mainly two tables – viz. (i) pages
and (ii) requests. Both tables have many parameters. We
retrieved approximately 3,00,000 records as a test case using
the query action available in Google Big Query for page
requests, loading time and size relevant parameters. Due to the
limitation of direct data download the resultant data is
exported into cloud storage of Google Developer Console. The
dataset obtained from the above source is pre-processed to
make it ready for statistical analysis.
B. Identification of critical performance factors using
Correlation between Parameters
To measure the strength of association involving two
variables, correlation coefficients method is used. It gives the
results including three types of value. It would show that the
variables are linearly positively related when value of
correlation coefficient is nearby 1, and straight line with
positive slope is shown in the scatter plot. The variables are
linearly negatively related when the result is nearby -1 and
straight line with negative slope on scatter plot. And weak
linear relationship between the variables is found then results
would be zero and generate negative slop on scatter plot [6].
To find the correlation between dependent and independent
parameters by the implementation of linear regression open
source tool R is used and sample calculations are
demonstrated below.
Cor.t = cor.test(big_query$reqTotal,big_query$reqJS)
(1)
Pearson’s product-moment correlation
data: big_query$reqTotal and big_query$reqJS
t = 536.7898, df = 290044, p-value < 2.2e-16
Alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
0.7041132 0.7077645
Sample estimates:
cor
0.7059435
As explained with the help of the above function we gain the
correlation between different parameters. We define these
performance parameters for the analysis of critical factors.
Total_http_request – total number of http request generated by
the website, Page_size – the size of underlying webpage
including images, scripts, css and multimedia, Page_loadtime
– a total time of page rendering in the browser are highly
dependent parameters on their relevant independent
parameters. So, we find correlation of these dependent
parameters with their relevant independent parameters which
is shown in table I, II and III. plot(mydata$img_size_KB, mydata$Page_size_KB)
(2)
Table I. CORRELATION BETWEEN REQUESTS PARAMETERS res = lm(mydata$Page_size_KB ~ mydata$img_size_KB)
(3)
Parameters Total_http_req
abline(res, col='red') (4)
Total_HTML_req 0.71
Total_Javascripts_req 0.71
Total_css_req 0.36
Total_images_req 0.91

Total_Multimedia_req 0.33

Table II. CORRELATION BETWEEN SIZE PARAMETERS

Page_size
Parameters (KB)

HTML_size (KB) 0.16

JS_size (KB) 0.33
CSS_size (KB) 0.19
Image_size (KB) 0.81

M_size (KB) 0.25

Fig. 1. Total Page Size ~ Image Size
Table III. CORRELATION BETWEEN LOADING PARAMETERS

Page_loadtime
Parameters (seconds)

HTML_loadtime (seconds) 0.53

Javascript_loadtime (seconds) 0.37
CSS_loadtime (seconds) 0.47
Images_loadtime (seconds) 0.71

Multimedia_loadtime (seconds) 0.51

From the above derived values we justify that

Total_http_request, Page_size and Page_loadtime are
dependent parameters and they strongly correlate with all
image relevant parameters. This analysis shows that images
are the critical affecting factor for web page performance.
Fig. 2. Total Page Size ~ Java Script Size
We can also show the relationship and correlation between
parameters graphically using plotting.

To carry out regression and parameter analysis of covariance

lm function is used. So, the result value of lm function is used
to draw an absolute regression line which specified the
variance of each variable and the correlation coefficient
between two variables.

The regression analysis graph plotting is created for image

size and total page size to established
relationship between these variables. The process of graph
creation is explained as follows.

Fig. 3. Total page size ~ HTML size
Fig. 4. Total Page Size ~ CSS Size
Fig. 5. Total Page Size ~ Multimedia Size
From the above graph plotting with regression line we can
easily derived that page relevant parameters- page size, page
requests and page load time are highly correlated with image
relevant parameters – image size, total number of images and
image load time.
C. Derivation of performance measurement matrix using
Confidence Interval
Performance of the web application depends mainly on three
measures – request value, size and load time of the parameters.
In this phase derivation of ideal values for performance
measurements is implemented for the parameters of load time,
size and requests. And, for that we have taken data of all those
parameters and apply confidence interval finding process to
fetch nearer ideal values. These ideal values are used to find
the confidence bounds for all the parameters.
The purpose of deriving ideal values is to indicate that
performance of the system does not meet its expected goals
and the performance limit lines are crossed. Based on this
compression suggestions will be given for the software
performance improvement. To find the upper bound and lower
bound values we calculate the confidence interval through
linear regression mining technique. In this paper we find the
confidence interval of independent variables using the normal
distribution.
First we loaded the data to calculate the confidence bounds we
have to find average values and standard deviation of each
parameter.
We use 95% confidence level to find the confidence interval.
qnorm(0.975)*s/sqrt(n) (5)
Where, qnorm returns the number whose cumulative
distribution matches the probability, s standard deviation, n
total number of observations
Then next we find the upper bound and lower bound of each
parameter.
mean ± (qnorm(0.975)*s/sqrt(n)) (6)
So, with the help of above formula we get the following
performance measurement matrix results.
Table IV. PERFORMANCE MEASUREMENT MATRIX
SR
.
No Parameter Name Confidence Interval
Lower Upper
Bound Bound
1 Page_load_time (sec) 5.79 6.87
2 HTML_loadtime(sec) 0.45 0.54
 3 Images_loadtime(sec) 1.74 1.81 Our next phase is to develop a new approached web
performance analyser which fetch the values of all above listed
4 javascript_loadtime(sec) 0.70 0.73
parameters and give suggestion for performance optimization
5 CSS_loadtime(sec) 0.54 0.56 of most affected parameters which are not given in other
6 Multimedia_loadtime(sec) 7.45 11.02 performance optimization tools.
7 page_size(KB) 1841.61 1859.90 REFERENCES
8 HTML_size(KB) 24.12 24.45
[1] X. Gan, “Software Performance Testing” Seminar paper University of
9 Image_size(KB) 1173.13 1187.09 Helsinki, Department of Computer Science (September 2006)
[2] C. Patel, R. Gulati, “Software Performance Testing Tools – A
10 Javascript_size(KB) 306.65 308.55 Comparative Analysis”, International Journal of Engineering Research
11 CSS_size(KB) 55.57 56.21 and Development, Volume 3, Issue 9, e-ISSN: 2278-067X, p-ISSN:
2278-800X, (September 2012).
12 Multimedia_size(KB) 255.10 259.44 [3] C. Patel, R. Gulati, “Software Performance Testing Measures”,
International Journal of Management & Information Technology Vol. 8,
13 Total_http_request 96.49 97.09 No. 2 ISSN 2278-5612 (January 2014)
14 Total_HTML_req 9.99 10.10 [4] “httparchive Dataset” [online] Available:
http://httparchive.org/about.php#bigquery
15 Total_images_req 55.91 56.30 [5] “HTTP Archive + BigQuery = Web Performance Answers”
[online] Available: https://www.igvita.com/2013/06/20/http-archive-
16 Total_Javascripts_req 18.91 19.03 bigquery-web-performance-answers/
17 Total_css_req 6.42 6.47 [6] “Regression and correlation analysis” [online] Available:
http://abyss.uoregon.edu/~js/glossary/correlation.html
[7] Rina, S. Tyagi , “A Comparative Study of Performance Testing Tools”
Now, these values are used as the upper and lower bounds of DCSA, Kurukshetra University,Haryana, India - International Journal of
the ideal value (Table IV). That is we can give the suggestion Advanced Research in Computer Science and Software Engineering,
to the developers that your parameters’ values should be ISSN: 2277 128X
[8] C. Smith, “Best Practices for Software Performance Engineering”,
between this specified ranges. If the values of the parameters Performance Engineering Services
lie within the specified range then the performance of the [9] D. Ushizima, “Data Analysis using the R Project for Statistical
application can be predicted to be of good quality. Computing”, NERSC Analytics/Visualization and Math Groups
Lawrence Berkeley National Laboratory
[10] B. Robinson and Joi Officer Advisor: Dr. Fred Bowers , “Data Mining:
V. CONCLUSION Predicting Laptop Retail Price Using Regression”
[11] M. Netto, S. Menon, H. Vieira, L. Costa, F. Oliveira, R. Saad, A. Zorzo,
To gain the overall quality of the system for a variety of “Evaluating Load Generation in Virtualized Environments for Software
industrial software systems performance requirement plays a Performance Testing”, ACM
significant role. To test the functionality of the system, [12] “Google cloud platform with BigQuery” [online] Available:
https://cloud.google.com/bigquery/what-is-bigquery
development team spend a great deal of resources but not [13] F. Vokolos ,E. Weyuker, “Performance testing of software systems”,
expand little bit for doing performance testing. With this AT&T Labs – Research, Proceeding WOSP '98 Proceedings of the 1st
approach, we can give the correlation between the parameters international workshop on Software and performance, ACM New York,
which will specify which parameters are more sensitive for NY, USA
[14] S. Puri, “Recommendation for performance benchmarking”, White
software performance and also give the ideal measurement’s Paper, Infosys
values to the developer at coding phase with an appropriate [15] C. Patel, R. Gulati, “Software Performance Analysis: A Data Mining
confidence bound value for each parameter. So performance Approach”, International Journal of Information Technology &
degradation issues can easily tracked at implementation phase. Management Information System (IJITMIS), ISSN 0976 – 6405(Print),
ISSN 0976 – 6413(Online), Volume 5, Issue 2, May - August (2014), pp.
To maintain the performance Service Level Agreement which 28-31 © IAEME
is bounded by the developing company with their customer, [16] G. Kanyongo, J. Certo, Brown I. Launcelot School of Education,
good quality of website with good performance optimization Duquesne University, "Using regression analysis to establish the
is must. So, to reach the software quality, developer requires relationship between home environment and reading achievement: A
case of Zimbabwe”, International Education Journal, 2006, 7(5), 632-
suggestions about the performance of an application formed 641. ISSN 1443-1475 © 2006 Shannon Research Press.
by him/her. In the phase of project maintenance, features of [17] S. Sharmila, E. Ramadevi, “Analysis of Performance Testing on Web
this concept can be utilized and expected outcomes can be Applications”, International Journal of Advanced Research in Computer
achieved in previously deployed project also. and Communication Engineering Vol. 3, Issue 3, March 2014, ISSN
(Online) : 2278-1021 ISSN (Print) : 2319-5940
[18] C. Babbar , N. Bajpai, D. Kapoor Sarmah, “Web Application
So, this paper basically presents the performance matrix – a Performance Analysis Based on Component Load Testing”, Proceedings
matrix of performance relevant parameters with their ideal of the 5th National Conference; INDIACom-2011 Computing For
value ranges to manage software quality. A software Nation Development, March 10 – 11, 2011
development and testing practice can be improved with this
theoretical analysis by considering the values of performance
matrix, correlation and dependence of various parameters
using regression analysis.
 Reducing Structured Big Data Benchmark Cycle
Time using Query Performance Prediction Model
Rekha Singhal
Performance Research Center Innovation Lab
Tata Consultancy Services
Mumbai, India
rekha.singhal@tcs.com
Abstract—The paradigm of big data demands either extension of
existing benchmarks or building new benchmarks to capture the
diversity of data and impact of change in data size and/or system
size. This has led to increase in cycle time of benchmarking an
application which includes multiple workloads executions on
different data sizes. This paper addresses the problem of reducing
the benchmark cycle time for structured application evaluation
on different data sizes. The paper propose an approach of
reducing Big data benchmark cycle time using prediction models
for estimating SQL query execution time with data growth. The
paper also proposes a model which could be used for efficient
tuning of benchmark queries before their executions, to speed up
the application evaluation process, on different data sizes. The
proposed model estimates structured query execution time for
large data size by exploiting data value distribution without
actually generating high volume data. The model is validated
against three lab implementation of real life applications and
TPC-H benchmarks.
Keywords—Data Volume; Model; Structured Big data; Query
tuning; Evaluation time
I. INTRODUCTION
The proliferation of big data applications, which are
dynamic in velocity, volume and variety of data, leads to
development of new benchmarks for evaluating their
performance. The complexity of defining benchmark for big
data has been discussed in [7]. Recently, BigDataBench [1] has
defined 19 workloads, based on diversity and type of data. This
kind of extensive benchmark suites increases the evaluation
time due to time spent in generation and loading of data of large
sizes and executing multiple workloads of different types and
sizes. Xion[5] has discussed PCA approach of reducing the
evaluation time by finding out a valid minimal subset of
workload representative which includes the same characteristics
as the benchmark suite without missing out any diversity.
However, these benchmarks still need to be executed multiple
times on different data sizes to evaluate performance sensitivity
to the data volume.
The big data application benchmarks could be classified
into four quadrants based on the variety of data and mode of
accessing it [4] – structured and SQL, structured and NOSQL,
unstructured and SQL and unstructured and NOSQL. The
conventional approach of evaluating benchmarks on different
data sizes will spent huge time generating and loading of data
before the actual execution which may contribute to large
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
percentage of the total evaluation cycle. In order to save time,
most of the benchmark queries may be tuned to execute fast on
large data sizes based on their execution on small data size.
However, the efficiency and accuracy of tuning may not be
guaranteed. PIQL [9] is a programming framework for writing
application‟s structured queries, over key/value store, which
ensure that the query execution time is at most SLO (service
level objective) with increase in data size i.e. query is scale
independent. The paper has classified queries into constant,
bounded, linear and non-linear depending on the relation
between increases in query execution time with increase in data
size. We shall focus, in this paper, on the constant, bounded and
linear structured queries for relational databases which fall in
the first quadrant [4] of the benchmarks for the efficient tuning
of their queries.
We propose an approach of reducing benchmark cycle time
which could be applied for any big data system architecture
centralized, distributed or parallel provided appropriate query
performance prediction models as function of data sizes are
available. In this paper, however we focus on reducing
benchmark cycle time for only structured centralized RDBMS
based Big data systems using SQL query prediction models
such as linear regression, for efficient tuning of benchmark
queries on large data sizes without actually generating it. An
intuitive approach is to estimate a query execution time as
linearly proportional to data size, however, in structured big
databases, a query‟s execution time depends on the query
execution plan including data access paths , as decided by the
optimizer; this execution plan may be different on different data
sizes for the same query. Here, the challenge is first to identify
which type of queries of an application will follow the constant
or linear model for different data sizes, and for what data size
performance should be measured which could be linearly
extrapolated for correct prediction. The main contribution of the
paper is proposing an approach to reduce benchmarking cycle
time by using query performance prediction models. The paper
also talks about an approach for identifying queries whose
execution time may be predicted using simple performance
prediction model such as linear regression. We show the
applicability of the approach on single RDBMS for large data
sizes.
The paper is organized as follows. The Section II talks
about the benchmarking process. The Section III discusses
efficient tuning of benchmark queries using a model to predict
query‟s execution time on large data sizes. Section IV presents
validation of the model on few of the queries from TPC-H
benchmark and three lab implementations of real life
applications – vehicle insurance, opinion and trends mining
from feedbacks and employee performance evaluation system.
Finally the paper is concluded in Section V.
II. BIG DATA BENCHMARKING PROCESS
Benchmarking of big data application goes through a cycle
of planning, generating data, generating tests, execution and
analysis and evaluation [6]. The benchmarking cycle is
repeated for different data sizes for complete evaluation. Time
spent in each cycle is largely governed by the time spent in data
generation, data loading and execution. Data generation and
loading time depends on the parallelism in the data generation
tool and data processing engine (e.g. Hive, RDBMS)
respectively. Execution time may be reduced by tuning the
benchmark queries to finish quickly on the system. The tuning
of a query is generally based on its execution on test data which
may not be accurate and may have longer execution time than
the best tuned query. An iterative process with a model to
estimate execution time for all possible tuning of a query may
help to choose the best tuned query. Moreover, if one has
knowledge of estimated time taken by the queries, the
scheduling of workload could be decided to speed up the
evaluation process.
III. TUNING BENCHMARK QUERY USING PERFORMANCE
PREDICTION MODEL
Benchmark evaluation could be for different data sizes or
different workload. A model predicting execution time of a
workload or queries could help in scheduling the queries
executions or workloads executions to reduce the benchmark
cycle time. Ganpathi[10] and Surajit[11] have proposed
statistical approach to predict a new query execution time in
isolation using past queries executions. Similarly Wento Wo
[12] and Duggan [13] have proposed model for predicting
execution time of a concurrent query and complete workload
time. These models may be applied in the approach discussed
in Fig 1. The approach is applicable for different architectures
of Big data systems. However, the availability of SQL
performance prediction models is a challenge. We have
proposed use of simple linear regression models based on data
distribution for predicting SQL query execution time for large
data sizes without actually generating and loading the data.
Figure.1. Use of Performance Prediction Model to Reduce Benchmark Cycle
Time.
Most of the existing proposed performance prediction
model works either for concurrent workload or a new query in
isolation on a given data size. They do not address the impact of
increase in data volume on query execution time. We shall also
outline a model as discussed below to address the impact of
data volume on query execution time.
Most of the popular relational databases are using cost
based optimizers which generate query execution plan before
its actual execution. These plans depend on the size and
distribution of the data in the database. These plans can be
viewed using query visualizers such as Picasso [8] to
understand different steps of the query execution. Similarly, a
HiveQL execution time depends on the input data distribution
which determine size of input to each „reduce‟ task in its map-
reduce executions. A HiveQL may have many Map-Reduce
jobs and we can use performance prediction model for map-
reduce jobs and HiveQL for tuning the Hive based queries [11].
A naïve tuning approach based on query execution plan on
small data size may not be the efficient tuning because the plan
may not be the same on the large data size. Another approach
may be to get the query execution plan on large data size using
database emulators [2], which uses the linear extrapolation of
database statistics to get query execution plan for large data
size. The tuning using only query execution plan may not give
optimized query since „cost‟ used in generating query plan
does not reflect correctly the query execution time. The „cost‟
indicates the unit of disk and CPU resource utilizations but do
not take into account of effect of various database and OS
caches. Moreover, for complex queries, the cardinality
estimates in the query plan may differ with the actuals; this can
impact the change in estimated execution time.
We propose an approach using mathematical based model
which can classify a query to have its execution time as
constant, bounded, linear and non-linear with increase in data
size. This classification has been discussed for PIQL[9]
framework. The query whose execution time is invariant to
increase in data size are classified as „constant‟, e.g. query with
„=‟ predicate or with predicate „<‟ on primary key. The query
whose execution time is bounded in certain limits and does not
go beyond that even with increase in data size are classified as
„bounded‟, e.g. single table query with operator „LIMIT‟. The
query whose execution time increases linearly with increase in
data size is classified as „linear‟ e.g. full table scan queries. The
query whose execution time increases non-linearly with
increase in data size is classified as „non-linear‟ e.g. complex
query involving many joins and large intermediate data.
The model shall not only generates the estimated execution
plan of a query but also predict its execution time within 10%
prediction error. The model for complex queries showing non-
linear increase in time with data size includes analytical
modeling of database cache and disk spill over operators. In this
paper we focus on applications having majority of their queries
classified as constant, bounded and linear. The model for non-
linear classified queries is not the focus of this paper.
A tuned query and database schema with projected data size
are given to the model to get its estimated execution time on the
projected data size. This exercise may be repeated for the
various tuning parameters of the same query and the one with
lowest estimated execution time could be chosen for actual
execution in the evaluation. Please note that this exercise may
be carried out in parallel to the data generation for the next
evaluation; this ensures better utilization of benchmarking cycle
time without any time overhead by the use of the proposed
model. The model for predicting query execution time is based
on data characteristics of the application, which may be
captured from the application requirement specifications.
A. Data Value Distributions
For an application (or benchmark) data volume increases
over a period of time which may affect the application
performance. We may perceive that the increase in data volume
will uniformly impact all functionalities (or queries) of the
application. However, execution time of different queries may
get impacted differently depending on the value distribution of
the data on which the query is executing.
To explain our model, we shall focus on only numeric
columns of table. Data distribution in a column is defined by
the range <min,max>, where min and max are the minimum
and maximum values respectively that the column can hold.
Increase in size of the table leads to change in any or all of the
min value, max value and the distribution of the values in the
column; this change may also be dependent on distribution of
data values in column of another table in the database; these are
referred as independent and dependent value distributions.
The column in a table which is not co-related with any other
column of tables in the database may exhibit independent value
distribution, where the distribution of the column values is
independent of any other column and depends only on the
increase in data size.
The data value distribution of a foreign column in a table
depends on the corresponding primary column value
distribution. If the primary column is invariant to the data size,
then the foreign column‟s min and max remain the same but the
average number of data items per value increases for all values
in the range <min max>. If the primary column values are also
increased with increase in data size then the max value of the
foreign column will change and the average number of data
items per value may remain the same.
B. Model to Predict Query Execution Time
The model uses the optimizer based database‟s capability of
generating a query execution plan before its actual execution.
Most of the commercial RDBMSs support this feature. This
feature could be exploited to get query execution plan for
content less databases as well by manipulating the database
statistics. For a given application, a database of small size
could be generated to measure underlying infrastructure and
application query‟s‟ performance. However, the corresponding
large size RDBMS is only emulated, to get query execution
plan, by changing the database statistics which are used by the
optimizer [2]. This is a mathematical model and could be
repeated for different data sizes to emulate the respective
instances.
The model uses the execution plan of an application‟s query on
large data size, its data distribution of the involved columns
and measurements on small data size to do prediction as shown
in Fig 2. The total execution time of a query is predicted as
summation of estimated time of each step in the query
execution plan on the emulated large sized database. Each step
of the query can be executed as an independent query with the
same plan as on large data size using RDBMS‟s optimizer
features e.g. hints in oracle.
The mathematical formulation of the model for measurements
done at small size 'S' and prediction of execution time for the
query on large size 'L' is as follows.
𝑃𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝐸𝑥𝑒𝑐𝑢𝑡𝑖𝑜𝑛 𝑇𝑖𝑚𝑒
= 𝑠𝑡𝑒𝑝 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑 𝑒𝑥𝑒𝑐𝑢𝑡𝑖𝑜𝑛 𝑡𝑖𝑚𝑒
𝑎𝑙𝑙 𝑠𝑡𝑒𝑝𝑠
𝑆𝑡𝑒𝑝 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑 𝑒𝑥𝑒𝑐𝑢𝑡𝑖𝑜𝑛 𝑡𝑖𝑚𝑒
𝐿
= × 𝑠𝑡𝑒𝑝 𝑚𝑒𝑎𝑠𝑢𝑟𝑒𝑑 𝑡𝑖𝑚𝑒
𝑆
A „constant‟ query execution time is invariant to data size
since it operates on fixed size of data and returns fixed amount
of data. A „constant‟ query execution time for any data size is
predicted same as measured on the small data size.
In PIQL[9] framework, a „bounded‟ query execution time for
key/value store is predicted using statistical model build by
various executions of the operator. However, in RDBMS,
actual number of operations may differ depending on the
qualified data blocks. In RDBMS, a bounded query may be a
query on single table with „LIMIT‟, on number of rows to be
returned, in where clause. The optimizer, in such case will read
only those number of blocks, referred as limit_blocks,
satisfying the number of rows specified by LIMIT. The upper
bound on the execution time of such queries may be linearly
extrapolated on limit_blocks using measurements of the query
on small data size.
A „linear‟ query execution is linearly extrapolated on number
of operations and number of blocks reads using measurements
on small data size. We have observed that majority of the
queries have either Full table Scan, Index Scan, Hash join,
Nested Loop join or Sort Merge Join. So, we will discuss
performance prediction of application‟s queries with these
operators. However, the approach is not limited to these
operators. The objective here is to emphasize that good
understanding of application helps in estimating its queries‟
execution time.
Figure.2. Model for Predicting Query Execution Time on Large Data Sizes
Full Table Scan
For a query/step accessing only single table with linear
increase in its data values with increase in data size and
involving only full table scan or index scan on columns with
independent data distribution, their execution time prediction
has been discussed in [3], which is basically a linear model . invariant to the data size, the join execution time will be linear
The model linearly extrapolates the query execution time using to increase in data size.
the measurements on the small data size; however the key point For queries with Hash join operation/Sort Merge join not
is to define how small the data size should be so that linear having fixed number of rows from the first/both table(s) will
model works because on very small data size the OS overheads not obey linearity because of disk spill over during
are high and the model may over extrapolate the query hashing/sorting. Similarly, for Nested loop with growing
execution time. number of rows filtered from the first table, the execution time
will be driven by the number of DB cache and OS cache misses
Index Scan which may not be linear. These complexities for „non-linear‟
For a query involving index scan on column having queries have been handled in our different model and are not in
independent data distribution, the data read and matching scope of this paper.
operations may increase linearly with uniform increase in data
size of the accessed table. IV. MODEL VALIDATIONS
For a query involving index scan on column having The model is validated for „constant‟ and „linear‟ classified
dependent data distribution, the repeated data block accesses queries of TPC-H benchmarks and three other lab
may not be served from DB and OS caches depending on the implementation version of real life applications - Vehicle
temporal gap between repeated accesses. This may lead to non- Insurance (VINS) application, Opinion and trends (OPTRA)
linear increase in query execution time with data growth. We mining application, and Employee performance evaluation
have discussed an analytical model for both types of dependent (GSPEED) application. The proposed model could be applied
data distributions (as discussed in Section 3.1) to predict DB for 100%, 75%, 50%, 10% queries of OPTRA (7 queries),
and OS cache misses and their impact on a query execution GSPEED (17 queries), VINS (10 queries) and TPC-H (22
time in one of our internal work[10]. In brief, the DB and OS queries) applications respectively. For validation, the data is
cache misses are analytically calculated based on data value generated using dbgen, an open source data generator for TPC-
distribution and data value placements in the table and data H and our lab implemented data generators for all other
value access pattern from the query; these are captured from the applications. The experiments are conducted on Intel dual core,
application requirement specifications and query design. The 4 GB RAM and 1TB SAN. The RDBMS considered is oracle
caches average latencies are measured on the small data size 11g for all applications and TPC-H benchmarks. The queries
system by running query three times and keeping each cache execution time measurements are collected for all applications
warm one by one– both OS and DB cache are cold, only OS on their respective small size databases (upto 1 GB size for
cache warm, both of them are warm. accessed table) and prediction is done for 40 GB database for
VINS and OPTRA, 128GB for TPCH application and 48 GB
Hash Join for GSPEED application limited by the availability of the
Hash join execution time is determined as execution time of storage space. The experiments are conducted after flushing the
accessing and hashing of the first table, execution time of database and operating system caches to ensure that data is read
accessing and matching of the second table. The join execution from the storage subsystem only.
time is primarily dependent on the number of rows filtered from In TPC-H most of the queries are „non-linear‟ except Q1
the first table (hashed table) in the join operation - If the and Q6 query templates which are on lineitem table. The table
number of rows filtered from the first table is invariant to the size increases linearly and hence their access time is also linear
size of the table, then the hash join operation could be linearly to increase in data size.
extrapolated from its measurements on the small data size. For TABLE 1: APPLICATIONS WITH THEIR QUERY SAMPLE
example, query of application GSPEED in Table 1 as discussed Query Execution Plan
Appln
in Section IV TPCH Select SELECT STATEMENT
. sum(l_extendedprice * SORT AGGREGATE
Nested Loop Join l_discount) as revenue TABLE ACCESS FULL LINEITEM
from lineitem
Nested loop join execution time depends on the first table‟s Where l_shipdate >=
access time, number of rows filtered from the first table and date '1994-01-01'
repeated index scan access time of the second table, where the and l_shipdate < date
'1994-01-01' + interval
access pattern in the second table is driven by the values '1' year
returned from the first table. If the filtered rows are invariant to and l_discount between
0.02 - 0.01 and 0.02 +
the size of the data, then query execution time will be invariant 0.01 and l_quantity <
if the index scan is on the column with independent data 25;
distribution or, the query execution time will be decided by the VINS SELECT SELECT STATEMENT
v.vin,v.regno,v.regdate, NESTED LOOPS
cache analytical model as mentioned before depending on the v.make,v.model,v.fuelt NESTED LOOPS
type of dependent data distribution. For example, query of ype,v.cc,v.purchasepric HASH JOIN
application VINS in Table 1 as discussed in Section IV. e FROM vehicle v, TABLE ACCESS FULL
customer c, CUSTOMER_DETAILS
customer_details cd TABLE ACCESS FULL
Sort Merge Join where v.vin=c.vin and CUSTOMER
In Sort Merge join, the execution time is access time of both c.traceid=cd.traceid INDEX UNIQUE SCAN PK_VEH
and cd.uname='1'; TABLE ACCESS
the tables and sorting of rows returned from each of the table. If BY INDEX ROWID VEHICLE
the number of rows accessed from the tables for sorting is
OPTRA SELECT distinct SELECT STATEMENT The query execution time prediction error percentage for
allwords.sen_id as SORT UNIQUE
sen_id FROM allwords HASH JOIN each query is and is plotted in Fig 3 for all these applications .It
WHERE HASH JOIN is calculated as:
allwords.sen_id IN HASH JOIN
(SELECT SEN_ID TABLE ACCESS FULL
FROM repoview1) DOCUMENT 100 ∗ (𝑎𝑐𝑡𝑢𝑎𝑙 𝑒𝑥𝑒𝑐𝑢𝑡𝑖𝑜𝑛 𝑡𝑖𝑚𝑒 − 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 𝑒𝑥𝑒𝑐𝑢𝑡𝑖𝑜𝑛 𝑡𝑖𝑚𝑒)
ORDER BY SEN_ID;
[CREATE VIEW
TABLE ACCESS FULL
FEEDBACK
𝑎𝑐𝑡𝑢𝑎𝑙 𝑒𝑥𝑒𝑐𝑢𝑡𝑖𝑜𝑛 𝑡𝑖𝑚𝑒
repoview1 AS TABLE ACCESS FULL
SELECT s.sen_id SENTENCES
FROM sentences s, TABLE ACCESS FULL We could see that linear model gives around 10% average error
feedback f, document d ALLWORDS
WHERE (((s.fb_id = for these queries. For some instances the error goes beyond
f.fb_id) AND 10% which may get reduced by averaging multiple
(f.docu_id = measurements on each instance.
d.docu_id)) AND
(d.repo_name
='repo_1'))]
GSPEED SELECT SELECT STATEMENT
Count(sequence_no) SORT AGGREGATE
cnt HASH JOIN
FROM TABLE ACCESS FULL
spd_appraisee_goal_sh SPD_APPRAISEE_GOAL_SHEET_MPG
eet_mpg a, TABLE ACCESS FULL
spd_emp_ipf b SPD_EMP_IPF
WHERE
a.goal_attr_sheet_no =
b.goal_attr_sheet_no
AND a.sequence_no =
1;

In VINS, most of the queries in the application have Nested

Loop join with first table returning constant number of rows as
shown in Table 1.

For the sample query in Table 1, the inner hash join

operation is on Customer_details and Customer, where fixed
number of rows are filtered from the first table independent of
the data size (because of cd.uname=1 and uname has
independent value distribution). The Hash join execution time
could be linearly extrapolated since customer table shows linear
growth. The Nested loop join always gets fixed number of rows
from the inner Hash join and perform unique index scan on
vehicle table. The execution time of this Nested loop join is
invariant to the data size. Therefore, the execution time of the
query is linearly extrapolated to the data size.data size.
OPTRA application is mining of trends and opinions from
feedbacks. All the queries in this application are based on view,
repoview, for which definition is shown in Table 1. The view
is a join of document, sentences and feedback tables. The inner
hash join with document table as the first table, is linearly
extrapolated since constant number of rows is filtered from the
document table. Therefore, the performance of the view query
is linear to the size of the database.
In GSPEED application, most of the queries are full table
scan or hash join – a sample query as shown in Table 1. In the
hash join operation, the spd_appriase_goal_sheet_mpg always
returns constant number of rows (because sequence_id=1 and
sequence_id has independent value distribution), therefore
linear model could be used to estimate the query execution time
for large data sizes. Q9 of GSPEED as shown in Fig 3 is based
on spd_home_page table having constant number of rows
therefore this query execution time is invariant to the size of the
database.
Figure.3. Query Execution Time Prediction Error using the Model
V. CONCLUSIONS
We have proposed an approach to reduce benchmark cycle
time for Big data applications for larger data sizes. We have
proposed a SQL query performance model using linear
regression to efficiently tune the benchmark queries while the
data is being generated, to reduce the cycle time for evaluation.
We have used query classification used in PIQL[9] and
presented the cases where a linear model can be used, for
structured relational database benchmark to estimate its query
execution time for different large data sizes without actually
generating the database. The proposed model makes use of
application data value distributions and query execution plan
for fitment of the linear model. We have validated the model
for TPC-H benchmark and three lab implementation of real life
applications on RDBMS where majority of their queries exhibit
data value distributions conducive for linear prediction model.
The average error in predicting application query execution
time for large data size is reported as 10%. In future, we plan to
build model for other Big data systems as well such as Hbase
and Hive.
REFERENCES
[1] Lei Wang, Jianfeng Zhan, Chunjie Luo, Yuqing Zhu, Qiang Yang,
Yongqiang He, Wanling Gao, Zhen Jia1, Yingjie Shi1, Shujie Zhang,
Chen Zheng,Gang Lu, Kent Zhan, Xiaona Li and Bizhu Qiu,
 “BigDataBench:A Big Data Benchmark Suite from Internet
Applications”, the 20th IEEE International Symposium On High
Performance Computer Architecture (HPCA-2014), February 15-19,
2014, Orlando, Florida, USA.
[2] R. Singhal and M. Nambiar, “Extrapolation of SQL Query Elapsed
Response Time at Application Development Stage”, Proceedings of
INDICON 2012, Kochi, India, December 2012.
[3] R. Singhal and M. Nambiar, “Measurement based model to study the
effect of increase in data size on query response time, Performance a nd
Capacity CMG 2013, La Jolla, California, November 2013.
[4] N. Raghunath, C. Ramesh and J. Ashok, “Data Management – A Look
Back and a Look Ahead”, Proceedings of WBDB 2012, Pune, India.
[5] Wen Xiong, Zhibin Yu*, Zhendong Bei, Juanjuan Zhao, Fan Zhang,
Yubin Zou, Xue Bai, Ye Li and Chengzhong Xu , “A characterization of
big data Systems”, Proceedings of IEEE International conference on Big
Data, 2013.
[6] Rui Han and Xiaoyi Lu, “On big data benchmarking”,
http://arxiv.org/abs/1402.5194, Cornell University, February 2014.
[7] Chaitanya Baru,Milind Bhandarkar, Raghunath Nambiar, Meikel Poess
and Tilmann Rabl, “Benchmarking big data systems and the big data top
100 list”, Journal of BIG DATA, Vol 1, No 1, March 2013, DOI:
10.1089/big.2013.1509.
[8] Jayant Haritsa,”The Picasso Database Query Optimizer Visualizer”,
Proceedings of the VLDB Endowment, Vol. 3, No. 2, 2010
[9] Michael Armbrust, Kristal Curtis, Tim Kraska Armando, Fox Michael, J.
Franklin and David A. Patterson,” PIQL: SuccessTolerant Query
Processing in the Cloud”, Proceedings of the VLDB Endowment, Vol. 5,
No. 3, August 2012, Istanbul, Turkey.
[10] R Singhal, M. Nambiar, Predicting “Cache Behaviour for Non-Unique
Index Access with Data Growth” , TACTics, TCS, India, Pune, March
2014.
[11] Zhang, Zhuoyao and Cherkasova, Ludmila and Loo, Boon Thau,
“Benchmarking Approach for Designing a Mapreduce Performance
Model”, Proceedings of the 4th ACM/SPEC International Conference on
Performance Engineering, ICPE '13, 2013
 Robust Blind Watermarking Technique for Color
Images using Online Sequential Extreme Learning
Machine
Ankit Rajpal Anurag Mishra Rajni Bala
Department of Computer Science Department of Electronics Department of Computer Science
Deendayal Upadhyay College, Deendayal Upadhyay College, Deendayal Upadhyay College,
University of Delhi, University of Delhi, University of Delhi,
New Delhi, India New Delhi, India New Delhi, India
ankit.cs.du@gmail.com anurag_cse2003@yahoo.com r_dagar@yahoo.com

Abstract—In this paper, a robust blind color image
watermarking technique using Online Sequential Extreme
Learning Machine (OS-ELM) is proposed. Blue channel is
utilized and transformed using DWT. Low frequency LL4 sub-
band is used for watermark embedding. A variant of mini-batch
machine learning algorithm i.e. OS-ELM is initially tuned with a
fixed number of training data used in its initial phase and size of
block of data learned by it in each step. The training data to OS-
ELM is constructed by combining the quantized and desired LL4
sub-band coefficients of the DWT domain. A random key decides
the starting watermark embedding position of the coefficients.
Two binary images are used as watermark. The robustness
towards common image processing attacks is enhanced using this
process. Experimental results show that the extracted
watermarks from watermarked and attacked images are similar
to the original watermarks. Computed time spans for embedding
and extraction are of the order of milliseconds, which is suitable
for developing real time watermarking applications.
Keywords—Blind Watermarking; Extreme Learning
Machine;Online Sequential; Normalized Correlation (NC); PSNR;
MSSIM; BER
I. INTRODUCTION
The explosive usage and growth of internet using advance
communication networks have led to a very alarming scenario
wherein illegal copying, reproduction and distribution of
original multimedia content has become easier and faster.
There are now significant opportunities to pirate the
copyrighted digital multimedia products. Therefore,
development of robust digital watermarking techniques has
acquired very special status in the research domain of
multimedia security [1]. Cox et al. [2] argued that a watermark
must be placed in perceptually significant components of the
signal for it to be robust to common signal processing
distortions and malicious attacks. Dey et al. [3] proposed a
robust biomedical content authentication system by
embedding different logo of the hospital or multiple electronic
patient records (EPR) within the retinal image using DWT-
DCT-SVD based approach of watermarking combined with
firefly algorithm. Mishra et al. [4] have proposed an
informed/non-blind watermarking scheme based on ELM. The
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
fact that watermark is distributed irregularly over the
transformed image makes it difficult for the attacker to remove
watermark or modify the watermarked content. The DWT
based watermarking algorithms have gained more popularity
as they end up giving better results in terms of visual
imperceptibility and robustness against common image
processing attacks [5-6]. Presently, the problem of
watermarking of images has been converged to be an
optimization problem wherein the twin requirements namely
visual quality of the watermarked image and robustness of the
embedding algorithm must be balanced out. Many soft
computing techniques have been employed to this end. The
visual quality of the watermarked and attacked images is
assessed by computing PSNR and MSSIM parameters while
multiple image processing attacks are implemented over
watermarked images to examine the issue of robustness. For
this purpose, usually two parameters - Normalized Correlation
(NC) and Bit Error Rate (BER) on case to case basis are used
to quantify and assess the degree of similarity between
embedded and recovered watermarks. Many research groups
have proposed different soft computing techniques, especially
meta-heuristic algorithms to develop robust watermarking
algorithms. Among these, adaptive meta-heuristic techniques
(AMHTs) are used to optimize the numerical values of scaling
factors or embedding strength while integrating the
watermarks with image coefficients in transform domain [7].
In addition to this, Artificial Neural Networks (ANNs) have
been used to embed and extract watermarks. Other techniques,
especially hybrid networks such as Adaptive Neuro-Fuzzy
Inference Systems (ANFIS) [8] and machine learning
techniques such as Support Vector Regression (SVR) and its
different variants are also used for image watermarking [9-10].
Pal et al. [11] proposed a reversible watermarking method
(Odd-Even Method) used for watermark insertion and
extraction in a bio medical image with large data hiding
capacity, security as well as high watermarked quality. Cheng-
Ri. Piao et al. [12] proposed a blind watermark embedding
and extraction algorithm using RBF Neural Network. The
interpolation method and few trigonometric functions can also
be used to embed secret bits into the gray planes of color
image [13]. Similarly, Der-Chyuan Lou et al. proposed new
healthcare image watermarking scheme based on HVS model
and back-propagation network (BPN) technique [14]. Their
experimental results indicate that this technique could survive
various image processing attacks including JPEG lossy
compression. Qun-Ting Yang et al. [15] have proposed a color
image oblivious watermarking scheme based on BPN and
DWT. We observe that primitive neural network training
algorithms such as gradient descent optimization based Back
Propagation Network (BPN) often suffer from various
drawbacks of long training time, multiple local minima etc.
[16]. Thus, it tends to waste time during training. But, they are
found to be adaptive in nature and operation. Although, the
fuzzy inference based schemes do not suffer from these
problems but they are not adaptive in nature. Therefore, it has
been suggested for this and other similar engineering
applications to make use of hybrid variants such as neuro-
fuzzy systems to achieve better results. These alternatives
have been tried and tested in a major way but they are proved
to be costly in terms of embedding and extraction time [17].
The traditional neural networks such as BPN or RBF are slow
in learning. Usually, it takes several minutes or several hours
to train neural networks for different applications by choosing
control parameters (i.e. learning rate, learning epochs,
stopping criteria and other pre-defined parameters) which
must be appropriately selected in advance. The inappropriate
values for these control parameters may either lead to
unsuccessful training of a neural network or the overfitting of
trained neural network. G. B. Huang et al. [18] have
implemented a batch learning algorithm called extreme
learning machine (ELM) for SLFNs which randomly chooses
the input weights and the hidden neurons’ biases for SLFN
with additive neurons, and then analytically computes the
output weights of SLFNs. This makes its training extremely
fast and is especially suitable for developing real time
applications.
In some online applications, the mini-batch sequential
machine learning algorithms may be preferred over batch
learning algorithms as they do not need to be re-trained for a
upcoming data [19, 20]. In this work, ELM is modified based
on recursive least-squares (RLS) algorithm, which is referred
to as Online Sequential Extreme Learning Machine (OS-ELM)
and is used for watermark embedding and extraction. OS-
ELM can learn the training data one-by one or chunk by chunk
(with fixed or varying size) and discard the training data as
long as the training procedure for those data is completed.
II. RESEARCH FOCUS & CONTRIBUTION
In this paper, the Online Sequential Extreme Learning
Machine (OS-ELM) is used to develop a novel watermark
embedding and extraction in three different color images. The
embedding is done in low frequency coefficients to ensure
robustness of binding the watermark with the coefficients of
the host image in transform domain. Visual quality of the
watermarked images is also found to be good after embedding.
The OS-ELM is used in the DWT domain of the host images,
which ensures better embedding process. More specifically,
two different binary watermarks have been embedded and
extracted from three host images of size 512  512  3 . The
OS-ELM algorithm is used to minimize time lapses to train
the neural network. This is contrary to the performance of
other gradient descent based neural architectures such as BPN.
This will be definitely helpful in extending and developing the
proposed scheme to real time moving multimedia data such as
video both in compressed and uncompressed form. The
processing time spans in milliseconds made our proposed
algorithm fit for developing real time image processing
applications such as color video watermarking. Four different
image processing operations over watermarked images as
attacks are carried out to examine the robustness of the
embedding scheme. These attacks are described in detail in
Section III. Visual quality of the watermarked and attacked
images is quantified by using PSNR and MSSIM metrics.
Note that both these metrics are full reference metrics which
require both the watermarked / attacked image along with the
original image to compute the visual quality. The watermarks
have been extracted in a blind manner by using values
predicted by the OS-ELM algorithm. The extracted
watermarks are matched with the embedded ones and the
degree of matching is evaluated by computing two different
parameters - Normalized Correlation or NC(X, X*) and Bit
Error Rate or BER(X, X*), X and X* being the original and
extracted watermarks. These correlation metrics have been
chosen in the present work as these are widely reported in
literature [23]. It is found that the embedding and extraction
processes are well optimized and the proposed DWT-OSELM
based watermarking scheme is robust enough against the
selected attacks.
III. OS-ELM FORMULATION AND EXPERIMENTAL
DETAILS
A novel color image watermarking scheme using a newly
developed SLFN known as Online Sequential Extreme
Learning Machine (OS-ELM) is implemented. This is done
with a view to achieve fast computation, good generalization
capability and accuracy and due to the reason that it offers a
solution in the form of a system of linear Eqn. H  Y . The
SVD method is used to calculate the Moore-Penrose
generalized inverse [21] of H according to the Eqn. (1).
ˆ  H T H  H T Y
1
(1)
where H is the hidden layer output matrix of the neural
network, ˆ is the estimate of the output weights and Y is the
expected output.
The activation function used in OS-ELM computation is
sigmoid activation function defined in Eqn. (2) is used to train
the OS-ELM.
1
g ( x)  (2)
1  e  x
where  is the gain parameter of sigmoid function.
The OS-ELM consists of two main phases. The first phase -
Boosting phase is to train the SLFNs using the primitive ELM
method with some batch of training data in the initialization
stage and these boosting training data will be discarded  C  (3)
as soon as boosting phase is completed. The required batch of Pi '  OSELM  Round  i  
 Q 
training data is very small, which can be equal to the
number of hidden neurons. In this work, the number of hidden This output is close to the desired output included in the
neurons has been set to 30, the number of initial training data dataset used to train the OS-ELM. The optimized numerical
to 75 and size of block of data learned by OS-ELM in each value of Q = 32 is used for all practical computations.
step to 5. In the second phase - Sequential Learning Phase, the 3. Select the starting location of watermark embedding
OS-ELM will learn the train data one-by-one or chunk-by- coefficient Ci using the random secret key.
chunk and all the training data will be discarded once the 4. Embed the watermark according to the Eqn. (4) which uses
learning procedure on these data is completed. the predicted output of the OS-ELM ( Pi ):
As stated above, a blind color image watermarking scheme Ci'key  Pi ' key    wi (4)
using the OS-ELM in DWT domain is implemented in this
work. For this purpose, LL4 sub-band coefficients are used to where wi is the watermark which is either binary or random
carry out embedding and extraction processes. For embedding
process, blue channel is extracted from the host image. Note
sequence,  is the embedding strength optimized to be equal
'
that the host image is of size 512  512  3 , therefore, the to 0.17 for both watermarks and Ci are the modified LL4 sub
extracted blue channel out of RGB image is of size band coefficients obtained after watermark embedding.
512  512  1. It has been observed that the best results are 5. Perform Inverse DWT to generate watermarked image.
obtained in case the blue channel is used for embedding and
extraction. The OS-ELM is trained with the LL4 sub-band B. Watermark Extraction Algorithm
coefficients. A random key is used to decide the initial Listing 2 gives the sequence of steps used to extract the
location of watermark embedding. Three standard color watermark from watermarked and attacked images in a
images - Airplane, Baboon, and Lena are used to embed complete blind manner.
binary watermarks after training the OS-ELM using the
quantized values of the LL4 sub-band coefficients. The size of Listing 2: Extraction Procedure
the input dataset is 1024 2 while it produces an output 1. Transform the watermarked blue image using 4-level
sequence of size 10241 whose coefficients are close to the DWT transform. Select LL4 ( 32 32 size) sub-band
desired LL4 sub-band coefficients. Two different 32  32 "
coefficients and set Ci to it.
sized binary watermarks are tested in this experimental work. "
2. Quantize Ci by Q, and use the already trained OS-ELM
The watermarked images are tested for visual quality by model to predict the output:
computing two full reference metrics:- PSNR and MSSIM
  C" 
[13]. The watermarked images are also subject to selected Pi"  OSELM  Round  i   (5)
image processing attacks to verify the issue of robustness.   Q 
These attacks are: - (a) JPEG (QF=50, QF=75 and QF=90), (b) '
3. Extract the watermark wi using the Eqn. (6) below, using
Scaling (resized to half and then restored to original size), (c) "
the output of the OS-ELM in Eqn. (5) and Ci and the secret
Gaussian Noise (5% and 10%), and (d) Salt and pepper (0.1 %
and 0.5%). Blind extraction of the watermarks from the key.
watermarked images is done before and after executing image wi'  ( Pi"  Ci" )  (1 /  ) (6)
processing attacks. Both embedding and extraction are done
using the same key and the same OS-ELM model. In this where  is the embedding strength.
scenario, only the watermarked or the attacked image is
required to recover the watermark (blind extraction) by The watermark extraction is carried out using the algorithm
predicting the output of the OS-ELM. A comprehensive given in Listing 2 and the normalized correlation NC (W, W’)
analysis of the results obtained in this simulation is given in is computed between the embedded and the extracted
section IV. watermarks. This formulation is given in Eqn. (7).
A. Watermark Embedding Algorithm
n
Listing 1 gives the sequence of steps used to carry out the  W (i)W ' (i) (7)
embedding process. NC (W , W ' )  i 1
n n

Listing 1: Embedding Procedure  W (i)  W ' (i)

i 1
2

i 1
2

1. Transform the extracted blue image using 4-level DWT

transform. Select LL4 ( 32 32 size) sub-band coefficients and
where W is the watermark of dimensions n n . The
set Ci to it.
2. Quantize Ci by Q, as the input value of OS-ELM and the magnitude range of NC varies between [0 and 1], the unity
value gives exact matching between the two watermarks [22].
desired output as Ci . The dataset supplied to the OS-ELM is
Another parameter Bit-Error Rate or Bit-Error Ratio (BER) is
of size 1024 2 . This OS-ELM constructs the model and
a unit less performance measure, is an approximated estimate
predicts the output of size 1024  1 as given by Eqn. (3):
of the bit error probability [23]. The formulation is depicted in
(d) (e)
Eqn. (8).
Fig. 1. Original host images- (a) Baboon (b) Airplane (c) Lena (d) Binary
Number _ of _ Bits _ in _ Error Watermark #1 (e) Binary Watermark #2
BER  (8)
total _ number _ of _ bits
The high PSNR value and MSSIM value near unity show that
The BER expression is depicted in Eqn. (9): the visual quality of these images is very good. Fig. 3(a-c)
depicts three binary watermarks recovered from images shown
mn
Wt  Wt '
BER   (9) in Fig. 2(a-c) respectively.
t 1 mn
PSNR= 47.190070 MSSIM= 0.9942

where, W denotes the original watermark and W’ denotes the

recovered watermark. Both W and W’ are of size m  n .

IV. RESULTS AND DISCUSSION

Fig. 1(a-c) represents three color host images – Airplane,
Baboon, and Lena. Fig. 1(d-e) shows the two 32  32 sized
binary watermark (#1 and #2). These binary watermarks are
embedded within chosen host images to obtain watermarked (a)
images depicted in Fig. 2(a-c). Their respective PSNR and
MSSIM values are mentioned above these images.
PSNR= 47.31025 MSSIM= 0.9994

(a)
(b)

PSNR= 47.111105 MSSIM=0.9996

(b)
(c)

Fig. 2. Watermarked Images using CS (Binary Watermark #1)- (a) Airplane

(b) Baboon and, (c) Lena.

(b) (c)
(a)

Fig. 3. (a-c) Extracted Watermarks from images of Fig. 4(a-c) respectively

(NC=1.0 in each case)
(c)
Fig. 4(a-c) depicts the Watermarked images- Airplane,
Baboon and Lena obtained after embedding the second binary
watermark in Fig. 1(e) into images depicted in Fig. 1(a-c)
respectively. Fig. 5(a-c) depicts three binary watermarks
recovered from images shown in Fig. 4(a-c) respectively.
PSNR= 44.065734 MSSIM= 0.9899
this, the OS-ELM training time, embedding and extraction
time spans are computed in this simulation. Again for Lena
these are respectively computed to be 125.00 milliseconds and
468.75 milliseconds and 140.62 milliseconds. The embedding
time in this case includes training time. These clearly indicate
that the proposed watermarking scheme is well suitable to
satisfy real time constraints. This makes it clearly suitable for
developing real time video watermarking applications.

The watermarked images are exposed to selected few image

processing attacks. A brief description of these attacks is given
in section III. The results are presented here:

A. JPEG Compression
JPEG Compression with Quality factor (Q.F.) =50, 75 and 90
(a) is used. In case of Binary watermark #1 and #2, the recovered
watermarks are depicted in Table I and the NC and BER
PSNR= 43.998204 MSSIM= 0.9989 values indicates the extracted watermarks are still
recognizable.

TABLE I. THE EXPERIMENT RESULT AFTER JPEG COMPRESSION

Attack Image PSNR NC Extracted

(dB)/ (W,W’)/ Watermark
MSSIM BER
(W,W’)
30.96/ 0.3286/
(b) Airplane 0.9175 0.3301
JPEG 24.82/ 0.3483/
PSNR= 44.091871 MSSIM= 0.9992 (QF=50) Baboon 0.8842 0.3262
Lena
31.43/ 0.2916/
0.9811 0.3564
30.86/ 0.3558/
Airplane 0.9148 0.3379
JPEG 24.79/ 0.3485/
(QF=50) Baboon 0.8833 0.3535
Lena
31.31/ 0.3352/
0.9806 0.3594
(c)
32.46/ 0.8190/
Airplane 0.9371 0.1436
Fig. 4. Watermarked Images using W (Binary Watermark #2) (a) Airplane JPEG 26.17/ 0.8471/
(b) Baboon and, (c) Lena (QF=75) Baboon 0.9126 0.1318
Lena
32.50/ 0.8095/
0.9849 0.1426
(a) (b) (c)
32.29/ 0.7170/
Fig. 5. (a-c) Extracted Watermarks from images of Fig. 4(a-c) respectively Airplane 0.9332 0.1592
JPEG 26.13/ 0.7177/
It is clear that NC values are quite high and is close to 1.0 in (QF=75) Baboon 0.9117 0.1445
case of both the binary watermarks. It indicates good 32.35/ 0.7189/
Lena
watermark recovery process. Note that the watermark 0.9845 0.1416
recovery is completely blind in the present work. Particularily 33.86/ 0.9583/
in case of Lena host image, an interesting observation in this Airplane 0.9539 0.1064
experimental work is the optimization of  . The numerical JPEG 27.90/ 0.9669/
values of PSNR, MSSIM, NC and BER are found to decrease (QF=90) Baboon 0.9375 0.0966
on either side of this value of  =0.17. Thus, this is taken as 33.90/ 0.9889/
Lena
the suitable embedding strength to carry out watermark 0.9888 0.0957
embedding in case of both binary watermarks. In addition to
 33.64/ 0.8673/ 30.50/ 0.9475/
Airplane 0.9496 0.0762 Airplane 0.7547 0.1064
JPEG 27.84/ 0.8903/ Gaussia 30.48/ 0.9446/
(QF=90) Baboon n Noise Baboon
0.9368 0.0645 10%
0.9656 0.1104
Lena
33.68/ 0.8988/ Lena
30.48/ 0.9475/
0.9883 0.0596 0.9770 0.1064

30.39/ 0.8356/
B. Image Scaling Airplane 0.7547 0.0850
Gaussia
The watermarked image is resized to the size 256 256 and n Noise 30.38/ 0.8574/
Baboon
then restored it to the original size i.e. 512  512 . In case of 10% 0.9651 0.0762
binary watermark #1 and #2 of size 32  32 , high NC (W, Lena
30.41/ 0.8734/
W’) value and low BER(W, W’) value show that the 0.9766 0.0742
watermark is successfully extracted as compiled in Table II. 34.69/ 0.9806/
Airplane 0.9568 0.0986
TABLE II. THE EXPERIMENT RESULT AFTER SCALING AND RESIZING Salt & 35.12/ 0.9889/
Pepper Baboon
Attack Image PSNR NC Extracted 0.9909 0.0977
0.1%
(dB)/ (W,W’)/ Watermark Lena
35.12/ 0.9861/
MSSIM BER 0.9932 0.0986
(W,W’)
30.75/ 0.9806/ 34.37/ 0.9117/
Airplane Airplane 0.9533 0.0557
0.9458 0.1006 Salt &
23.06/ 0.9699/ 34.74/ 0.8881/
Scaling Pepper Baboon
Baboon 0.8081 0.1045 0.9905 0.0674
0.1%
32.26/ 1/ Lena
34.68/ 0.8938/
Lena 0.9924 0.0596
0.9839 0.0938
30.65/ 0.8974/ 27.73/ 0.7865/
Airplane Airplane 0.8185 0.1582
0.9421 0.0635 Salt &
23.04/ 0.8753/ 28.27/ 0.8348/
Scaling Pepper Baboon
Baboon 0.8066 0.0732 0.9579 0.1523
0.5%
32.11/ 0.9228/ Lena
28.08/ 0.8070/
Lena 0.9664 0.1465
0.9835 0.0527
27.78/ 0.7519/
Airplane 0.825 0.1289
C. Image Noising Salt &
Gaussian noise with noise amount=5% and 10% respectively is 28.12/ 0.7138/
Pepper Baboon 0.9569 0.1514
added to the watermarked images. In case of Binary watermark 0.5%
#1 and #2 of size 32  32 , the recovered watermarks are still Lena
28.13/ 0.7450/
0.9670 0.1318
recognizable and shown in Table III.
The results compiled in this section indicate that OS-ELM
TABLE III. THE EXPERIMENT RESULTS AFTER ADDING NOISE algorithm has been quite successful in developing color image
PSNR NC Extracted watermarking scheme which can produce results by fulfilling
Noise (dB)/ (W,W’)/ Watermark real time constraints. The algorithm successfully handles the
Image
Type MSSIM BER selected image processing attacks. We observe that watermark
(W,W’) recovery is good. This is indicated by high computed values
36.87/ 1/ of NC. The corresponding BER values are the lowest as per
Airplane 0.9213 0.0938
Gaussia our expectation. The visual quality of signed and attacked
n Noise 36.90/ 1/ images is also very good. This is indicated by high computed
Baboon 0.9918 0.0938
5% PSNR and MSSIM values. Thus, the OS-ELM algorithm is
Lena
36.86/ 1/ capable to minimize the tradeoff between robustness and
0.9946 0.0938 imperceptibility at a fast processing speed. This makes it
36.47/ 0.9206/ particularly suitable for extending this work to video
Airplane 0.9188 0.0547 watermarking applications.
Gaussia 36.46/ 0.9206/
n Noise Baboon V. CONCLUSIONS
5%
0.9913 0.0537
36.46/ 0.9206/ Copyright protection and image authentication are crucial
Lena
0.9942 0.0537 domain of contemporary multimedia research. To this end,
numerous soft computing techniques based watermarking
schemes have been proposed worldwide. However, they do
not touch upon the crucial requirement of minimizing the
processing time spans. This requirement is found to be an
important one for developing secure multimedia applications
such as image or video watermarking. In this paper, a novel
blind watermarking scheme using single layer feed-forward
neural network (SLFN), commonly known as Online
Sequential Extreme Learning Machine (OS-ELM) is proposed
for color images. We train the OS-ELM by using quantized
LL4 sub-band coefficients of the blue channel of the host
image by taking its 4–level DWT transform. A dataset of size
1024  2 is prepared by using LL4 sub-band coefficients. This
dataset is used to train the OS-ELM network which produces a
output sequence of size 10241 used to carry out embedding.
A random key decides the beginning position of the
coefficients where the watermark is embedded. Two different
binary watermarks of size 32  32 are used in this work. Blind
extraction of the watermarks is carried out. We conclude that
the proposed DWT-OS-ELM based watermarking scheme is
efficient to minimize the tradeoff between the imperceptibility
and robustness. Experimental results show that the extracted
watermarks from watermarked and attacked images are similar
to the original watermark. The computed time spans for
embedding and extraction are in millisecond which makes it
suitable for developing real time watermarking applications.
Overall, the watermarking scheme is well optimized for visual
quality of images and robustness and may be successfully used
to carry out watermarking of uncompressed video sequences.
REFERENCES
[1] Hartung F., Kutter, M.: Mulimedia Watermarking Techniques.
Proceedings of the IEEE (1999) 1079-1094
[2] I. J. Cox, Joe Kilian, F. Thomson Leighton and Talal
Shamoon, “Secure Spread Spectrum Watermarking for
Multimedia”, (1997), IEEE Transactions on Image Processing,
vol. 6(12), pp. 1673-1687
[3] Dey, N., Samanta, S., Chakraborty, S., Das, A., Chaudhuri, S.
and Suri, J., “Firefly Algorithm for Optimization of Scaling
Factors during Embedding of Manifold Medical Information:
An Application in Ophthalmology Imaging,” (2014) , Journal of
Medical Imaging and Health Informatics, 4, pp. 384-394
[4] Anurag Mishra, Amita Goel, Rampal Singh, Girija Chetty,
Lavneet Singh, "A novel image watermarking scheme using
Extreme Learning Machine," (2012), International Joint
Conference on Neural Networks (IJCNN), pp.1-6
[5] Dey, N.; Das, P.; Roy, A.B.; Das, A.; Chaudhuri, S.S., "DWT-
DCT-SVD based intravascular ultrasound video watermarking,"
(2012), World Congress in Information and Communication
Technologies (WICT), pp.224-229
[6] Wang Zhenfei, Zhai Guangqun, Wang Nengchao, “Digital
watermarking algorithm based on wavelet transform and neural
network,” (2006), Wuhan University Journal of Natural
Sciences, 11(6), pp. 1667-1670
[7] N. Dey, S. Samanta, X.-S. Yang, A. Das, and S. S. Chaudhuri,
“Optimisation of Scaling Factors in Electrocardiogram Signal
Watermarking using Cuckoo Search,“ (2013), International
Journal of Bio-Inspired Computation, 4(5), pp. 315-326
[8] Charu Agarwal, Anurag Mishra, Arpita Sharma, “A novel gray-
scale image watermarking using hybrid Fuzzy-BPN
architecture,” (2015), Egyptian Informatics Journal, 16(1), pp.
83-102
[9] Pan-Pan Zheng, Jun Feng, Zhan Li, Ming-quan Zhou, “A novel
SVD and LS-SVM combination algorithm for blind
watermarking,” (2014), Neurocomputing, 142 (22), pp. 520-528
[10] Mehta, R.; Mishra, A.; Singh, R.; Rajpal, N., "Digital Image
Watermarking in DCT Domain Using Finite Newton Support
Vector Regression, (2010), Sixth International Conference on
Intelligent Information Hiding and Multimedia Signal
Processing (IIH-MSP), pp.123-126
[11] Pal, A.K.; Dey, N.; Samanta, S.; Das, A.; Chaudhuri, S.S., "A
hybrid reversible watermarking technique for color biomedical
images," (2013), IEEE International Conference on
Computational Intelligence and Computing Research (ICCIC),
pp.1-6
[12] Cheng-Ri Piao, Seunghwa Beack, Dong-Min Woo, and Seung-
Soo Han, “ A Blind Watermarking algorithm Based on HVS and
RBF Neural Network for Digital Image”, (2006), Springer
Verlag Berlin Heidellberg, pp. 493-496
[13] Chakraborty, S.; Maji, P.; Pal, A.K.; Biswas, D.; Dey, N.,
"Reversible Color Image Watermarking Using Trigonometric
Functions," (2014), International Conference on Signal
Processing and Computing Technologies (ICESC) in Electronic
Systems, pp.105-110
[14] Der-Chyuan Lou, Ming-Chiang Hu, and Jiang-Lung Liu,
“Healthcare Image Watermarking Scheme Based on Human
Visual Model and Back-Propagation Network”, (2008), Journal
of C.C.I.T, Vol 37(1) , pp. 151-162
[15] Qun-Ting Yang, Tie-Gang Gao, Li Fan, "A novel robust
watermarking scheme based on neural network," (2010),
International Conference on Intelligent Computing and
Integrated Systems (ICISS), pp.71-75
[16] Charu Agarwal, Anurag Mishra, Arpita Sharma, "Digital image
watermarking in DCT domain using Fuzzy Inference System,"
(2011), 24th Canadian Conference on Electrical and Computer
Engineering (CCECE), pp.822-825
[17] Charu Agarwal and Anurag Mishra, “A Novel Image
Watermarking Technique using Fuzzy-BP Network”,
(2010), Proceedings of 6th International Conference on
Intelligent Information Hiding and Multimedia Signal
Processing, pp. 102-105
[18] G-B Huang (2004), The Matlab code for ELM is available
on: http://www.ntu.edu.sg/home/egbhuang
[19] G.-B. Huang, N.-Y. Liang, H.-J. Rong, P. Saratchandran, and N.
Sundararajan, “On-line sequential extreme learning machine,”
IASTED International Conference on Computational
Intelligence (CI 2005), Calgary, AB, Canada, Jul. 4–6, 2005.
[20] N.-Y. Liang, G.-B. Huang, P. Saratchandran, and
N. Sundararajan, “A Fast and Accurate Online Sequential
Learning Algorithm for Feedforward Networks”, (2006), IEEE
Transactions on neural networks, VOL. 17(6), pp. 1411-1423.
[21] D. Serre (2002), “Matrices: Theory and Applications”,
Springer Verlag, New York Inc.
[22] Zhou Wang; Bovik, A.C.; Sheikh, H.R.; Simoncelli, E.P.,
"Image quality assessment: from error visibility to structural
similarity," Image Processing, IEEE Transactions on , vol.13,
no.4, pp.600,612, April 2004
[23] Fung, V.; Rappaport, T.S.; Thoma, B., "Bit error simulation for
π/4 DQPSK mobile radio communications using two-ray and
measurement-based impulse response models," (1993), Selected
Areas in Communications, IEEE Journal on , vol.11(3), pp.393-
405
 A Detection Mechanism of DoS Attack using
Adaptive NSA Algorithm in Cloud Environment
Sumana Maiti Chandan Garai
Department of C.S.E Department of C.S.E
NITTTR, Kolkata NITTTR, Kolkata
Kolkata, India Kolkata, India
sumana.maiti.27@gmail.com chandangarai@hotmail.com
Abstract—Security of any distributed system is not only
complex in nature, it also needs much more attention as most of
the applications being used and developed in recent past are on
distributed platform. Denial of Service (DoS) attack causes drop
in quality of service and may also reach to entire absence of
service for some ‘real’ users. Identifying some users as attackers
also need appropriate algorithm. Negative selection algorithm
(NSA) is a very effective approach in identifying some user as
attacker. However declaring some ‘real’ user as an attacker is a
very common limitation of these types of algorithms unless and
until the mechanism of detection is updated at regular intervals.
In this research work we have modified NSA algorithm to take
into account the necessity of updating the detector set from time
to time. We have introduced a second detection module to
accommodate the updation. Both the algorithms are implemented
on common data set and comparative study is presented. Our
proposed algorithm comes out with much improved results and
significantly reduces false positive (false alarm) cases.
Keywords—NSA; DDoS; Feature Vector; IP Spoofing
I. INTRODUCTION
Negative Selection Algorithm (NSA) is one of the very
effective approaches [1-3] to build an Artificial Immune
System (AIS). T-cells of thymus gland of human body, by
eliminating body own cells, detects foreign molecules. This is
a well-accepted approach to detect malicious traffic from
normal traffic. In case of Distributed Denial of Service
(DDoS), both the server (host) and the bandwidth of the
network are consumed by the attacker(s) and as a result the
regular and genuine users are not getting the services to the
desired and committed level. As such it is very hard to detect
between these two types of traffic (genuine and malicious)
correctly. Conventional firewalls mechanism to drop all UDP
packets during UDP flood attack causes drop of genuine (non-
attack) UDP packets also and obviously is not a desirable
solution as it cannot guarantee the interest of the genuine
users. More accurate mechanisms for anomaly detection, fraud
detection etc. are thus required to prevent any loss of interest
of genuine users.
The Novel Neighborhood Negative Selection Algorithm
(NNNSA) [1-6] takes care of some of the limitations of NSA
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Ranjan Dasgupta
Department of C.S.E
NITTTR, Kolkata
Kolkata, India
ranjandasgupta@ieee.org
by introducing better mechanism for neighborhood
representation, novel matching technique etc. However, this
approach lacks in adaptability of the self-learning mechanism
and needs periodical update of the detector set to eliminate
unnecessary presence of the old and irrelevant dataset. This
augmentation will improve the performance of the overall
system by removing false positive cases.
II. REVIEW WORK
A. Denial of Service (DoS) Attack
In Denial of Service attack [20-25], attacker sends huge
numbers of requests to a server that the server fails to respond
to an authorized user in time. In case of Flooding attack like
TCP SYN flood, UDP flood and ICMP flood, attacker sends
excessive amount of packets to waste different resources.
Whereas for Logic attack (such as Ping of death, Land,
Teardrop), attacker triggers some error in software system.
In TCP SYN flood [15-18], UDP flood [18] and ICMP
flood [18], the attacker sends a large number of SYN requests
to a server, UDP packets to random ports of remote host and
ICMP echo requests respectively. In most cases the attacker
uses spoofed IP addresses. As a result, a large number of half
open connection floods the ports capacity.
In Ping of death, the attacker sends a malformed ping
message to a computer [18]. This oversized and malformed
message can cause a buffer overflow which results in system
crash. In land attack [18], attacker sends TCP SYN packets to
a target system by using spoofed IP of the target system. So
the system reply itself continuously.
Teardrop attack [18] uses fragmented IP packets, sent by
the attacker with errors in offset field. Offset field indicates
starting position of the fragments. At the time of reassembling
of the packets, due to error in offset, it cannot be done
correctly and are overlapped.
B. Distributed Denial of Service (DDoS) Attack
Distributed Denial of Service (DDoS) attack [19] [25] is a
type of Denial of Service attack where multiple sources are
used to attack one single system. The targeted systems and all
the other systems which are maliciously used by the attacker
are the victims of DDoS attack. In cloud environment DDoS Traffic pattern based system was proposed in [8-10]. This
attack generally focus on cloud resources which are provided process consists of two parts. Flow header detection and
by cloud service provider. traffic pattern detection. In „flow header detection‟ part it
checks the fields of flow header as shown in Fig. 2[9].

Fig. 2 Flow header

This part mainly detects different logic based attacks. Table
I [9] shows the flow header detection. In this part the fields of
flow header are compared with predefined static threshold.

TABLE I. FLOW HEADER DETECTION

Transport Satisfying Condition Attack
protocol
ICMP (i)echo request(ii)destinationIP Smurf
=broadcast
(i)flow size/packet count is too high Ping of death
(i)packet count=large(ii)flow size =large ICMP
Fig. 1DDoS attack in cloud environment flooding
TCP (i) source IP=destination IP Land
Fig.1 describes the overall process of DDoS attack in (ii) source port=destination port
cloud environment. In the total procedure four layers are (i)packet count=large(ii)flow size=large TCP flooding
present. Attacker is present in Layer 1. Zombie computers, UDP (i)destination port=reflecting port Fraggle
(ii)destination IP= broadcast
Reflector computers and victims are present in Layer 2, Layer
(i)packet count=large(ii)flow size=large UDP flooding
3 and Layer 4 respectively. Attacker initiates this process by
„Traffic pattern detection‟ aggregates related flows and then
sending commands to zombie computers. Zombie computers
generate two hash tables namely source IP based hash table
are those computers which are present in same botnet. In Fig.1
and destination IP based hash table and then different features
Z1, Z2, Z3 are zombie computers. In the next step the zombie
are calculated from those hash tables. Fig. 3[9] shows the
computers send requests to many innocent computers which
traffic pattern detection.
are present in different clouds by spoofing the IP address of
victim computer. Those innocent computers are called
Reflector computer. In Fig.1 R11, R1n, R21, R2n etc. are
reflector computers. As the IP address used here is spoof IP
address the reflector computers think that the requests are
coming from victim IP. So in the next phase all the reflector
computers give reply to the victim which is present in another
cloud. As a result the victim‟s computer is flooded with
traffic.
C. Different DoS attack detection systems Fig. 3 Traffic pattern detection
Detection mechanism for DoS attack works on three major
technical aspects – namely traffic pattern based system [8-10], The features which are used in traffic pattern detection system
sketch based system [11-13] and flow-level system [1] [14]. In are shown in Table II [9].
case of traffic pattern based detection system both the flooding In traffic pattern detection also the features are compared
and logic attack [see sec 2.1] can be detected by using some with predefined static threshold and different flooding based
predefined threshold values. Sketch based detection system attacks are detected. For threshold calculation a time series
mostly detects TCP SYN flood from a combination of TCP analysis is done which is a discussed in [9] detail. Traffic
SYN flood and port scans. Flow-level system on the other pattern based detection system can detect different types of
hand keeps track of the flow of traffic and any sudden or logic and flooding attacks whose methodology is known, but
abrupt change in flow characteristics has been identified as an if attacker attacks with a new methodology then for this type
attack declaring any abrupt change as an attack is an of attacks thresholds are not present, so this type of attack
oversimplified approach and in our proposed work, we cannot be detected by this system. If an attack does not
augment the flow-level detection mechanism and our solution influence network traffic, it is difficult to be detected and
provides a more precise mechanism for detection of attacks static thresholds are used which may not be suitable for every
and thus reduces false positives. network environment.
We provide a brief description of all the three types of
detection mechanism with a comparison of all the above three
mechanisms for sake of completeness.
TABLE II. DESCRIPTION OF FEATURES REQUIRED IN TRAFFIC PATTERN method is used which is given in [13] in details. This method
DETECTION
averages the data in a way that gives less weight to data as
Hash Notation Description they are removed in time to time. Mo(t) represents current
tables #SYN-#SYN/ACK at time interval t, Mf(t) represents the
n(flow) The number of flows with same destination IP forecasted #SYN-#SYN/ACK at the time interval t.
avg(flow_size Average of flow size with same destination IP
) Mf(t)= αMo(t) + (1-α)Mf(t-1) t>2
avg(n_packet) Average of packet count with same destination Mf(t)= Mo(1) t=2
Destina IP
tion IP n(D_port) The total number of destination port with same
based destination IP It is observed in [13] that α varies within 0.1 to 0.9. The
n(ACK) The total number of ACK packets with same difference between the forecasted value and actual value is
destination IP used for detection in sketch based system. It is observed in
n(SYN) The total number of SYN packets with same [12] that most of the DoS attacks are TCP SYN flood. In
destination IP
sketch based system except TCP SYN flood and different
Source sum(n_packet Summation of packet count with same source
IP ) IP types of port scans no other attacks are discussed.
based sum(flow_siz Summation of flow size with same source IP In flow level information system [1] [14], flow information
e)
are stored in either hash tables or tree structure and then
features are extracted to detect anomaly in feature extraction
In case of Sketch based system [11-12] TCP SYN flood is module. Such as in [1] the flow information are stored in tree
detected from combinations of TCP SYN flood and different structure and then many vertical and horizontal features are
port scans. Port scan is a technique that sends client request to extracted and then neighborhood negative selection algorithm
a port and listening for the answer. The response indicates port is used to detect anomaly in detection module. This system
status. It is generally done for future attack. Port scans are can detect unknown attack. The flow level detection system
basically of two types. In vertical scan, attacker sends requests discussed in [1-3] has two modules. Feature extraction module
to several destination ports on a single host. Vertical scanning and detection module. In the feature extraction module a tree
is done to know which port is open and which port is not. As structure as shown in Fig. 4[1] is used to store flow
the attacker knows vulnerability of a particular port and wants information. For TCP, UDP the tree height is six and for
to know in which machine this port is open, in case of ICMP tree height is four. For TCP and UDP first level is the
horizontal scanning attacker targets the same port on several root and the other lower levels are protocol, IP address of
hosts. In this process different combinations of source IP, inside host, port no of inside host, IP address of outside host,
destination IP, and destination port are taken as a key and port number of outside host respectively and for ICMP first
(#SYN-#SYN/ACK) is calculated and then this difference is level is root and other lower levels are protocol, IP address of
compared with threshold. #SYN represents the number of inside host, IP address of outside host respectively as shown in
SYN packets. #SYN/ACK represents the number of Fig. 4[1]. From this tree structure different vertical and
SYN/ACK packets. FlOODING_DIP_SET represents the horizontal features as shown in Table IV [1] are extracted.
attacker‟s IP and FLOODING_SIP_SET represents the Vertical features are those features which are extracted from
victim‟s IP. The overall sketch based detection process is one single flow and horizontal features are those features
shown in following Table III [11-12]. which are extracted from more than one flow.
TABLE III. SKETCH BASED DETECTION SYSTEM
St Key If(#SYN-#SYN/ACK)> threshold
ep
1 {dest The dest IP is included into FLOODING_DIP_SET
IP, dest
port}
{sourc Check if dest IP is If „Yes‟ the source IP is
2 e presents in included into
IP,dest FLOODING_DIP_SE FLOODING_SIP_SET
IP} T? If „No‟ mark it as vertical scan
Fig. 4 Tree structure to store flow information
{sourc Check if source IP If „Yes‟ mark it as non-spoofed
3 e presents in TCP SYN flood In the detection module negative selection algorithm [1-4]
IP,dest FLOODING_SIP_SE If „No‟ mark it as horizontal is used. This module has two phases – training and detection.
port} T? scan.
At the training phase a mature detector set is built and at
Uniqueness of key is defined as the capability of
detecting phase the incoming new data is compared with
differentiating an attack from combinations of different types
mature detector set and anomaly is detected from incoming
of attacks. The key which can detect only one type of attack is
new data. Before training starts, feature vectors are extracted
called best key. E.g. - uniqueness of {destination IP,
from normal traffic in feature extraction module. The feature
destination port} is 1 as this key can detect only TCP SYN
vectors which are fully adjacent but mutually disjoint are
flood. Similarly uniqueness of {source IP, destination IP} is
called neighborhoods that means vectors which belong to
1.5 as 1 for vertical scanning and 0.5 for non-spoofed TCP
same type are called neighborhood. The union of these
SYN flooding. For the threshold calculation in sketch based
neighborhoods is called shape space.
system Exponential Weighted Moving Average (EWMA)
 here the “self-set” in which the total detection process depends
on may not be accurate. Comparisons between different
TABLE IV. VERTICAL AND HORIZONTAL FEATURES detection systems are shown in Table V.
Vertical features Horizontal features
1.Type of protocol, 1. Number of packets which have the same inside IP
III. SCOPE OF WORK
2.Length of flow address with current flow within 2 seconds Flow level system [1-6] is more accurate than sketch based
3.Number of packets 2.Number of bytes which have the same inside IP system [11-12] and traffic pattern system [8-10]. But several
from inside to address with the current flow from outside within 2
outside seconds
challenges still remain. The self-set which is used in the
4.Number of bytes 3.Number of flows which have the same inside IP training phase of Detection Module is not accurate, as the
from inside to address with the current flow within 2 seconds elements of self-set are taken randomly [6]. Suppose an
outside 4.Number of flows which have the same inside IP anomaly unfortunately matches with predefined self-set, the
5.Number of packets address but different protocol with the current flow feature vectors of this anomaly flow are not entered into
from outside to within 2 seconds detector set. So, at the time of detection, this type of attacks
inside 5.Number of flows which have same protocol but
6.Number of bytes different IP address with the current flow within 2 cannot be detected. In other side, if any normal traffic flow
from outside to seconds does not match with predefined elements of self-set, then this
inside flow is entered into detector set. Then this type of normal
Self-set is built by randomly taking neighborhoods. At the traffic flow is detected as an attack. To address these types of
time of training first a neighborhood is selected randomly and problems an adaptive algorithm is required, which will
then affinity of this neighborhood with the every elements of periodically update the self-set. In our proposal we improve
self-set is calculated. The neighborhood which is taken is the self-set with the help of IP Address Database (IAD) which
called candidate detector. This candidate detector is eliminated is discussed in [7]. Here good and bad packets are filtered by
if the affinity is more than training threshold otherwise the comparing them with their previous history. It is observed in
candidate detector is added into detector set. This process is [7] that attack traffic generally contains a false source address,
done iteratively until a mature detector set is built. At the which is normally generated randomly. Our proposal is to
detecting phase, the incoming new feature vector which lasts build an IP Address Database (IAD) with the frequent IP
over 2 seconds is compared with every sample of mature addresses and this IAD is used to build self-set in
detector set. If the affinity is greater than the detection Neighborhood Negative Selection Algorithm. Self-set is thus
threshold, then incoming vector is attack otherwise this is improved than previous. The possible attacks detected by the
treated as a normal traffic. We use this algorithm in our system are filtered with the help of IAD to find real attack.
proposed solution with some modification.
IV. PROPOSED SOLUTION
D. Comparisons between different system
A. System block diagram
In traffic pattern detection system different types of
flooding and logic based attacks are detected by comparing Fig. 5 shows the proposed system block diagram, which
with their predefined static threshold. consists of Feature Extraction Module (FEM), Build IP
Address Database Module (BIADM), Detection Module1
TABLE V. COMPARISON OF DIFFERENT SYSTEMS (DM1), and Detection Module 2(DM2). All the traffic come to
FEM and BIADM. FEM receives all the traffic coming to the
Traffic pattern Sketch based system Flow level system
based system
network and generates the feature vectors which include
1. Different types of 1. TCP SYN flood is 1. Can detect several horizontal and vertical features as indicated in Table
flooding and logic detected from combinations unknown attacks as IV. In BIADM Module the frequent source IP of traffic will be
attacks are detected. of TCP SYN flood and well as real attack. stored into IAD. This is a continuous process. The outcome of
2. Only the known different port scans. 2. The elements of FEM goes to DM1. Two phases are present in DM1 namely
attacks can be 2. Except TCP SYN flood self-set which is Training phase and Detecting phase. After initial t1 time of
detected. If attacker no other attack is detected. used in training
use a new method to 3. Cannot detect unknown phase of detection Training phase mature detector set is built with the help of
attack then cannot attacks. phase are taken BIADM module and at the Detecting phase DM1 detects the
be detected. 4. Exponential Weighted randomly, so not feature vectors as either possible attack or normal traffic. The
3. The threshold Moving Average accurate. possible attack detected by DM1 goes to DM2, where with the
which is used is method(which gives less 3. Some attacks may help of BIADM, we have further filtered the possible attack as
static, which may weight to the data as they be detected as
not be suitable for are removed time to time) is normal traffic and
either attack or normal traffic.
every network used to calculate the some normal traffic
environment threshold may be detected as
anomaly.
In sketch based system threshold calculation is improved
by using EWMA (Exponentially Weighted Moving Average)
formula. This system detects attack from different
combinations of attacks. It detects TCP SYN flood and
different port scan. Flow level system is improved than traffic
pattern system and sketch based system. Flow level system
discussed in [1] not only detects different known attack but
also different unknown attacks and also real attack flow. But

Fig. 5 System Block Diagram
B. Methodology
It has been observed [7] that there exists three types of traffic
namely normal traffic, flash crowd and denial of service
attack. Flash crowd is a type of normal traffic where many
legitimate users start to access a site at the same time.
1) Feature Extraction Module: Feature extraction module
receives all the traffic coming to the network as input and
generates feature vectors. The feature vectors consist of
different vertical and horizontal features. The feature vectors
which are generated from this module go to DM1 to detect it
as either possible attack or normal traffic. A detail description
of FEM was reported in [1] and here we provide a summary to
explain the augmentation proposed by us in connection with
improvement in elimination of false positive cases. As shown
in Fig. 4[1] for TCP and UDP the tree height is six and for
ICMP the tree height is four. For TCP and UDP the first level
is root, second level is protocol, third level is IP address of
inside host, fourth level is port number of inside host, fifth
level is IP address of outside host and sixth level is port
number of outside host. For ICMP protocol the first level is
root, second level is protocol, third level is IP address of inside
host, fourth level is IP address of outside host. Inside host
refers to the system present inside the network and outside
host refers to the systems present outside of the protected
network. In the tree structure parent node points to its first
child with double-sided arrow and all the other child nodes
point to their parent node with one sided arrow and a node
points to its brother with a dotted arrow. Two types of flow
features are present. Those are vertical features which refer to
the features extracted from one single flow and the horizontal
features are those extracted from more than one flow. The lists
of different types of vertical features and horizontal features
which are extracted from the tree structure are shown in Table
IV [1].
2) Build IAD Module: Creation of IAD for detection of
attack was proposed in [7]. We here use this concept with
modification and utilize the modified version in both DM1 and
elimination of false positive cases, if any, in DM2. The IP
database is continuously updated to include new IP addresses
which are frequent (as defined below). This will help in
identifying possible new attacks. At the same time IP
addresses which are very infrequent with respect to time stamp
are removed from the database so that the volume of the
database can be managed without loss of quality. Two
different rules are employed in [7] to define frequent IP. P1
refers to the collection of unique IP addresses which appeared
at least d days within last D days (D>>d). P2 refers to
collection of unique IP addresses each have at least u number
of packets. In our proposal F represents the frequent IP
addresses of the network. We define F= P1UP2. Where „U‟
represents union. By combining these two rules defined in [7]
IAD is built. IAD contains source IP and time stamp. To keep
the size of IAD within manageable volume, we remove the old
IP addresses which do not appear in last P days on anticipation
that they are nosafer. Algorithm for this updation is presented
in section „Build IAD Module‟ of the „Algorithm‟ section.
3) Detection Module 1: The features vectors, extracted in
„feature extraction module‟ come to „Detection module 1‟ as
input. All the feature vectors of traffic flow which are lasts
more than 2 second in this network are input into Detection
Module1. We use NNS algorithm [1] [2] [3] with some
modification using Build IAD Module. This algorithm has two
phases namely training phase and detecting phase. At the
training phase, a mature detector set is built and at the
detecting phase this mature detector set is used to detect
possible attack. Before training starts feature vectors are
extracted from normal traffic to build fully adjacent but
mutually disjoint neighborhoods. Within eleven features, if
nine features are matched then we call these two feature
vectors as neighborhood. The union of these neighborhoods is
called shape space. From this shape space neighborhoods are
taken randomly to build „Self set1‟. For the initial t1 time of
training phase we first select a neighborhood randomly and
then we calculate the affinities between this neighborhood and
every element of „Self set 1‟.The neighborhood which is taken
is called candidate detector. The candidate detector is
eliminated if the affinity is more than training threshold;
otherwise this candidate detector is added into detector set.
After the initial t1 time of training phase, the corresponding
feature vectors of the source IP present in IAD are extracted.
Then the new self set is Self set2. Self set2= Self set 1U
Feature vectors of source IPs present in IAD. After this, again
neighborhood (candidate detect) is taken at randomly and then
affinity of this neighborhood and every element of Self set 2 is
calculated. If the affinity is greater than training threshold then
this candidate detector is eliminated, otherwise this is added
into detector set. Iteratively the process is going on until we
get one mature detector set. The traffic flows which exist more
than 2 seconds in the network will be checked in every two
seconds at the detecting phase. At the detecting phase, we
match feature vectors of the flows which lasts over two
seconds with the every element of mature detector set. A new
incoming feature vector is treated as „possible attack‟ if the
affinity between it and the mature detector set is more than
detecting threshold otherwise this is treated as a „normal‟
traffic.
4) Detection Module 2: Detection Module 2 filters out the
real attack traffic from the possible attack detected in
Detection Module 1. The „possible attack‟ detected by
„Detection module 1‟ comes to this module as input.
„Detection module 2‟ checks whether the source IP of
„possible attack‟ traffic is present in IAD or not. If the source
IP is not present in IAD, then this possible attack is treated as
real attack flow, otherwise this is normal traffic flow.
Algorithm: Total Attack Ac Fa Tota Actu Fals Total posit
INPUT: Traffic flow coming to the network (as per tu lse l al e detecte ive
UCLA al po dete attac posit d redu
OUTPUT: {attack, normal traffic} databa att siti cted k ive possibl ction
se) ac ve attac e
Feature Extraction Module: k k attack
1. Extract feature vectors from traffic to build 800 47 17 25 42 17 10 27 60%
neighbourhood. (0)
900 50 44 10 150 45 55 100 48.1
Build IAD Module: 6 (+1) 1%
2. Add frequent source IP into IAD and remove the IP 1000 57 36 28 64 39 19 58 32.1
from IAD which does not appear in last P days (+3) 4%
1800 102 47 16 215 53 130 183 22.6
Detection module1: 8 (+6) %
3 The neighbourhoods are taken randomly to build 2000 105 62 17 238 63 95 158 46.0
6 (+1) 2%
Self-set 1.
From the result it has been observed that there exists a
(Training Phase) quite significant number of cases which are declared as attack
4 Select a neighbourhood randomly by existing negative selection algorithm, particularly done in
5 If (time < t1) then GO TO 6 [1] while it is much less in our proposed algorithm. Also it has
Else GO TO 9. been observed that some new cases have been declared as
6 Calculate affinity between this neighbourhood and actual attack (0, 1, 3, 6, 1) in our algorithm which are passed
every member of Self-set1. as safe in the other. Both these changes indicate the adaptive
nature of our algorithm.
7 If (affinity > training threshold)
Then eliminate For instance in row 2 of Table VI, we take 900 feature
Else add it into detector set. vectors from UCLA database [26] in which the number of
8 If detector set is mature then EXIT actual attack is 50 as declared in UCLA database. We see the
Else GO TO 4 number of attacks detected by existing negative selection
9 S= neighbourhoods corresponding to source IP algorithm is 150, in which the number of actual attack is 44
present in IAD and the number of false positives is 106, while in our
10 Self set2= Self set1 U S algorithm the number of detected possible attack is 100, in
11 Calculate affinity between selected neighbourhood which the number of actual attack is 45 and the number of
and every member of Self set2 false positive cases is 55. In this case our algorithm reduces
48.11% false positives and one extra actual attack is detected
12 Go to 7
by our algorithm which cannot be detected by existing
13 End of training phase
negative selection algorithm.
(Detection Phase)
14 Calculate affinity of the feature vectors of the flows
of traffic (checks every 2 seconds until they expire)
with every sample of mature detector.
15 If (affinity > detection threshold) then incoming
vector is treated as „possible attack‟
Else treat it as a normal.
Detection module 2:
16 If (the source IP of „possible attack‟ is present at that
time IAD) Fig. 6 Detected false positives vs number of feature vectors
17 Then treat it as „normal In row 5 of Table VI, we take 2000 feature vectors from
Else mark it as an „attack‟. UCLA database in which number of actual attack is 105 as
declared in UCLA database. The number of attacks detected
V. IMPLEMENTATION by existing NSA is 238 in which the number of actual attack is
The proposed algorithm has been implemented and compared 62 and the number of false positive cases is 176, while in our
with existing negative selection algorithm [1] with UCLA [26] proposed algorithm the number of detected possible attack is
dataset taken at randomly. 158 in which the number of actual attack is 63 and the number
of false positive cases is 95. In this case our algorithm reduces
TABLE VI. IMPLEMENTATION RESULT 46.02 % false positive cases.
No of feature Existing NSA Proposed algorithm Perc As mentioned before, the system proposed in this paper
vectors from algorithm enta focuses on reducing false positives. Fig. 6 shows that for each
UCLA database (Wang) ge of cases, existing NSA detects significantly large number of false
chosen randomly false positives whereas, in those cases, the proposed algorithm
detects much less number of false positives. For the test cases
shown in Table VI, the reduction in false positive detection
ranges from 22.6% to 60%.
Fig. 7 shows that in case of actual attack detection the
proposed algorithm is slightly better than existing NSA and in
some cases it outperforms the existing negative selection
algorithm.
Fig. 7 Detected actual attacks vs number of feature vectors
VI. CONCLUSION
Detection of anomaly from a set of data is always a very
critical issue for all types of solution approaches. It becomes
harder when the system is very dynamic and anomaly may
come from different sources with various modifications. Any
method which works in a static approach might work excellent
at the beginning but naturally will lose its sheen with passage
of time. As initial negative selection algorithm was not
inherently adaptive, it will naturally declare more genuine
users as attacker and thus may cause problem for users. We
have taken care of the missing adaptability of the algorithm by
introducing periodic updation of the self-set after initial t1
time, comparing the source IP of the possible attacks with the
IAD and modifying the IAD by incorporating the frequent IPs‟
in the system, and the modified algorithm provides a much
better result on both accounts, i.e. reduction of false positive
and detection of actual attack.
References
[1] D. Wang, L. He, Y. Xue, Y. Dong, “Exploiting Artificial Immune
systems to detect unknown DoS attacks in real-time”, 2nd International
Conference on CCIS 2012 IEEE vol. 02, 978-1-4673-1855-6,pp. 646-
650.
[2] Stephanie Forrest, Alan s. Perelson, Lawrence Allen, Rajesh Cherukuri,
“Self-Nonself Discrimination in a Computer” IEEE Computer Society
Symposium on Research in Security and Privacy, 1994, 0-8186-5675-1,
10.1109/RISP.1994.296580,pp. 202-212.
[3] Dawei Wang, Yibo Xue ,Yingfei Dong, “NNS：A Novel Neighborhood
Negative Selection Algorithm”, World Automation Congress, 2012
IEEE,978-1-4673-4497-5, pp. 453-457.
[4] F. Gonzalez, D. Dasgupta, “Anomaly detection using real-valued
negative selection”, Genetic Programming and Evolvable Machine, vol.
4(4), pp. 383-403.
[5] D. Wang, Y. Xue and D. Yingfei, 'Anomaly Detection Using
Neighborhood Negative Selection', Intelligent Automation & Soft
Computing, vol. 17, no. 5, pp. 595-605, 2011.
[6] Qi Jin ,Ma Ming, “A method to construct self set for IDS based on
negative selection algorithm” ,International Conference on Mechatronic
Science, Electric Engineering and Computer (MEC), 2011, 978-1-
61284-719-1, pp.1051 – 1053.
[7] T. Peng, C. Leckie, K. Ramamohanarao, “Protection from distributed
denial of service attacks using history-based IP filtering” IEEE
International Conference on Communications 2003. ICC‟03, vol. 1, 0-
7803-7802-4,pp. 482-486.
[8] Theerasak Thapngam , Shui Yu ,Wanlei Zhou ,S. Kami Makki,
“Distributed Denial of Service (DDoS) detection by traffic pattern
analysis” Peer-to-Peer Netw. Appl. (2014), Springer.
[9] M.S. Kim, H.J. Kong, S.C. Hong and et al, “A flow-based method for
abnormal network traffic detection”, Network Operation and
Management Symposium,NOMS 2004,IEEE IFIP vol.1,0-7803-8230-7,
pp. 599- 612.
[10] H.N. Wang, D.L. Zhang and K.G. Shin, “Detecting SYN Flooding
Attacks”, In Proc. 2002 IEEE INFOCOMM, pp. 1530-1539.
[11] Y. Gao, Z. Li and Y. Chen, “A dos resilient flow-level intrusion
detection approach for high-speed networks”, in Proc. 2006 ICDCS, pp.
39.
[12] Schweller R, Zhichun Li, Yan Chen, Yan Gao and et al, “Reversible
Sketches: Enabling Monitoring and Analysis Over High-Speed Data
Streams”, in 2007 Networking, IEEE/ACM Transactions, vol.15, pp.
1059-1072.
[13] Cisar P, Maravic Cisar. S, “EWMA Statistic in Adaptive Threshold
Algorithm”, 11th International Conference on IEEE Intelligent
Engineering Systems,2007, pp. 51-54.
[14] V. Paxson, “Bro: A system for detecting network intruder in real-time”,
Computer Networks, vol. 31, no.23-24, 1999, pp. 2435-2463.
[15] R. Shea, J. Liu, “Understanding the impact of Denial of Service attacks
on Virtual Machines”, 20th International Workshop on Quality of
Service (IWQoS), 2012 IEEE ,978-1-4673-1296-7, pp.1-9.
[16] R. Shea and J. Liu, 'Performance of Virtual Machines Under Networked
Denial of Service Attacks: Experiments and Analysis', IEEE Systems
Journal, vol. 7, no. 2, pp. 335-345, 2013.
[17] S.H.C Haris, R.B Ahmad,M.A.H.A Ghani, “Detecting TCP SYN Flood
Attack Based on Anomaly Detection”, Second International Conference
on Network Applications Protocols and Services (NETAPPS), 2010,
978-1-4244-8048-7, pp.240 - 244.
[18] S.S.Chapade, K.U Pandey, D.S Bhade, “Securing Cloud Servers Against
Flooding Based DDOS Attacks”, International Conference on
Communication Systems and Network Technologies (CSNT), 2013,
978-1-4673-5603-9, pp.524 – 528.
[19] Rejimol Robinson, R.R. Thomas, C, “Evaluation of mitigation methods
for distributed denial of service attacks”, 7th IEEE Conference on
Industrial Electronics and Applications (ICIEA), 2012, 978-1-4577-
2118-2, pp.713 – 718.
[20] T.Gunasekhar, K.Thirupathi Rao,P.Saikiran,P.V.S Lakshmi , “A Survey
on Denial of Service Attacks”, International Journal of Computer
Science and Information Technologies, Vol. 5 (2),2014, pp.2373-2376.
[21] F. Sabahi, “Virtualization-level security in cloud computing”, 3rd
International Conference on Communication Softwareand Networks
(ICCSN) 2011 IEEE, 978-1-61284-485-5, pp.250 – 254.
[22] X. Luo, L. Yang, L. Ma and et al, “Virtualization Security Risks and
Solutions of Cloud Computing via Divide-Conquer Strategy”, 2011
Third International Conference on Multimedia Information Networking
and Security (MINES),978-1-4577-1795-6, pp.637-641.
[23] A. Sperotto, G. Schaffrach, R. Sadre and et al., “An overview IP flow-
based intrusion detection”, IEEE Communication surveys and tutorials,
vol.12(3), 2010, pp. 343-356.
[24] S Alarifi, S.D.Wolthusen, “Mitigation of Cloud-Internal Denial of
Service Attacks”,8th International Symposium on Service Oriented
System Engineering (SOSE), 2014 IEEE, 10.1109/SOSE.2014.71,
pp.478 – 483.
[25] R. Buyya, J. Broberg and A . Gościński, Cloud computing. Hoboken,
N.J.: Wiley, 2011.
[26] 'UCLA CSD Packet Traces'. [Online]. Available:
http://www.lasr.cs.ucla.edu/ddos/traces/. [Accessed: 05- Jan- 2015]
 Vehicle Security and Forensics in Mauritius and
Abroad
Pravin Selukoto Paupiah
Amity Institute of Higher Education
Mauritius
selukoto@gmail.com
Abstract—Mauritius as a fast developing and upper middle
income country has witnessed an important expansion in the
number of vehicles imported to the country. During the past two
decades the number of modern vehicles has been rapidly
increasing on the Mauritian road. Today’s motor vehicle are
equipped with various digital components working through
interior vehicular networks instead of only mechanical
mechanisms. Obviously the innovations in vehicles technologies
have brought forward many advantages for drivers and
passengers in terms of efficiency and safety. These major changes
in in the automobiles system structures are not without
consequences. Researchers, through various experimental
assessments, have discovered a number of potential threats. This
paper outlines the number of possible issues, and how Mauritius
a country with an increasing fleet of high tech vehicles, is exposed
to new potential risks.
Keywords— vehicles; digital components; threats.
I. INTRODUCTION
Vehicles are being equipped with high technology systems
and sophisticated internet connections or satellite navigation
system. Cars are expensive and hence it is becoming more
vulnerable to theft. Cars are expensive and just like any other
expensive asset, a car brings with its secondary cost that is the
risk of theft. It is enough secured when the locking the doors to
ward off the threats however in some places it is better to guard
the car with some more security.
With against theft for example before anyone tries to open
the car a message is sent to the owner informing the latter
about the security threat. There are a number of ways used to
protect against threat, hacking and even accidents. Attacks on
vehicles are fatal and need to be addressed. Automobiles
manufacturers have bought huge evolution in vehicles during
the last three decades. They have set the path for the future
automobiles. Often to abide to certain rules and regulations set
by government, they are obliged to come with innovative
concept. On the other hand with the rapid changing
technologies, car industries need to reorganize themselves.
Information can be accessed everywhere and every time. Even
in vehicles, you can be connected to global village. With the
shifting towards IPv6, the number of connected IP devices will
be enormous. Ericsson predicts that by 2020 there will be 50
billion IP connected devices [1].
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
II. LITERATURE
The purpose of this project is to gain a better understanding
of the modern automobile system and also to give an idea of
vehicles in Mauritius. Nowadays due to the advancement of
technology, lots of gadgets are becoming available and these
electronic devices are being equipped in vehicles. However, the
manufacturers have to become more cautious while designing
these vehicles in [22], Thatcham Research, who is best known
by automobilists for examining the safety of automobiles and
was initiated by Motor insurance industry in 1969 says that
since they are familiar with current security threat associated
with the theft of vehicles, where attackers are trying to gain
access to vehicles without keys hence they are working
incorporation with the police and manufacturers of automobiles
to come up with ideas to protect vehicles and owners of
vehicles against these kinds of attacks. The research institute is
aware of this problem for around two years, and has countered
for this by updating the insurance vehicle security standard
whereby the manufacturers of automobile must abide for
insurance purposes.
III. THE MODERN AUTOMOBILE STRUCTURE
A wide system constituting subsystems all working together for
a particular purpose is what actually defines the modern
automobile system. Some of these parts have come from the
development of the new technology like electronic computer,
high-strength plastics, and new alloys of steel and nonferrous
metals [2]. There is a rapid rise in the growing of electronic car
systems and hence the lines of embedded codes in software in
cars have now increased. Moreover, the basic model of the
evolution of the structure in car electronic is changing. The
designs of the modern automobile consist of sub-systems,
where each function is utilized to the Electronic Control Unit
(ECU), however due to the increase in the amount of ECUs,
sub-systems, buses in vehicles, and the growing of linkage of
functions makes these systems difficult to be examined and
approved [3]. This modern era, Vehicles are connected to
alternative units and bodies via wireless communications [4].
A. The Connected Car
This can be explained as the presence of devices in a
vehicle that enable the device to connect to other devices in
other vehicles or any gadget, network or service that is outside
the particular car for example another vehicle or any
infrastructure [5]. A connected vehicle extends the in-vehicle
network to allow communication to take place between mobile
devices through cellular networks if it is equipped with
portable communication device exchanging messages with
neighboring vehicles and Road Side Units (RSUs) through
Vehicular Ad-hoc Networks (VANETs) if equipped with Wi-Fi
communication devices, and exchanging messages with
personal devices– e.g., Personal Digital Assistances (PDAs),
iPods, tablets, and smart cellphones–if equipped with short
range wireless communication devices–e.g., Bluetooth, Near
Field Communication (NFC), and Infrared (IR) devices [4].
there were about 465, 052 vehicles being recorded at the
National Transport Authority (NTA). Comparing to the year
2013 where there has been an increase of 21, 557 that is by 4.9
% and hence reaching the number of registered vehicles is of
443,495.
Below is given the statistics of the number of vehicles
(cars, dual purpose and double cab pickup) and their relative
age groups [10]:

IV. ECU (ELECTRONIC CONTROL UNITS)

Electronic Control unit also known as embedded electronic
components are a very important part in a car’s architecture.
The role of ECU is that it manages the fuel injection system,
ignition timing, and the idle speed control system. The ECU
consist of an 8-bit microprocessor, random access memory
(RAM), read only memory (ROM), and an input and output
interface [6]. ECU’s are arranged together on one or more
busses based on controller area network standard [8]. To be
noted here is that the CAN is an international standardization
organization (ISO) defined serial communications bus
originally developed for the automotive industry to replace the
complex wiring harness with a two-wire bus [7]. The ECU’s
communicate to each other by sending packets. These packets
are then transferred to all the components on the bus and each
component decide whether it is sent for them, even if
segmented CAN networks do exist. There is no source
identifier or authentication built on CAN packets [8]. The
Age group 2013 2014
typical role of ECU’s is to communicate between them to
manage properly the different vehicular subsystem and hence (Years)
Number % Number %
forming an automotive network. Like the computer system, the
automotive network is also vulnerable to risks being exploited <5 92, 976 43.9 99, 134 44.0
by intruders. Nowadays with the rise in technology, modern 5 < 10 52, 389 24.8 56, 596 25.1
vehicles possess the capability to allow for communication 10 < 15 24, 418 11.5 27,746 12.3
wired (USB) and wireless (Bluetooth, Wifi, 3G...). Therefore, ≥ 15 41,803 19.8 42, 045 18.6
knowing that the computer systems embedded in cars
nowadays can be no longer be considered a closed network Total 211,586 100.0 225,522 100.0
hence this increases cyber-attacks [9] .
Some of the reasons for attacks are listed as follows [9]:
Fig 1. Statistics [10]
1. Theft: This could be one of the most common reason
behind attacks. An intruder could exploit a vulnerability
present in a wireless communication protocol to secretly open A. Road Traffic accident in Mauritius
the particular car and deactivate immobilizer. The Number of road accident that has been recorded for the
2. Sabotage: Here the aim of the attacker is to deactivate year 2014 was 26, 400 however for the year 2013 the number
the ECUs, the alteration of the software or a denial of service of road accidents was 23,563 and this shows that there has been
on the network to decrease the vehicle’s efficiency. a slight increase in the number of road accident from the year
2013 to 2014 [10].
3. Privacy Breach: A vehicle may sometimes contain
confidential information such as driver’s call history, GPS Causes of Road accidents in Mauritius:
coordinates history and here the goal of the attacker is to Security is to be free from threats and for this we need to
retrieve these information to use it in an illegal manner. consider the main reasons of the occurrence of accidents and
below is listed some of the causes of accidents in Mauritius
[21]:
V. AUTOMOBILES IN MAURITIUS 1. Speed on the road: Drivers usually do not follow the
As per the Road Transport and Road Traffic Accident speed limit that has been put on the road to ensure their safety.
Statistics Mauritius, by the culmination of December 2014, However, the government has come with initiative such as
putting speed cameras on road whereby ensuring everyone is
following speed limits otherwise the latter may be prone to pay
for the crime committed of not following speed limit.
2. Drink-driving: The drink- drive limit is of
80mg/100ml, most drivers are unable to concentrate after
consuming alcohol.
3. Low Road safety awareness: Some precautions that
road drivers need to take into consideration are that there could
be motorist driving too close, pedestrians choosing to cross the
road at an unsafe place. For this, some awareness programs
need to take place to bring more light regarding this subject.
VI. SECURITY ISSUES IN MODERN AUTOMOBILES:
Automobiles manufacturers need to be creative to stay
competitive. The modern automobiles consist of a variety of
novel features. For instance users can make his or her mobile
devices to be wirelessly connected to the vehicle. Recent
vehicle have incorporated smart phones in vehicles itself. The
challenges of automobile manufacturers are immense. The cars
are highly equipped with more and more computer chips.
Nowadays, the value of the car is based on the number of
embedded electronic devices. The modern is much more a
sophisticated intelligent apparatus which processed large
quantity of data and respond accordingly. The large number of
sensors and various computer units made the car more efficient
to fuel consumption. The modern car is even equipped with
black box which logged all occurrences. In cases of accident
forensic examiner retrieve information to understand the
accident. A permanent internet connection supply real time
information’s such as weather conditions, traffic jams and
news [11]. In 2011, the computer security enterprise, Mc Afee
published a report and evokes the threat that modern cars are
facing. The increasing numbers of computer chips in
automobiles makes them more vulnerable to attacks [1]. The
modern car is fitted with various hardware which requires
software’s to operate. The software require millions of lines of
codes to function.
A. Tire Pressure Monitoring System - TPMS
A well-known fact is that tire pressure influence the
vehicle’s performance in terms of its stability, fuel
consumption and flexibility. A low tire pressure indicates a tire
failure this can be considered as a serious issue as this may
cause accidents. Intelligent Vehicle Tire Pressure Monitoring
System (TPMS) is the current innovation which has the ability
to present real-time tire pressure, temperature and automatic
tire pressure monitoring to ensure traffic safety. TPMS is
comprised of two types, one that is called the indirect TPMS
and the other is called direct TPMS. The role of indirect TPMS
is that it differentiates the speed between the tire through car
ABS wheel speed sensors to monitor tire pressure and direct
TPMS makes use of lithium-ion battery-powered pressure
sensor which is placed in each tire to estimate the tire pressure
[12]. A research paper evoking TPMS vulnerabilities was
published in November 2011 by various researchers of two
universities namely, South Carolina and Rutgers. They
documented how TPMS could be used to manipulate ECUs of
automobiles. The scientists were capable of injecting spoofed
messages in loop that low tire pressure alert started blinking.
The driver believed that the tire was deflated which was false.
To counter the issue, scientists propose to apply best practices
when designing software. Furthermore, applying basic
cryptography security mechanism will not affect data
communication performance. A level of security will be at least
implemented [13].
B. Keyless Entry System
The RFID (Radio Frequency Identification) is based on radio
frequency technology and is an automatic identification
technique. RFID is the wireless use of electromagnetic fields to
be able to carry data, for the purpose of automatically
recognizing and following tags attached to objects. The tags
holds electronically stored information [14]. The entry system
of automobiles utilized RFID to provide access from a
distance. The vehicle releases signal cyclically in a low
frequency range of 120 to 135KHz. When the key is this range,
the car would be able to unlock provided the response is valid.
The reply is in the range of 315 to 433, an ultra-high frequency
[15].
C. Global Positioning system in vehicles and security
The Global Positioning system in vehicle is a worldwide
navigation and positioning system developed by US
department of Defense for both military and civilian use. GPS
is used to identify where an object is actually on earth with an
accuracy from ten of meters down to millimeters, depending on
the receiver and signal processing technology [16]. This new
technology, allows to locate a vehicle’s location automatically.
By merging GPS position technology with those systems that
can display geographic information, a new dimension in
surface transportation is realized [17].
D. The need for GPS (Global positioning system)
People usually think that a GPS is useful only when they have
to reach a destination whereby the route is unknown, that is
moving from point x to y whereby y is unknown. However, a
GPS has many other tracking systems that may prove to be
useful. Nowadays, the GPS tracking system has the ability to
track anything and anyone. Any individual processing a vehicle
will want to place a GPS tracing system under their dash or
groove compartment and this way if ever the vehicle gets
stolen, it can be re-located fast. The GPS satellite gives the
exact location of the device which is found in the car. The
device is connected to the local GSM (Global System for
mobile communication) service provider through GSM
network as it contains a SIM card, hence the GPS parameter
which the device has are sent to the tracking server which
possess a static IP address through a GPRS network [18]. This
can be illustrated in figure below:

Fig 2. GPS Tracking System [19].
The GSM is a standard which has been developed by the
European Telecommunication standard institute to describe
protocols for 2G digital cellular networks used by mobile
phones [19].
VII. CAR BLACK BOX
The Black box is considered as a digital video-audio event
data recorder. Its purpose is that it records all information
related to the particular car for example, speed and temperature
of the engine. It also records the time and location before and
after an accident so that the results gained can be used to
analyze the accident properly. The black-box also includes a
web-camera, a microphone, a real-time clock and other
electronic components, al managed by a microprocessor with
embedded software and fully self-contained in a tiny, rugged
black box that installs unobtrusively on the dash board.
The application of a car’s black box include [21]:
1. Collision data for research, the data obtained is used
to improve the vehicle’s design both internally and externally.
2. Its aim is also to try to reduce possible accidents by
minimizing the speed of the vehicle in case of an accident.
3. Better crash reports and this helps to produce
improved driver education programs, safer road designs and
improve highway safety.
VIII. AN INTRODUCTION TO VEHICLE FORENSICS.
The forensic analysis of a vehicle is important as whenever,
there is an occurrence of an unwanted event like crimes, car
accidents it helps to determine the real cause behind the
mishap. Forensic inspectors and researchers are the one who
conduct an in-depth study on vehicles by obtaining user’s
information, iVe is a vehicle system forensic tool that helps
these officers to accomplish this particular duty. An
automobile infotainment and telematics system gathers a large
amount of data for example, the history of the routes that the
vehicle has visited, call archives, videos and pictures and much
more. iVe has the ability to support over more than 4,000
models of vehicles comprising Chevrolet, BMW, Toyota
vehicles, Jeep [23].
IX. CONCLUSION
All in all this research papers aim was to present the state of
security and forensics as they relate to vehicles. The subject is
very vast and exhaustive and is continually evolving with state
of the art technologies. Experts point out the electronic gadget
known as electronic control unit that require millions of lines
of codes to manage interconnected systems. The same wireless
technology that power cells phones and Bluetooth headsets in
cars are becoming vulnerable. A malicious hacker could, in
theory, disable the vehicle, re-route GPS signals or otherwise
put the driver in danger. The forensic expert is the one who has
the responsibility to investigate in any case of accident, theft
and hacking in vehicular systems. Finally, there is a need to
focus on all the vulnerable entry points. Car manufacturers
need to urgently update the security of automotive computer
systems.
REFERENCES
[1] M. Wolf, C. Paar, S. Winfried and S. Goss, "Caution: Malware Ahead,
An analysis of emerging risks in automotive system security", McAfee
Report, 2011. Online: http://www.mcafee.com/mx/resources/reports/rp-
caution-malware-ahead.pdf. [Accessed 15 August 2015].
[2] G. C. Cromer, "Alternative titles: auto; car; motorcar", 2015. Online:
http://www.britannica.com/technology/automobile[ Accessed 29 August
2015].
[3] E. Frank, R. Wilhelm, R. Ernst, A. Sangiovanni-Vincentelli and M. Di
Natale, "Methods, Tools and Standards for the Analysis, Evaluation and
Design ofModern Automotive Architectures", IEEE Xplore, [Design,
Automation and Test in Europe, Munich, pp. 659 - 663, 2008].
[4] B. L. Othmane, A., Al-Fuqaha, B. E. Hamida, M. van den Brand,
"Towards extended safety in connected vehicle", IEEE Xplore,
Intelligent Transportation Systems for All Modes, [16th International
IEEE Conference on Intelligent Transportation Systems - (ITSC),
Netherlands, pp. 652 - 657, 2013].
[5] L. Walford, "Definition of Connected Car – What is the connected car?
Defined", AUTO Connected Car, Aproprose Publishing, USA, 2015.
Online: http://www.autoconnectedcar.com/definition-of-connected-car-
what-is-the-connected-car-defined/ [Accessed 28 August 2015].
[6] T. C. Vincent , "Electronic Control Unit (ECU)", About.com, Auto
Repair, 2000. Online:
http://autorepair.about.com/cs/generalinfo/l/bldef_160.htm [Accessed 27
August 2015].
[7] S. Corrigan, "Introduction to the Controller Area Network (CAN)",
Application Report, Texas Instruments Incorporated, 2008.
[8] C. Valasek and C. Miller, "Adventures in Automotive Networks and
Control Units", Comprehensive Information Security, IOActive, USA,
2014.
[9] I. Studnia, V. Nicomette, E. Alata, Y. Dewarte, M. Kaâniche, Y.
Laarouchi, "Survey on security threats and protection mechanisms in
embedded automotive networks", IEEE Explore, [2nd Workshop on
Open Resilient Human-aware Cyber-Physical Systems, Budapest,
Hungary, pp. 1 - 12, 2013].
[10] Statsmauritius, "Road Transport and Road Traffic Accident Statistics,
(Island of Mauritius)". Online:
http://statsmauritius.govmu.org/English/StatsbySubj/Documents/ei
%201167/Road%20Transport%20Year%202014.pdf [Accessed 27
August 2015].
[11] P. Nisch, "Security Issues in Modern Automotive Systems", 2012.
Online: http://www.panisch.com/wp-
content/uploads/2012/06/Security_Issues_in_Modern_Automotive_Cars.
pdf [Accessed 27 August 2015].
[12] Z. Jiang, H. Liu and Q. Dai, "A New Intelligent Tire Pressure
Monitoring System", IEEE Xplore, ICM [International Conference of
Information Technology, Computer Engineering and Management
Sciences, China, pp. 332 - 335, 2011].
[13] I. Rouf, R. Miller, H. Mustafa, T. Taylor, S. Oh, W. Xu, M. Gruteser, W.
Trappe, and I. Seskar. "Security and privacy vulnerabilities of incar
 wireless networks: a tire pressure monitoring system case study", ACM [18] S. Mohol , A. Pavanikar and G. Dhage, "GPS Vehicle Tracking
Digital Library, [19th USENIX conference on Security, USA, pp. 21– System", in the International Journal of Emerging Engineering Research
21, 2010]. and Technology, vol. 2, no. 7, pp. 71-75 2014.
[14] Wikipedia, "Radio-frequency Identification", 2015. Online: [19] Wikipedia, "GSM", 2015. Online: https://en.wikipedia.org/wiki/GSM
https://en.wikipedia.org/wiki/Radio-frequency_identification [Accessed [Accessed 29 August 2015].
27 August 2015]. [20] Traffic Management and Road Safety Unit, "Road Accidents Mauritius:
[15] A. Francillon, B. Danev, and S. Capkun. "Relay Attacks on Passive An engineering perspective", 2002. Online:
Keyless Entry and Start Systems in Modern Cars", Cryptology ePrint http://publicinfrastructure.govmu.org/English/Publication/Documents/ra
Archive, Report 2010/332, 2010. m.pdf [Accessed 29 August 2015].
[16] GPS, "What is GPS?". Online: [21] D. Pawar and P. Poddar. "Car Black Box with Speed Control in Desired
http://www.physics.hmc.edu/research/geo/gps.html [Accessed 29 Areas for Collision Avoidance", in Engineering, Technology & Applied
August 2015]. Science Research, vol. 2, no. 5, pp. 281-284, 2012.
[17] J. N. Shane, "The promise of GPS technology for increasing safety and [22] M. Briggs, "Theft Without Keys", Online:
security, reducing congestion, and improving efficiency are limitless. https://www.thatcham.org/files/pdf/Theft_Without_Keys.pdf [Accessed
Quite simply, GPS hasbecome the enabling technology for 29 August 2015].
transportation", 2006. Online: http://www.gps.gov/applications/roads/ [23] Berla, "Enabling forensic acquisition and analysis of vehicle
[Accessed 29 August 2015] infotainment and telematics systems". Online:
https://berla.co/downloads/ive_datasheet.pdf [Accessed 29 August
2015].
 Performance Evaluation and Analysis of Layer 3
Tunneling between OpenSSH and OpenVPN in a
Wide Area Network Environment
Irfaan Coonjah Pierre Clarel Catherine
Faculty of Engineering School of Innovative Technologies and Engineering
University of Mauritius University of technology, Mauritius
Réduit, Mauritius La Tour Koenig, Pointes aux Sables
irfaan.coonjah@umail.uom.ac.mu c.catherine@ieee.org
Abstract—Virtual Private Networks (VPNs) provide a secure
encrypted communication between remote networks worldwide
by using Internet Protocol(IP) tunnels and a shared medium
like the Internet. End-to-end connectivity is established by
tunneling. OpenVPN and OpenSSH are cross-platform, secure,
highly configurable VPN solutions. The performance comparison
however between OpenVPN and OpenSSH VPN has not yet
been undertaken. This paper focuses on such comparison and
evaluates the efficiency of these VPNs over Wide Area Network
(WAN) connections. The same conditions are maintained for a
fair comparison. To the best knowledge of the authors, this is
the first reported test results of these two commonly used VPN
technologies. Three parameters, namely speed, latency and jitter
are evaluated. Using a real life scenario with deployment over the
Linux Operating System, a comprehensive in-depth comparative
analysis of the VPN mechanisms is provided. Results of the
analysis between OpenSSH and OpenVPN show that OpenSSH
utilizes better the link and significantly improves transfer times.
Keywords—Tunneling, OpenSSH, OpenVPN, VPN.
I. I NTRODUCTION
A Secure Shell(SSH) protocol was designed by Tatu
Ylonen in 1995, so as to secure data sent over unprotected
network like the Internet. It does so by encrypting the traffic
[1]. SSH software was open source and free, it was therefore
very popular on the market. Unfortunately in December 1995
Tatu made the software proprietary and it was no longer
freely available to users [2]. In 1999, Open source developers
set up a research community to develop a free version of the
Open Secure Shell (OpenSSH), deriving from the version of
SSH [3]. In 2000, around more than 2 million people were
using the version of OpenSSH for free. OpenSSH developers
advance that the application is more secure than the original
SSH protocol due to their policy of producing clean and
audited code under the BSD license known as a family
permissive free software licenses [4]. As such, Open SSH
became one of the most popular security implementation,
resulting by default in a large number of operating systems
[5]. Open source VPN is another OpenVPN developed
by a different research group. The advantage of using
OpenVPN over OpenSSH is that OpenVPN is deployable
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
K. M. S. Soyjaudah
Faculty of Engineering
University of Mauritius
Réduit, Mauritius
ssoyjaudah@uom.ac.mu
on practically all platforms whereas OpenSSH can only be
installed on operating systems which is identical to UNIX
[6]. However OpenSSH forms part of a big protocol family
unlike OpenVPN which can only be used for VPN tunneling.
Contribution to the research in SSHVPN has been very slow
and new version has fixed patches of security flaws and as
well as cryptography. Since OpenVPN releases have more
innovations, it is currently the choice of companies, for
their VPN needs [7]. The performance comparison however
between OpenVPN and OpenSSH VPN has not yet been
undertaken.
This document is organized into the following sections:
• Section 2 will cover the basics of VPN and give some
insight into the concepts associated with OpenVPN and
OpenSSH.
• Section 3 describes the Physical framework.
• Section 4 details the experimental testing.
• The last section entitled ’Performance measures’, per-
forms some exhaustive test between OpenVPN and
OpenSSH VPN.
• We then conclude by giving the results and observations.
II. V IRTUAL P RIVATE N ETWORK (VPN)
Virtual Private Networks (VPNs) [8] are nowadays be-
coming the most worldwide method used for remote access.
Companies tend to expand to multiple different locations and
have office branches in many countries. VPNs securely convey
information (file sharing, video conferencing, etc...) across
the Internet connection to remote users, office branches and
business partners into an extended corporate network. A VPN
is created using dedicated connections, virtual tunneling pro-
tocols or traffic encryption to establish a virtual point-to-point
connection. OpenSSH [9] (OpenBSD Secure Shell),based on
the SSH protocol is a collection of secure network-level
services which protect network communications by encrypting
network traffic and providing secure tunneling capabilities.
OpenVPN [6] is open-source and implements virtual private
network (VPN) techniques for creating secure point-to-point
or site-to-site connections in Wide Area Network. It uses a
custom security protocol and SSL/TLS is the key exchange
for the protocol.
III. PHYSICAL FRAMEWORK DESCRIPTION
In this article, the authors took into consideration the
physical topology layout, by using an existing live connection
from a company which shows the real layout how the router
is connected to the firewall and other devices like the switch
and server. This real-life scenario will give more accurate
values for the performance evaluation and analysis between
OpenVPN and OpenSSH. The open source VPN software,
OpenVPN has been the choice for many companies and
OpenSSH is less widely used VPN. According to technical
discussions, VPN technologies incorporating OpenSSH or
SSH Tunnel are predicted to provide fast speeds when
compared to using OpenVPN using TCP as base. OpenVPN
is usually slower than other protocols when used via TCP,
although TCP offers advantages in restricted networks [10].
When compared to stability, OpenVPN is prognosticated
to be a very stable protocol on all kinds of connection
(WLAN, wired, mobile) and OpenSSH is deemed to be
unstable, susceptible to disconnect with short network
issues. Thus, security features for both VPNs must be
carefully considered and the same encryption mechanism will
be applied in order to have the same parameters on both sides.

The server is located in country A. The two clients are

located in two different countries B and C. A VPN will be
setup between the clients and the server A, and two clients B
and C will then be on the same subnet and can communicate
with each other.

In this work, port-forwarding will be used. Port forwarding

[11] allows mapping of specific ports on remote or local hosts
to the SSH port (usually 22) on the network and port 1194 for
OpenVPN. Port forwarding, allows individual services in the
network to be accessed from the internet though ssh servers
by opening ports to inbound traffic for a specified service.
If the administrator for example hosts a web server inside a
LAN, he could configure a port forwarding rule for HTTPs
traffic. When Internet users send HTTPs requests to the WAN Fig. 1. Physical topology of VPN
public IP address, the router opens the specified port to direct
traffic to the server. For both OpenVPN and OpenSSH, port-
forwarding will be used. Once the tunnel is set up, the client
will have direct access the local network 10.1.2.0/24. In other to its processing power. The tests are performed through a
words, the client would be connected to the server through ssh Wide Area Network and the compression for both VPNs has
tunnel. Tunneling [12] is a way to transform data frames to been disabled. During the measurements, the size of packets
allow them pass networks with incompatible address spaces will be increased and different graphs will be plotted with
or even incompatible protocols [13]. axes; latency, jitter and speed of the VPN tunnel vs size of
packets. Iperf [14] will be used for monitoring. Iperf is the
IV. EXPERIMENTAL TESTING DESCRIPTION network testing tool which is often used to generate User
The experimental testing procedure is divided into three Datagram Protocol (UDP) and Transmission Control Protocol
main parts: evaluation of latency, speed of transfer, and jitter (TCP) data streams and measures the throughput of a network.
when sending UDP traffic in both VPN under a congested Parameters set using Iperf allow the user to test/ optimize or
network. All measurements took place at the server node due tune a network. Iperf has a client and server functionality,
and can measure either unidirectionally or bi-directionally
the throughput between the two ends. During the tests, data
packets with different sizes, ranging from 1Mb to 10Mb are
sent in the OpenSSH and OpenVPN tunnel successively. Data
is sent from the client to the server and the speed in Kbits/Sec
and Jitter in microseconds is carefully noted. During the
experiment, different text files with sizes from 1Mb to 10Mb
will be sent to the server using Iperf.

V. PERFORMANCE MEASURES
Using a set of metrics, the performance of a network can be
measured. For OpenVPN and OpenSSH, the speed of transfer,
jitter and latency, are used to represent the performance of
the VPN. Latency [15] is defined as an expression of the
amount of time a packet takes to move from one designated
point to another. The machine hardware, the link speed, and
the encapsulation time affect the latency through a VPN
tunnel. Jitter [16] is the variation in the time between packets
arriving, caused by network route changes or congestion.

The snapshot below shows how the tests are performed on

iperf and how the values have being recorded.
root@theo62:˜# iperf -s
------------------------------------------------------------
Server listening on TCP port 5001
TCP window size: 64.0 KByte (default)
------------------------------------------------------------ Fig. 2. Comparison of speed for OpenSSH and OpenVPN
[ 4] local 10.0.0.100 port 5001 connected with 192.168.1.4..
[ ID] Interval Transfer Bandwidth
[ 4] 0.0-18.1 sec 896 KBytes 405 Kbits/sec
[ 5] local 10.0.0.100 port 5001 connected with 192.168.1.4..
[ 5] 0.0-18.5 sec 896 KBytes 397 Kbits/sec
[ 4] local 10.0.0.100 port 5001 connected with 192.168.1.4..
[ 4] 0.0-15.8 sec 768 KBytes 399 Kbits/sec

root@PC:˜/Downloads$iperf -c 192.168.1.4 -F 8mb.txt

------------------------------------------------------------
Client connecting to 10.8.0.1, TCP port 5001
TCP window size: 64.0 KByte (default)
------------------------------------------------------------
[ 4] local 10.8.0.6 port 52373 connected with 10.8.0.1 port 5001
[ ID] Interval Transfer Bandwidth
[ 4] 0.0-11.5 sec 2.50 MBytes 1.82 Mbits/sec

This section provides a comparative view of the conducted

results and also an attempt to give a better explanation of
the experiment results has been made. However, important
characteristics of the LAN that can affect the performance of
the connection should be discussed first. The Cisco switch is
configured on a 10Mbps duplex setting. The maximum file size
has therefore been limited to 10Mbps for this test. All network
metrics for OpenSSH are remarkably concentrated. Standard
deviation values rendered in both graphs confirm this remark.
Figure 2, demonstrates that SSH VPN mechanism gives highly
increased speed in Kbits per seconds when compared to
OpenVPN. For instance, there is a big difference for a file of
9M and 10M. When the size of the file increases, the gap in
speed increases proportionally. Figure 3 depicts a comparison
of the latency when sending different file size inside the tunnel.
This plot gives a clearer idea about the achieved network
performance. When testing Jitter, UDP traffic was sent through Fig. 3. Comparison of latency between OpenSSH and OpenVPN
the tunnel using iperf features. There is negligible change
in performance when it comes to Jitter. In short, OpenSSH

Fig. 4. Comparison of jitter between OpenSSH and OpenVPN
scenarios visibly have the best performance in having a lower
latency and higher speed. The exponential increase in latency
can be explained by the fact that when the IP datagrams are
lost, duplicated or reordered, the sequence numbers are used to
reassemble the stream. The acknowledge number notifies the
sender if a segment was lost. If an acknowledge for a recently
sent segment does not reach in a certain amount of time,
the sender assumes a lost packet and resends that segment
[17]. The time it takes to re-send lost TCP packets within the
data transfer create a high latency on the transfer time. When
the TCP packet is larger, the frequency of drops is higher,
therefore a larger packet experience more latency.
Network administrators who needs a secure and reliable VPN
connection based on TCP can make use of OpenSSH VPN.
The experiments performed in this paper demonstrate that
OpenSSH is faster than OpenVPN in a scale or 1M to 10M.
In the Internet or Wide Area Network however, parameters
like bandwidth, delay and loss rate differ from one connection
another and the parameters change over time on a single
connection. When sending packets the timeout in seconds
does not have a fixed range neither on a fast LAN nor on a
congested international link. OpenSSH which is lighter will be
more convenient than OpenVPN. One potential disadvantage
of SSH is that it is over TCP, not over UDP.
VI. C ONCLUSION
This paper addresses performance comparison OpenVPN
and OpenSSH VPN connections. Two distinct scenarios
were used to test the two VPN mechanisms. The results
disclose that OpenSSH utilizes the link better and thus
provides radically improved transfer times and speed when
compared with OpenVPN. In conclusion, the OpenSSH is a
cost-effective solution for setting up a VPN for remote access
over the Internet. OpenSSH is throughput-limited to 10Mbps,
but is sufficient for most Internet connections.
As future work it would be good to expand this study
by investigating the performance of VPN on mobile devices.
However mobile devices cannot afford batteries with unlimited
capacity, so the amount of energy required for this type of
secure connections should be detected.
R EFERENCES
[1] R. Seggelmann, M. Tuxen, and E. Rathgeb, “Ssh over sctp x2014;
optimizing a multi-channel protocol by adapting it to sctp,” in Commu-
nication Systems, Networks Digital Signal Processing (CSNDSP), 2012
8th International Symposium on, pp. 1–6, July 2012.
[2] A. .B, D. .F, and N. .B, “Linux operations and administration,” vol. 1,
pp. 399–400, August 2012.
[3] A. Jacoutot, “About openssh.” http://www.openssh.com/, May 2014.
[4] A. Jacoutot, “Openssh.” http://www.openssh.com/portable.html, May
2014.
[5] D. Jim, “Ssh.” http://www.openssh.com/, 2014.
[6] Y. Yang, J. Ding, Q. Wen, and H. Zhang, “Research and design of the
pmi-based access control model for openvpn,” in Advanced Intelligence
and Awarenss Internet (AIAI 2010), 2010 International Conference on,
pp. 77–80, Oct 2010.
[7] D. Meng, “Implementation of a host-to-host vpn based on udp tunnel and
openvpn tap interface in java and its performance analysis,” in Computer
Science Education (ICCSE), 2013 8th International Conference on,
pp. 940–943, April 2013.
[8] D. Wood, V. Stoss, L. Chan-Lizardo, G. S. Papacostas, and M. E.
Stinson, “Virtual private networks,” in Private Switching Systems and
Networks, 1988., International Conference on, pp. 132–136, Jun 1988.
[9] Z. Dai, L. Wu, H. Xiao, H. Wu, and X. Chi, “A lightweight grid mid-
dleware based on openssh - sce,” in Grid and Cooperative Computing,
2007. GCC 2007. Sixth International Conference on, pp. 387–394, Aug
2007.
[10] S. vpnsecure, “Vpn secure.” http://support.vpnsecure.me/articles/
tips-tricks/comparison-chart-openvpn-pptp-ssh-tunnel, 2009.
[11] C. S. Community, “Setting up port forwarding and port
triggering on rv0xx series routers.” http://www.cisco.com/c/en/us/
support/docs/long-reach-ethernet-lre-digital-subscriber-line-xdsl/
asymmetric-digital-subscriber-line-adsl/12905-827spat.html, 2012.
[12] J. O. Fajardo, J. Pico, and A. Munoz, “New tunneling capabilities for
bgp/mpls ip vpn in gnu/linux,” in Networking, 2008. ICN 2008. Seventh
International Conference on, pp. 324–329, April 2008.
[13] L. Foundation, “Ipip tunnels.” http://www.linuxfoundation.org/
collaborate/workgroups/networking/tunneling, 2012.
[14] O. Olvera-Irigoyen, A. Kortebi, and L. Toutain, “Available bandwidth
probing for path selection in heterogeneous home networks,” in Globe-
com Workshops (GC Wkshps), 2012 IEEE, pp. 492–497, Dec 2012.
[15] S. Y. Park, H.-S. Ahn, and W. Yu, “Round-trip time-based wireless
positioning without time synchronization,” in Control, Automation and
Systems, 2007. ICCAS ’07. International Conference on, pp. 2323–2326,
Oct 2007.
[16] Y. Cai, S. A. Werner, G. Zhang, M. J. Olsen, and R. D. Brink, “Jitter
testing for multi-gigabit backplane serdes - techniques to decompose and
combine various types of jitter,” in Test Conference, 2002. Proceedings.
International, pp. 700–709, 2002.
[17] O. Titz, “Why tcp over tcp is a bad idea.” http://sites.inka.de/bigred/
devel/tcp-tcp.html, 2001.
 Overview of Data Quality Challenges in the Context
of Big Data
Suraj Juddoo
School of Science and Technology
Middlesex University(Mauritius Branch Campus)
Vacoas,Mauritius
s.juddoo@mdx.ac.mu

Abstract—Data quality management systems are Big Data quality dimensions, methodologies and activities
thoroughly researched topics and have resulted in many tools would be discussed to result with research questions most
and techniques developed by both academia and industry. relevant for data quality in Big Data context.
However, the advent of Big Data might pose some serious
questions pertaining to the applicability of existing data quality
concepts. There is a debate concerning the importance of data
quality for Big Data; one school of thought argues that high
data quality methods are essential for deriving higher level
analytics while another school of thought argues that data
quality level will not be so important as the volume of Big Data
would be used to produce patterns and some amount of dirty
data will not mask the analytic results which might be derived.
This paper aims to investigate various components and
activities forming part of data quality management such as
dimensions, metrics, data quality rules, data profiling and data
cleansing. The result list existing challenges and future
research areas associated with Big Data for data quality Figure 1 Data Quality components
management.
Data quality dimensions are the ways to express the
Keywords—Big Data, Data quality, Data profiling, Data notion of data quality; the major investigation concerning
cleansing, data quality rules, dimensions, metrics.
dimension for this paper concerns whether the same
dimensions which have been applied for traditional data
I. INTRODUCTION quality strategies such as completeness, relevancy,
We are currently living at the beginning of a Big Data consistency amongst others, could also be relevant for Big
era. This Information Technology based concept might Data quality strategies? Or could there be new dimensions
prove to radically change the ways organizations operate, only applicable in the context of Big Data?
and at a larger scale how human society operates. Big Data In order to quantify dimensions so as to bring some
is interesting not only to IT people, but also to physicists, notion of measurability and comparison to the dimensions,
mathematicians, politicians, law/security officials and also metrics need to be applied. Thus, there is the need to
people involved in the tourism and hospitality sectors investigate whether the same metrics which have been
amongst many others[6]. The term „Big Data‟ in itself is a traditionally applied to measure data quality dimensions
poor definition of its representation; it often only conveys could still be applicable for dimensions determined
the idea of a large volume of data too large to be handled by applicable to Big Data, and whether existing metrics and
current processing power of computers[12][13]. However, their formulas would need to be upgraded.
Big Data does concern the large volume of data but also
includes the capacity to search, process, analyze and present The execution part of a data quality strategy would
valuable information coming from large, diverse and rapidly consist of several activities whose common goal is to
changing datasets. This leads to Big Data being defined by improve the quality of data based upon the dimensions
volume, variety and velocity [13]. Veracity is another identified in a particular context. Some of the most cited
characteristic of Big Data which is growing in popularity data quality activities are:
and discusses the rising issue of certainty involved with
 Data profiling: examination of data sources to
using data.
generate information about the data and the
The principal goal of the paper is to analyze and present datasets.
the data quality techniques which would be more
appropriate for Big Data in a general context; issues such as

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

  Data cleansing: consists of techniques whose main Believability Extent to which data is regarded as
purpose is to enhance or transform data based upon true and credible
certain business rules. Completeness Extent to which data is not missing
 Construction of data quality rules(DQRs): the rules and is of sufficient breadth and
which would be used to determine whether there depth for the task at hand
are issues with a given datasets and which would Consistent Extent to which data is presented in
be used for the cleansing activity. representation the same format
Ease of Extent to which data is easy to
This paper is architected in the following structure; manipulation manipulate and apply to different
Literature review section would discuss in more detail about
tasks
related and connected earlier literature involved around data
quality methods, tools, measures, metrics which have Free-of-error Extent to which data is correct and
traditionally been used and new data quality methods, tools, reliable
measures, metrics being proposed for big datasets. Analysis Interpretability Extent to which data is in
section will discuss main findings via the use of valid appropriate languages, symbols and
scientific methods. The last section would be the wrap-up of units, and the definitions are clear
the report with the main results being emphasized. Objectivity Extent to which data is unbiased,
unprejudiced and impartial
II. LITERATURE REVIEW Relevancy Extent to which data is applicable
and helpful for the task at hand
Surprisingly, there are not many scholarly debated works
Reputation Extent to which data is highly
defining data quality; the definition which is being
regarded in terms of its source and
considered for this paper is “reliability and application
content
efficiency of data”[2]. Consequences of improper data
Security Extent to which access to data is
quality have been reported as costing US organizations
restricted appropriately to maintain
around 600 billion US dollars a year and also been blamed
its security
for the explosion of the US space shuttle “challenger”[1].
Timeliness Extent to which data is sufficiently
There are argumentations that data quality involved with
up-to-date
Big Data could propose different challenges due to the large
volumes and several varieties of data sources which need to Understandability Extent to which data is easily
be processed in a very short timeframe[18]. Furthermore, comprehended
compared to a „closed world‟ scenario with very specific Value-added Extent to which data is beneficial
data needs of users for data in a data warehouse, the needs and provides advantages from its
of data analytics for Big Data might vary according to the uses
context of their use. Thus, some data which might be
irremediably considered as „dirty‟, that is, not of appropriate Batini and Scannapieca(2006) discuss the idea that each
quality for a data warehouse application might in some dimension is due to cover specific aspects which might fall
situations be considered very useful in other situations. The under the general idea of data quality. With each dimension,
main point here is that data quality for big datasets depends there should be several metrics which could be applied to
upon their context of use[4].There are such a number of quantify a given dimension, and for each metric, there could
ways data are used and applied that expressing quality has be more than one measurement method. The following table
been expressed in terms of dimensions[1][4]. summarizes the dimensions discussed by the above named
authors:
A. Big Data quality dimensions
Throughout literature, there have been lots of different sets Dimension Definitions
of dimensions being considered by several authors. A brief Accuracy Closeness of representation of a real-
comparison is given hereunder: life phenomenon that a data value
tries to represent; measured by edit
In [16], the authors have investigated how to measure or distance comparison functions
assess level of quality of data. They argue that some Correctness Also termed as syntactic accuracy,
assessments of data quality could be task independent, which refers to the closeness of a data
therefore not restrained by the context of application while value with respect to a domain
others are task dependent. The following table depicts the
Completeness Measure of missing values for a
main dimensions they thought were worthy of discussion:
specific column in a table; often
Dimensions Definitions illustrated via the NULL value and
Accessibility Extent to which data is available, or which could represent facts as value
easily and quickly retrievable not existing, value existing but
Appropriate Extent to which volume of data is unknown and not knowing if value
amount of data appropriate for the task at hand exists. Influenced by the closed and
 open world assumptions.
Currency Concerns how promptly data are
updated; can be measured by the
lastupdated metadata
Volatility Characterizes frequency with which
data varies in time; metric given by
length of time data remains valid
Timeliness Expresses how current the data is for
the task at hand; involves currency
measurement and check whether data
is available before planned usage time
Consistency Concerns the violation of semantic
rules defined over data items and
usually expressed an integrity
constraints
Accessibility Ability for a user to access data from
his own culture, physical status and
technologies available.
Believability Whether a certain source provides
data which can be considered as true,
real and credible
Reputation Considers how trustable is an
information source
Objectivity Takes into account impartiality of
sources
Value added How beneficial data is and advantages
derived from their use
Relevancy How applicable is the data for the
current task
Ease of How much data is clear, without
understanding ambiguity and easily comprehended
A mere comparison of the dimensions mentioned by the
two set of authors cited above clearly indicate a high level
of correlation and similarity between what are the principal
ways that data should be assessed for quality purposes.
However, we raise the question concerning how far those
dimensions would still apply in a Big Data context? What
are those dimensions which are judged to be more
important for data quality in a Big Data use context?
B. Dimensions of data quality for Big Data
The high volume and velocity properties of Big Data entails
that segregating correct and erroneous data for further data
analysis is more important. Also, due to data coming from
multiple sources, there is a need of higher method of data
integration to harmonize the semantics of the data being
used[17]. However, according to [18], the importance of
improving data quality for Big Data might not be so
important as the amount of incorrect data is deemed to be
negligible to affect the final outcome after data have been
analysed. Thus, which of those two completely contrasting
schools of thought is relevant seems to depend upon the
amount and impact of the erroneous or „dirty‟ data as part of
a big dataset. This increases the importance of
understanding which dimensions are more relevant for Big
Data.
Caballero et al posit that the main data quality dimension to
be considered for Big Data is consistency, which they
explain as the capability of information systems to ensure
uniformity of datasets when data are being transferred
across networks and systems[4]. Their main hypothesis is
that the business value of a dataset can only be estimated in
its context of use. They further subdivide consistency into
three subsequent parts, as discussed here under. However,
they have connected many of the traditional data quality
dimensions with the three consistency subdomains as
follows:
 Contextual consistency refers to how far big
datasets are used within same domain of interest
independently of data format, size and velocity of
production of data. Thus, relevancy, credibility,
ease of understanding, accuracy and confidentiality
are considered to be very important for contextual
consistency to occur.
 Temporal consistency conveys the idea that data
needs to be understood in a consistent time slot,
such that the same data might not be comparable if
they are not from the same time slot. Time
concurrency, availability and currency are deemed
to be essential for temporal consistency.
 Operational consistency brings in the operational
influence of technology upon the production and
use of data. Availability, portability, precision,
completeness and traceability are considered the
main connected dimensions here.
Caballero et al have mapped how the 3v‟s of big data
affect the 3Cs of data quality as follows:
Figure 2: Matrix of 3Cs relative to the 3Vs
However, the ways the dimensions have been mapped to the
3v‟s were based solely upon hypothesis and no actual
research method has been applied to generate this mapping.
Thus, there is a need in terms of research in the area of Big
Data quality to deeper investigate which dimensions are
more important in the context of Big Data.
C. Measurements and metrics
As briefly outlined in the previous section, the measurement
of data quality is closely linked with measuring the
dimensions of data quality. Most metrics used for
measurement of data quality are normally within a range
from 0 to 1, with 0 representing incorrect value and 1
representing a correct value [3]. Many dimensions such as
accuracy, completeness and consistency amongst others are
calculated by the following function:
D = 1 – (Ni/Nt) (1)
Where D is the metric for a given dimension, Ni is the
number of incorrect values and Nt is the total amount of
values for the dimension concerned. This measurement and
associated metric would definitely still hold even for Big
Data, but it could be quite difficult to derive both Ni and
Nt in situations where there is constant input of data. Thus,
the velocity aspect of Big Data could be the most
problematic property in terms of Data Quality
measurement; but if this velocity aspect has been mastered,
there is no reason why the same metrics won‟t be
applicable for Big Data.
D.Data cleansing
Data cleansing is a well cited process and potentially
involve the most data transformations with respect to data
quality activities. The need to transform or edit some data
source to meet certain data quality standard is an important
dilemma when it comes to Big Data quality; as the same
data could be used or analyzed towards different use cases
with Big Data, transforming the original dataset according
to the business rules for one use case might negatively
impact the Big Data activities for another use case with the
same dataset[12].
Some researchers from IBM-research India have
identified four stages as part of data cleansing process for
large enterprise datasets and are summarized in the
diagram below[9].
Figure 3: Main stages of data cleansing process
To accomplish those four main stages, the above named
authors cite different tools and methods which should be
applied as part of or supporting the different activities. The
following gives a brief summary of some of them:
 Classification: building classes from the data in a
dataset is explained to be the first step to build rule sets.
An example of classification for address standardization
is that the „street‟ value could be linked with
corresponding values such as „Road‟, „avenue‟ or
„lane‟; thus all those values could be assigned the same
classification label.
 Patterns: they give a generalized view of how the data
is formatted; it involves parsing the data and classifying
all tokens into appropriate classes and replacing those
classes with pre-defined labels. An example of a
resultant pattern could be N/N+B+II where N
represents numbers, B is a label for city names and I
represent a single alphabet. Rules are then written to
process the patterns.
 Dictionaries: standardized data could be validated
against some domain to ensure proper identification.
For example, a dictionary of available cities in a
country could be used to validate city names stored as
part of addresses in a dataset.
 Discovering variants of a term: this a very elaborated
sub-stage which would involve the use of reference sets
for each token or value of data; use of syntactic
clustering concerning data which possess the same set
of terms in a sequence except some minor differences;
use of resemblance measure for detecting groups of
similar data which would use a formula to denote; use
of a diff-utility to find the difference between groups of
similar records.
Thus, [9] have proposed a data-driven tool relying upon
detecting characteristics of „dirty‟ data in a given dataset.
The dependence of using domain experts has been
minimized. However, the question as to whether this same
proposed data cleansing methodology could be applied to
Big Data datasets is more than ever relevant. How feasible
will it be to devise reference sets and creating rule sets given
the 3v‟s characteristics of Big Data? What would be the
performance of this data cleansing method in order to
generate dictionaries and applying resemblance measures
and diff-utility? Some serious scientific research is
warranted to uncover those aspects.
Dealing with user defined functions (UDFs) has been
reported as one of the challenge in implementing data
cleansing when scaling to Big Data. Thus, Khayyat et al has
proposed a new architecture to incorporate more efficiently
the application of UDFs and named it „BigDansing‟[10].
The typical data cleansing steps as described by the above
named authors are (1) specifying quality rules, (2) detecting
errors w.r.t data quality rules and (3) repairing detected
errors. However, detecting and repairing data quality issues
are reported to face some issues, namely:
a) High complexity of rules leads to intractable
computations over large datasets and thus limiting the
applicability of data cleansing systems.
b) Effective parallelization is hard to achieve with UDFs
when the latter is specified through the use of
procedural languages.
„BigDansing‟ is reported to deal with those issues by (1)
abstracting and simplifying the process of rules specification
for UDFs and (2) to enable the application of distributed
repair algorithms. „BigDansing‟ have been benchmarked
with other systems which could support some level of data
cleansing routines such as Spark SQL, Shark, NADEEF and
PostGreSQL. The results show that „BigDansing‟
outperforms the other systems using measures such as time
to scale data quality activities upon large datasets, higher
efficiency in deduplication of large datasets and
improvements of repair efficiency.
However, it could be argued that the process of rule
specification being on the shoulders of users could be one of
the limiting factors of a system such as „BigDansing‟. Thus,
questions such as whether statistical processes in terms of
supervised or unsupervised learning models could automate
the process of rule specifications and further enhance the
efficiency of such systems?
Another recent method to help improve data cleansing for
Big Data has been with the application of Bayesian
networks and has been termed BayesWipe [20]. The authors
emphasize that traditional data cleansing techniques such as
outlier detection, noise removal, entity resolution and
imputation cannot provide effective solutions in the context
of Big Data. The fact that techniques such as CFDs depend
upon clean external reference sets to learn data quality rules
is one of the major drawbacks in devising effective data
cleansing solution with Big Data. Even devising rules from
the „dirty‟ data is not judged to be a satisfactory enough
solution[20]. Thus, the authors posit that a statistical process
underlies the generation of both clean and dirty data with
the data source and error models used to undertake the
detection and repairing stages of data cleansing. Algorithms
are generated from the statistical process are coupled with
updated query rewriting techniques; the fact that BayesWipe
could also be applied in an online scenario where only the
top-k data portion of the data are considered and the
cleansing process is performed while the data is being
retrieved adds its improved applicability in the context of
Big Data. Empirical evaluations performed over both
synthetic and real datasets tend to show improvements in
terms of amount of data cleansing ratios when BayesWipe is
compared with CFDs and Amazon Mechanical Turk, but
there is still a very large portion of dirty data not cleansed.
For example, the offline version cleans only 40 % of the
data in a synthetic car database. Another question about
BayesWipe concerns the efficiency of the data source and
error models which is the foundation of this method. The
evaluation results denote that those models could be
improved to lead to higher data cleansing ratios.
All the above discussed techniques apply data cleansing
techniques upon the data itself, but do not attempt to correct
the cause which create the errors and dirtiness in the data; as
most of those data quality issues are reported to be
systematic, thus inherent to the process of data creation, it
is quite reasonable to find meaningful ways to cure the
causes of data quality errors as an efficient data cleaning
process [22]. Xiaolan et al explain that diagnosing data
quality errors in Big Data environments raise some issues
for traditional methods such as provenance analysis, feature
selection and causal analysis in terms of Massive Scale,
System complexity and High error rates.
Data X-Ray proposes to overcome the above issues by (1)
finding a hierarchical structure of features which best
represent erroneous elements, (2) using Bayesian analysis to
estimate the causal likelihood of features being associated
with potential causes of errors and diagnose those causes
using conciseness, specificity and consistency dimensions.
E. Data profiling
Profiling of data is basically about examining data available
in a given data source and collecting/producing statistics
and information such as metadata, relationships,
dependencies, patterns and cardinalities [14]. The results out
of a data profiling job usually proves to be very useful as
they contribute towards creation of constraints and rules
which could be applied during data cleansing. The
traditional use cases of profiling include query optimization,
data cleansing, data integration, scientific data management
and data analytics; out of those, data cleansing and data
analytics seem to be the more relevant when we consider
data quality for Big Data.
Big Data is reported to raise three main impacts/challenges
when it comes to data profiling:
1. Profiling could be very useful in assessing
usefulness of known and unknown data; this could
help deployment of future Big Data use cases.
2. High variety and volume of data challenges
existing data profiling tools and techniques in
terms of computational complexity and memory
requirements amongst others.
3. New data management and architectures are
involved with Big Data, such key-value and
document based stores, high usage of parallel and
distributed systems and so on. Thus, there is the
need to re-think how data profiling need to be
carried out in the context of Big Data[14].
The following chart maps the upcoming research challenges
associable between Big Data and data profiling.
 Figure 4: Research challenges for Big Data profiling
Online profiling is triggered by the fact that users might
need to wait a substantial amount of time before viewing the
results if profiling is handled by traditional data profiling
methods; hence, the idea of displaying intermediate results
with proven level of confidence might help the user to take a
decision whether they would want to continue to work for a
specific use case. Given the huge amount for Big Data, this
would certainly be an appreciable feature for potential Big
Data users such as data analysts and data scientists.
Furthermore, it is imperative to understand that profiling
would be an activity to be performed both before and after
data cleansing as after the cleansing process, the information
about the data and the data source would be updated.
Incremental profiling is linked with the velocity
characteristic of Big Data. As data is expected to be updated
quite often, there should means to perform profiling upon
changing data with periodic timespans. Re-using past
profiling results to improve the computational response
times of profiling might be an avenue to explore!
Continuous profiling is almost the same idea as
incremental profiling with the exception that we are now
expecting profiling to be performed upon data while it is
being created or updated.
One of the aim to profile Big Data would be to determine
the common properties or level of heterogeneity of different
datasets; thus, two types of heterogeneity is of particular
importance when profiling for Big Data, namely syntactic
heterogeneity, which is mainly about finding inconsistent
formatting between data, and structural heterogeneity,
which is about unmatched schemata. Consequently, the
information generated by profiling would be very useful in
the integration aspect of different sources of a Big Data use
case. This is an open area of research as existing solutions
simply appears to be ineffective for Big Data.
Recognizing the domain of yet unknown data is the aim of
topical profiling. With Big Data use cases, there could be
the use of unknown datasets such as social media data.
There should be some thorough research upon how topical
profiling could be implemented efficiently for Big Data.
F. Data quality rules
Enforcing Data quality rules(DQRs) are integral activities to
improve the cleansing part of data quality. Prior research in
the field show that DQRs are being enforced via various
methods such as Functional dependencies(FD), Conditional
Functional Dependencies(CFDs), Dedupalog, Integrated
Constraints(ICs) and Bayesian networks amongst
others[5][23][24]. All of those techniques have been
evaluated through several researches with the common
purpose of improving the efficiency of the cleaning or
repairing activities. Another common theme through all
those researches is the fact rules are discovered out of the
characteristics present in the data.
Yeh and Puri aimed at increasing consistency in datasets by
discovering rules for more efficient CFDs[24]. Current
challenges with increasing consistency are (1) it is a labor
intensive process and (2) rule discovery is largely a manual
process which relies heavily upon subject matter experts.
Furthermore, current methods for discovering CFDs have
been reported to have difficulties to scale for relations
having a large number of attributes and they are not robust
with datasets having a high level of dirty data. Thus, Yeh
and Puri have developed an approach called CFinder which
follows the following main steps to automatically generate
better CFDs:
Figure 5: Main steps of CFinder
Even if CFinder outperforms CFD-TANE in terms of the
recall and precision metrics, Yeh and Puri are investigating
ways to improve CFinder with (1) the use of heuristics to
improve scalability (2) application of industry ontologies to
determine which attributes are related for the pruning
process and (3) exploration of other metrics to eliminate
weak CFDs.
There could be pitfalls relying upon automated data repair
solutions based on DQRs, especially in use cases dealing
with critical data. However, the involvement of human users
to validate data repairs means the response time of data
quality tools degrade considerably. Thus, an interactive
approach which performs some proportion of automatic
repairs while allowing users to validate repairs have been
proposed [23]. It involves generating repairs to only the top-
k most important violated rules and ranking the most
beneficial repairs from the users‟ perspectives for their
validation.
Experiments comparing the rules ranking method with other
techniques such as the Greedy and Random algorithms tend
to demonstrate that the rule ranking method is more
efficient. However, in the context of Big Data, there are
several questions which would be raised by the current
research:
1. How practical is it to have the user validating repairs,
even for the top-k rules, when there could be the
assumption of a huge size of top-k repairs to be
undertaken?
2. There should be some measures for the computational
complexity of the rule ranking algorithm as it involves
nested loops and user interactions. Furthermore, the
stopping condition of one of the loop equals to the fact
that there is no more dirty tuples in a given dataset. In a
Big Data scenario, with the high velocity of data
production, this could well result in the algorithm
generating infinite loops!!
3. The top-k repairs would invariably be linked with the
use case for which analytics are being applied in a Big
Data dataset. As it is already questioned in this
research, there is a legitimate question whether to
transform the data repairs according to one particular
use case and thus update the original dataset OR create
a copy of the corrected data while keeping the original
dataset for other use cases.
RULEMINER is a system to discover DQRs addressing the
main limitations of existing rules discovery methods [3]:
 Existing rule discovery algorithms are usually
designed for a single rule language, thus unable to
discover many useful rules for a dataset.
 Most existing algorithms generate a large amount
of rules, where many of those rules are not
adequate.
 Manual evaluation of the output of rules is a time
consuming process.
RULEMINER discovers rules expressed as Denial
Constraints (DCs) which is supposed to subsume FDs and
CFDs. However, it‟s quite unclear whether DCs would
subsume more elaborate rules such as for semantic
interpretation of data as discussed in the data cleansing
section [9] of this current research, thus the first limitation
listed above is unsure to be addressed by RULEMINER.
Another issue with RULEMINER remains the fact that there
is a dependence upon users to validate repairs in terms of its
Negative Example-Positive Examples pairing; with Big
Data, this could result in computationally too costly repairs.
However, this method seems to be a very user friendly with
a front end interface which allows by easily allowing the
user to specify the maximum amount of errors to display for
a given discovered rule(similar to the top-k notion) and a
filtering option allows the user to focus upon certain rules
depending upon a given use case.
III. FINDINGS
A. Dimensions
Concerning data dimensions which provide the basis of
measuring data quality, the above discussions sourcing from
existing literature point to two important conclusions:
Applicability of traditional dimensions: the dimensions
which are normally cited for measuring data quality for
normal datasets are on the whole still relevant in the context
of Big Data. Thus, consistency, completeness, accuracy and
credibility/believability have been ranked among the most
useful dimensions for big datasets. Further empirical
research is recommended to validate the assumptions of
caballero et al. to link the traditional dimensions with the
“v‟s” of Big Data.
Dimensions tightly coupled with use case: the area of
application or use case of big datasets highly influence
which dimensions would be more relevant to ensure data
quality. For example, the use of big datasets coming from
social media to generate sentiment analysis systems could
require higher degree of consistency and credibility
compared to completeness. On the other hand, big datasets
used in the context of power consumption and sensor/RFID
based data might demand a more important component of
accuracy and/or completeness to ensure a better data quality.
However, whatever the use case, it has been discussed the
data quality activities are slowly growing in importance to
ensure that the results of analytics could be trusted.
B. Measurement and metrics
Most measurements and metrics applicable in normal data
quality practices would still hold for Big Data, even if the
velocity inherent with Big Data might make metrics
calculation more cumbersome. Getting interim accurate
results from data profiling would be highly useful.
However, the major area of research would most certainly
relate to the meaning of the measurement, e.g can it be
argued that a 90% consistency measure of data in a data
warehouse system be equivalent to a 90% consistency
measure of data in a Big Data context? This seems highly
unlikely as due to the 3v‟s of Big data, there is bound to be
higher tolerance of dirty data.
C. Data quality rules
Rules discovery normally involve several expressions such
as CFDs and DCs. However, due to the heightened decision
making exigencies associated with Big data analytics, there
is a need to create new methods which would efficiently
deal with multiple and complex rules.
Automatic or user-defined rules: The involvement of
users during rules generation would be a limiting factor with
Big Data, but the rules could be potentially so complex that
there could be no other alternative than relying upon user
validation for certain type of highly complex repairs. On the
other hand, automatic discovery and application of rules to
the repair process might be quite dangerous, especially in
use cases dealing with critical data. Thus, this research
recommends discovering DQRs for Big Data with the use of
algorithms which should be robust enough to automatically
discover a certain level of repairs potentially via heuristics
or learning with dirty data in terms of Bayesian networks;
human validations should be restricted to only repairs which
are occurring to very ambiguous and complex rules.
Computational complexity: The most recent research point
towards aiming at the most important or top-k rules which
must be taken into account to generate enough confidence
level in the use of Big Data. This could largely be left upon
subject or domain specialists to identify and denote those
top-k rules; ultimately, the weak spot of complex and
efficient data quality management systems for Big Data
could once again be the capability of the human specialists
to correctly denote the top-k rules for particular use cases.
D. Data profiling
Profiling is one of the most important groundwork to enable
a proper data quality management system and is central to
the proper generation of DQRs and performing data
cleansing. However, the traditional outputs of profiling are
seriously challenged by Big Data in the following terms:
New Big Data models: The fact that many Big Data models
such as HBase do not use indexes imply that some of the
outcomes of traditional profiling such as foreign key
dependencies no longer hold. Thus, even if traditional
profiling results such as number of NULL values could still
be relevant, there is a growing need for data profiling to
provide other statistics in view of the new Big Data models.
On the other hand, the fact that most architecture of Big
Models makes use of in-memory processing could resolve
some of the disk related computational issues generally
involved with data quality activities.
Online profiling: it needs to be investigated deeper as the
intermediate results of profiling could allow both decision
takers and data quality management developers take the
most suitable actions in a context of high velocity of data.
E.g, it could ease the formulation of the top-k rules much
more rapidly or identify inconsistency patterns for certain
attributes which are more repetitive, and therefore support
the data cleansing activity on the fly.
Variety: Harnessing the power of yet unknown data is one
of the benefits of implementing Big Data solutions. Those
data are very often in improperly defined or totally
unstructured format which is a major obstacle to properly
carry out profiling activities and could ultimately result in
vastly incorrect profiling outputs. Thus, there is a high need
to develop new profiling techniques taking into account
heterogeneity of data sources to consolidate the integration
level required for Big Data implementation.
E. Data cleansing
Cleansing or repairing activities are normally the
culmination of any proper data quality management process.
The following are the major challenges which Big Data
seems to bring for data cleansing:
Repair upon original data: How some datasets are
intended to be used is critical towards understanding how
cleansing activities would need to be carried out. Thus, in
situations where the same dataset could be subject to
different types of analysis as it has been seen in many Big
Data use cases, repairing or providing edits on the original
dataset might prove very beneficial for a particular use case
but could cause the dataset to be not suitable for other use
cases. Thus, there should be the development of
mechanisms which denote original and edited data for Big
Data cleansing activities.
Symptoms v/s causes: Unfortunately, many data quality
management systems correct the data which is judged to be
dirty according to some rules rather than the actual cause or
triggers of those errors. Thus, with the high velocity of Big
Data, we could face a situation where a previously cleansed
dataset could be filled with new errors very rapidly. E.g,
machine generated data from sensors which are improperly
calibrated might be the cause of flooding datasets with
continuously improper or dirty data. Instead of wasting
resources to continuously detect and correct the data, it
makes more sense to re-calibrate the sensors accordingly.
Volume issues: Typical data cleansing methods rely upon
reference sets or dictionaries of clean data in order to
compare and clean a particular dataset. The availability of
this clean reference set can prove to be impossible to
produce with Big Data use cases; thus, existing research are
moving towards the use of clean reference sets towards the
use of heuristics or statistical processes to detect dirty data,
but those techniques still need to be improved for the
efficiency of the amount of dirty data corrected. E.g,
Bayeswipe is reported to clean only 40% of dirty data in a
synthetic and not so huge dataset. With real life big dataset,
this ratio should be worse. Hence, there is a need to have
more efficient data cleansing methods.
Computational complexity: some data cleansing activities
might involve huge processing power. E.g, performing
resemblance measure might perfectly be achievable in
datasets of millions of records with tens of attributes, but
when it is being scaled to Big Data proportions of billions of
records across thousands of attributes, it could result into
disastrous response times if the data quality solution is not
being properly powered by due processing capabilities.
IV. CONCLUSION AND FUTURE WORK
Data quality has been of major importance since the
growing importance of information systems for decision
making in the past decades. The fact that Big Data promises
to unlock largely untapped data sets the scene for the
importance of data quality once again; it would be foolish to
be able to process and analyze large quantity and variety of
data if those data cannot be guaranteed to be „fit for use‟,
thus potentially the benefits which data analytics promises
to bring might be jeopardized by improper handling of the
data quality process. The latter has largely been researched
specially in the context of data warehousing and produced
very positive results in terms of methodologies, tools and
techniques to improve the quality of data in known data
environments. However, Big Data is largely up to now a
still burgeoning and developing data environment being
characterized mostly by the volume, velocity and variety
aspects. Many researchers are starting to investigate the
veracity or data quality aspect relative to Big Data. Hence,
how relevant are the known and used methodologies, tools
and techniques in the context of Big Data?
Throughout this paper, there has been the investigation of
main elements pertaining to data quality. The first general
learning from and consensus amongst experts in Big Data
quality circle is that many traditional data quality processes
are highly relevant or would need some minor updates to fit
the Big Data environment. Still, there are certain areas
which need to be largely updated for Big Data. Hence, the
dimensions relevant for Big Data seems to be more relevant
to particular use case and thus, it would seem highly
improbable to come up with some general dimensions
which could be argued to be always applicable. The
measures and metrics being developed so far tends to
indicate an overall applicability to Big Data, with doubts
remaining only to the velocity or ever changing statistics of
the data in Big Data context. Thus, we could argue that
measurements and metrics for Big Data would have a highly
important temporal aspect to be able to be used for
comparisons and decision making. In terms of data quality
rules, there is a big challenge associated with Big Data when
it concerns user-defined functions and the human
validations of rules discovered; the volume, velocity and
variety of the data might make it impossible for human
validation of rules to be established so that they might be
applied for data cleansing processes on time. Thus, the
whole method of rule discovery needs to be investigated in a
deeper basis, as it could also prove to be computationally a
very expensive process. Data profiling for Big Data might
also require new methods of being performed; online
profiling seems to be one of the demand of profiling experts
which could be applied for Big Data and there definitely
need to be new methods to resolve data profiling issues
related to heterogeneous datasets. Finally, there are lots of
challenges set by Big Data for the data cleansing activity;
the volume of data might just make it computationally too
expensive to repair all dirty data detected, whether to make
corrections on original dataset or make a copy of corrected
dataset depends upon the characteristics of the use case of
the Big Data scenario, and determining the exact causes of
dirty data and eliminating those causes are major research
challenges.
Future work relative to this research would involve carrying
more empirical evaluations and testing of the main research
questions set out in the previous paragraph; solving those
different questions would gradually bring a new
nomenclature towards the creation of data quality processed
which could be more successful in the context of Big Data.
REFERENCES
[11] Loshin David, 2014. Understanding Big Data Quality for Maximum
Information Usability, s.l.: s.n.
[12] Loshin, D., 2013. Big Data Analytics: from strategic planning to
enterprise integration with tools, techniques, NoSql and Graph.
Elsevier.
[13] Malik, P., 2013. Governing Big Data: Principles and Practices. IBM
Journal of Research and Development.
[14] Naumann, F., 2013. Data Profiling Revisited. s.l., s.n.
[15] Obhyung Kwon, N. L. S., 2014. Data quality management, data usage
experience and acquisitionintention of big data analytics. International
Journal of Information Management, pp. 387-394.
[16] Pipino, L., Yang, L. & Wang, R., 2002. Data Quality Assessment.
Communications of the ACM.
[17] Saha, B. & Srivastava, D., 2014. Data Quality: The other face of Big
Data. s.l., s.n.
[18] Soares, S., 2012. Big Data quality. In: Big Data Governance: An
emerging imperative. s.l.:MC Press, pp. 101-112.
[19] Steve lavalle, E. l. R. S. H. N. K., 2011. Big Data, Analytics and the
path from insights to value. MIT Sloan Management Review.
[20] Sushovan, D., Yuheng, H., Yi, C. & Subbarao, K., 2014. BayesWipe:
A Multimodal System for Data Cleaning and Consistent Query
Answering on Structured BigData. s.l., s.n.
[21] UNECE, 2014. A suggested framework for the quality of Big Data, s.l.:
s.n.
[22] Xiaolan, W., Xin Luna, D. & Alexandra, M., 2015. Data X-Ray: A
diagnostic tool for data errors. Melbourne, ACM.
[23] Yakout, M., Elmagarmid, A. K. & Neville, J., 2010. Ranking for data
repairs. s.l., IEEE.
[24] Yeh, P. Z. & Puri, C. A., 2010. An efficient and robust approach for
discovering Data Quality Rules. s.l., IEEE.
[25] Youngwa Lee, K. K. R. L., 2003. The Technology acceptance
model:Past, present and future. Communication for the association of
information systems.

[1] Anon., 2006. Data Centric Systems and Applications. In: Data Quality.
s.l.:Springer.
[2] Anon., n.d. data quality. [Online]
Available at:
http://searchdatamanagement.techtarget.com/definition/data-quality
[Accessed 4 May 2015].
[3] Blake, R. & Mangiameli, P., 2011. The effects and interactions of Data
Quality and Problem Complexity on Classification. ACM Journal of
Data and Information Quality, 2(2).
[4] Caballero, I., Serrano, M. & Piattinni, M., 2014. A data quality in Use
model for Big Data. ER workshops, pp. 65-74.
[5] Chu, X., Ilyas, I. F., Papotti, P. & Ye, Y., 2014. RULEMINER: Data
Quality Rules Discovery. s.l., IEEE.
[6] danah boyd, K. c., 2012. Critical questions for Big Data. Information,
Communication and Society, 10 May, pp. 663-679.
[7] Davenport, T. H., 2013. At the Big Data crossroads: turning towards a
smarter travel experience, s.l.: Amadeus IT group.
[8] Gretzel, U., 2013. Technology and tourism: building competitive
digital capability. s.l.:s.n.
[9] Hima, P. K. et al., 2011. Data Cleansing techniques for Large
Enterprise datasets. s.l., IEEE Computer Society.
[10] Khayyat, Z. et al., 2015. BigDansing: A system for Big Data
Cleansing. Melbourne, ACM.
 A Classification Method to Classify
High Dimensional Data
Amit Gupta Naganna Chetty
Department of Computer Science Department of Computer Science
and Engineering, Graphic Era Hill and Engineering, Mangalore Institute of
University, Technology and Engineering,
Dehradun Uttarakhand, India Karnataka, India nsc.chetty@gmail.com
amitgupta7920@gmail.com
Abstract—The rapid computerization and advancement in
the technology has led to huge amount of data in the databases.
Research has shown that the amount of data in the world doubles
in every 20 months. However, this available data consists of large
number of noise values and thus, cannot be directly used. The
extraction of information from the vast pool of data has emerged
a major challenge.
Machine learning techniques have emerged as an effective
tool to overcome this challenge. Several machine learning
algorithms (like SVM, K-means etc.) are effectively applied in
data mining.
In this paper author have applied classification and
clustering techniques on different datasets and have proposed a
model for enhancing the performance of K-means data clustering
method and Naïve Bayes data classification method. The
efficiency of the proposed model is calculated based on general
parameters like accuracy, precision, recall, F-measure and
number of iterations.
Keywords—Data Mining; K-Means; Naïve Bayes
Classification; Clustering; preprocessing; WEKA tool
I. INTRODUCTION
Due to the increased availability of computer hardware
and software and the rapid computerization of business, large
amount of data has been collected and stored in databases.
Researchers have estimated that amount of information in the
world doubles for every 20 months.
However raw data cannot be used directly. Its real value
is predicted by extracting information useful for decision
support. In most areas, data analysis was traditionally a
manual process. When the size of data manipulation and
exploration goes beyond human capabilities, people look for
computing technologies to automate the process [1].
Data mining is one of the research activities in the field of
computing science and is defined as an extraction of
interesting (non-trivial, implicit, previously unknown and
potentially useful) patterns or knowledge from huge amount of
data. Data mining is applied to gain some useful information
out of bulk data. There are number of tools and techniques
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Shraddha Shukla
School of Computing Science and
Engineering, Galgotias University,
Greater Noida,
Uttar Pradesh, India
shraddha52d27@gmail.com
provided by researchers in data mining to obtain the pattern
out of data. Different patterns can be mined by classification,
clustering, association rules, regression, outlier analysis, etc.
[2]. There are abundant tools available for data mining. Some
of them are Rapid Miner, R, Knime, Own Code, Weka or
Pentaho, Statistica, Sas or Sas Enterprise Miner, Orange,
Tanagra, And Matlab.
Classification [9] is the process of finding a model (or
function) that describes and distinguishes data classes or
concepts, for the purpose of being able to use the model to
predict the class of objects whose class label is unknown. The
derived model is based on the analysis of a set of training data
(i.e., data objects whose class label is known).
Different types of classification algorithms [7,8] are:
 Decision tree induction
 Naïve Bayesian Classification
 Rule-Based Classification
 Classification by Back propagation
 Support Vector Machines
 Associative Classification
Clustering plays an important role in data mining by
categorizing objects into groups based on their similarity.
Cluster analysis or clustering is not an algorithm. It is a task
which includes several different algorithms. Clustering is very
essential in data mining.
Different types of clustering algorithms are:
 Connectivity based clustering (hierarchical
clustering).
 Centroid based clustering (K-means clustering).
 Distribution based clustering (expectation
maximization).
 Density based clustering [3].
 A good clustering algorithm always results in accurate
number of clusters with high intra class similarity and low
inters class similarity. Also which is scalable, interpretable,
usable and deals with different types of attributes. Algorithm
should be able to deal with noisy data and outliers [4]. If the
algorithms do not have these parameters then there will be no
optimal solution or clusters. There will be always a chance of
non-similar clusters. If a single cluster contains data points
having dissimilar properties then that cluster is not a final
cluster. That has to be re-clustered by the same process.
The basic approach of objects clustering is as shown in
the Figure 1. The approach is applied on random data points to
form three different groups of clusters. Each cluster consists of
similar type of datasets.
Figure 1 : Basic Clustering Approach
Approaches for Data Preprocessing
The following two different approaches are used for data
preprocessing.
Normalisation:
Normalisation is the process in which the data is reduced
to a form ensuring data integrity and eliminating data
redundancy.
 Data integrity - entire data in the database is consistent,
and satisfies all integrity constraints.
 Data redundancy – if data in the database can be found in
two different locations (direct redundancy) or if data can
be calculated from other data items (indirect redundancy)
then the data is said to contain redundancy.
Randomization:
Randomization is the process of making something
random means that generating a random alternative of a
sequence (such as shuffling cards). Selecting a random sample
of residents (important in statistical sampling). Allocating
experimental units via random assignment to a management or
control condition.
Approaches for Attribute Selection
A framework for attribute selection is shown in Figure 2.
For selecting features we used WEKA [6] data mining tool.
Initially we applied preprocessing approaches on the original
data to remove noise from data. A feature selection approach
by selecting appropriate feature evaluator and search
technique has been applied on the preprocessed data. This
approach results in reduced set of features.
We have used classifier subset evaluator as an attribute
evaluation and best first search as a search method. Classifier
subset evaluation takes naive bayes classifier as its base of
classification. For abalone dataset we have taken greedy
approach and filtered subset evaluation. For wine dataset,
attribute selection is done followed by remove filtration.
Figure 2: Framework for Attribute Selection
The produced results after the attribute selection are given
below. In the Table 1, we have listed the total initial features
and the elected features. After applying attribute selection we
have reduced the number of features on the basis of their
importance. In abalone dataset, initially the features were 9
including its class after applying feature selection reduced to
6. In iris dataset, initially the features were 5 including the
class attribute after applying feature selection reduced it to 4.
Similarly, for wine dataset, we have 15 attributes initially and
then it is reduced to 9.
 Table I: Initial and Selected Features for Different Datasets The results produced by classification using naïve bayes
classifier are as follows. These results are obtained in 3 steps.
Datasets Initial Features Selected No. of No.of First the naïve bayes is applied to the original dataset,
Features Initial Selected
secondly to the preprocessed dataset and finally to the selected
Features Features
Sex Sex attributes of the datasets. The results obtain for all three stages
Length - are compared in Table 2 for abalone dataset.
Diameter -
Height Height Table II: Abalone Classification
Abalone Whole wt. whole wt
Shucked wt. - 9 6
Viscera wt. viscera wt Misclassi F- Classifie
Shell wt. shell wt fied Rec Mea d
Abalone Time accuracy Precision all sure Accuracy
Rings Rings
Classification (sec) (%) (%) (%) (%) (%)
sepal length sepal length
sepal width - Original Data 0.2 47.8334 50 52 47 52.16
Iris petal length petal length
petal width petal width 5 4 After
class class Randomization 0.1 47.8334 50 52 47 52.16
Alcohol Alcohol
Malic acid - After Attribute
Ash Ash Selection 0.2 45.9899 52.3 54 74.1 54
Alcalinity of ash Alcalinity of
Magnesium ash
Total phenols Magnesium
Wine Flavanoids - From the Table 2 we observed the values obtained after
Nonflavanoid Flavanoids 15 9
Nonflavanoid
classification. The execution time for the original dataset was
phenols
Proanthocyanins phenols 0.2 sec which was reduced to half its value for the randomized
Color intensity - dataset. The no. of incorrectly classified attribute was
Hue - originally 47.8334% and it was reduced to 45.9899% after
OD280/OD315 Hue
-
attribute selection. Similarly, values of precision, recall, F-
of diluted wines
Proline - measure and accuracy was 50%, 52%, 47% and 52.16%
Proline respectively which was improved to 52.3%, 54%, 74.1 and
54% respectively. The Figure 3 shows the graphical
II. DATA CLASSIFICATION METHOD representation of the results produced. It can easily show the
Classification Approach using Naïve Bayes improvements in the results.

Naïve bayes classification assumes that each attribute is 80

independent and contributes equally to the model. The general Original Data
methodology for naïve bayes classifier is presented in Figure 60
3. We used WEKA data preprocessing tool for applying the
naïve bayes classifier on the datasets used. 40 After
Randomization
D : Set of tuples 20
After Attribute
 Each Tuple is an „n‟ dimensional attribute vector 0 Selection
 X : (x1,x2,x3,…. xn) (sec) (%) (%) (%) (%) (%)
Let there be „m‟ Classes : C1,C2,C3…Cm
Naïve Bayes classifier predicts X belongs to Class Ciiff
 P (Ci/X) > P(Cj/X) for 1<= j <= m , j <> i Figure 3: Abalone Classification Graph
Maximum Posteriori Hypothesis
 P(Ci/X) = P(X/Ci) P(Ci) / P(X) Table III: Iris Classification
 Maximize P(X/Ci) P(Ci) as P(X) is constant
With many attributes, it is computationally expensive to Iris Ti F- Classificati
Classification me Misclassifie Precisi Reca measur on
evaluate P(X/Ci). (se d accuracy on ll e accuracy
Naïve Assumption of “class conditional independence” c) (%) (%) (%) (%) (%)
P (X/Ci)=nπk=1 P (xk/ Ci) Original Data 0.1 4 96 96 96 96
P(X/Ci) = P(x1/Ci) * P(x2/Ci) *…* P(xn/ Ci) After
Normalization 0.1 2.6667 97.3 97.3 97.3 97.3
Figure 3: Algorithm of Naïve Bayes Classifier After
Attribute
Selection 0.1 3.33 96.7 96.7 96.7 96.7
 From the Table 3 we observe the values obtained after the The results obtained above shows significant
classification. The time was constant in all the results. The no. improvements in the values of various parameters used for
of incorrectly classified attribute was originally 4%, it was naïve bayes classification. The classification accuracy for
reduced to 2.6667% after normalization and 3.33% after abalone, iris and wine dataset is increased by 1.84%, 0.7% and
attribute selection. Similarly, values of precision, recall, F- 1.69% respectively for selected attributes. The precision,
measure and accuracy was 96% each which was improved to recall and F-measure values are increased for all three
97.3% each after normalization and 96.7% each after attribute datasets. The misclassified accuracy of all the three datasets is
selection. The Figure 4 shows the graphical representation of also decreased
the results produced. It can easily show the improvements in
the results. III. CLUSTERING METHOD
Clustering Approach Using K-Means
150
100 K-means clustering is most widely used clustering
Original Data algorithm which is used in many areas such as information
50 retrieval, computer vision and pattern recognition. K-means
0 After Attribute clustering assigns n data points into k clusters so that similar
(s) (%) (%) (%) (%) (%)
Selection data points can be grouped together. It is an iterative method
which assigns each point to the cluster whose centroid is the
Time
Misclassified accuracy
Precision Recall F-Measure
Classified accuracy
nearest. Then it again calculates the centroid of these groups
by taking its average. The algorithm 1 shows the basic
approach of K-means clustering [5].
Figure 4: Iris Classification Graph
We have used preprocessing tool WEKA which performs
Table IV: Wine Classification the K-means clustering in the following derivative manner. It
Wine Classi works on the given algorithm and evaluates the results.
Classificati F- fied
on Ti Misclassified Precisi Measur accur
me accuracy on Recall e acy
(s) (%) (%) (%) (%) (%) Algorithm: k-means. The k-means algorithm for
Original partitioning, where each cluster’scenter is represented by the
Data 0 3.3708 96.7 96.6 96.6 96.62 mean value of the objects in the cluster.
After
Attribute
Input:
Selection 0 1.6854 98.3 98.3 98.3 98.31 K: the number of clusters,
D: a data set containing n objects.
Output: A set of k clusters.
From the Table 4 we observe the values obtained after the Method:
classification. The time was constant in all the results. The no. (1) Arbitrarily choose k objects from D as the initial
of incorrectly classified attribute was originally 3.3708%, it cluster centers;
was reduced to 1.6854% after attribute selection. Similarly, (2) Repeat
values of precision was 96.7% now it is 98.3%, recall value (3) (Re) assign each object to the cluster to which the
was 96.6% now it is 98.3%, F-measure value was 96.6% now object is the most similar,based on the mean value of the
it is 98.3 and accuracy was 96.62% now it is 98.31%. The objects in the cluster;
Figure 5 shows the graphical representation of the results (4) Update the cluster means, i.e., calculate the mean
produced. It can easily show the improvements in the results. value of the objects foreach cluster;
(5) Until no change;
120 Figure 6: K-Means Clustering Algorithm
100
80
60 The K-means clustering algorithm discussed in Figure 6 is
40 original data
20 explained with the help of an example presented in Figure 7
0
after attribute
selection

Figure 5: Wine Classification Graph

 10 10

8
9

8
200
7

6
7

6
150
5

4
5

4 100 original data

3 3

1
2

1
50
0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
0 after
randomization
10 10

7
9

7
after attribute
6

5
6

5
selection
4 4

3 3

2 2

1 1

0 0
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10

Figure 8: Abalone Clustering Graph

Figure 7: Working Of K-Means Clustering Algorithm

Table VI: Iris Clustering

Figure 7 shows the graphical representation for working Iris Clustering No. Of SSE Time Incorrectly Accuracy
of K-means algorithm. In the first step there are two sets of Iterations (sec) Clustered (%)
objects. Then the centroids of both sets are determined. (%)
According to the centroid again the clusters are formed which Original Data 6 6.998 0.1 11.33 88.67
After 6 3.348 0.1 4 96
gave the different clusters of dataset. This process repeats until Normalization
the best clusters are achieved. After 6 4.31 0.1 9.33 90.67
Attribute
The results produced by K-means clustering are as Selection
follows. The results obtained after applying K-means
clustering to the three datasets are discussed is the following In the above Table 6 we can observe the values noted
table. after the clustering. The no. of iterations is constant
throughout the iris clustering experiment. The SSE was 6.998
Table V: Abalone Clustering initially then it is reduced to 3.348 after normalization and
4.31 after attribute selection. The time is constant. . The
Abalone No. Of SSE Time Incorrectly Accuracy incorrectly clustered attribute was 11.33% then it was reduced
Clustering Iterations (sec) Clustered (%)
(%)
to 4% after normalization and 9.33% after attribute selection.
Original Data 57 172.694 0.7 79.70 50.3 Hence the accuracy is improved by 88.67% to 96% after
After 26 172.694 0.33 49.70 50.3 normalization. The Figure 9 shows the graphical
Randomization representation of the results produced. It can easily show the
After Attribute 17 96.539 0.16 46.2054 53.79 improvements in the results.
Selection

120
In the above Table 5 we can observe the values noted 100
80 original data
after the clustering. The no. of iterations was 57 then it was 60
reduced to 26 after randomization and 17 after attribute 40
20
selection. The SSE was reduced from 172.694 to 96.539. The 0 after
time was 0.7 sec initially then it was improved to 0.33 and normalization
0.16 after randomization and attribute selection respectively.
The incorrectly clustered attribute was 79.70% then it was after attribute
reduced to 49.70% after randomization and 46.2054% after selection
attribute selection. Hence the accuracy is improved from
50.3% to 53.79%. The Figure 8 shows the graphical
representation of the results produced. It can easily show the
improvements in the results. Figure 9: Iris Clustering Graph
 Table VII: Wine Clustering [3] A. K. Jain, prof. S. Maheshwari, “Survey of recent
clustering techniques in data mining”, international journal
Wine No. Of SSE Time Incorrectly Accuracy of computer science and management research, vol 1 issue
Clustering Iterations (sec) Clustered (%) 1 Aug 2012
(%)
Original 8 48.970 0 5.618 94.382 [4] D. Sisodia, L. Singh, S. Sisodia, K. Saxena, “Clustering
Data
Techniques: A Brief Survey of Different Clustering
After 7 27.972 0 8.988 91.012
Attribute
Algorithms”, International Journal of Latest Trends in
Selection Engineering and Technology (IJLTET), Vol. 1 Issue 3
September 2012
For the wine clustering a separate preprocessing have not
[5] E. Kijsipongse, S. U-ruekolan, "Dynamic load balancing
been performed. In the above Table 7 we can observe the
on GPU clusters for large-scale K-Means clustering," 2012
values noted after the clustering. The number of iterations is IEEE International Joint Conference on Computer Science
reduced from 8 to 7. The SSE is improved from 48.970 to and Software Engineering (JCSSE), vol., no., pp.346, 350,
27.972. The time is constant. The Figure 10 shows the May 30 2012-June 1 2012.
graphical representation of the results produced. It can easily
show the improvements in the results. [6] http://www.cs.waikato.ac.nz/ml/weka/downloading.html

100 [7] Zhong, N.; Zhou, L.: “Methodologies for Knowledge

80 Discovery and Data Mining”, The Third Pacific-Asia
60 Conference, Pakdd-99, Beijing, China, April 26-28, 1999;
40 original data Proceedings, Springer Verlag.
20
0
[8] Raj Kumar, Dr. Rajesh Verma, “Classification Algorithms
after attribute for Data Mining: A Survey” International Journal of
selection Innovations in Engineering and Technology (IJIET).

[9] Serhat Özekes and A.Yilmaz Çamurcu,”Classification and

Prediction in a Data Mining Application “Journal of
Marmara for Pure and Applied Sciences, 18, pp. 159-174,
Figure 10: Wine Clustering 2002

IV. CONCLUSION
The results discussed here shows that the K Means
clustering algorithm applied on the selected attribute set
produces significant improvements in the values obtained for
the original dataset. The results obtained for abalone and iris
datasets shows an improvement in the accuracy by 3.49% and
2% respectively for the selected attribute set. For randomized
dataset the accuracy remains same for the abalone dataset
while it is increased by 8.67% for the iris dataset using
normalization. The number of iterations and execution time is
also decreased for all three datasets. The percentage of
incorrectly attributes and SSE is also decreased for all three
datasets.

REFERENCES
[1] E. A. Khadem, E. F. Nezhad, M. Sharifi, “Data Mining:
Methods & Utilities”, Researcher2013; 5(12):47-59.
(ISSN: 1553-9865).

[2] Y. S. Patil, M.B. Vaidya, “ A Technical Survey on cluster

analysis in data mining”, International Journal of
Emerging Technology and Advanced Engineering, ISSN
2250-2459, Volume 2, Issue 9, September 2012.
 Experimental Performance Comparison between
TCP vs UDP tunnel using OpenVPN
Irfaan Coonjah Pierre Clarel Catherine
Faculty of Engineering School of Innovative Technologies and Engineering
University of Mauritius University of technology, Mauritius
Réduit, Mauritius La Tour Koenig, Pointes aux Sables
irfaan.coonjah@umail.uom.ac.mu ccatherine@umail.utm.ac.mu
Abstract—The comparison between TCP and UDP tunnels
have not been sufficiently reported in the scientific literature.
In this work, we use OpenVPN as a platform to demonstrate the
performance between TCP/UDP. The de facto belief has been
that TCP tunnel provides a permanent tunnel and therefore
ensures a reliable transfer of data between two end points.
However the effects of transmitting TCP within a UDP tunnel
has been explored and could provide a valuable attempt. The
results provided in this paper demonstrates that indeed TCP in
UDP tunnel provides better latency. Throughout this paper, a
series of tests have been performed, UDP traffic was sent inside
UDP tunnel and TCP tunnel successively. The same tests was
performed using TCP traffic.
Keywords—TCP, UDP, OpenSSH, VPN, Tunneling.
I. I NTRODUCTION
An IP tunnel [1] is defined an Internet Protocol (IP) commu-
nications medium between two networks. It encapsulates its
own network protocol within the TCP/IP packets carried by
the Internet. IP tunnels connect separate IP networks which
are not directly connected to each other. Tunneling protocol
allows access to a network service which is not supported. The
nature of the traffic sent through the tunnel can be hidden by
encryption standard which repackage traffic data into different
form. It makes use of a layered protocol model such as those
of the Open Systems Interconnection model (OSI) or TCP/IP
protocol suite. Tunneling is used in all VPNs; one common
open source application layer solution available is OpenVPN
(Open Virtual Private Network). The popularity of VPNs has
increased due to its low cost and the security it provides. The
trade-offs between TCP and UDP regardless of VPN usage is
always said to be the same: Speed is sacrifice for reliability
as UDP is connectionless and the server sending the data
theoretically does not ensure if it reaches the destination or
not. UDP is claimed to be faster but TCP is meant to be
more reliable. This paper focuses on such comparison and
evaluates the efficiency between UDP tunnel and TCP tunnel
using OpenVPN. Throughout this paper, a series of tests have
been performed, UDP traffic was sent inside UDP tunnel and
TCP tunnel successively. The same tests was performed using
TCP traffic.
This document is organized into the following sections:
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
K. M. S. Soyjaudah
Faculty of Engineering
University of Mauritius
Réduit, Mauritius
ssoyjaudah@uom.ac.mu
• Section II covers Virtual Private Network, with different
layers in the OSI model.
• Section III covers User Datagram Protocol (UDP) and
Transmission Control Protocol (TCP).
• Section IV makes a comparison between TCP and UDP.
• Section V details the experimental testing.
• Section VI describes the Physical framework.
• Section VII details the performance measures.
• We then conclude by giving the results and observations
in section VIII.
II. V IRTUAL P RIVATE N ETWORK (VPN)
A virtual private network (VPN) [2] makes use of a public
network to connect multiple remote locations. A VPN spread
a private network by making use of a public network (Internet)
to set up a point-to-point connection and virtual tunneling
protocols. Using a VPN, a computer can communicate across
public networks as if it is connected in the same LAN of the
private network. A VPN is a logical network on top of an
already existing network. Different VPN solutions work on
different layers in the Open System Interconnect (OSI) model
[3]. In the tunnels, the traffic is encrypted and sent through
using the lower layers in the OSI model. The VPN traffic
is split up from any other network traffic by using encrypted
tunnels between the VPN hosts. Inside a tunnel, the forwarded
traffic is encapsulated into a special packet format on which
a block cipher is used to encrypt the traffic [4]. As mentioned
in the previous paragraph, a VPN can work on different
layers of the OSI model. Three common types of VPNs are
Application Layer VPNs, Network Layer VPNs and Datalink
Layer VPNs. Secure Shell (SSH), Secure Sockets Layer (SSL),
and OpenVPN are VPNs that work on the Application layer
of the OSI model [5]. The tunneled traffic is encapsulated into
application specific headers before being sent to the other side
using the available Transport Layer Protocol [6], such as User
Datagram Protocol (UDP) or Transmission Control Protocol
(TCP).
III. T RANSMISSION C ONTROL P ROTOCOL (TCP) AND
U SER DATAGRAM P ROTOCOL (UDP)
TCP is the main protocol in TCP/IP networks. The IP
protocol process data packets while TCP allow two hosts
to exchange data streams and establish a connection. TCP
guarantees that packets will arrive their destination in the same
order in which they were sent [7].
UDP provides unreliable, minimum, best-effort, message
delivery to upper-layer protocols and applications. UDP do
not setup a permanent connection between two end points [8].
IV. C OMPARISON BETWEEN TCP AND UDP
The adjustments between TCP and UDP regardless of VPN
usage is always said to be the same: Speed is sacrifice for
reliability as UDP is connectionless and the server sending
the data theoretically does not ensure if it reaches the des-
tination or not. TCP is a connection-oriented protocol [9],
which implies that end-to-end communications is set up using
handshaking. Once the connection is established, data can be
transferred bi-directionally over the link. UDP is a connec-
tionless protocol and therefore less complex message based
[10] when compared to TCP, which includes that the point-to-
point connection is not dedicated and data is transferred uni-
directional from the source to its destination without checking
whether the receiver is active. TCP regulate retransmission,
message acknowledgment, and timeout [11]. TCP deliver lost
messages along the way upon multiple attempts. In TCP, there
is no missing data, and if ever there are multiple timeouts, the
connection is dropped. When a UDP message is sent there
is no guarantee that the message will reach its destination;
it could get drop along the way. There is no retransmission,
timeout and acknowledgment. When two data packets are
sent in sequence, the first message will reach the destination
first. When data segments arrive in the wrong order, TCP
buffers hold the data until all data are re-ordered before being
transmitted; when using UDP the order in which messages
arrive cannot be predicted.
The TCP protocol has extensive algorithms to ensure correct
delivery of the data. Having two TCP connections stacked
together will thus force the algorithms of both TCP con-
nections to work in parallel [12]. TCP was not designed to
work this way and problems are likely to occur in different
situations. The retransmission problems, TCP meltdown and
double retransmit, are problems caused by tunneling TCP in
TCP. The problems can occur when both of the stacked con-
nections are retransmitting packets. In previous work, related
to TCP in TCP tunneling, it is not entirely clear, how severe the
retransmission problems really are. TCP protocol suite allows
automatic recovery of any dropped or lost data even if a host
goes down in the network. When TCP packets are transmitted
from one end to a remote end across the network, the data
packets are reordered in the same sequence created by the
sender. The protocol notifies when segments of the data stream
have been corrupted, reordered, discarded or duplicated by
the network [11]. TCP is a reliable protocol as the sender can
retransmit damaged segments. However retransmission creates
latency.
TCP was designed to make an efficient protocol with little
overhead, a protocol set having a basic amount of ’extra’
data being transferred. This extra data is named as overhead,
and they package the data that needs to be transferred. TCP
tunnel is a technology that combines and transmits packets
sent between end hosts as a single TCP connection. When
using a TCP tunnel, the fairness among aggregated flows
can be improved and several protocols can be transparently
transmitted through a firewall.
V. E XPERIMENTAL TESTING
The experimental testing procedure consists of four parts
in the evaluation of latency. Two tunnels will be created,
TCP tunnel and UDP tunnel using OpenVPN as base. In the
TCP tunnel, TCP traffic will be sent and the latency will be
recorded in seconds. Then UDP traffic will be sent in the TCP
tunnel and the same value will be recorded. The same test
will be performed but using UDP tunnel. All measurements
will take place at the server node because of its processing
power. The tests will be performed through a Local Area
Network and Wide Area Network environment successively.
The compression for OpenVPN will be disabled since it is
not part of the tests. During the measurements, the size of
packets will be increased and two different graphs will be
plotted with axes; latency of the VPN tunnel vs size of packets.
Iperf will be used for monitoring latency. Iperf, is the network
testing tool that will be used to measure the parameters during
the experiments. Since iperf can generate TCP and UDP data
streams, this make iperf suitable for this test. Iperf is also
flexible since the user can set different parameters that can be
used in the test or alternatively optimize or tuning the network.
Iperf has a client and server functionality, which makes it
possible to measure the throughput between the two ends, bi-
directionally and uni-directionally.
VI. E XPERIMENTAL FRAMEWORK DESCRIPTION
In this paper, the authors took into consideration the physi-
cal topology layout, setting up a complete topology by using
two computers, switch and router. The setup is identical to a
real live scenario and is expected to give more accurate values
for the performance evaluation and analysis between TCP and
UDP tunnel based on OpenVPN. The open source VPN soft-
ware, OpenVPN has been the choice for many companies and
according to technical discussions, VPN technologies using
UDP tunnel as base are predicted to provide fast speeds when
compared to using OpenVPN using TCP as base. OpenVPN
is said to be slow than other protocols when used via TCP.
When compared to stability, TCP is prognosticated to be more
stable protocol on all kinds of connection and UDP is deemed
to be less stable, which makes it susceptible to disconnection
with intermittent network issues. Compression is a built-in
feature in OpenVPN and it has not been tested yet whether
compression has an impact on transfer time due to processing
power in CPU. In this experiment therefore, compression has
been disabled.
The server is connected to the switch via an Ethernet
cable and physical ip address of 10.1.24.19/24. The client is
connected to another port on the switch and has IP address
10.1.100.21/24. The switch is connected to the router, which
has the gateway IP address for the network. The server and
the client are in two different subnets, therefore they cannot
communicate making use of their physical IP address. A VPN
Fig. 1. Latency comparison between UDP and TCP inside a TCP tunnel
will be setup between the client and the server, and two
computers will have a logical IP address on the same subnet
and can then communicate with each other. The setup will
be used for measuring latency in the LAN, but for measuring
traffic in the WAN; the router will be connected to the internet
and have port-forwarding configured, and the client will be
on a remote network with internet access. The client will
therefore access the private network through port-forwarding
and OpenVPN.

VII. PERFORMANCE MEASURES

The performance measurement section details a comparative
study of the conducted results. The authors attempt to give a
better explanation of the experiment results. One important
characteristics of the LAN that may disturb the performance
of the connection must be discussed. The switch is configured
on a 10Mbps duplex setting. The maximum file size has
therefore been limited to 10Mbps for this test. All network
parameters for the tunnels are concentrated. Standard
deviation values rendered in both graphs confirm this remark. Fig. 2. Latency comparison between UDP and TCP inside a UDP tunnel
Tunneling performance is measured using a set of criteria or
metrics. The latency is used to characterize the performance
of both tunnels. Latency is defined as the time taken it takes increased transfer speed compared to TCP tunnel when the
packet size increases. The latency gap in UDP tunnel becomes
for a data packet to get from one destination point to another.
Latency through a tunnel is dependent on the machine wider when the packet size increases from 4Mb to 10Mb. This
hardware, the link speed, and the encapsulation time. can be explained with the fact that when the VPN uses TCP
tunnel, TCP connections will use IP packets sent through the
The snapshot below shows how the tests are performed on VPN, therefore creating the TCP overhead twice. The UDP
iperf and how the values have being recorded. tunnel base VPN thus has the potential for slightly better
C:\iperf-2.0.5-2-win32>iperf.exe -s performance. TCP provides a bidirectional tunnel for data,
------------------------------------------------------------
Server listening on TCP port 5001 but relies on packets, so there will be some ”administrative”
TCP window size: 64.0 KByte (default) packets, e.g. acknowledges: this is the TCP overhead. For
------------------------------------------------------------
[ 4] local 172.16.1.1 port 5001 connected with 172.16.1.6 portinstance,
65410 if the server A sends 10 MB to the client B, client
[ ID] Interval Transfer Bandwidth B will also send some packets to A confirming reception.
[ 4] 0.0-10.1 sec 21.6 MBytes 17.9 Mbits/sec
When doing VPN over TCP, the VPN has its own TCP-based
Figure 1 demonstrates the latency comparison between UDP overhead, and transports the administrative packets for any
and TCP inside a TCP tunnel in a LAN environment, measured connection within the VPN. The graph therefore confirm the
in seconds and packet size in megabyte. The UDP tunnel gives TCP stacking problem when transmitting TCP traffic over TCP

Fig. 3. Latency comparison between UDP and TCP inside a TCP tunnel in
WAN environment
Fig. 4. Latency comparison between UDP and TCP inside a UDP tunnel in
WAN environment
tunnel. The TCP stacking starts when the message size reaches
2 MB, since the difference in latency before 2MB is very small
when compared to the latency gap after 2MB. On the other
hand, figure 2 shows the latency comparison between UDP
and TCP when using a UDP tunnel in a LAN environment.
When using UDP tunnel, we do not have the double stacking
problem. Compared to figure 1, the latency is smaller for
both TCP and UDP traffic. It was said that encryption, slow
down UDP traffic since when bits of data is missing the
entire message may need to be re-sent again, causing latency.
During the tests, it is shown that, UDP message is not affected
by encryption mechanism when the size is between 1MB to
10MB. Sending TCP traffic inside a TCP tunnel does not
create the double TCP stacking problem, this is proved in
figure 2. When we compare the two graphs for example, the
time taken to transmit a TCP message size of 2MB is twice
when using TCP tunnel; the latency in TCP tunnel is around
10 seconds whereas in UDP tunnel, the latency is around 5
seconds. Figure 3 is a replication of the test perform in figure
1, that is, latency comparison between UDP and TCP inside a
TCP tunnel in a WAN environment. The latency gap increases
proportionally when the message size increases. The pattern is
similar to figure 1, except that the latency is higher. The test
in figure 4 has been performed in a WAN environment, that
is latency comparison between UDP and TCP inside a UDP
tunnel in a WAN environment. In a WAN environment, the
latency graphs for both TCP and UDP are two straight lines
and the difference in latency is constant.
VIII. C ONCLUSION
This paper addresses performance comparison between TCP
and UDP tunnel connections. Two distinct scenarios were
used to test the two VPN tunneling mechanisms. The results
conclude that UDP tunnel utilizes the link more efficiently
and provide a radically improved transfer times and speed
compared with TCP tunnel. The results also demonstrates that
indeed TCP in UDP tunnel provides better latency.
It would be good to investigate the performance of the tunnels
on Mobile network. OpenVPN can be installed on mobile
devices, therefore another direction is to test the VPN on
mobile device, and measure the amount of energy which is
required to secure connections, since mobile devices have
batteries with limited capacity.
R EFERENCES
[1] S. Zhou and J. Luo, “A novel ip over udp tunneling based firewall traver-
sal for peer-to-peer networks,” in Service Operations and Logistics, and
Informatics (SOLI), 2013 IEEE International Conference on, pp. 382–
386, July 2013.
[2] R. Bush and T. Griffin, “Integrity for virtual private routed networks,” in
INFOCOM 2003. Twenty-Second Annual Joint Conference of the IEEE
Computer and Communications. IEEE Societies, vol. 2, pp. 1467–1476
vol.2, March 2003.
[3] Y. Li, W. Cui, D. Li, and R. Zhang, “Research based on osi model,”
in Communication Software and Networks (ICCSN), 2011 IEEE 3rd
International Conference on, pp. 554–557, May 2011.
[4] A. Mayer, B. Collini-Nocker, F. Vieira, J. Lei, and M. Castro, “Analytical
and experimental ip encapsulation efficiency comparison of gse, mpe,
and ule over dvb-s2,” in Satellite and Space Communications, 2007.
IWSSC ’07. International Workshop on, pp. 114–118, Sept 2007.
[5] M. Mimura and H. Tanaka, “Behavior shaver: An application based
layer 3 vpn that conceals traffic patterns using sctp,” in Broadband,
Wireless Computing, Communication and Applications (BWCCA), 2010
International Conference on, pp. 666–671, Nov 2010.
[6] D. Sarkar and H. Narayan, “Transport layer protocols for cognitive net-
works,” in INFOCOM IEEE Conference on Computer Communications
Workshops , 2010, pp. 1–6, March 2010.
[7] I.-S. Yoon, S.-H. Chung, and J.-S. Kim, “Implementation of lightweight
tcp/ip for small, wireless embedded systems,” in Advanced Information
Networking and Applications, 2009. AINA ’09. International Conference
on, pp. 965–970, May 2009.
[8] T. Le, G. Kuthethoor, C. Hansupichon, P. Sesha, J. Strohm, G. Hadynski,
D. Kiwior, and D. Parker, “Reliable user datagram protocol for airborne
network,” in Military Communications Conference, 2009. MILCOM
2009. IEEE, pp. 1–6, Oct 2009.
[9] Y.-L. Chang and C.-C. Hsu, “Connection-oriented routing in ad hoc
networks based on dynamic group infrastructure,” in Computers and
Communications, 2000. Proceedings. ISCC 2000. Fifth IEEE Symposium
on, pp. 587–592, 2000.
[10] E. Bethel and J. Shalf, “Grid-distributed visualizations using connec-
tionless protocols,” Computer Graphics and Applications, IEEE, vol. 23,
pp. 51–59, Mar 2003.
[11] M.-Y. Park and S.-H. Chung, “Distinguishing the cause of tcp retrans-
mission timeouts in multi-hop wireless networks,” in High Performance
Computing and Communications (HPCC), 2010 12th IEEE International
Conference on, pp. 329–336, Sept 2010.
[12] D. Lu, Y. Qiao, P. Dinda, and F. Bustamante, “Modeling and taming
parallel tcp on the wide area network,” in Parallel and Distributed
Processing Symposium, 2005. Proceedings. 19th IEEE International,
pp. 68b–68b, April 2005.
 6to4 Tunneling Framework using OpenSSH
Irfaan Coonjah Pierre Clarel Catherine
Faculty of Engineering School of Innovative Technologies and Engineering
University of Mauritius University of technology, Mauritius
Réduit, Mauritius La Tour Koenig, Pointes aux Sables
irfaan.coonjah@umail.uom.ac.mu ccatherine@umail.utm.ac.mu
Abstract—6to4 tunneling enables IPv6 hosts and routers to
connect with other IPv6 hosts and routers over the existing
IPv4Internet. The main purpose of IPv6 tunneling is to maintain
compatibility with large existing base of IPv4 hosts and routers.
OpenSSH VPN tunneling is said to have limitations with numer-
ous IPv6 clients and therefore it is advisable to use OpenVPN.
To the best knowledge of the authors, this is the first reported
successful implementation of 6to4 tunneling over OpenSSH with
more than one client. This proof-of-concept positions OpenSSH
therefore as a potential alternative to conventional VPNs.
Keywords—Tunneling, OpenSSH, VPN, IPV6.
I. I NTRODUCTION
In 1995, Tatu Ylonen, designed the Secure Shell (SSH)
protocol to protect data by encrypting traffic before sending
it over an unsecured network such as the internet [1]. Since
the SSH Software was open source and freely available to
the public [2], it gained its popularity quickly. In December,
1995, the founder of SSH made the software proprietary,
hence, it was no longer accessible to the public and developers
[3]. However, in 1999, the open source developers founded
a research community and created a free version of Open
Secure Shell (OpenSSH), which is a derived version of the
SSH [4]. More than 2 million users in 2000, use the free
version of OpenSSH and by 2013, the number of users reached
10 million. The Open SSH developers are convinced that their
application is more secure than the original SSH protocol as
they respect their policy that is to produce clean and audited
code. So, a family of permissive free software licenses are
released under the BSD license [5]. The OpenSSH is currently
one of the most used security implementation, implemented in
a large number of operating systems and has been a reliable
solution for the internet security vendor’s equipments.
OpenVPN, developed by another research group is another
well-known Open source VPN. OpenSSH is limited to UNIX
like operating systems whereas OpenVPN can be deployed
on almost all platforms. Unlike OpenSSH, OpenVPN does
not form part of a protocol family and has been limited to
VPN tunnelling. The OpenSSH group research in SSH VPN
has been very slow and the recent versions have fixed patches
of security flaws, therefore the 6to4 tunneling feature will
undoubtedly be of great help to OpenSSH users[6].
This document is organized into the following sections:
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
K. M. S. Soyjaudah
Faculty of Engineering
University of Mauritius
Réduit, Mauritius
ssoyjaudah@uom.ac.mu
• Section II covers Virtual Private Network.
• Section III describes OpenSSH.
• Section IV covers IPv6 and IPV4 Private and Public
addressing.
• Section V describes the requirements an prerequisites for
buiding 6to4 tunneling on OpenSSH.
• Section VI describes the scenario and section VII details
the steps to create 6to4 tunneling.
• Section VIII performs a series of tests.
• We then conclude by giving the results and observations
in section IX.
II. V IRTUAL P RIVATE N ETWORK (VPN) AND IP TUNNEL
An IP tunnel [7] is an Internet Protocol (IP) communi-
cations connection between two networks used to transport
another network protocol. It does this by encapsulating its
own network protocol within the TCP/IP packets carried by
the Internet. IP tunnels are used to connect two separate IP
networks that are not directly connected to each other. A user
can access any network service not supported in its local area
network using tunneling protocol; one example is to run IPv6
over IPv4. A second use is to provide unworkable services
for example, giving a different subnet IP address to a remote
user whose physical network address is not part of the LAN.
Tunneling protect the type of traffic that is run through the
tunnel [8], using encryption standard to repackage traffic data
into a different form. A virtual private network (VPN) [9]
makes use of a public network to connect multiple remote
locations. A VPN extend a LAN using a public network, such
as the Internet by establishing a point-to-point connection and
virtual tunneling protocols. It allows a computer to commu-
nicate across public networks as if it is connected directly to
the same LAN. A VPN is a logical network on top of an
already existing network. Different VPN solutions work on
different layers in the Open System Interconnect (OSI) model
[10]. In the tunnels, the traffic is encrypted and sent through
using the lower layers in the OSI model. IPsec VPN operates
on the Network layer of the OSI model [11]. The tunneling
is transparent to the applications working on the higher 6
layers. Support for Network Layer tunneling is implemented
by the operating system of the host that is an end point of the
tunnel. Data-link layer VPNs [12] are working on the Data-
link Layer of the OSI model. They are most commonly used
on top of PPP in order to secure modem based connections,
although PPP actually encrypts the traffic. The most popular
encapsulation protocols used for Data Link layer VPNs are
PPTP, L2TP and L2F. In this paper, the authors will make use
of an application layer VPN which is OpenVPN.
III. O PEN S ECURE S HELL (O PEN SSH) VPN FEATURE
The open source software OpenSSH is created by the
OpenBSD group under BSD license cite[13].It uses the SSH
protocol to encrypt and secure communication over a computer
network. The latest version of OpenSSH v6.6, released on
March 15, 2014 has additional features and patches [14].
SSH Transport Layer Protocol uses distinctive encryption al-
gorithms [15], and the session keys are obtained from a shared
secret and an exchange hash. OpenSSH VPN is not famous
since it is supported on only OpenBSD, NetBSD, FreeBSD,
AIX, HP-UX, IRIX, Linux, NeXT, SCO, SNI/Reliant Unix,
Solaris, Digital Unix/Tru64/OSF, Mac OS X, and Cygwin [13].
IV. IP V 6 AND IPV4 P RIVATE AND P UBLIC ADDRESSING
Internet Protocol version 6 (IPv6) is the most recent version
of the Internet Protocol (IP) intended to replace IPv4. The
Internet Engineering Task Force (IETF) developed the IPv6 to
cater for the long-anticipated problem of IPv4 address exhaus-
tion [16]. With the growth of the Internet, more IPv4 address
space is needed to connect new devices in the future. IPv6 uses
a 128-bit address more than 7.9 times than IPv4, which makes
use of 32-bit addresses and has approximately 4.3 billion
addresses. The two protocols are not created to work together
therefore complicating the transition from IPv4 to IPv6. For-
tunately, many IPv6 transition mechanisms have been devised
to allow IPv4 to communicate with IPv6. IPv6 addresses are
represented as 8 groups of 4 hexadecimal digits separated by
colons, example 2001:0db5:0000:0042:0000:7a2e:0380:7334
[17]. creates a new packet format to limit packet header
processing by routers. Since both IPv4 IPv6 packets head-
ers are not the same, the two protocols are therefore not
interoperable.[18].
V. R EQUIREMENTS AND PREREQUISITES
Root access to the client and the server is needed. The
client and the server are linked to the internet and both have
public IP addresses. The aim is to use OpenSSH to connect
to the server from a remote computer and set up a virtual
tunnel with ipv6 ip addresses to allow the client computer to
connect virtually in the server network.
1) SSH version 6.0 or higher on both ends of the VPN:
2) Standard Linux or Ubuntu:
3) Two computers:
VI. S CENARIO
The SSH server has ip address 10.1.2.24/24 and is
connected to the internet with public ip address 202.123.2.13.
The client computer has private IP address 192.168.1.2 and
will connect to the server, using a simple ADSL line. Both
server and client is running Ubuntu operating system 13.04.
Fig. 1. Physical topology
VII. C ONFIGURATION OF O PEN SSH VPN
A. Server Configuration
1) Setup root access on both server and client [19]:
# sudo passwd
Enter new UNIX password:
Confirm new UNIX password:
2) Give root access by using the ”PermitRootLogin yes” in
etc /ssh/sshd config [20]:
PermitRootLogin yes
The commands do not permit root to log in without a
password, but it forces root to use the ssh key.
3) Using any editor, open /etc/ssh/sshd config add the
”PermitTunnel” line:
PermitTunnel yes
4) Allow NAT. These commands will enable NAT without
the need to reboot[21]:
# sudo sysctl -w net.ipv4.ip_forward=1
To set as default, using any editor, open /etc/sysctl.conf and
add:
# Needed to add for forwarding
net.ipv4.ip_forward = 1
5) Configure iptables to allow masquerade(NAT) [22]:
# sudo iptables -t nat -A POSTROUTING -s 10.0.0.0/24
-o eth0 -j MASQUERADE
6) The iptables settings will be lost when reboot unless it
is configured a way of saving the settings. Open /etc/rc.local
and add this line (above the exit 0 line) [23]:
# iptables-restore < /etc/iptables.rules
7) Configure the tunnel (tun0):
# sudo ssh -w 0:0 202.123.2.11
This command creates a tunnel interface named tun0 on both
machine, server and client.
If there is no errors, then the tun0 interface will be seen on
both systems, but not configured:
# ip addr show tun0
tun0: <POINTOPOINT,MULTICAST,NOARP,UP,LOWER_UP>
mtu 1500 qdisc pfifo_fast state UNKNOWN qlen 500 link/none
Sometime the MTU of tun0 interface need to be changed gateway fe80::2
to1500:
ifconfig tun0 mtu 1500 up
C. Bring up the VPN connection
B. Client Configuration ifup and ifdown commands are automated process for
1) Make a ssh key, call it VPN [24]: bringing up the tunnels.
# ssh-keygen -f VPN -b 1024 Bringing the connection up:
When generating the key, enter a password when asked to sudo ifup tun0
increase security. Shutting down the connection down:
2) On the client, put the private key (VPN) in /root/.ssh and sudo ifdown tun0
set permissions [25]: If needed, make the directory /root/.ssh Configuring the interfaces manually by giving them each
first. an IP address (do this as root):
# sudo mkdir /root/.ssh
# sudo cp VPN /root/.ssh/VPN On the server:
# sudo chown root:root /root/.ssh/VPN
ip link set tun0 up
# sudo chmod 400 /root/.ssh/VPN
ip -6 addr add fe80::2/64 dev tun0
3) Configure the key on the server: After transferring
the public key (VPN.pub) to the server, first put it in On the client:
/root/.ssh/authorized keys. If needed, make the directory
/root/.ssh first [24]. ip link set tun0 up
ip -6 addr add fe80::1/64 dev tun0
# sudo mkdir /root/.ssh !
# sudo bash -c "cat VPN.pub >> /root/.ssh/authorized_keys" route add -net 10.1.2.0 netmask 255.255.255.0 dev tun0

Now, edit /root/.ssh/authorizedkeys adding a forced com-

mand (just before the ssh-rsa AAA ...)
tunnel="0",command="/sbin/ifdown tun0; /sbin/ifup tun0",
no-portforwarding,no-ptty,no-X11-forwarding,
no-agent-forwarding ssh-rsa AAAA ....
(long sequence of random numbers / letters )
Note: The options ”no-port-forwarding, no-ptty, no-X11-
forwarding, no-agentforwarding” further enhances security.
Keep the forced command and the rest of the key all on a
single line.
4) Configure tun0 on the client:
# sudo vi /etc/network/interfaces
iface tun0 inet static
pre-up ssh -f -w 0:0 202.123.2.11
’ifdown tun0; ifup tun0’
pre-up ssh -i /root/.ssh/VPN -S
/var/run/ ssh-vpn-tunnel-control
pre-up sleep 5
address fe80::1
netmask 64
pointopoint fe80::2
up ip route add 10.0.0.0/24 via fe80::1
up ip route add 202.123.2.11/32 via 192.168.1.1
down ip route del 10.0.0.0/24 via fe80::1
down ip route del 202.123.2.11/32 via 192.168.1.1
post-down ssh -i /root/.ssh/VPN -S
/var/run/ssh-vpn-tunnel-control -O exit 202.123.2.11
address fe80::1 # sets the tunnel ipaddress
pointtopoint fe80::2 # establishes the point-to-point connection
(VPN). netmask = 64 #netmask for tunnel
pre-up commands are run before the interface (tun0) is enabled
post-down ssh ...#brings down the tunnel
5) Configure tun0 on the server:
# sudo vi /etc/network/interfaces
iface eth0 inet6 static
pre-up modprobe ipv6
address fe80::1
netmask 64
The command ’netstat -rn’ can be used to view the routing
table.
VIII. T ESTING
After configuring each interface, the VPN status will be up
and ping from the Client to the Server will be successful.
On the Server:
Ping test the remote tunnel address on the client.
SERVERPC:˜# ping6 -I tun0 fe80::1
PING fe80::1(fe80::1) from fe80::2 tun0: 56 data bytes
64 bytes from fe80::1: icmp_seq=1 ttl=64 time=12.4 ms
64 bytes from fe80::1: icmp_seq=2 ttl=64 time=98.2 ms
64 bytes from fe80::1: icmp_seq=3 ttl=64 time=12.6 ms
64 bytes from fe80::1: icmp_seq=4 ttl=64 time=15.1 ms
64 bytes from fe80::1: icmp_seq=5 ttl=64 time=103 ms
ˆC
--- fe80::1 ping statistics ---
5 packets transmitted, 5 received, 0% packet loss, time 4006ms
rtt min/avg/max/mdev = 12.473/48.375/103.298/41.853 ms
On the Client:
Ping test the remote tunnel address on the Server.
CLIENTPC:˜# ping6 -I tun0 fe80::2
PING fe80::2(fe80::2) from fe80::1 tun0: 56 data bytes
64 bytes from fe80::2: icmp_seq=1 ttl=64 time=10.2 ms
64 bytes from fe80::2: icmp_seq=2 ttl=64 time=14.2 ms
64 bytes from fe80::2: icmp_seq=3 ttl=64 time=12.6 ms
64 bytes from fe80::2: icmp_seq=4 ttl=64 time=17.1 ms
64 bytes from fe80::2: icmp_seq=5 ttl=64 time=106 ms
ˆC
--- fe80::1 ping statistics ---
5 packets transmitted, 5 received, 0% packet loss, time 4006ms
rtt min/avg/max/mdev = 74.029/74.424/75.208/0.554 ms
Ping test the file server ip address on the Server’s private LAN
10.1.2.0/24.
# ping 10.1.2.44
PING 10.1.2.44 (10.1.2.44) 56(84) bytes of data.
64 bytes from 10.1.2.44: icmp_seq=1 ttl=64 time=75.2 ms
64 bytes from 10.1.2.44: icmp_seq=2 ttl=64 time=74.0 ms
64 bytes from 10.1.2.44: icmp_seq=3 ttl=64 time=74.0 ms
--- 10.1.2.44 ping statistics ---
 3 packets transmitted, 3 received, 0% packet loss, time 2002ms[17] J. Govil, J. Govil, N. Kaur, and H. Kaur, “An examination of ipv4
rtt min/avg/max/mdev = 74.029/74.424/75.208/0.554 ms
IX. C ONCLUSION
VPN is an advanced feature in OpenSSH technologies
which unfortunately people do not use. IPv6 tunneling main-
tain compatibility with large existing base of IPv4 hosts and
routers. It is tedious to gather articles or procedures how
to build and setup a VPN based on OpenSSH. It takes
time to gain sufficient knowledge and experience in Linux,
OpenSSH, and networking. This article is a complementary of
all knowledge required and contains all information necessary
to build a VPN system on IPv6 network. This technology has
therefore many cost benefits which can be extended to other
applications to be used in other situations. It will be good to
test the LibreSSL under the VPN feature of OpenSSH.
and ipv6 networks : Constraints and various transition mechanisms,” in
Southeastcon, 2008. IEEE, pp. 178–185, April 2008.
[18] J. Pickard, A. Patrick, and A. Robinson, “Analysis of enterprise ipv6
readiness,” in SoutheastCon 2015, pp. 1–7, April 2015.
[19] U. documentation, “Rootsudo.” https://help.ubuntu.com/community/
RootSudo, 2013.
[20] M. Stahnke, Pro OpenSSH. USA: Apress, 2006.
[21] Admin, “Enable ip forwarding on ubuntu 13.04.” http://www.
networkinghowtos.com/howto/enable-ip-forwarding-on-ubuntu-13-04,
2013.
[22] O. Andreasson, “Iptables tutorial 1.2.2.” http//iptables-tutorial.frozentux.
net iptables-tutorial.html.pdf, 2008.
[23] O. Titz, “Why tcp over tcp is a bad idea.” http://sites.inka.de/bigred/
devel/tcp-tcp.html, 2001.
[24] P. G. Sery, Ubuntu Linux For Dummies. 2006.
[25] B. Zazen, “Vpn over ssh.” http://bodhizazen.net/Tutorials/
VPN-Over-SSH, 2009.

R EFERENCES
[1] R. Seggelmann, M. Tuxen, and E. Rathgeb, “Ssh over sctp x2014;
optimizing a multi-channel protocol by adapting it to sctp,” in Commu-
nication Systems, Networks Digital Signal Processing (CSNDSP), 2012
8th International Symposium on, pp. 1–6, July 2012.
[2] J. Sarkinen, “An open source(d) controller,” in Telecommunications
Energy Conference, 2007. INTELEC 2007. 29th International, pp. 761–
768, Sept 2007.
[3] A. .B, D. .F, and N. .B, “Linux operations and administration,” vol. 1,
pp. 399–400, August 2012.
[4] A. Jacoutot, “About openssh.” http://www.openssh.com/, May 2014.
[5] A. Jacoutot, “Openssh.” http://www.openssh.com/portable.html, May
2014.
[6] D. Meng, “Implementation of a host-to-host vpn based on udp tunnel and
openvpn tap interface in java and its performance analysis,” in Computer
Science Education (ICCSE), 2013 8th International Conference on,
pp. 940–943, April 2013.
[7] S. Zhou and J. Luo, “A novel ip over udp tunneling based firewall traver-
sal for peer-to-peer networks,” in Service Operations and Logistics, and
Informatics (SOLI), 2013 IEEE International Conference on, pp. 382–
386, July 2013.
[8] P. Rawat, J. Bonnin, and L. Toutain, “Designing a tunneling header
compression (tucp) for tunneling over ip,” in Wireless Communication
Systems. 2008. ISWCS ’08. IEEE International Symposium on, pp. 273–
278, Oct 2008.
[9] R. Bush and T. Griffin, “Integrity for virtual private routed networks,” in
INFOCOM 2003. Twenty-Second Annual Joint Conference of the IEEE
Computer and Communications. IEEE Societies, vol. 2, pp. 1467–1476
vol.2, March 2003.
[10] Y. Li, W. Cui, D. Li, and R. Zhang, “Research based on osi model,”
in Communication Software and Networks (ICCSN), 2011 IEEE 3rd
International Conference on, pp. 554–557, May 2011.
[11] M. Zhuli, L. Wenjing, and G. ZhiPeng, “Context based deep packet
inspection of ike phase one exchange in ipsec vpn,” in Innovative Com-
puting Communication, 2010 Intl Conf on and Information Technology
Ocean Engineering, 2010 Asia-Pacific Conf on (CICC-ITOE), pp. 3–6,
Jan 2010.
[12] M. Khan, “Quantitative analysis of multiprotocol label switching (mpls)
in vpns,” in Students Conference, 2002. ISCON ’02. Proceedings. IEEE,
vol. 1, pp. 56–65 vol.1, Aug 2002.
[13] D. Djim, “Openssh, keeping your communiques secret,” January 2011.
http://openbsd.das.ufsc.br/openssh/.
[14] D. Jim, “Openssh portable openssh.” http://www.openssh.com/portable.
html, 2014.
[15] G. J. Caceres Forte, Next Generation SSH2 Implementation: Securing
Data in Motion. Burlington, MA: Syngress Publishing, Inc., second ed.,
2011.
[16] J. Jayanthi and S. Rabara, “Ipv6 addressing architecture in ipv4 net-
work,” in Communication Software and Networks, 2010. ICCSN ’10.
Second International Conference on, pp. 461–465, Feb 2010.
 Comparative Study of Wireless Sensor Network
Standards for application in Electrical Substations
Fabrice Labeau Akash Agarwal
McGill University Indian Institute of Technology, Patna
Montreal, Canada India
fabrice.labeau@mcgill.ca
Abstract—Power utilities around the world are modernizing their
grid by adding layers of communication capabilities to allow for
more advanced control, monitoring and preventive maintenance.
Wireless Sensor Networks (WSNs), due to their ease of
deployment, low cost and flexibility, are considered as a solution
to provide diagnostics information about the health of the
connected devices and equipment in the electrical grid. However,
in specific environments such as high voltage substations, the
equipment in the grid produces a strong and specific radio noise,
which is impulsive in nature. The robustness of off-the-shelf
equipment to this type of noise is not guaranteed; it is therefore
important to analyze the characteristics of devices, algorithms
and protocols to understand whether they are suited to such
harsh environments. In this paper, we review several WSN
standards: 6LoWPAN, Zigbee, WirelessHART, ISA100.11a and
OCARI. Physical layer specifications (IEEE 802.15.4) are similar
for all standards, with considerable architectural differences
present in the higher layers. The purpose of this paper is to
determine the appropriate WSN standard that could support
reliable communication in the impulsive noise environment, in
electrical substations. Our review concludes that the
WirelessHART sensor network is one of the most suitable to be
implemented in a harsh impulsive noise environment.
Keywords—Wireless Sensor Networks; 6LoWPAN; Zigbee;
WirelessHART; ISA100.11a; OCARI; impulsive noise
environment; reliable communication.
I. INTRODUCTION
Today, electrical substations require real-time information for
adaptive energy allocation to the end users, for efficient
delivery of power to the customers and to increase profitability.
To address such needs, advanced wireless devices with low
power and mobile sensors are emerging. Such sensor devices
monitor the equipment in the substation and provide adaptive,
self-healing electric automation system for tele-maintenance,
tele-protection and tele-control [1]. Several Wireless Sensor
Network (WSN) standards currently exist, such as 6LowPAN,
Zigbee, ISA100.11a, WirelessHART and, more recently,
Optimization of communication for Ad-hoc reliable industrial
networks (OCARI) [2]. However, their implementation in the
electrical substation is an open area for research. In the rest of
the paper all the above mentioned standards will be collectively
termed as „WSN standards‟
The impact of WSN in electrical substation depends on the
reliable communication in the harsh and complex impulsive
noise environment (INE) of the electric substation [3]. In order
to deploy WSN in smart grids, the knowledge of parameters
such as wireless channel model or link quality information in
This work was supported by Hydro-Québec, the Natural Sciences and
Engineering Research Council of Canada and McGill University in the
framework of the NSERC/Hydro-Québec Industrial Research Chair in Interactive
Information Infrastructure for the Power Grid.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Basile Agba
Hydro-Québec (IREQ)
Varennes, Canada
such an environment is essential.
This paper provides the description and study of the above
mentioned WSN standards along with their stack structures
and network architectures. Also, an overview of the
measurement and characteristics of impulsive noise such as
impulse rate, amplitude, duration and rising time etc. in the
electric power systems environment (400/275/132-kV) is
provided.
The rest of the paper is organized as follows: In Section II,
an overview of the estimation and characteristics of impulsive
noise environment is provided. In Section III, we study the
protocol stack structure of WSN standards. In Section IV, we
compare the WSN standards, to figure out the most suitable
for application in the INE. Finally, we conclude in Section V.
II. IMPULSIVE NOISE ENVIRONMENT IN ELECTRICAL
SUBSTATION
It is known that the noise environment in electrical
substations is adverse and typically is impulsive [4]. Such
impulsive nature of noise degrades the communication carried
out on operating frequency band of the wireless network,
deployed in such an environment [3]. The major sources of
impulsive noise in an electricity substation are (1) Partial
Discharge (PD), which typically is caused due to imperfect
insulation, and (2) Sferic Radiation (SR), which results from
operation of circuit breakers and isolators etc. Both of these
processes produce radio waves that can be measured using
UHF antennas. Apart from the electrical substations,
atmospheric noise and other man-made noises also are the
sources of impulsive noise.
The estimation of impulsive noise can help in assessment of
difficulties in the deployment of a wireless network. Much
effort has already been laid towards measurement of impulsive
noise in a variety of physical environments. For instance, the
statistical characterization of the wireless channel such as
shadowing deviation and path loss, has been studied under
various environments within the substation, such as 500 kV
substations, an industrial power control room, and an
underground network transformer vault [5].
Various other experiments for the estimation of impulsive
noise in different frequency bands have also been conducted.
According to the measurement set up in [6], four types of
antennas are used for monitoring the partial discharge. Two
quasi-TEM half horns, designed to capture signals in
frequency bands 0.716 – 1.98 GHz and 1.92 – 5 GHz, a high
band (HB) horn, that covers the range 2 – 6 GHz and a low
band (LB) horn to cover the range 0.7 – 2 GHz. The fourth
antenna is a di-cone antenna to collect data below 700 MHz. It
is observed that some external interferences are encountered
during the on-site partial discharge measurements such as
discrete spectral interferences, periodic pulse shaped
interferences from power electronics or other periodic
switching etc., and Sferic radiation [4].
Similarly, in [7], the directional wideband antenna is used,
which covers the ISM band from 800 MHz to 2.5 GHz, to
evaluate the mean pulse duration, maximum amplitude and
mean number of discharges, in different voltage areas.
According to the measurement results in [7], there is a
correlation between the amplitude and the pulse duration
distributions.
In [8], the discrete wavelet packet transform technique is
used to remove the partial discharge from other interferences
in [6]. With this separation we observe the various
characteristics of the impulsive noise, such as impulse rate,
amplitude, duration of impulse and rise time of impulse.
Different distributions are observed from all the above
mentioned four antennas used. Of which the di-cone and HB
horn antennas show Gaussian normal distribution for
impulsive noise amplitude [6]. But LB horn antenna shows a
sharp glitch in the PDF close to 0 mV, which suggests the
presence of two impulsive noise processes, having a strong
periodic or quasi-periodic component [6]. Also, the PDF of the
rise times and impulse durations from all the antennas show
peaks at nearly 100 ns and 50 ns respectively, which
demonstrates that the rise times and durations of impulsive
noise are extremely short.
Hence, accurate estimation of the impulsive noise channel
model specifically for electrical substation could enable a
system designer to appropriately design the physical and MAC
layer specifications of a wireless network, to achieve more
reliable throughput.
III. WSN STANDARDS SPECIFICATION OVERVIEW
A. 6LoWPAN
6LoWPAN is an IPV6 based low powered Wireless
Personal Area Network (WAN) [9]. The nodes in 6LoWPAN
are connected in a star or mesh topology and usually support
the data rate of 20- 250 kbps to a distance of nearly ten meters.
This standard is developed so that wireless sensor devices can
connect to existing IP networks, such as IPV4 networks,
without the need of translation gateways or proxies [9].
2128 The 6LoWPAN stack structure is illustrated in Fig.1.a. In
the protocol stack, the link layer is divided into IEEE 802.15.4
MAC sub-layer and 6LoWPAN adaptation layer. [9].
The 6LoWPAN standard physical layer is based on IEEE
802.15.4-2006 (PHY) with 868/914 MHz or 2.4 GHz radio
[10]. The 6LoWPAN MAC sub-layer is fully compliant with
IEEE 802.15.4-2006 MAC [10]. The IEEE 802.15.4 MAC
superframe structure is bounded by beacons sent by
coordinators (first slot of superframe) and is divided into 16
time slots. Optionally the superframe can be divided into
active region and inactive region. During the inactive region
the coordinator may enter the low power mode or sleep mode.
The active period is subdivided into the contention access
period (CAP) and the contention free period (CFP) or the
guaranteed time slots (GTS). When a device wishes to
transmit in the CAP period then it has to compete with other
devices using a slotted CSMA-CA mechanism. The
6LoWPAN adaptation layer accomplishes the task of
fragmentation, header compression and reassembly of the
IPV6 packets to fit in IEEE 802.15.4 specified frame. This
standard also supports the low power listening mode (LPL) to
achieve low power mode operation in the network and to
access the channel in completely distributed and
unsynchronized way, through appropriate selection of sleep
interval for the devices [10][11][12].
In terms of IP layer routing, 6LoWPAN support protocols
such as Routing Protocols of Low power and Lossy networks
(RPL) [13], that mitigates problems such as non-deterministic
link statistics and lack of visibility into physical topology.
6LoWPAN supports only link layer security through 128-
bit Advanced Encryption Scheme (AES) encryption, which
can provide shielding from foreign attackers and outside
networks.
Co-existence with other devices such as Wi-Fi, is not efficient,
because of the use of same channel by all the devices. Some
advantage can be taken of using the 2.4 GHz radio, but even
then only three (15, 20 and 25) out of sixteen channels will
avoid interference [14].
Due to its internet compatibility and the use of IP based
network management tools such as Simple Network
Management Protocol (SNMP), 6LoWPAN is easily
integrated in other environments. 6LoWPAN also offers
interoperability with other wireless 802.15.4 devices and with
any other IP network link.
Application Protocols Application Upper
Layer (Application
Object & ZDO)
UDP & ICMP
Application
Network Layer (IPV6)
Support Sub-Layer
6LoWPAN Adaptation Network Layer
IEEE 802.15.4 MAC IEEE 802.15.4 MAC
IEEE 802.15.4 PHY IEEE 802.15.4 PHY
(a) (b)
Fig.1 (a) 6LoWPAN protocol stack (adapted from Z. Shelby and C. Bormann,
[9]), (b) Zigbee Stack Structure (adapted from Zigbee Document 053474r17,
[15])
B. Zigbee
Currently, in smart grids the most widely used WSN
protocol is Zigbee. It basically deals with the upper network
and application layers. The nodes in Zigbee are connected in a
star, mesh and tree topology [15]. Zigbee is designed for local
networks in home environments, and it cannot directly
communicate to servers on internet [15]. As the size of the
link layer address is 16 bits, therefore a total of 216 devices
can be supported in a network. Zigbee network architecture
supports three types of devices namely the Zigbee coordinator
(ZC), Zigbee Router (ZR) and Zigbee end device [15].
The Zigbee standard physical layer is based on IEEE
802.15.4-2003 (PHY) with 868/914 MHz or 2.4 GHz radio.
The data link layer is the IEEE 802.15.4-2003 MAC sub-layer
as defined in [16]. IEEE 802.15.4-2003 MAC superframe
structure is the same as that of IEEE 802.15.4-2006. Mostly,
this standard works in non-beaconed mode [15]. The stack
structure defined for Zigbee standard is shown in Fig.1.b.
The network layer provides services to the application support
sub layer, and supports Ad-hoc on-demand distance vector
routing protocol (AODV) and Tree routing protocol [18]. The
transport layer protocol is not defined in Zigbee. This standard
supports Network and Application layer security through 128-
bit AES encryption, where 128-bit link key is used for end-to-
end security between two devices and 128-bit network key is
shared by all devices in the network. According to [15],
though frequency agility is supported to mitigate interferences
from other non-Zigbee devices, for a large interval of time (in
hours) devices in the network use the same channel. Hence,
the performance of Zigbee devices degrades in presence of
other RF devices, such as Wi-Fi.
Zigbee offers interoperability with other Zigbee devices
with the same profile, but it needs an application layer
translator/gateway to communicate with internet.
C. WirelessHART
WirelessHART technology provides a robust wireless
protocol for the full range of process measurement, control,
and asset management applications [17]. It is based on HART
communication protocol [18], in which the two-way
communication can be carried over legacy 4-20 mA wire.
WirelessHART network supports star topology, mesh
topology and a combination of both. In Fig.2.a, the stack
structure of WirelessHART is depicted. As the size of the link
layer address is 16 bits, thus up to 216 devices within a
network are supported. But the high packet latency and power
consumption in the network could limit the device scalability
[18].
Application Layer Application Layer
UDP & TCP UDP & TCP
Network Layer Network Layer
Logical Link Upper Data Link Layer
Control
MAC Extension
IEEE 802.15.4 MAC
IEEE 802.15.4 MAC
IEEE 802.15.4 PHY IEEE 802.15.4 PHY
(a) (b)
Fig.2.(a) WirelessHART Stack Structure (adapted from S. Petersen and S.
Carlsen, [18]), (b) ISA100.11a stack model (adapted from S. Petersen and S.
Carlsen, [18])
The WirelessHART network architecture consists of field
devices for field sensing and actuating functions, routers (all
devices have routing capability), Gateway to translate
protocols between the application layer of the device
following this standard and other devices, Access Points to
connect the wireless mesh network to the gateway, and
Handheld device and Adapter device to provide physical and
logical connection to the external devices to the wireless
network [17].
This standard is based on IEEE 802.15.4–2006 (PHY) 2.4
GHz radio [10]. The data link layer is divided into Logical
Link Control (LLC), and MAC sub-layer fully which is based
on IEEE 802.15.4-2006 MAC specification [10]. This
standard uses TDMA mechanism to allocate fixed time slots
of 10ms each, which is allocated for device-to-device
communication without waiting for other devices. The channel
sharing is carried out through the CSMA-CA mechanism,
specifically through carrier sense only (CSO) mode. A unique
channel per time slot (frequency hopping) is assigned to the
communicating devices for interference free communication,
thereby making devices immune against the interference from
other RF devices, operating in the same band.
In this standard, the mandatory protection in data link and
network layer is provided through Symmetric AES 128-bit
keys. Three types of keys namely: join keys, network keys and
session keys are used. Session keys are allotted for device to
device communication, and network and join keys are used by
all devices.
WirelessHART offers interoperability with devices using
the same „HART‟ communication protocol, but it does not
support compatibility with internet, however many companies
manufacture WirelessHART based devices such as ABB,
Pepperl + Fachs, Emerson, MACTeK etc. [18].
D. ISA100.11a
ISA100.11a is an industrial project, a part of ISA100, which
belongs to a family of standards for wireless systems for
industrial automation, process control and related applications.
ISA100.11a supports star topology, mesh topology and the
combination of the two. The standard is based on IPV6.
ISA100.11a network supports star topology, mesh topology
and the combination of the two. This standard resembles
WirelessHART in many aspects. The architectural and
analytical differences between both are studied in [18] and
[20]. This standard is compliant with the 6LoWPAN standard,
as its network and transport layers are based on it. Fig.2.b
shows the stack structure of ISA100.11a standard.
This standard is based on IEEE 802.15.4–2006 (PHY) 2.4
GHz radio [10]. In ISA100.11a the data link layer is divided
into IEEE 802.15.4-2006 MAC sub-layer, Upper Data Link
Sub-layer (DLL) and Data link shim layer between MAC and
Upper DLL. The functions of upper DLL sub-layer is TDMA,
channel hopping and mesh routing. The MAC sub-layer is
responsible for transmission and reception of individual
frames, using the CSMA-CA mechanism. The CSMA-CA
back-off and retry mechanism is different in ISA100.11a. It
involves the use of spatial diversity, time diversity and
frequency diversity as the retry mechanism. For this standard
the use of CSMA technique, as supported by IEEE 802.15.4
PHY is optional. Depending on the system configuration, the
physical layer shall disable CSMA as requested by data link
layer.
This standard also uses TDMA and channel hopping with
ARQ mechanism for interference suppression from, and
coexistence with other RF devices. As compared to
WirelessHART, ISA100.11a defines three types of channel
hopping mechanisms namely, slotted hopping, slow hopping
and hybrid operation. In slotted hopping, equal length time
slots are used. Each time slot uses a different radio channel in
a hopping pattern. Timeslot scale is between 10 to 12 ms per
hop. This kind of hopping pattern requires quite tight time
synchronization between time slots. Slow hopping on the other
hand, allows a set of adjacent timeslots to be combined on a
single radio channel. In this case, timeslot scale is typically
100 to 400ms per hop. Time synchronization is relaxed in this
type of hopping. The hybrid operation is the combination of
the two for a particular arrangement.
 In ISA100.11a, and according to [20], the manufacturers are
not interested in being interoperable. As this standard is
compliant with 6LoWPAN to handle IPV6 traffic, it is also
compatible with internet. Some companies manufacture
devices based on this protocol such as Honeywell and
Yokogawa [20]. This standard supports symmetric AES 128-
bit encryption.
E. OCARI
This is another wireless sensor network protocol which is
aimed at providing reliable communication in harsh
environments such as Power Grids, Warships etc. The goal of
this protocol is to provide maximum life time to the network
devices, and hence the network. This standard is still under
development, and it is been promoted as an open source
standard for industrial wireless technology [2]. In OCARI, the
concentration is towards the improvement of Zigbee standard.
Fig.3.depicts the protocol stack structure of OCARI. This
standard supports tree topology. The OCARI network consists
of an OCARI end device, OCARI cell coordinator, and a
workshop coordinator.
The OCARI standard is based on IEEE 802.15.4–2006
(PHY) 2.4 GHz radio [10]. The data link layer of OCARI
standard is based on a synchronized tree based MaCARI
protocol [21] instead of IEEE 802.15.4 MAC. This link layer
protocol ensures regular sleep schedules to the network
devices and bounded end–to-end delay [21]. Unscheduled
activities periods are similar to the inactivity period of IEEE
802.15.4 superframe structure. During this period, the
coordinators can communicate among themselves using the
CSMA-CA mechanism. The messages which do not require
bounded end-to-end delay can be sent during this time period.
The main aim of this protocol is to increase the network
lifetime by utilizing low power of the network devices. For
this purpose, the network layer uses energy efficient routing
algorithms along with the node activity scheduling algorithms.
Application Upper
Layer
Zigbee Application Support Sub-
Layer
NwCARI (OCARI Network
Layer)
MaCARI (OCARI MAC
Layer)
IEEE 802.15.4 PHY
Fig.3. OCARI stack structure (adapted from K. Al Agha, M. Bertin et al, [2])
The properties like co-existence with other RF devices,
interoperability and market availability and support are not as
easy to assess given the novelty of this proposed protocol.
IV. COMPARISON
In industrial applications the factors such as the network
lifetime, coexistence with other RF devices, interoperability,
market availability, network security and reliability, are of
prime importance to decide the deployment of a WSN in a
specific environment.
WirelessHART supports fixed time slots with channel
hopping for inter-device communication, which prevents from
interference with other RF devices and reduces power
consumption. In ISA100.11a, the configuration of time slots is
flexible and hence, interference and utilization of power
increases. Zigbee on the other hand, supports almost no
frequency hopping, which enhances the message collisions,
and hence the power consumption. Also ZC and Zigbee
routers need to be continuously awake to listen to the channel,
which increases utilization of power. In 6LoWPAN, devices
operate on the same frequency band, due to which interference
with other existing RF devices increases, and hence the power
consumption. In OCARI, the MaCARI protocol provides low
power mode operation to maximize the network lifetime.
In WirelessHART and ISA100.11a, coexistence with other
existing RF devices in the network is somewhat better than
other WSN standards, as discussed.
The network security is supported and mandatory in
WirelessHART than in ISA100.11a. The benefit of data-link
and network/transport layer security is to have in-network and
external network protection. Zigbee on the other hand supports
layer security on demand. 6LowPAN just supports data-link
security.
The market support of WirelessHART is comparatively
better as compared to other WSN standards. Zigbee on the
other hand is a widely used standard in smart grid
applications, but its non-compatibility with internet makes it
more complicated to integrate with other systems.
As both ISA100.11a and 6LoWPAN support IPV6 based
traffic, therefore the networks formed by them are quite
scalable. WirelessHART and Zigbee on the other hand, being
non-compatible with internet are not so scalable.
Except ISA100.11a, all the other standards support the
property of interoperability.
If we study the application of Zigbee and 6LoWPAN in
smart grids then, we observe that the application of
6LoWPAN in smart grids is more robust than that of Zigbee
[22]. The reason is that the IP compatibility and architecture of
6LoWPAN makes it more immune to the network
breakdowns, whereas the total network breakdown will occur
if the Zigbee controller (ZC) managing Zigbee network
collapses. Similarly in ISA100.11a, only one router carries all
the routing operations. If that fails, network breakdown could
occur. Contrary to that, in WirelessHART all the devices
possess routing capability, thus in case of failure of any one,
another could route data.
Now, comparing the protocol stack structures and
characteristics of WSN standards, it could be seen that for
WirelessHART, the properties such as 2.4 GHz band
operation, lower power consumption, time slotted channel
hopping, channel black listing, fixed time slot communication
among devices, network layer graph and source routing,
interoperability, and security at both data link and network
layers, propose it to be a good option among all WSN
standards, to be tested in the impulsive noise environment, in
the electrical substation.
But, in [14] and [23] it has been shown that the performance
of WirelessHART degrades when deployed in coexistence
with IEEE 802.11 network. It has been shown in [14] that,
with the increase in the traffic on IEEE 802.11 network, the
packet loss increases significantly in WirelessHART network.
Also, it is shown in [23] that, if the duty cycle of the nodes in
WLAN network increase, then it will deteriorate the
performance of WirelessHART network (packet loss)
significantly. In 2.4GHz band, the IEEE 802.15.4 PHY
channels that do not interfere with the IEEE 802.11 b/g PHY
channels are 15, 20 and 25. In order to deploy WirelessHART
in coexistence with 802.11 b/g, the radio planning for the site
is essential. Also, considering the security conditions, in
WirelessHART standard, the security on the network and the
data link layer is mandatory and requires extra processing time
and energy [18]. Thus, there could be an issue of extended
battery life of the nodes and hence the network life time.
V. CONCLUSION
In this paper, the specifications of standards such as
6LoWPAN, Zigbee, WirelessHART, ISA100.11a and OCARI
has been reviewed. The stack comparison of these standards
along with their major properties such as coexistence,
interoperability, network lifetime, market availability and
support etc. has been done. Also, an overview of the impulsive
noise in the electricity transmission substation has been
provided. The channel hopping capability is carried through
the MAC layer and higher layers so that the users of 802.15.4
radios have the ability to select the best available channel for
operation. Through this selection, routing protocols can select
the appropriate routes, and also the interference by other
devices present in the nearby environment can be mitigated.
Through our review it could be concluded that WirelessHART
network with the above mentioned properties could be the
most suitable for implementation and testing for packet error
rates in the impulsive noise environment present in electrical
substations.
As the ISA100.11a protocol suite resembles that of
WirelessHART, with non-compulsory layer security
requirements, it would be interesting in future work to study
and analyze the performance of ISA100.11a in the impulsive
noise environment, as this property of ISA100.11a could
provide additional power reduction and reduced processing
time and memory consumption. Also, performance tests of
ISA100.11 in coexistence with other RF devices in the
industrial plant such as WLAN could be an interesting area to
study.
As 6LoWPAN network is based on IPV6 addressing,
therefore even if the edge router or intermediate router fails,
the information can still be extracted from 6LoWPAN end
nodes, as each node has an IPV6 address. The application of
this idea has been proposed for smart grids in [22], but has not
been tested for packet error rates in the impulsive noise
environment. Hence, the performance analysis of 6LoWPAN
protocol suite with appropriate node duty cycles would also be
an interesting study to analyze.
REFERENCES
[1] V. C. Gungor and F. C. Lambert, “A survey on communication networks
for electric system automation,” Comput. Netw., vol. 50, no. 7, pp. 877–
897, May 2006.
[2] K. Al-Agha, M.-H. Bertin, T. Dang, A. Guitton, P. Minet, T. Val, and J.-
B. Viollet, “Which Wireless Technology for Industrial Wireless Sensor
Networks? The Development of OCARI Technology,” IEEE Trans. Ind.
Electron., vol. 56, no. 10, pp. 4266–4278, 2009.
[3] G. Madi, F. Sacuto, B. Vrigneau, B. L. Agba, Y. Pousset, R. Vauzelle,
and F. Gagnon, “Impacts of impulsive noise from partial discharges on
wireless systems performance: application to MIMO precoders,”
EURASIP J. Wirel. Commun. Netw., vol. 2011, no. 1, p. 186, Nov. 2011.
[4] S. A. Bhatti, I. A. Glover, R. Atkinson, Q. Shan, Y. Yang, J. M. R. de S
Neto, and J. S. da Rocha Neto, “Vulnerability of Bluetooth to impulsive
noise in Electricity transmission substations,” in IET International
Conference on Wireless Sensor Network, 2010. IET-WSN, 2010, pp. 53–
58.
[5] V. C. Gungor, B. Lu, and G. P. Hancke, “Opportunities and Challenges of
Wireless Sensor Networks in Smart Grid,” IEEE Trans. Ind. Electron.,
vol. 57, no. 10, pp. 3557–3564, 2010.
[6] Q. Shan, I. A. Glover, R. C. Atkinson, S. A. Bhatti, I. E. Portugues, P. J.
Moore, R. Rutherford, M. de Fátima Queiroz Vieira , A. M. N. Lima, and
B. A. de Souza, “Estimation of Impulsive Noise in an Electricity
Substation,” IEEE Trans. Electromagn. Compat., vol. 53, no. 3, pp. 653–
663, 2011.
[7] F. Sacuto, F. Labeau, J. Beland, M. Kassouf, S. Morin, S. Poirier, and B.
L. Agba, “Impulsive Noise Measurement in Power Substations for
Channel Modeling in ISM Band,” presented at the 2012 CIGRE Canada
Conference, Montreal, 2012.
[8] L. Satish and B. Nazneen, “Wavelet-based denoising of partial discharge
signals buried in excessive noise and interference,” IEEE Trans. Dielectr.
Electr. Insul., vol. 10, no. 2, pp. 354–367, 2003.
[9] Z. Shelby and C. Bormann, 6LoWPAN: The Wireless Embedded Internet.
Wiley, 2009.
[10] “Wireless Medium Access Control (MAC) and Physical Layer (PHY)
specifications for Low-Rate Wireless Personal Area Networks
(WPANs),” IEEE 802.15.4, 2006.
[11] B. Krishnamachari, Networking Wireless Sensors. Cambridge University
Press, 2005.
[12] E. Toscano and L. Lo Bello, “Comparative assessments of IEEE
802.15.4/ZigBee and 6LoWPAN for low-power industrial WSNs in
realistic scenarios,” in 2012 9th IEEE International Workshop on Factory
Communication Systems (WFCS), 2012, pp. 115–124.
[13] J. W. H. <jhui@archrock.com>, “The Routing Protocol for Low-Power
and Lossy Networks (RPL) Option for Carrying RPL Information in
Data-Plane Datagrams.” [Online]. Available:
http://tools.ietf.org/html/draft-ietf-6man-rpl-option-06. [Accessed: 01-
Dec-2013].
[14] L. Angrisani, M. Bertocco, D. Fortin, and A. Sona, “Experimental Study
of Coexistence Issues Between IEEE 802.11b and IEEE 802.15.4
Wireless Networks,” IEEE Trans. Instrum. Meas., vol. 57, no. 8, pp.
1514–1523, 2008.
[15] “ZigBee Document 053474r17.” Zigbee Alliance, Jan-2008.
[16] “Wireless Medium Access Control (MAC) and Physical Layer (PHY)
specifications for Low-Rate Wireless Personal Area Networks
(WPANs).” IEEE 802.15.4, 2003.
[17] “Industrial communication networks – Wireless communication network
and communication profiles – WirelessHART.” International
Electrotechnical Commission, Apr-2010.
[18] S. Petersen and S. Carlsen, “WirelessHART Versus ISA100.11a: The
Format War Hits the Factory Floor,” IEEE Ind. Electron. Mag., vol. 5, no.
4, pp. 23–34, 2011.
[19] “ISA100.11a.” International Society of Automation, May-2011.
[20] M. Nixon, “A Comparison of WirelessHART and ISA100.11a.” Emerson
Process Management.
[21] G. Chalhoub, A. Guitton, and M. Misson, “MAC specifications for a
WPAN allowing both energy saving and guaranteed delay,” in Wireless
Sensor and Actor Networks II, A. Miri, Ed. Springer US, 2008, pp. 221–
232.
[22] C.-W. Lu, S.-C. Li, and Q. Wu, “Interconnecting ZigBee and 6LoWPAN
wireless sensor networks for smart grid applications,” in 2011 Fifth
International Conference on Sensing Technology (ICST), 2011, pp. 267–
272.
[23] S. Petersen and S. Carlsen, “Performance evaluation of WirelessHART
for factory automation,” in IEEE Conference on Emerging Technologies
Factory Automation, 2009. ETFA 2009, 2009, pp. 1–9.
 Classification Performance Analysis of MNIST
Dataset utilizing a Multi-resolution Technique
Ramesh Kumar Mohapatra, Banshidhar Majhi, and Sanjay Kumar Jena
Department of Computer Science and Engineering
National Institute of Technology Rourkela
Odisha, 769008, India
{mohapatrark, bmajhi, skjena}@nitrkl.ac.in

Abstract—Here, we propose a method for recognition of hand-
written English digit utilizing discrete cosine space-frequency
transform known as the Discrete Cosine S-Transform (DCST).
Experiments have been conducted on the publicly availabe
standard MNIST handwritten digit database. The DCST features
along with an Artificial Neural Network (ANN) classifier is
utilized for solving the classification issues of written by hand
digit. The Discrete Cosine S-Transform coefficients are extracted
from the standard images of MNIST handwritten isolated digit
database. The database consists of a total of 70000 including
60000 training samples and 10000 test samples. To overcome
the computational overhead, we have normalized all the images
of the MNIST dataset from 28 × 28 to 20 × 20 image size
by eliminating the unsought boundary pixels up to width four.
Further, the classification of digits has been made by using a back
propagation neural network (BPNN). This work has achieved
precisely 98.8% of success rate for MNIST database.
KeywordT‰
Classification;
MNIST
dataset;
Discrete
Cosine

S-Transform;
BPNN.
I. I NTRODUCTION
The problem of optical scanned character recognition [1]
is to construe as well as translate apprehensible handwritten
input automatically, which is of great interest as well as a
gainsay in the field of object identification. It has numerous
down to earth applications in real time e-processing of data,
such as automatic address reading, postal mail sorting through
ZIP/PIN code, bank check processing, etc. As a benchmark,
the Modified National Institute of Standards and Technology
(MNIST) dataset has been used often times to design afresh
digit recognition system [2], [3]. In the literature study, we
found a great amount of works based on MNIST dataset,
suggesting many different methods [4]–[6]. Lauer et al. [7]
have proposed trainable feature extractors for handwritten digit
recognition. A comprehensive examination by Liu et al. [8]
compares the performance of previously proposed classifiers
including linear and polynomial classifier, Nearest Neighbor
classifier, and different neural networks. Normally individuals
don’t generally compose the same digit in the very same way
at a given purpose of time. Due to this within class variance
of the shape of a character, it is a major challenge for the
researchers to classify the character. Many feature extraction
approaches such as biologically inspired features [9], higher
order singular value decomposition [10], GA-based feature
selection approach [11], Fuzzy model based recognition [12]
have been proposed to specify the shape invariance within
a class to improve the discrimination ability. It is observed
that the accuracy and efficiency of many classifiers could be
improved substantially by extracting direction features, local
structural or curvature feature [13]. Kumar et al. [22] used
mathematical morphology techniques to divide the dataset into
two groups and further classify those considering the structural
features of the digits. The overall recognition rate recorded
as 92.5%. The system fails to identify the digits due to the
broken lines with large gap, incomplete digits and the digits
with uneven strokes. Also, many multi resolution techniques
are used for the classification of English digit. S-transform is
one of the multi resolution techniques that extracts multi-scale
resolution like discrete wavelet transform [17]. It has been
widely used in image restoration and texture analysis [16].
One of the major limitations of S-transform is its exponential
computational overhead. As of late a speedier adaptation of
the S-transform,in particular, discrete orthonormal Stockwell
transform (DOST) has been intended for feature extraction. We
have utilized the discrete cosine transform (DCT) rather than
fast Fourier transformation in DOST which prompts Discrete
Cosine S-Transform (DCST).
The remaining piece of the paper is composed as takes after.
Section II gives a full depiction of the database utilized. The
discrete cosine S-transform examined briefly in Section III-A.
The proposed method is depicted in Section IV. Section V
depicts the outcomes took after by finishing up concluding
remark in Section VI.
Fig. 1: 100 sample digit images of MNIST dataset

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

 TABLE II: Total number of samples for each digit in the test set.
Test set 0 1 2 3
Frequency 1001 1127 991 1033
II. DATABASE U SED
The MNIST database was gotten from NIST’s Special
Database 3 (SD-3) and Special Database 1 (SD-1) [2] that
comprises of binary images of written by hand digits ranging
from 0 to 9. These are collected by high school students
and employees of the United States Census Bureau. A total
of 30,000 patterns from SD-3 and 30,000 from SD-1 are
gathered aimlessly to have 60000 samples in the training set.
Test set comprises of 10,000 samples out of which 5,000 are
chosen from SD-3 and the rest are from SD-1. The MNIST
dataset contains gray-scale images of size 28 × 28. Thus, the
dimensionality of each image sample vector is 784. Hundred
sample images from the MNIST database are appeared in
Fig. 1. Since the samples are as of now standardized along
these lines, a little measure of preprocessing is required. For
our experiment we have normalized further all the images
of the MNIST dataset to 20 × 20 image size by killing the
unsought boundary pixels up to width four. Accordingly, the
dimensionality of every image test vector gets to be 400
lengths. The MNIST database is publicly available at the
TABLE I: MNIST dataset files and their sizes
Files Size in Bytes
Training set images 9912422
Training set labels 28881
Test set images 1648877
Test set labels 4542
homepage of LeCun [3] and these are shown in Table I along
with their sizes in bytes. The distribution of each digit of the
test set is shown in Table II.
III. F EATURE E XTRACTION
4 5 6 7 8 9
980 862 1014 1070 944 978
to the Discrete Stockwell Transform (DST) and is given by,
N
 −1
2π 2 m2 i2πmk
S[k, n] = e− n2 H[m + n]e N (2)
m=0
where k is the time translation and n is the index of frequency
shift with n = 0. Here H(.) is the DCT of h(.). The 2D-
DST of an image of size N × N has the growth function
O[N 4 + N 4 log(N )]. Due to high complexity and redundant
information of S-transform, it is not frequently used in many
applications. Notwithstanding, DCST can be utilized to speak
to information effectively, which is the orthonormal variant of
the DST, creating N point time recurrence representation for a
signal of length N. In this way, DCST give features with zero
data redundancy. The voice frequencies (vx , vy ) are obtained
within the bandwidth of 2px −1 × 2py −1 . If we represent the
voice frequencies as a complex image where, vx corresponds
to real and vy corresponds to imaginary part  then the mag-
nitude and phase angle is obtained by Mv = vx2 + vy2 and
v
θv = tan−1 ( vxy ) respectively. Hence, for each sample, we
obtained 400 DCST coefficients where the size of the image
is 20×20. So, our feature matrix size for m samples is m×400.
B. Classification Phase
We have considered the back propagation neural network
(BPNN) for the classification of the test set. The Artificial
Neural Network (ANN) used has three layers as shown in the
Fig. 2. The DCST features are extracted from each sample in
the training set. The network is trained considering the samples
from the training set. We fixed a total of 400 (i.e., n = 400)


 


 

The proposed scheme uses a multi-resolution technique i.e.,

2D discrete cosine S-transform (DCST) for feature extraction, 
which is a variation of Stockwell transform (ST) [18]. Each

 

image from MNIST database is resized to 20 × 20, further the
DCST coefficients are calculated at each pixel to get a feature



length of 400.

 
A. Discrete Cosine S-Transform 

The ST gives a full time decomposition of a signal and 

retains absolutely referenced phase information. Likewise, the

ST is evinced in the Cosine domain as,
∞ Fig. 2: General structure of a three layer ANN of size (n-k-m)
(2πψ)2
− i2πψτ
s(τ, f ) = C(ψ + f )e 2f 2 e dψ (1)
neurons in the input layer and since we have ten classes to
−∞
classify, the ANN has 10 neurons in the output layer (i.e., m
where C(f ) is the Cosine spectrum of c(t) and other param- = 10). For better convergence performance, the neurons in the
eters have their usual meaning. The discretization of (1) leads hidden layer are appositely fixed during the experiment.
    
 
rate by the following equations.

   
 
TP
Sensitivity (or T P R) = (3)
  

 
"  


 
TP + FN
TN


 Specif icity (or T N R) = (4)




   TN + FP
     

FP
F all out (or F P R) = (5)
TN + FP
 
 

!"   !"  

FN


 

# $
 # $

M iss rate (or F N R) = (6)
TP + FN
%   #
  %  '$

 

 &   & TP + TN
Accuracy = (7)
TP + FN + FP + TN

 The accuracy of the network can be calculated from the

Equation 7.


TABLE III: Confusion values for MNIST database

(

!

 & False False True True
Negative Positive Positive Negative
Digit Rate Rate Rate Rate
! 
0 0.0009 0.0090 0.9910 0.9991
1 0.0015 0.0107 0.9893 0.9985
2 0.0024 0.0232 0.9768 0.9976
Fig. 3: Proposed scheme 3 0.0028 0.0156 0.9844 0.9972
4 0.0017 0.0193 0.9807 0.9983
5 0.0016 0.0208 0.9792 0.9984
6 0.0009 0.0118 0.9882 0.9991
IV. P ROPOSED METHOD 7 0.0015 0.0149 0.9851 0.9985
8 0.0024 0.0202 0.9798 0.9976
9 0.0020 0.0154 0.9846 0.9980
Our proposed method is illustrated in the Fig. 3. The 400
DCST feature vector of all normalized images in the training
set is extracted and stored in a feature matrix, leading to a V. R ESULTS AND D ISCUSSION
size of 60000 × 400. The feature matrix along the associated
labels is trained using the neural network tool in MATLAB to In the supervised learning of the MNIST dataset [19] we
produce a network ‘nn’. Now, the process is repeated for each utilize 10,000 specimens out of 70000 accessible examples
sample from test set and a class is assigned to each sample to decrease the computational expense. Out of these, 7000
by performing a pattern matching over the network i.e., nn. examples are decided for training, 1500 are for validation and
To perform simulation, we have assumed the following: the rest are utilized for testing. Experimentally, the number
of neurons in the hidden layer of the network is observed
1) The dividerandom() method is used to split the data set to be eighty for better convergence characteristics (Fig. 6).
into three sets. Subsequently the ANN classifier has the structure of 400-80-
2) For faster training, scaled conjugate gradient method 10. The performance parameters are recorded in the Table
trainscg is used. III demonstrating the sensitivity, specificity, fall-in and fall
3) The train of the neural network stops when either the out separately and the curve between the cross-entropy and
gradient reaches a value of 10−6 or the number of number of epochs is plotted and appeared in Fig. 4. The
consecutive validation checks reaches 6, i.e., there is no
better improvement of the gradient for six consecutive TABLE IV: Confusion matrix for MNIST dataset
iterations.
Digit 0 1 2 3 4 5 6 7 8 9
4) In testing dataset, 70% samples are used to train the 0 993 0 0 1 1 1 1 1 1 2
network, 15% are used for validation and the rest are 1 0 1114 5 0 2 2 0 0 2 2
2 1 5 969 4 1 3 0 3 5 0
used for testing the network. 3 2 0 8 1008 0 6 0 4 4 1
4 0 0 4 0 965 1 2 0 2 6
The performance of the classification is evaluated with the 5 1 0 0 5 0 847 5 0 4 0
assistance of confusion matrix that use to assess the quality of 6 3 1 1 0 3 0 1006 0 0 0
7 0 3 3 3 1 0 0 1057 0 3
the output of a classifier on a given data set. The performance 8 2 2 2 2 0 4 4 5 922 1
measures like true positive (T P ), false positive (F P ), true 9 0 1 0 1 11 1 0 3 1 960
negative (T N ), and false negative (F N ) are ascertained which
thus assesses the sensitivity, specif icity, f all out, and miss receiver operating characteristic (ROC) curves for all ten
classes are shown in one place (see Fig. 5). The region under 1
the ROC curve gives the precision too. The more the curve
towards the upper left corner the more is system’s exactness. 0.98
The confusion matrix for testing patterns is appeared in the
Table IV. It is observed that proposed scheme gives an overall

True Positive Rate

0.96
accuracy of 98.8%. The portion of samples misclassified is
0.012 i.e., 1.2% error. We have also compared our scheme 0.94
with some of the existing schemes that is portrayed in Table
V. 0.92

TABLE V: Other results on MNIST.

0.9
Scheme Over all Error rate(%)
Structural features [22] 7.5
Nearest neighbor classifier (Euclidean) 3.09 0.88
Higher order singular value decomposition [10] 2.0 0 0.02 0.04 0.06 0.08 0.1 0.12
Convolutional Net LeNet-1 [20] 1.7 False Positive Rate
Our scheme 1.2
Polynomial SVM [20] 1.1 Fig. 5: ROC curves for 10 classes of MNIST dataset
Direction gradient [21] 0.42
Human [19] 0.2
1
10

0
10
Train 10
0

Validation
Test
Cross−Entropy (crossentropy)

Best
gradient

−1
−1
10
10

−2
10

−2
10
−3
10
0 20 40 60 80 100
Number of Epochs

−3
10
0 10 20 30 40 50 60 70 80 90 Fig. 6: Training convergence characteristics of MNIST dataset
Number of epochs

Fig. 4: Performance curve of MNIST dataset

6

VI. C ONCLUSION 5

Our preliminary result shows that the discrete cosine S-

4
transform features alongside back propagation neural network
validation fail

is adequately descriptive for the purpose of classification

3
of handwritten English digits of the MNIST dataset. This
demonstrates viability of the proposed method over some of
2
the existing schemes. Still, further improvement is conceivable
to accomplish more exactness than recommended in this paper.
1
ACKNOWLEDGMENT
Special thanks to LeCun for being the pioneer of making 0
0 20 40 60 80 100
the MNIST dataset and made it available for research purpose. Number of Epochs

R EFERENCES
Fig. 7: Validation checks during the training of MNIST dataset
[1] Richard O Duda, Peter E Hart, et al. Pattern classification and scene
analysis, volume 3. Wiley New York, 1973.
[2] Li Deng. The mnist database of handwritten digit images for machine
learning research [best of the web]. Signal Processing Magazine, IEEE, [3] Yann Lecun and Corinna Cortes. The MNIST database of handwritten
29(6):141–142, Nov 2012. digits.
 [4] Ernst Kussul and Tatiana Baidyk. Improved method of handwritten
digit recognition tested on {MNIST} database. Proceedings from the
15th International Conference on Vision Interface. Image and Vision
Computing, 22(12):971 – 981, 2004.
[5] Zhihong Man, Kevin Lee, Dianhui Wang, Zhenwei Cao, and Suiyang
Khoo. An optimal weight learning machine for handwritten digit image
recognition. Special issue on Machine Learning in Intelligent Image
Processing. Signal Processing, 93(6):1624 – 1638, 2013.
[6] Ping Zhang, Tien D. Bui, and Ching Y. Suen. A novel cascade ensemble
classifier system with a high recognition performance on handwritten
digits. Pattern Recognition, 40(12):3415 – 3429, 2007.
[7] Fabien Lauer, Ching Y. Suen, and Grard Bloch. A trainable fea-
ture extractor for handwritten digit recognition. Pattern Recognition,
40(6):1816 – 1824, 2007.
[8] Cheng-Lin Liu, Kazuki Nakashima, Hiroshi Sako, and Hiromichi Fuji-
sawa. Handwritten digit recognition: benchmarking of state-of-the-art
techniques. Pattern Recognition, 36(10):2271 – 2285, 2003.
[9] Angelo Cardoso and Andreas Wichert. Handwritten digit recognition
using biologically inspired features. Neurocomputing, 99(0):575 – 580,
2013.
[10] Berkant Savas and Lars Eldn. Handwritten digit classification us-
ing higher order singular value decomposition. Pattern Recognition,
40(3):993 – 1003, 2007.
[11] C. De Stefano, F. Fontanella, C. Marrocco, and A. Scotto di Freca. A
ga-based feature selection approach with an application to handwritten
character recognition. Frontiers in Handwriting Processing. Pattern
Recognition Letters, 35(0):130 – 141, 2014.
[12] M. Hanmandlu and O.V. Ramana Murthy. Fuzzy model based recogni-
tion of handwritten numerals. Pattern Recognition, 40(6):1840 – 1854,
2007.
[13] Jie Zhou, Adam Krzyzak, and Ching Y. Suen. Verificationa method
of enhancing the recognizers of isolated and touching handwritten nu-
merals. Handwriting Processing and Applications. Pattern Recognition,
35(5):1179 – 1189, 2002.
[14] Ying Wen and Lianghua He. A classifier for bangla handwritten numeral
recognition. Expert Systems with Applications, 39(1):948 – 953, 2012.
[15] Hossein Khosravi and Ehsanollah Kabir. Introducing a very large dataset
of handwritten farsi digits and a study on their varieties. Pattern
Recognition Letters, 28(10):1133 – 1141, 2007.
[16] Robert Glenn Stockwell. A basis for efficient representation of the s-
transform. Digital Signal Processing, 17(1):371–393, 2007.
[17] Rafael C Gonzalez. Digital image processing. Pearson Education India,
2009.
[18] Yanwei Wang and Jeff Orchard. The discrete orthonormal stockwell
transform for image restoration. In 16th IEEE International Conference
on Image Processing (ICIP), pages 2761–2764. IEEE, 2009.
[19] Yann LeCun, LD Jackel, L Bottou, A Brunot, C Cortes, et al. Com-
parison of learning algorithms for handwritten digit recognition. In
International conference on Artificial neural networks. volume 60, pages
53–60, 1995.
[20] Lecun, Y. and Bottou, L. and Bengio, Y. and Haffner, P. Gradient-based
learning applied to document recognition. In Proceedings of the IEEE.
volume 86(11), pages 2278–2324, 1998.
[21] Liu, Cheng-Lin and Nakashima, Kazuki and Sako, Hiroshi and Fujisawa,
Hiromichi Handwritten digit recognition: benchmarking of state-of-the-
art techniques. Pattern Recognition. volume 36(10), pages 2271–2285,
2003.
[22] Kumar, V Vijaya and Srikrishna, A and Babu, B Raveendra and Mani,
M Radhika. Classification and recognition of handwritten digits by
using mathematical morphology. Sadhana. volume 35(4), pages 419–
426, 2010.
 Improving User Cognitive Processes in Mobile
Learning Platforms through Context-Awareness
Brita Curum
Faculty of Engineering
University of Mauritius (UoM)
Réduit, Mauritius
brita_curum@hotmail.com
Abstract—The use of mobile devices for learning is prominent
nowadays. Mobile learning has become an essential part to
enhance today’s learning style. However, technologies still lack
abilities to fully understand human reaction and comprehend
their cognitive processes to be able to interact with them and
deliver the right amount of contents. Therefore, one research
scope is having devices which can sense and acknowledge user
context information according to the individual’s cognitive load.
The purpose is to seriously engage a learner with learning
materials with an appropriate speed of information flow, within
the mobile device without constraints of time, locations, and
device restrictions. In this paper, a detailed evaluation of context-
aware mechanisms to improve cognitive load of users in mobile
learning platforms is carried out. Different evaluation criteria
are identified and discussed followed with a detailed assessment
on the recent existing works of context-aware algorithms.
Identified open challenges and research directions for using
context-aware information to improve mobile learning are
presented.
Keywords—mobile learning; context-awareness; cognitive load;
context-aware mechanisms; mobile learning platforms
I. INTRODUCTION
Technology has brought about many changes in our daily
life. The dominating characteristic of mobile devices
permeates our routine with its permanent accessibility to
information and communication. Being an ubiquitous asset,
mobile devices have shown its potential to support learning.
Mobile learning refers to the use of mobile devices whereby
learning is supported while on the move, leaving behind
limitations of traditional educational environments. Reference
[1] defines mobile learning as the process to teach and learn
with the aid of mobile technologies such as smart phones,
PDAs, media players and tablet computers which can be
available anywhere instantly. In recent years, mobile learning
has become widespread [2] and has attracted many learners.
As a result, a clear growth of the demands for adapted delivery
of educational contents is seen [3]. A platform such as Moodle
is an integrated open Learning Management System (LMS)
which assist learners and educators in creating personalized
educational materials. Other similar systems exist which are
gradually moving on mobile. Examples are MOMO, MLE-
Moodle and many others. The key reason behind many
researchers‟ interest about mobile learning stems from its
portable, customizable and situated characteristics [4-7].
However, with the vast quantity of learning resources
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Kavi Kumar Khedo
Faculty of Engineering
University of Mauritius (UoM)
Réduit, Mauritius
k.khedo@uom.ac.mu
available, studies have also pointed out that too much
information causes cognitive load [8-9]. Mobile learning on
education is restricted to two characteristics. These are the
effectiveness of m-learning and the design of these systems
[10]. Researchers have identified m-learning as efficient using
learning outcomes such as motivations, perceptions, attitudes,
academic records and satisfaction of the student, rather than
the learning processes [11-12]. Learning processes which
cover cognitive load need to be looked into to address this
gap.
II. MOBILE LEARNING
Mobile learning is a focused and challenging area for
enhancing learning approaches in the next few years. Mobile
devices which are used for information sharing and
communication are produced with different hardware and
software specifications [13]. This new form of learning will
provide us with an interface based on our actual situation and
helps us to strengthen our knowledge in any instance [14].
Wisdom will come from mobile devices and be propagated
amongst peers in a social context system as supported by
reference [15], who also reported that m-learning supports
instant communication with peers. Many authors viewed
mobile learning as an add-on to e-learning which allows a
wireless connection [16]. This additional feature has
modernized the lifestyles of people, allowing various
functions which were not present earlier. As a result, mobile
technologies are reshaping the routine of users in many ways.
A. Recent advances in Mobile learning
Mobile devices have impacted many other sectors,
regardless educational sectors such as politics and are seen to
be lucrative to productivity [17]. Given that the reach of
mobile technologies is further expanded, with its low
competitive prices and newly added features, more and more
people are allowed to own one or many such device. Since
people carry mobile devices with them most of the time,
learning can occur anytime and anywhere.We are no more
restricted to traditional classrooms to study. The flexibility
supported by mobile learning is sustained. Furthermore, the
concept of Bring Your Own Device (BYOD) in workplaces is
increasing productivity. Richer and dynamic learning is
encouraged in this way. The portable characteristic together
with the power of applications (apps) in mobile devices are
influencing another look upon the schools‟ own policies
regarding mobile devices. Mobile phones are seen as a
principal feature of BYOD environments [18].
To solve challenges in a mobile environment, cloud
computing was introduced. Unlimited resources offered by
cloud computing in a resource-constraint mobile environment
is used [19]. In newly developed mobile cloud architecture,
mobile devices can access cloud services either through ad-
hoc mobile network or access points [20]. Although
smartphones come with powerful hardware specifications, the
device energy, network connectivity, bandwidth and mobility
remains the chief limitations. However, the rapidly evolving
technology of cloud computing could assist smartphones to
cope with these critical factors, by running and storing
applications in cloud [21]. Cloud computing assists mobile
devices to overcome constraints in terms of data storage,
bandwidth, heterogeneity, scalability, availability, reliability,
and privacy [22].
Gamification is a growing form of digital content in the
form of games. They are integrated to entertain users while
learning is supported simultaneously. Reference [23] reported
gamification as the provider of interactive edutainment,
increasing student engagement and alternative ways of
extending the learning process. These edu-games can be
accessed on mobile devices at any time.
One noticeable recent trend in learning using new
technologies is the rise of Massive Open Online Courses
(MOOCs). Many universities have adopted these courses as a
way to increase the number of students‟ admissions. Mobile
technologies will enable MOOCS in the coming years to
propose more customized learning [24].
Although recent advancements in technology present
different options and capabilities for learning through mobile
devices, it is very difficult to grasp full spectrum of the
possibilities provided by the device. Digital technologies
development has been limited to social communication in the
past. Only few people centered mobile learning as an
academic activity in higher institutions [25]. Indeed, there is a
growing interest to cater for user context change and provide
improved learning experiences as stated by Reference [26].
B. Mobile learning platforms
Investigations have shown that students like the flexibility
that mobile learning is providing them and with experiences
they are developing positive perceptions towards this
technology [27]. There exist a number of commercial mobile
learning platforms already implemented and in use today.
Moodle is one model as seen in section I. which is widely
adopted internationally. In this section, other mobile learning
platforms are presented and discussed.
Mobl21 [29] is an awarded mobile learning platform that
supports dynamic and unconventional learning. Available as a
mobile apps, desktop widgets or web applications, educators
develop customizable contents that are available through
mobile devices. It allows study at a more convenient pace with
an instant access to valuable learning material anytime and
anywhere. Mobl21‟s additional features ease educators‟ tasks
in controlling content access and check test performances.
WizIQ [28] is one of the world‟s largest cloud-based
education platforms. Deployed in mobile devices, this
platform allows students to participate in interactive classes
anytime and anywhere. Connection with people having similar
interests is permitted. Moreover, notifications of important
deadlines, tests and class schedules are made available to its
users. Being the first system capable of delivering live classes,
the real potentials of mobile learning are unlocked.
The above mentioned platforms have proved how efficient
mobile technologies enhanced teaching and learning practices.
Global organizations, like UNESCO is investing significant
resources into exploring the fast expansion of mobile
technologies to encourage learning [24]. These nonstop efforts
will promote prolific conditions of mobile learning platforms
to learners and educators.
III. CONTEXT AWARENESS IN MOBILE LEARNING
A system is considered as “context-aware” if it can extract,
read and use contextual information and adapt its
functionalities to the current context in use [29]. According to
researchers, the three types of information to be collected to
define context are places, people and things [30]. To define
these three entities, four categories of context are identified
which are clearly described below [38]:
 Identity: characterizes the entity with an explicit
identifier, unique in the name domain of the
application.
 Location: includes spatial and geographical data.
 Status or Physical context: contains properties
distinguished by the user. It includes noise level,
temperature, lighting and many others.
 Time: can be time of the day, month, year, date.
A. Types of context
A common approach to classify context instances is to
distinguish between the context dimensions. References [31]
and [32] broadly refer to these dimensions as external or
extrinsic and internal or intrinsic, while Reference [33]
grouped them into physical and logical context. External or
physical dimensions refer to context that can be measured by
hardware sensors such as light, sound, touch, temperature and
many others. Internal or logical dimension is specified by the
user or captured by monitoring user interactions, such as the
learner‟s goals, tasks to complete and emotional state.
Importance is given to external context instances most of the
time. Location, activity, identity and time are identified as
primary contexts. Nonetheless, some efforts to use other
context information were carried out in recent years [34].
Additionally, contextual information is vital for effective
content adaptations in m-learning platforms. An adaptive
engine is required to provide adaptive learning [35]. The
engine obtains data and outputs relevant adaptation results.
Context is based on the learner‟s state, educational state,
activity‟s state, infrastructure‟s state and the surrounding‟s
state [36]. The learner‟s state consists of the following
dimensions: cognitive skills, intentions, learning styles,
preferences and interactions with the system [37]. Depending
on the individual‟s contextual information, these dimensions
may be constant or may vary over time. In this paper, focus is
mainly on the learner‟s state since it enfolds the cognitive
abilities of a learner.
On the other hand, Reference [38] stated that contextual
information is captured using sensors which are of three basic
categories. These are the Activity sensors, Bio sensors and the
Environmental sensors. These sensors study and record
efficiently the situation and current location of the user.
Activity sensors are supportive to users while they perform
registration of the work currently in progress or saving
activities that have already been completed. It principally
keeps tracks of the tasks the user is trying to solve. An
example is the DYONIPOS [39], which allows capturing of
activities that the learner undertakes and suggests appropriate
learning materials to the learner. Bio sensors on the other
hand, capture temperature, respiration, heart rate and so on.
An example is the accelerometer, a device which functions as
a compass. It detects motion and calculates the acceleration
based on this data. Environmental sensors use photo, audio
and proximity sensors to extract information about the users‟
surroundings. Proximity sensor as used in smartphones these
days turns off display of the phone device when the user is on
call [40]. These sensors remain very limited in terms of
usability and power consumption. With the large brands of
mobile devices available in the market today, not all of them
possess all the sensors required to perform an evaluation of the
cognitive processes of users.
The challenge is to maintain the use of different contexts
combined together to deliver accurate learning materials to the
learner [41-42][39]. User performance based on recorded data
obtained from their independent thinking, the contextual
information and their interaction with the learning system as
combined will bring m-learning to another level. It is
supported that users will eventually boost up their motivations
and maintains good learning performance [43].
IV. CONTEXT-AWARE MECHANISMS TO IMPROVE COGNITIVE
LOADS IN MOBILE LEARNING
Cognitive load is defined as the load generated with an
assigning work to an individual cognitive system. A higher
cognitive load brings a lower user satisfaction towards
learning [44]. Hence, it is significant to find ways to lower the
cognitive loads in teaching activities. Improper design of
learning elements results in an increase in cognitive burden of
students, causing them to overload their working memory
[45]. Researchers are still looking for answers in the new
proposed technology as to whether the learners are well
prepared to face the pressure and difficulties they are
displayed with. When employing new technologies to provide
richer content, factors that will affect the learning
performances of the learner need to be thoroughly investigated
[46].
Basically, cognitive load stems from three types that
directly affect learning which are intrinsic, extraneous and
germane. For learning to occur, the total cognitive burden
should never go beyond an individual‟s working memory
capability. Generally, the addition of extraneous and germane
cognitive load is believed to be equivalent to the cognitive
burden without considering the intrinsic cognitive burden.
Therefore, instructional designing‟s aim is to reduce the
extraneous load as the germane cognitive burden is increased
[48]. Elaborated hereunder are the three types of cognitive
load:
 Intrinsic cognitive load refers to the complexity that is
involved in learning materials. It is related to the
amount of information the working memory deals with
at a simultaneous time [49].
 Extraneous cognitive load consists of non-relevant,
unimportant materials causing learners to make
additional use of their mental processes. It can be
defined as the unimportant processing, caused by
instructional design [50].
Germane presents elements that enable learners to use
proper cognitive mental resources to learning process.
Moreover, it encourages the growth of a learner‟s knowledge
level. In other words, it is the load which allows learners‟
conscious focus of attention to intentionally recall and
comprehend the learning contents they come across [51].
The picture below depicts the cognitive load types where
the dashed lines represent the loads that can be manipulated by
purposeful instructional design [47].
Fig. 1. Depiction of Cognitive Load types [47].
Cognitive Load Theory (CLT) recommends the extraneous
load to be reduced by re-engineering learning activities when
the intrinsic complexity of a task stay fixed [52]. When
extraneous cognitive load is low, the released cognitive
resources can be re-assigned as germane aspect of automation
to balance the load and eventually ease learning [53].
The current situation of learning through a mobile device
causes negative impact on a user due to the high cognitive
load from improper design of contents delivered. Thus, it is
vital to center the problem of adaptation to the cognitive level
required for a particular context. Reference [54] came up with
a theory called “Split Attention principle” which proves that
people learn more when the information presented is in more
than one form. Learning material can be textual or pictorial,
animation or verbal presentation. In order to provide better
mobile learning experience, it is important to produce
instructional context to the learner, where he feels at ease with
the frequent change of the surrounding condition and
simultaneously be immersed in contents of the device.
Another example of instructional design principle based on
cognitive load theory was identified [55]. This is called the
goal free problem effect. In geometry problem-solving we
need to find a value for a particular angle from a diagram
whereas in goal-free problems students are required to find
values of as many angles as they could [56].
Reference [57] introduced the redundancy effect.
Redundant information increases the extraneous cognitive
load and forms part of the learner‟s working memory capacity.
It processes unnecessary information which is not even vital
for learning. Therefore, to support multimedia learning,
display of text in pictorial or animation format should be
presented as narration instead of static text.
Expertise reversal effect [58] can be described as the
phenomenon where instructional techniques are advantageous
to inexperienced learners and on the other side, it causes a
negative impact on experienced learners. The prime
recommendation for this is that instructional design methods
need adjustments as learners require more knowledge in a
specific field of study. The table below evaluates theories with
the respective cognitive load type.
TABLE I. EVALUATION OF COGNITIVE LOAD THEORY
Cognitive Load Type of cognitive load
Theory
Intrinsic load Extraneous load Germane load
Split Attention
Principle [54]
Constant High
Goal free Constant Low
problem effect
[55]
Redundancy Constant High
effect [57]
Expertise Constant High
Reversal effect
[58]
Building an adaptive educational system to adapt to
different context as previously described is not an easy task.
Contract-based Adaptive Software Architecture (CASA) [59]
is a framework which allows adaptive applications
developments. It dynamically adapts their functionality in
response to changes during execution time. Reference [60]
proposed the 5R adaptation concept based on location
information to allow learning at a suited time, in the right
environment, using the proper device while providing the
appropriate materials to the learner. Likewise, reference [61]
developed the Facilitation Activities Technology (FAT)
framework. It guides the design of contextual and location-
based learning. These frameworks are evaluated for their
situated characteristics of mobile learning, mobile application
design or contextual influences on learning contents. Similar
to the above described frameworks, many more have been
built to cater for adaptive learning materials.
Sensing technologies such as Global Positioning System
(GPS), Radio-Frequency Identification (RFID) and Quick
Response (QR) codes has enabled learning systems to focus
on and detect real-world locations and contexts of learners
[62]. However, those efforts are very limited, not exploiting
the full capabilities of the devices available, with narrow
content adaptation algorithms and little focus on the context
information [39]. Existing context-aware adaptive system also
use very limited context information [63]. For example,
JAPELAS [64] is a context-aware learning system equipped
with GPS which assists in practising Japanese in the real
world. It is supported by JEDY, an online dictionary which
considers the intrinsic context of the system while TANGO
[64] helps Japanese students to recognize English words using
mobile phones through RFID tag reader. It includes six
modules to select suitable English words based on learner
models. Learner models can be the learner‟s profile such as
name, gender, interests and knowledge level. English
vocabulary learning [65] makes use of WLAN positioning
technologies which help to locate the current position of the
learner. In this way, it encourages the student‟s interests and
allows for good performance. Additionally, collected data
Low such as the learner‟s location and the time available for
High
learning pre-select and adapt learning contents according to
the individual‟s preferences. Some researchers further
established DYONIPOS [40] as described in section III where
Low sensors are used to capture appropriate intrinsic contexts to
provide suitable learning materials to learners‟ preferences. Its
Low environment context captures the nature of the location the
user is currently in.
Mobile Learning Support System (MLSS) is a tool
designed by Reference [41] to ease students‟ learning in
outdoor settings. It supports multimedia learning through the
content learning function, the tags searching function and the
database connection function. With the MLSS, students scan
tags attached to corresponding objects and a variety of related
material is displayed on screen. It uses GPS and 2D barcode
technologies which assist students to get information by
interacting with the surroundings they are in. This system
enables the combination of real-world and digital resources. It
provides additional adaptive learning activities in a variety of
outdoor and indoor settings.
 TABLE II. EVALUATION OF CONTEXT-AWARE MECHANISMS

Context-aware Types of context

systems Internal External
DYONIPOS [40] Temperature, heart
rate,learners‟ Environment
behaviour
JAPELAS / GPS technologies for
Learners‟ knowledge
JEDY [64] location
Uses RFID tag
TANGO [64] Comprehensive level
reader/writer
English Individual abilities WLAN positioning
Vocabulary information such as technologies
Learning [65] knowledge and skills
MLSS [41] Allows combination of Uses GPS technologies
knowledge through 2D to detect location
barcode technologies
and content learning
Fig. 2. The system architecture of the MLSS [41]. function
AICHE [66] No intrinsic context Time, location, activity
Ambient Information Channels (AICHE) [66] system
CAMCLL[67] Learner levels on Time, location, activity
helps in developing context-aware e-learning applications. It previous knowledge
consists of an information channel which carries information
SAMCCO [68] Learning method Environment, Activity
that acts as a context-aware medium. This platform
concentrates on only extrinsic context information which is
time, location and activity. It does not make use of any As seen from the above table, internal context has not been
intrinsic context information such as the learner‟s motivation, exploited as such compared to external context.
his past knowledge or his attentiveness on the related topic. Environmental or location context is the most expanded in
Adaptation requires both the extrinsic and intrinsic context systems analysed. No system provided proofs of an evaluation
information to provide an appropriate learning degree. Relying made across the user adaptation with their cognitive processes
on only one type of information does not provide an effective on the materials delivered. Several attempts to improve mobile
mobile learning experience. learning experience have been made, however the problem of
content adaptation based on user context and cognitive load
requirements has not been solved [69]. Context-aware mobile
CAMCLL [67] is a context-aware m-learning system
learning is seen as a promising learning system but a fully
which supports communications in Chinese language. functional system has not been developed yet. Moreover,
Contexts of CAMCLL include time, environment, activities References [70-72] argued that one main challenge which we
performed and the past knowledge of the learner. The system must provide continuous support to is the generation of the
proposes appropriate sentences based on the learners‟ right learning contents that can be adapted to the needs of the
contexts. Besides, it is supportive to learners by providing user.
feedback with regards to the individual‟s previous knowledge.
Its adaptive engine is based on ontology and rule-based V. FUTURE RESEARCH DIRECTIONS
matching. In the near future, researchers will have to focus their
attention to the problem of cognitive load of users in mobile
Another context mobile learning system is the SAMCCO learning platforms. It is equally important to support the
[68], French abbreviation for contextual and collaborative growth of dynamic content adaptation while taking
mobile learning system for professional fields. The system is consideration in the fact that cognitive mechanism will
based on Electronic Performance Support System (EPSS), eventually ease learning. An evaluation to measure the
which groups storage of technical working and learning data, cognitive processes of users needs to be carried out and an
aiming for instant guidance, just enough information, tools attempt to design a system with personalization and contextual
and help to repair equipment, appliances or products learning needs to be done. Moreover, cognitive load remains
distributed in the smart city environment. It edits and the central concern since instructional design requires
organizes learning contents stored in the EPSS information cognitive theories for proper functioning. It will therefore
database, from which well-structured contents for learning are bring about changes in learners‟ behaviour for conducting
retrieved. Based on AM-LOM metadata, which is an ontology seamless learning activities.
used in the learning unit model, it describes and identifies
contextual mobile learning contents in a particular specialized VI. CONCLUSION
field. As seen in existing literatures, mobile learning has
emerged into a flourishing perspective during the past decade.
Users can now take control over time and location to learn.
Though, m-learning appears to be thriving, it still has a long
way to go. Newly geared technologies which are constantly
being integrated among the existing tools will help mobile
learning to mature. This paper demonstrated how personalised
learning while considering contextual information is not yet
supported, instead every learner are presented with static
identical contents based on contextual information. Devices
are not ready content-wise. Richer contents display creates
cognitive strain on users. As identified, cognitive load plays an
important role in delivering materials in m-learning. We can
derive from cognitive levels, the appreciation and
achievements of a learner. Some initial existing systems that
attempt to customise learning materials based on user contexts
has been surveyed and analysed in this paper.
REFERENCES
[1] G. Duman, G. Orhon, and N. Gedik.. “Research trends in mobile
assisted language learning from 2000 to 2012”. ReCALL, 27, 2015, pp
197-216.
[2] J. Yau, and M. Joy, “A context-aware and adaptive learning schedule
framework for supporting learners‟ daily routines”. Department of
Computer Science, University of Warwick, UK, 2007.
[3] I. J. CHEN, C. C CHANG, and J. C, YEN. “Effects of presentation
mode on mobile language learning: A performance efficiency
perspective”. Australasian Journal of Educational Technology, 28(1),
2012, pp.122-137.
[4] L. Motiwalla, “Mobile learning: A framework and evaluation”. Journal
of Computer and Education, 49, 2007, pp.581-596.
[5] B. Patten, I. Arnedillo Sanchez, I., and B. Tangney. “Designing
collaborative, constructionist and contextual applications for handheld
devices”. Computers & Education, 46(3), 2006, pp.294-308.
[6] P. L. P. Rau, Q. Gao, and L. M. Wu. “Using mobile communication
technology in high school education: Motivation, pressure, and learning
performance”. Computers & Education, 55(1), 2008, pp.1-22.
[7] H. Ryu and D. Parsons. “Risky business or sharing the load? Social flow
in collaborative mobile learning”. Computers & Education, 58, 2012,
pp.707-720.
[8] A. Efklides. “Metacognition. Defining its facets and levels of
functioning in relation to self-regulation and coregulation”. European
Psychologist, 13, 2008, pp.277-287.
[9] P. Salonen, M. Vauras, and A. Efklides. “Social interaction-What can it
tell us about metacognitive and coregulation in learning?”, European
Psychologist, 10, 2005, pp.199-208.
[10] W. Wu, Y. J. Wu, C. Chen, H. Kao, C. Lin, and S. Huang, “Review of
trends from mobile learning studies: A meta-analysis”. Computers &
Education, 59(2), 2012, pp.817-827.
[11] G. D. Chen, C. K. Chang and C. Y. Wang. “Ubiquitous learning
website: Scaffold learners by mobile devices with information-aware
techniques”. Computers & Education, 50(1), 2008, pp.77–79.
[12] G. J. Hwang and C. C. Tsai. “Research trend in mobile and ubiquitous
learning”. A review of publications in selected journal from 2001 to
2010. British Journal of Education Technology, 42(4), 2011, pp.E65–
E70
[13] E. Soykan, and H. Uzunboylu. “New trends on mobile learning area:
The review of published articles on mobile learning in science direct
database”. World Journal on Educational Technology, 7(1), 2015, pp.
31-41
[14] R. V. Kranenberg. “The Internet Things: A Critique of Ambient
Technology and the All-Seeing Network of RFID”. Amsterdam, The
Institute of Network Cultures, 2008, pp. 12-16.
[15] J. Lam, J. Yau, and S. Cheug. “A review of mobile learning in the
mobile age”. In: Proc. 3rd International Conference on Hybrid Learning,
Lecture Notes in Computer Science Springer, Berlin, 6248, 2010, pp.
306–315
[16] A. Mostakhdemin-Hosseini, and J. Tuimala. “Mobile Learning
Framework”. Proceedings IADIS International Conference Mobile
Learning 2005, Malta, 2005, pp. 203-207.
[17] C. Shuler, N. Winters and M. West. “The Future of Mobile Learning:
Implications For Policy Makers And Planners”. United Nations
Educational, Scientific and Cultural Organization (UNESCO), 2013, pp.
7-35.
[18] Cisco. The Mobile Learning Phenomenon in Education: Accelerating the
delivery of personalized learning, 2013, pp. 3-34.
[19] Q. Liu, X. Jian, J. Hu, H. Zhao, and S.Zhang. “An optimized solution for
mobile environment using mobile cloud computing”. InWireless
Communications, Networking and Mobile Computing, 2009.
WiCom'09. 5th International Conference, IEEE, 2009, pp. 1-5
[20] A. R. Khan, M. Othman, S. A. Madani, and S. U. Khan. “A survey of
mobile cloud computing application models”. Communications Surveys
& Tutorials, IEEE, 16(1), 2014, pp.393-413.
[21] F. Muhammad, T. Kechadi, and N. An Le-Khac. The State of the Art
Forensic Techniques in Mobile Cloud Environment: A Survey,
Challenges and Current Trends. IJDCF, 7.2, 2015, pp.1-19.
[22] H.T. Dinh, C. Lee, D. Niyato, and P. Wang. “A survey of mobile cloud
computing: architecture, applications, and approaches”. Wireless
communications and mobile computing, 13(18), 2013, pp.1587-1611.
[23] IBIS Capital. Global e-Learning Investment Review, January 2013, pp.
5-100.
[24] C. Shuler, N. Winters, and M. West. “The Future of Mobile Learning:
Implications For Policy Makers And Planners”. United Nations
Educational, Scientific and Cultural Organization (UNESCO), 2013, pp.
7-35.
[25] M. O. M, El-Hussein, and J. C. Cronje. “Defining Mobile Learning in
the Higher Education Landscape”. Educational Technology & Society,
13 (3), 2010, pp.12–21.
[26] Y.-L, Jeng, T.-T. Wu, Y.-M.Huang, Q. Tan, S. J. H. Yang. “The add-on
impact of mobile applications in learning strategies: a review study”.
Educational Technology & Society, 13 (3), 2010, pp.3–11.
[27] P. Pollara, and K. Broussard. “Student perceptions of mobile learning: A
review of current research”. M. Koehler & P. Mishra (Eds.), Proceedings
of Society for Information Technology & Teacher Education
International Conference, 2011, pp. 1643–1650.
[28] Wikipedia, Emantras, 2015. Available from:
https://en.wikipedia.org/wiki/Emantras
[29] H. Singh. E-learning queen, March 21, 2013. Available from:
http://elearnqueen.blogspot.com/2013/03/interview-with-harman-singh-
wiziq.html
[30] H.E., Byun, and K., Cheverst. “Utilizing context history to provide
dynamic adaptations”. Journal of Applied Artificial Intelligence, 18 (6),
2004, pp. 533–548.
[31] A., Dey, G., Abowd, and D., Salber, „A conceptual framework and a
toolkit for supporting the rapid prototyping of context-aware
applications‟, Human-Computer Interaction 16(2-4), 2001, pp. 97–166.
[32] P., Prekop, and M., Burnett. „Activities, context and ubiquitous
computing‟. Computer Communications, 26, (11), 2003, pp. 1168-1176.
[33] R. M., Gustavsen. „Condor – an application framework for mobility-
based context-aware applications‟, Proceedings of the Workshop on
Concepts and Models for Ubiquitous Computing, Goeteborg, Sweden,
2002.
[34] T., Hofer, W., Schwinger, M., Pichler, G., Leonhartsberger, and J.,
Altmann. „Context-awareness on mobile devices – the hydrogen
approach‟, Proceedings of the 36th Annual Hawaii International
Conference on System Sciences, 2002, pp.292–302.
[35] M., Baldauf, S., Dustdar, and F., Rosenberg. A survey on context-aware
systems. Int. J. Ad Hoc and Ubiquitous Computing. 2(4), 2007, pp. 263-
277.
[36] A. A., Economides. „Adaptive context-aware pervasive and ubiquitous
learning‟, International Journal of Technology Enhanced Learning, 1(3),
2009, pp. 169-192.
[37] A. A., Economides. „Requirements of mobile learning applications‟,
International Journal of Innovation and Learning, 5(5), 2008, pp. 457-
479.
[38] A., Zafar, and H. S., Hasan. “Towards Contextual Mobile Learning”.
International Journal of Modern Education and Computer Science
(IJMECS), 6(12), 2014, pp. 20-25.
[39] S. A., Rath, N., Weber., M., Kroll, M., Granitzer, O., Dietzel, and N. S.,
Lindstaedt. “Context-Aware Knowledge Services”. Workshop on
Personal Information Management, CHI '08, Florence, Italy, 2008.
[40] D. N., Lane, E. Miluzzo, H. Lu, D. Peebles, T. Choudhury, and T. A.,
Campbell. A survey of mobile phone sensing, IEEE Communications
Magazine, 48 (9), 2010, pp. 140-150.
[41] K.-Y. Chin, and Y.-L. Chen. “A Mobile Learning Support System for
Ubiquitous Learning Environments”. Procedia – Social and Behavioral
Sciences, 73, 2013, pp.14-21.
[42] B. Zhang, B. David., C. Yin, and R. Chalon. Contextual Mobile
Learning for Professionals working in the “Smart City”. Interaction
Design and Architecture(s) Journal – IxD&A, November, 17, 2013, pp.
67-76.
[43] H. C., Chu, G. J., Hwang, and C. C., Tsai. A Knowledge Engineering
Approach to Developing Mindtools for Context-Aware Ubiquitous
Learning. Computers & Education, 54(1), 2010, pp. 289-297.
[44] N. Segall, T. L., Doolen, and J. D., Porter. A usability comparison of
PDA-based quizzes and paper-and-pencil quizzes. Computers &
Education, 45(4), 2005, pp. 417-432.
[45] H.-C., Chu. Potential Negative Effects of Mobile Learning on Students‟
Learning Achievements and Cognitive Load – A format Assessment
Perspective. Educational Technology & Society, 17(1), 2014, pp. 332-
344.
[46] J.-L., Shih, C.-W., Chuang, and G.-J., Hwang. “An Inquiry-based
Mobile Learning Approach to Enhancing Social Science Learning
Effectiveness”. Educational Technology & Society, 13 (4), 2010, pp.
50–62.
[47] Wang, C. X. Victor. Advances in Educational Technologies and
Instructional Design, 2014.
[48] P. W. M. Van Gerven, Paas, F. Paas, and H. K., Tabbers.Cognitive aging
and computer-based instructional design: Where do we go from here ?
Educational Psychology Review, 18(2), 2006, pp. 141-157.
[49] G. J., Hwang, P. H., Wu, Y.Y Zhuang, and Y. M., Huang. Effects of the
inquiry-based mobile learning model on the cognitive load and learning
achievement of students, Interactive Learning Environments, 21(4),
2013, pp.338–354.
[50] G., Cierniak, K., Scheiter, and P., Gerjets. Explaining the split-attention
effect: Is the reduction of extraneous cognitive load accompanied by an
increase in germane cognitive load? Computers in Human Behavior, 25,
2009, pp. 315–324.
[51] J. Sweller, J. Van Merriënboer, and F. Paas, F. "Cognitive architecture
and instructional design". Educational Psychology Review 10 (3), 1998,
pp. 251–296.
[52] J. Sweller, P. Ayres, and S. Kalyuga. Cognitive load theory. New York:
Springer 2011.
[53] F. Paas, A. Renkl, and J. Sweller. Cognitive load theory: Instructional
implications of the interaction between information structures and
cognitive architecture. Instructional Science, 32, 2004, pp. 1–8.
[54] R. E., Mayer, and R. Moreno. A Split-Attention Effect in Multimedia
Learning: Evidence for Dual Processing Systems in Working Memory.
Journal of Educational Psychology, 90 (2), 1998, pp. 312-320.
[55] J. Sweller, and G. A, Cooper. The use of worked examples as a
substitute for problem solving in learning algebra. Cognition and
Instruction, 2, 1985, pp. 59-89.
[56] F. Paas and F. Kirschner. The goal-free effect. In N. M. Seel (Eds.),
Encyclopedia of the Sciences of Learning (2), 2012, pp. 1375-1377.
Netherlands: Springer.
[57] P. Chandler, and J. Sweller. Cognitive load theory and the format of
instruction. Cognition and Instruction, 8, 1991, pp.293-332.
[58] S. Kalyuga, P. Ayres, P. Chandler, and J. Sweller. The Expertise
Reversal Effect. Educational Psychologist, 38(1), 2003, pp. 23-31.
[59] A. M., Mukhija, and Glinz. A Framework for Dynamically Adaptive
Applications in a Self-organized Mobile Network Environment.
Proceedings 4th International Workshop on Distributed Auto-adaptive
and Reconfigurable Systems (DARES 2004), Tokyo, 2004, pp. 368-374.
[60] Q. Tan, X. Zhang, Kinshuk and R. McGreal. The 5R adaptation
framework for location-based mobile learning systems. 10th World
Conference on Mobile and Context Learning (M-Learn 2011), October,
2011, pp.18-21, Beijing, China.
[61] E. Tan. and H.-J, So. Location-based collaborative learning at a
Geography trail: Examining the relationship among task design,
facilitation and discourse types. In H. Spada, G. Stahl, N. Miyake, & N.
Law. (Eds), In the proceedings of CSCL 2011 conference: Connecting
computer-supported collaborative learning to policy and practice.
International Society of the Learning Sciences, 2011, pp. 41-48.
[62] G.-J., Hwang. Definition, framework and research issues of smart
learning environments – a context-aware ubiquitous learning
perspective. Smart Learning Environments, 2014, 1(4).
[63] P. Zervas, S. Go´mez, D. Sampson, R. Fabregat. Delivering adaptive and
context-aware educational scenarios via mobile devices. 12th IEEE
International Conference on Advanced Learning Technologies (ICALT),
2012, pp. 197–201.
[64] H. Ogata, and Y. Yano. Context-Aware Support for Computer-
Supported Ubiquitous Learning. Proceedings of the 2nd IEEE
International Workshop on Wireless and Mobile Technologies in
Education, 2004, pp. 27-34.
[65] C.-M., Chen, Y.-L., Hsieh, S.-H., Hsu, Mining learner profile utilizing
association rule for web-based learning diagnosis, Expert Systems with
Applications: An International Journal, 33(1), pp.6-22, July, 2007.
[66] M. Specht, S., Ternier, and W., Greller. Mobile augmented reality for
learning: A case study. Journal of the Research Center for Educational
Technology, 7(1), 2011, pp. 117–127.
[67] K. Al-Mekhlafi, X., Hu, and Z. Zheng. An Approach to Context-Aware
Mobile Chinese Language Learning for Foreign Students. Eighth
International Conference on Mobile Business, Washington, DC, USA,
2009.
[68] C. Yin. SAMCCO: a System of Contextual and Collaborative Mobile
Learning in the Professional Situations. Thesis of Ecole Centrale de
Lyon, Computer Science Department, 25 January 2010.
[69] T.-C., Yang, G.-J., Hwang, and S. J.-H., YANG. Development of an
adaptive learning system with multiple perspectives based on students'
learning styles and cognitive styles. Educational Technology & Society,
16 (4), 2013, pp.185–200.
[70] Y., Akbulut, and C. S., Cardak. Adaptive educational hypermedia
accommodating learning styles: A content analysis of publications from
2000 to 2011. Computers & Education, 58(2), 2012, pp.835-842.
[71] P., Cristea, and C., Stewart. Automatic authoring of adaptive educational
hypermedia. In Z. Ma (Ed.), Web-based intelligent e-learning systems:
Technologies and applications, 2006, pp. 24-55.
[72] J. G. K., Foss, A. I., Cristea, and M. Hendrix. Continuous use of
authoring for adaptive educational hypermedia: A long-term case study.
Paper presented at the 10th IEEE International Conference on Advanced
Learning Technologies, ICALT 2010. Available from
http://dl.acm.org/citation.cfm?id=1846121.
 Outputs$nalysisand'ual,mmersion
0ethodfor&haotic6ystems
K.A.A. Langueh, O. Datcu, J-P Barbot, G. Zheng and K. Busawon ∗
$EVWUDFW²When secure data transmission is
implemented through chaotic systems, the choice of the
output is a preliminary problem. In this paper, the
quality of the transmitted information is analysed with
respect to the observability concept, for each potential
output. More-over, in order to overcome observability
loss, a dual immersion technique is proposed.The use of
high order sliding mode observer on a well known Lorenz
system, allows to highlight the well founded of the
proposed analysis and method.
 Keywords—C Chaotic systems; Observability;
Singularity; Immersion, Secure data transmission
1. Introduction
Since the works of Pecora and Caroll [19], it is well
known that two chaotic systems can be synchronized.
Following this fact, many authors have proposed secure
data transmission schemes (see for example [5, 1, 17])
based on the synchronization of chaotic systems. Nev-
ertheless, the output choice with respect to the synchro-
nization, at the best of our knowledge, is less analyzed
(see for example [13, 9]). In [18] Marels and Nijmei-
jer draw the link between unidirectional synchroniza-
tion and observation. This allows to use the control
system theory to analyze the synchronization of chaotic
systems. Nevertheless, as chaotic systems are always
nonlinear systems, linear control system theory can’t be
directly applied to such systems. This is due to the fact,
for example, basic properties as stability, controllability
and observability are generally local for nonlinear sys-
tems. It is the reason why in [11] local observability
concepts and criteria are introduced. These lead to de-
termine an observability singularity set [9]. This set is
∗ K.A.A Langueh is with QUARTZ Laboratory, IPGP, ENSEA,
6 Avenue du Ponceau, 95014 Cergy-Pontoise, O. Datcu is with Po-
litehnica University of Bucharest, Faculty of Electronics, Telecom-
munications, and Information Technology and QUARTZ Laboratory,
IPGP, ENSEA, 6 Avenue du Ponceau, 95014 Cergy-Pontoise, J-P.
Barbot is with QUARTZ Laboratory, IPGP, ENSEA, 6 Avenue du
Ponceau, 95014 Cergy-Pontoise, EPI Non-A, INRIA. G. Zheng is
with EPI Non-A INRIA, Lille, Nord-Europe and K. Busawon is with
Northumbria University, Newcastle, UK.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
only constituted of ‘forbidden’ states (i.e. state which
are unobservable). In this paper, starting from the re-
sults introduced in [11] and other recent works [13, 9]
an analysis of the “best” output for secure data trans-
mission is realized on the basic Lorenz circuit. It is im-
portant to mention that the observability analysis with
respect to the output choice is only a preliminary step
in the design of data secure transmission scheme. Af-
ter that, many other problems occur, such as the choice
of the ciphering method, the input choice,the method
for retrieving information (see for example [6, 12]),...
Moreover, when an output is chosen, it may exist an ob-
servability singularity set, which leads to lose informa-
tion in some part of the state space. Hereafter it is pro-
posed a solution, based on the immersion technique, to
overcome this problem under very weak conditions. It
is important to mention that, since the work of [22, 21],
the immersion technique was extensively used in the ob-
server design context. Nevertheless, this is usually used
to recover the linearity property by diffeomorphism and
output injection [3, 4, 24, 23], but generally with only
a local diffeomorphism. In the majority of the men-
tioned papers, immersion was accomplished by adding
a dynamic by means of output integration [3]. The sta-
bility of such extra dynamics can be problematic and an
elegant solution is proposed in [23]. Nevertheless the
problem of observability singularity was not tackled. A
dual immersion technique using only extra differentia-
tions is proposed in this paper. This method is close to
the one proposed in [2], in another context. Moreover,
for stability arguments it is chosen, in this paper, to im-
pose exponentially stable dynamics instead of constant
dynamics. The proposed approach is feasible thanks
to the finite time differentiators as for example the one
proposed in [7] (but for this method, the delay appears
due to the data acquisition frame) or High Order Sliding
Mode (HOSM) [15, 8].
The paper is organized as follows: in the next sec-
tion some observability concepts, symbolic observabil-
ity index and observability singularity definitions are
recalled and the problem statement is explained in the
framework of Lorenz system. In section III, the dual
immersion method is presented. After that, some re-
calls on HOSM differentiator are given in section IV. In
section V, simulation examples, with respect to Lorenz with ⎛ ⎞
system, are given. The paper ends with some conclu- dh
⎜ .. ⎟
sions and perspectives. dO(n)|x = ⎝ . ⎠ (5)
0
dLn−1
f h |x0
2. Some recalls and problem statement
Since dO(n)|x is only of full rank n around x0 , this
0
Consider the following autonomous system: implies that there might exist some x¯ ∈ Rn such that
rank dO(n)|x¯ < n. Due to this fact, let us define the fol-
ẋ = f (x) lowing observability singularity set:
y = h(x) (1)
Sn = {x ∈ Rn : rank {dO(n)|x } < n} (6)
where x ∈ Rn is the state, and y ∈ R is the output. f and
In this case, for the purpose of removing the singulari-
h are supposed to be C∞ vector fields. Rougthly speak- ties in O(n)|x defined in (6), it is necessary to increase
ing, in [11] in order to determine if or not the system (1)
the dimension of (5) by involving more derivatives of
is observable (weakly locally observable) Hermann and the output. This will be the purpose of the next section,
Krener have introduced the following condition: but before, an example of the output choice is given
hereafter.
Proposition 1 The system (1) is weakly locally observ-
able if Example 1 Let us consider the Lorenz system
Rank{dO(∞)} = n (2)
ẋ1 = σ (x2 − x1 )
where ⎛ ⎞ ẋ2 = rx1 − x2 − x1 x3 (7)
dh
⎜ .. ⎟ ẋ3 = −bx3 + x1x2
⎜ ⎟
⎜ . ⎟
⎜ dLn−1 h ⎟ Where x = [x1 , x2 , x3 ]T ∈ R3 is the state. Moreover, the
⎜ f ⎟
dO(∞)|x =⎜ ⎟ (3) parameter are chosen as follow σ = 10, r = 28 and
⎜ .. ⎟
0
⎜ . ⎟ b = 38 .
⎜ ⎟
⎜ dLn+k h ⎟
⎝ f ⎠ Case 1: The output y = x1
..
. | For this output dO(3)|x is equal to
x0
⎛ ⎞
∂ Li−1 h
1 0 0
In (3) Lif h = ∂f x f is the usual Lie derivative and dO(3)|x = ⎝ −σ σ 0 ⎠
 i
∂ L f h ∂ Lif h ∂ Lif h σ (r + σ − x3 ) −σ (σ + 1) −σ x1 |
dL f h = ∂ x , ∂ x , ..., ∂ xn is the associated 1-form.
i x
(8)
1 2
and then
Remark 1 If the system (1) is linear, then the weakly S3 = {x ∈ R3 : x1 = 0}
locally observability implies the classical global ob- On the figure 1 it is shown that the set of observabil-
servability. Moreover, due to the Cayley Hamilton theo- ity singularity intersect the strange attractor. In fact the
rem, it is sufficient to compute the n first rows of dO(∞) observability property is lost because only two deriva-
(denoted dO(n)). tives of the output y = x1 are considered. Neverthe-
less, if the third derivative is also considered dO(4)|x
Starting, from this linear argument and the fact that it becomes:
is natural to design an observer of the same dimension ⎛ ⎞
1 0 0
that the original system in [20] the concept of regularly ⎜ − σ σ 0 ⎟
dO(4)|x = ⎜
⎝ σ (r + σ − x3 )
⎟
weakly locally observability was introduced. Roughly −σ (σ + 1) −σ x1 ⎠
speaking, only the n − 1 output derivatives are consid- A B C |
x
ered for recovering all the state. This leads to the fol- (9)
lowing condition: where A = −σ [σ 2 + rσ (2σ + 1) + 2x1x2 − (2σ + b +
1)x3 ], B = σ [σ 2 + σ (r + 1) + 1 − x21 − σ x3 ] and C =
Proposition 2 The system (1) is regularly weakly lo- σ [(2σ + b + 1)x1 − σ x2 ]. Then, the observability singu-
cally observable if larity set is:

Rank{dO(n)} = n (4) S4 = {x ∈ R3 : x1 = x2 = 0}
 from the following injective function:
50
T
40
O(4) = y, ẏ, ÿ, y(3) (10)
30

20

10 Remark 2 By extension O(n + k) is defined as follow:

0
30

T
O(n + k) = y, ẏ, ÿ, ..., y(n+k)
20

10 (11)
0

−10 20

Setting z = O(4), the extended dynamics becomes

15
10
−20 5
0
−5
−10
−30 −15
−20
⎛ ⎞
Figure 1. Lorenz attractor and the map S3 z2
∂ O(4) ⎜ z3 ⎟
ż = f (x)|x = ⎜
⎝
⎟
⎠
∂x z4
L3f h(x)

for which a traditional high-gain observer [10] (or slid-

ing mode observer [15]) can be easily designed to esti-
50
mate z, noted as ẑ. Then, one can obtain the estimation
40

30
of x, denoted x̂ by:
−40
20

x̂ = arg min||ẑ − φ (x̂)||

10
0 x̂∈Rn
20 15
10 5 0
Figure 2. Laurenz attractor and the line S4
this set does not intersect the strange attractor, see fig-
ure 2. Moreover, it is important to note that on S4 , the
system (7) is stable and consequently its behaviour is
not chaotic.
Case 2: The output y = x2
In the same way as the case 1, it is obtained
S3 = {x ∈ R3 : 2x2 x21 − brx1 + rσ x2 − σ x2 x3 = 0}
Case 3: The output y = x3
Finally, for the third case, it is obtained
S3 = {x ∈ R3 : 2σ x22 − 2rx21 + 2 ∗ x21x3 = 0}
Comparing the set of observability singularities, for the
sake of simplicity and space in the next only the firs case
is considered.
The previous example, case 1, has shown that, even
if the O(3) contains singularities, i.e. its rank is not
equal to 3 for some x ∈ Rn , it is still possible to obtain
a higher dimensional map: O(4) which will not con-
tain singularity on the strange attractor. Thus it may be
interesting to design an observer of greater dimension
than 3 which is equivalent to use a state space represen-
tation greater than 3. More precisely, (9) was obtained
−20
0
20
−5 −10 40
−15 −20
which can be seen as an optimization problem. For
the above optimization framework, the essential point is
that the Jacobian of O(4) is a matrix of 4 by 3, thus non
square, and this will cause difficulties when applying
numerical methods to solve this optimization problem.
Motivated by this consideration, for a given O(n +
k) defined in (11), hereafter, it is proposed to find a dif-
feomorphism φ by increasing the original state space.
In the literature, this technique is called immersion and
many authors use this immersion technique in order to
obtain a specific normal form [23]. Most often, the im-
mersion is realized by means of output integration. In
this paper, it will be realized by a dual method which
will be presented in the next section.
3. Dual immersion technique
3.1. Preliminary results
For a given dO(n) containing the singularities be-
longing to Sn , this section will propose a constructive
way to deduce an injective function O(n + k) defined in
(11) by increasing the output derivative order. Then the
following result is trivial.
Lemma 1 Consider the observability singularity set
Sn defined in (6), then the generated set Sn+i for i ≥ 1
via differentiation is always included in Sn+i−1 , i.e.
Sn+i ⊆ Sn+i−1
Proof. For i ≥ 0, the following equation z = O(n + k) is a global injective function. Then let

us present the following procedure to compute a global
O(n + i − 1)
O(n + i) =
Ln+i
f
1
h will be defined hereafter).
is verified, hence
rank{dO(n + i)} ≥ rank{dO(n + i − 1)}
which implies that
Sn+i ⊆ Sn+i−1
Before presenting the dual immersion technique,
the following assuption is reclame
Assumption 1 It is assumed that system (1) is globally
observable for x ∈ Rn , or at least observable in the do-
main of interest D, i.e. there exists a least an integer
k ∈ Z + such that
rank{dO(n + k)} = n (12)
for all x ∈Rn , or at least for all x ∈ D, where O(n + k)
is defined in (11).
With the above assumption, we can state the fol-
lowing result.
Lemma 2 For system (1), if Assumption 1 is satisfied,
then there exists at least an integer k ∈ Z + such that
the following inclusion is satisfied:
0/ = Sn+k ⊆ · · · ⊆ Sn+1 ⊆ Sn
on Rn , or at least in D. Moreover, z = O(n + k) is such
that rank{dO(n + k)} = n, on Rn , or at least in D.
Proof. The above result is a direct consequence of the
previous lemma.
The above results show that one can reduce the ob-
servability singularity set by just increasing the number
of the derivative of the output. Consequently, one can
find at least an integer k ∈ Z + such that z = O(n + k)
has a Jacobian of rank n on Rn , or at least in D. How-
ever, its Jacobian is not square ( a matrix of (n + k) × n).
In what follows, we will propose a constructive way to
calculate a new diffeomorphism φ (x, ξ ) from the de-
duced O(n + k) in Lemma 2 via the technique of dual
immersion, whose Jacobian d φ (x) will be square and
invertible, at least on D.
3.2. Dual immersion approach
Consider again system (1), and suppose that we
have already found a least integer k ∈ Z + such that
(or at least in D) diffeomorphism φ (x, ξ ) (where ξ ∈ Rk
Dual immersion algorithm
1. Initialization: Set z1 = φ1 (x) = y ;
2. Compute z j = φ j (x) = y( j−1) until a singularity of
( j−1)
observability appears in the row dL f h;
3. Define φ j = y( j−1) + ξ1 where ξ1 is an additional
state, with the following dynamics:
ξ̇1 = −ε1 ξ1 (13)
initialized at ξ1 = 0.
( j+1)
4. • If L f h is not considered in (11), define
( j)
φi, j+1 = L f h − ε1 ξ1
( j+1)
• If L f h is considered in (11), define
( j)
φi, j+1 = L f h − ε1 ξ1 + ξ2 , with
ξ̇2 = −ε2 ξ2 + ε̇1 ξ1 − ε12 ξ1
5. repeat the operation until the last line of (11) is
reached. The resulting matrix d{φ } becomes a
square matrix and if there exist ε j such that the
matrix is regular for all x then the algorithm 3.2
converges.
This algorithm gives
Theorem 1 For system (1), if Assumption 1 is satisfied
and Algorithm 3.2 converges then φ (x, ξ ) with ξ ∈ Rk
is a global, or at least on D, diffeomorphism. More-
over, the Jacobian of the generated φ (x, ξ ) via the dual
immersion satisfies:
rankd φ (x, ξ ) = n + k
for all x ∈ Rn , or at least in D.
Proof. The proof follows from the algorithm, and as-
sumption 1 is a necessary condition for the convergence
of algorithm 3.2.
Roughly speaking, the dual immersion algorithm
works as follow: considering that for the order r = n+1,
O(r) generates an ”acceptable” (see [9] for chaotic sys-
tems) observability singularity set Sr and the singular-
ity problem arrives in the last line of O(n), it is pos-
sible to define a new change of coordinate z = φ (x, ξ )
as follows: z j = y( j−1) until j = n − 1 (i.e. until such
a problem of observability singularity appears in the
 (n−1)
row dL f h). and for j = n, z j are assigned equal to 5. Lorenz example and immersion.
y( j−1) + ξ1 where ξ1 is an extra state, which verifies the
following dynamics: Consider again the system (7) with y = x1 , with ini-
tial conditions x1 (0) = 0, x2 (0) = 0 and x3 (0) = 0. It
ξ̇1 = −ε1 ξ1 (14) was shown, in the second section that S4 = {x ∈ R3 :
x1 = x2 = 0} and that this set of observability singular-
initialized at ξ1 = 0. Finaly for j = n+1, z j are assigned ity is not on the strange attractor, thus as D (the consid-
equal to y( j) + ξ̇1 = żn . ered domain) is the strange attractor, then it is possible
to use the dual immersion method.
Remark 3 From the stability and the initialization of Then O(n + k)
the dynamics (14) it is not necessary to simulate this
dynamic in the observer design because it is equivalent ...
φ (x) = [y, ẏ, ÿ, y ]T
of adding zero at y( j−1) . In [2] the proposed solution
consists to adding a constant in order to be able to de- has a regular Jacobian on D. So, the immersion is im-
sign a High gain observer in the original coordinates. plemented as follows:
⎛ ⎞
y
4. Recalls on high-order sliding-mode ⎜ ẏ ⎟
z = φ (x, ξ ) = ⎜
⎝ ÿ + ξ ⎠
⎟
...
In this paper, the proposed method is based on the y + ξ̇
so-called real-time exact robust HOSM differentiator
[15, 16], which is recalled in the following. with ξ̇ = −εξ and ξ (0) = 0 as initial condition. Then
Consider a signal y(t) ∈ Ck (at least k times dif- d φ (x, ξ ) is a square matrix:
ferentiable), let us suppose (y, · · · , y(k) ) = (z1 , · · · , zk+1 ). ⎛ ⎞
1 0 0 0
The HOSM robust differentiator proposed in [15] takes ⎜ −σ σ 0 0 ⎟
⎜
d φ (x, ξ ) = ⎝ ⎟
the following form: σ (r + σ − x3 ) −σ (σ + 1) −σ x1 1 ⎠
A B C −ε
1 k
ẑ˙1 = −λ0 M k |ẑ1 − y| k+1 sign(ẑ1 − y) + ẑ2
1 k−1
with A = −σ [σ 2 + rσ (2σ + 1) + 2x1 x2 − (2σ +
ẑ˙2 = −λ1 M k−1 |ẑ2 − v1| k sign(ẑ2 − v1 ) + ẑ3 b + 1)x3 ], B = σ [σ 2 + σ (r + 1) + 1 − x21 − σ x3 ] and
.. C = σ [(2σ + b + 1)x1 − σ x2 ].
. The ε parameter is chosen such that d φ (x, ξ ) = 0
1 1
˙ẑk = −λk−1 M 2 |ẑk − vk−1| 2 sign(ẑk − vk−1 ) + ẑk+1 (i.e. ε x1 = (2σ + b + 1)x1 − σ x2 ) on D. Thus, as
ẑ˙k+1 = −λk Msign(ẑk+1 − vk ) no derivative of ε is considered in this exemple, this
parameter is chosen as follows:
where M is chosen to be bigger than the k-th deriva- ε is equal to 104 if det{d φ¯(x, ε )|ε =104 } = 0 else ε = 10,
tive of y(t), λi are positive design parameters, and the this ensuring that det{d φ¯(x, ε ) is equal to zero only if
adjustment of those parameters is described in detail in x1 = x2 = 0.
[15]. Define the observation errors as: ei = zi − ẑi , thus
the observation errors dynamics is given by: As the HOSM observer gives ẑ = z in finite time,
the estimated of x is also obtain in finite time as follows:
e1 = ẑ1 − y
x = φ −1 (z)|ε =
e2 = ė1 = λ0 M |e1 |
1
k
k
k+1 sign(e1 ) ⎡ ⎤
z1
.. ⎢ σ z1+z2 ⎥
. ⎢  σ

−1 
⎥
⎣ x3 −σ y 1 L1 ⎦
=
1
ėk−1 = λk−1 M 2 |ek−1 | 2 sign(ek−1 )
1
=
ek ξ σ [(2σ + b + 1)y − σ x2 ] −ε L2
ek+1 = ėk = λk Msign(ek ) ...
with L1 = ÿ + (σ + 1)ẏ − σ (r − 1)y and L2 = y − [σ 2 +
It has been proven in [14] that there exists t0 such that σ (r − 1) + 1]ẏ + [σ 2 (r − σ − 1) + σ r]y + (σ y + ẏ)y2
∀t > t0 we have
The simulation was done with initial conditions
ei = zi − ẑi = 0 for 1 ≤ i ≤ k + 1 chosen randomly in [0, 1] for the system and the ob-
server, moreover on the HOSM observer the paramer M
In the next section, a Lorenz example case 1 will is equal to 106 . In Fig. 3, the exact values of is obtained
be simulated and commented. after 1s.
 60
50
40
30
20
10
0
−10
0 2 4 6 8 10 12 14 16 18 20
Figure 3. The stat x3 and its estimate x̂3
6. Conclusion
In this paper an analysis of the observability, with
respect to different outputs, was done. After that, a dual
immersion method was proposed, in order to overcome
some problems due to observability singularity set.
This method is based on an immersion with a stable
fictitious dynamics initialized at the equilibrium point.
The inversion of the obtained diffeomorphism is not,
generally, as simple as that in the Lorenz system with
x1 as output. In a future work, a numerical solution
of the inverse diffeomorphism will be investigated
by taking into account the particular structure of the
obtained diffeomorphism.
Acknowledgements: The work has been funded
by the Sectoral Operational Programme Human Re-
sources Development 2007-2013 of the Ministry of Eu-
ropean Funds through the Financial Agreement POS-
DRU/159/1.5/S/132395.
References
[1] G. Alvarez, F. Montoya, M. Romera, and G. Pastor,
“Cryptanalysis of an ergodic chaotic cipher,” Physics
Letters A, vol. 311, pp. 172–179, 2003.
[2] V. Andrieu and L. Praly, “Note on diffeomorphism ex-
tension for observer design,” hal-00957648, 2014.
[3] J. Back and J. Soe, “Immersion of nonlinear systems
into linear systems up to output injection: characteris-
tic equation approach,” International Journal of Control,
vol. 77, no. 8, pp. 723–734, 2004.
[4] J. Back, K. Yu, and J. Soe, “Dynamic observer error lin-
earization,” Automatica, vol. 42, no. 12, pp. 2195–2200,
2006.
[5] M. Baptista, “Cryptography with chaos,” Physics Letters
A, vol. 240, pp. 50–54, 1998.
[6] V. Bijalwan, P. Kumari, J. Pascual, and V. B. Semwal,
“Machine learning approach for text and document min-
ing,” arXiv preprint arXiv:1406.1580, 2014.
[7] M. Fliess, C. Join, and H. Sira-Ramirez, “Nonlinear es-
timation is easy,” International Journal of Modelling
Identification and Control, vol. 4, no. 1, pp. 12–27,
2008.
[8] L. Fridman, J. Moreno, and R. Iriarte, Sliding Modes
after the first Decade of the 21st Century: State of the
Art. Series: Lecture Notes in Control and Information
Sciences , V 412, Springer, 2011.
[9] M. Frunzete, J.-P. Barbot, and C. Letellier, “Influence
of the singular manifold of observable states in recon-
structing chaotic attractors,” Physical Review E, vol. 86,
2012.
[10] J.-P. Gauthier, H. Hammouri, and S. Othman, “A sim-
ple observer for nonlinear systems with applications to
bioreactors,” IEEE Transactions on Automatic Control,
vol. 37, no. 6, pp. 875–880, 1992.
[11] R. Hermann and A. Krener, “Nonlinear controllabil-
ity and observability,” IEEE Transactions on Automatic
Control, vol. 22, no. 5, pp. 728–740, 1977.
[12] P. Kumari and V. Abhishek, “Brainwave based user iden-
tification system: A pilot study in robotics environ-
ment,” Robotics and Autonomous Systems, vol. 65, pp.
15–23, 2015.
[13] C. Letellier and L. A. Aguirre, “Symbolic observabil-
ity coefficients for univariate and multivariate analysis.”
Physical Review E, vol. 79, no. 066210, 2009.
[14] A. Levant, “Higher-order sliding modes, differentiation
and output-feedback control,” International Journal on
automatic, vol. 76, no. 9/10, pp. 924–941, 2003.
[15] ——, “Homogeneity approach to high-order sliding
mode design,” Automatica, vol. 41, pp. 823–830, 2005.
[16] A. Levant and M. Livne, “Exact differentiation of sig-
nals with unbounded higher derivatives,” IEEE TAC,
vol. 57, no. 4, pp. 1076–1080, 2012.
[17] M. L’Hernault, J.-P. Barbot, and A. Ouslimani, “Feasi-
bility of analogue realization of a sliding mode observer:
application to data transmission,” IEEE Transaction on
Circuit and System −I, vol. 55, no. 2, pp. 614–624,
2008.
[18] H. Nijmeijer and I. Mareels, “An observer looks
at synchronization,” IEEE Transactions on Circuits
and Systems-1: Fundamental theory and Applications,
vol. 44, no. 10, pp. 882–891, 1997.
[19] L. Pecora and T. Carroll, “Synchronization in chaotic
systems,” Physical Review Letters, vol. 64, no. 8, pp.
821–824, 1990.
[20] W. Perruquetti and J.-P. Barbot, Sliding Mode Control in
Engineering. Marcel Dekker, 2002.
[21] A. Rapaport and M. Maloum, “Design of exponen-
tial observers for nonlinear systems by embedding,” In-
ternational Journal of Robust and Nonlinear Control,
vol. 14, pp. 273–288, 2004.
[22] A. Rapaport and V. Maloum, “Embedding for exponen-
tial observers of nonlinear systems,” Proceedings of the
39th IEEE CDC, Syndey, 2000.
[23] R. Tami, D. Boutat, and G. Zheng, “Extended output
depending normal form,” Automatica, 2013.
[24] A. Ticlea, “Extended output depending normal form,”
Ph D Thesis INPG 22 September 2006, 2006.
 Comparison of Different Antenna Arrays for the
Channel Capacity
Shu-Han Liao
Smart Network System Institute, Institute for Information
Industry
Taipei, Taiwan, R.O.C.
e-mail:shliao@iii.org.tw
Chien-Ching Chiu
Electrical Engineering Department, Tamkang University
New Taipei City, Taiwan, R.O.C.
e-mail:chiu@ee.tku.edu.tw
Abstract—Three types of antenna arrays such as uniform
linear, uniform rectangular arrays and uniform cube arrays are
used in the transmitter and their corresponding channel capacity
on several paths in the indoor environment are calculated.
Numerical results show that uniform linear arrays is better than
that for uniform rectangular arrays and uniform cube arrays
system with and without interference.
Keywords—uniform linear array; uniform rectangular arrays;
uniform cube arrays
I. INTRODUCTION
Wireless data rates continue to increase by different
techniques to date, and one of these techniques is antenna
array that deploys multiple antennas on transmitter or receiver.
In general, antenna array can be called single-input multiple-
output (SIMO) while single antenna on transmitter and
multiple antennas on receiver are deployed, and called
multiple-input single-output (MISO) while multiple antennas
on transmitter and single antenna on receiver are deployed.
Furthermore, multiple-input multiple-output (MIMO) also
belongs to one kind of the antenna arrays, which multiple
antennas on both transmitter and receiver are deployed.
Compared with single-input multiple-output (SISO), MISO
and SIMO utilize transmitting antenna diversity and receiving
antenna diversity respectively to yield the corresponding
diversity gains, and the gains can enhance average receiving
power and obtain more channel capacity further. In contrast
with MISO and SIMO, MIMO breaks a multipath channel into
several individual spatial channels, then provide an additional
spatial dimension for communication and yield a spatial
degree of freedom gain. These additional spatial degrees of
freedom can be exploited by spatially multiplexing several
data streams onto the MIMO channel, and lead to an increase
in the capacity [1], [2].
MIMO in theory is possible to break a multipath channel
to several un-correlation sub-channels and transmit multiple
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Chien-Hung Chen
Department of Information Technology and Mobile
Communication, Taipei College of Maritime Technology
New Taipei City, Taiwan, R.O.C.
e-mail: f1092@mail.tcmt.edu.tw
Min-Hui Ho
Institute for Information Industry
Taipei, Taiwan, R.O.C.
e-mail: katrina650@hotmail.com
data streams simultaneously through the channel to obtain
high channel capacity, because it can provide more spatial
degrees of freedom [1], [2]. However, the MIMO capacity is
affected largely by the number of paths and correlation
between sub-channels in the realistic environment, which both
the number of paths and correlation between sub-channels
affect spatial degree of freedom provided by MIMO. It is well-
known that a very complicated environment can offer large
paths and little correlation between sub-channels to obtain
high MIMO capacity, and the opposite results will be obtained
when a very simple environment is considered. The question
is whether a communication environment is complicated
sufficiently to support spatial degree of freedom provided by
MIMO.
Spatial degree of freedom depends on not only
environmental complication but also the number of antennas
of both transmitter and receiver. In other words, for getting
necessary MIMO capacity, environmental complication and
suitable antenna deployment have to be treated together.
Hence, how to deploy suitable the number of antennas of both
transmitter and receiver in the corresponding environment is
an important research topic. UP to now, most researches about
MIMO capacity are based on symmetric deployment which
expresses the numbers of transmitting antennas is equal to that
of receiving antennas. However, asymmetric deployment
which expresses the numbers of transmitting antennas is
unequal to that of receiving antennas is worth to investigate,
because the capacity for asymmetric deployment is better than
for symmetric deployment under some conditions.
For wireless communication systems, two main sources of
performance degradation are the thermal noise present in the
channel or generated in the receiver and unwanted signals
emanating from the same or nearby stations. Co-channel
interference (CCI) is one of the unwanted signals and it
appears due to frequency reuse in wireless channels. CCI
reduction has been studied and used in a very limited form in
wireless networks for many years [1], [2], [3]. The use of
directional antennas and antenna arrays has long been channel matrix denotes the complex channel gain from the ith
recognized as an effective technique for reducing CCI, since
the differentiation between the spatial signatures of the desired interference antenna to the xth receiving antenna.
signal and CCI signals can be exploited to reduce the A matrix representation of MIMO-NB system with single
interference when multiple antennas are used. Although using
multiple antennas (SIMO and MISO) to reduce CCI has been CCI is shown in Fig. 2. In this figure, the desired signal can
proven effective in many literatures, it is not very sure that still be fed into several sub-channels by corresponding signal
MIMO can be used to reduce CCI as well. Furthermore,
capacity affected by CCI is still largely an unsolved problem processing and SVD. However, the signal processing and
[3] and corresponding literatures are few. As a result, it is
SVD is useless for the interference signal, since the
worth to investigate whether MIMO can effectively reduce
CCI on channel capacity calculation. interference channel matrix Hi is unknown to the receiver. As
The 60 GHz band which provides 7 GHz of unlicensed a result, the result contributed by the interference can be
spectrum with a potential to develop wireless communication
systems with multi Gbps throughput as part of the fifth- expressed as U* H i X i . Finally, an equivalent architecture of
generation (5G) system [4]. For wireless communication
MIMO-NB system with single CCI based on CSI-B is shown
systems, CCI [5] is one of the unwanted signals and it appears
due to frequency reuse in wireless channels. in Fig. 3. It is seen that the received signal vector Ŷ can be
Reference[6] analysis of the MIMO capacity of WLAN
systems with single CCI at 60GHz band have been seen as the combination of the desired signal vector X d
investigated. However, to our knowledge, a comparative study propagating through several sub-channel plus the result
about the performances of uniform linear, uniform rectangular
contributed by the interference Xi and the zero mean additive
arrays and uniform cube arrays when applied to Channel
Capacity analysis problems has not yet been investigated for Ŵ .
MIMO-WLAN system. In this paper, we compare the three white Gaussian noise vector The system can be expressed
types of antenna arrays such as uniform linear, uniform as follows:
rectangular arrays and uniform cube arrays are used in the
transmitter and their corresponding channel capacity on ˆ  DX  SX  W
Y ˆ
d i (3)
several paths in the indoor environment are calculated.
where S  U H i denotes a N r  N i equivalent channel matrix
*

II. SYSTEM DESCRIPTION AND CHANNEL MODELING

for the interference.
A. System description
A sketch of time-invariant MIMO-NB system is shown
in Fig. 1, and the received signal is described as follows:
Y  H d Xd  H i X i  W

(1)

where Y , X d , X i and W denote the N r  1 received signal

vector, the N d  1 desired transmitted signal vector, the N i  1

interference signal vector and the N r  1 zero mean additive

white Gaussian noise vector at a symbol time, respectively. In

the equation, H d is the N r  N d channel matrix for desired

signal, and the element hxy of the channel matrix denotes the
complex channel gain from the yth transmitting antenna to the

xth receiving antenna. Moreover, H i is the N r  N i channel

matrix for interference signal, and the element hxi of the

 Xi S

Xi Hi  xi,1   s11 s12 s13  s1Ni 

   
 xi,2   s21 s22 s23  s2 Ni 
 xi ,1  h11 h12 h13  h1N i  x  s
    s32 s33  s3 Ni 
 xi , 2  h21 h22 h23  h2 N i   i ,3   31 
h h h  h          
x   
 i ,3   31 32 33 3Ni
   s N r 2 s N r 3  s N r Ni 
 s N r 1
           x i , N i 
   
 xi , N i  hN r 1 hN r 2 hN r 3  hN r N i  Ni × 1 Nr × Ni

Ni × 1 Nr × N i Xd D Ŷ

 xd ,1   1  00 

0 0 0
  yˆ1 
Xd Hd Y    0 2 0  0 00   
 xd , 2     yˆ 2 
 xd ,1  h11 h12 h13  h1N d   y1  x   0 0 3  0 00   yˆ 
       d ,3     3 
 xd , 2  h21 h22 h23  h2 N d         0 00
 y2   
x  h h h  h   y3     0 0 0 0 m 00  yˆ 
 d ,3   31 32 33 3N d
    xd , N m   
          0 0  Nm 
      0 0 0 0 0
    y     
 xd , N d  hN r 1 hN r 2 hN r 3  hN r N d   Nr              
 xd , N d   0 0 0 0 0 0  0   yˆ N r 
Nd × 1 Nr × Nd Nr × 1
Nd × 1 Nr × Nd Nr × 1

W Ŵ

 w1 
 wˆ 1 
   
 w2   wˆ 2 
 w3   wˆ 
   3 
    
w   wˆ 
 Nr 
 Nm 
 
Nr × 1  
 wˆ N r 
Figure 1. A sketch of MIMO-NB system with single CCI
Nr × 1

Figure 3. A matrix representation of MIMO-NB system with

Xi MIMO channel of interference signal single CCI
B. Channel modeling
Hi By using these images and received fields, the channel
frequency response can be obtained as following [7-14]
Np
H  f    a p ( f )e
jθ p ( f )
(4)
MIMO channel of desired signal p 1

C. Simulation versus measurement

V* D U W By measurement is the most direct method to predict
channel characteristic which a lot of unpredictable variations
are filled in the realistic environment, and it can always get the
correct results by proper and exact instrument. One important
V̂
Transmitting Receiving Û problem is that such instrument is often especially expensive
pre-processing post-processing when it is applied to the future or uncommon systems. For the
reasons mentioned above, few instruments can be used to
Xd
measure channel characteristic of WLAN systems and they are
Ŷ
very expensive. Additionally, a system equipping MIMO
requires expensive specialized hardware to measure [15], and
Figure 2. A matrix representation of MIMO-NB system with the measuring process is very time-consuming in general. As a
single CCI result, it is difficult to predict channel characteristic of MIMO-
WLAN system by measurement in the realistic environment.
Furthermore, measurement is always limited by its
measurement distance and some natural restrictions. In other
words, measurement could not be applied arbitrarily to desired
environments. Also instrumental setting can not be changed
when required. Fortunately, the same process by simulation
mechanism can effectively overcome these drawbacks as long 120
8X8-Linear with CCI

Average Capacity (bits/sec/Hz)

as accuracy of the simulation mechanism has been proven
before use. 100 8X8-Linear without CCI
8X8-Rectangular with CCI
III. NUMERICAL RESULTS 80 8X8-Rectangular without CCI
Layout of a small personal communication environment is 8X8-Cube with CCI
shown in Figure 1. 60 8X8-Cube without CCI
Three types of antenna array such linear arrays, uniform
rectangular arrays and uniform cube arrays consist of eight 40
dipole antennas are used in the transmitter and receiver, as
shown in Fig. 2. 20
2.5 m 0.2 m 2.5 m

Y 0
0 20 40 60 80
door door SNRt(dB)
X Figure 3. The average capacities of WLAN systems for both MIMO-Linear,
Z
MIMO-Rectangular and MIMO-Cube with and without single CCI

4m
VRc for linear arrays, uniform rectangular arrays, uniform
Desired transmitter CCI transmitter cube arrays and SISO are shown in Figure 4.
partition
(1.25m, 2m, 1.2m) (3.95m, 2m, 1.2m)

desk
100
metallic
cabinet
90
Room 1 partition Room 2

Figure 1. Layout of a small personal communication environment 80

VRC(%)

0.0025m 70

60

50 1X1
8X8-Linear
Y 40 8X8-Rectangular
8X8-Cube
30
X 0 20 40 60 80
Z SNRt(dB)
(a) (b) (c)
Figure 4. VRc for MIMO-Linear, MIMO-Rectangular, MIMO-Cube
Figure 2. Layouts of different arrays. (a) Linear Array (b) Rectangular and SISO
Array (c) Cube Array
The numerical results for the outage probability with and
without inference are shown in Fig. 5 and Fig. 6, respectively.
The average capacities of WLAN systems calculated from Here SNRt is 35dB. Note that the average SNRr (receiving
236 receiving locations for both linear arrays, uniform signal power to noise power ratio) is SNRt subtracting about
rectangular arrays and uniform cube arrays with and without
single CCI are shown in Figure 3. 25dB in our cases and the value is about 10dB.
 without CCI (SNRt=35dB) REFERENCES
1 [1] G. D. Durgin, Space-Time Wireless Channels. New Jersey: Prentice
Hall PTR, 2003.
[2] D. Tse and P. Viswanath, Fundamentals of Wireless Communication.
0.8
outage probability Po

United Kingdom: Cambridge University Press, 2005.

[3] A. J. Paulraj, R. Nabar and D. Gore, Introduction to Space-Time
Wireless Communication. U.K.: Cambridge Univ. Press, 2003.
0.6 [4] “5G radio access: requirements, concepts and technologies,” White
Paper, NTT DOCOMO, Inc., Jul. 2014. [Online].Available:
https://www.nttdocomo.co.jp/english/binary/pdf/corporate/technology/w
0.4 hitepaper 5g/DOCOMO 5G White Paper.pdf
[5] W. Nam, D. Bai, J. Lee, and I. Kang, “Advanced interference
8X8-Linear management for 5G cellular networks,” IEEE Commun. Mag., vol. 52,
0.2
8X8-Rectangular May 2014, pp.52–60.
8X8-Cube [6] C. T. Hsieh, S. H. Liao, C. C. Chiu and M. H. Ho, “Capacity Analysis of
0 MIMO-WLAN Systems with Single Co-Channel Interference,”
0 5 10 15 Automatika ‒ Journal for Control, Measurement, Electronics,
transmission rate R (bit/sec/Hz) Computing and Communications, Vol. 55, no. 2, Apr. 2014, pp. 143-152.
[7] P. C. Richardson, W. Xiang and W. Stark, “Modeling of Ultra-
Figure 5. outage probability for 8×8 without CCI Wideband Channels within Vehicles,” IEEE J. Sel. Areas Commun., vol.
24, no. 4, 2006, pp. 906-912.
with CCI (SNRt=35dB)
[8] S. H. Liao, C. C. Chiu and M. H. Ho, “Location Optimization for
1 Antennas by Asynchronous Particle Swarm Optimization,” IET
Communications, Vol. 7, Issue: 14, pp.1510-1516, Sept. 2013.
[9] C. C. Chiu, C. L. Liu and S. H. Liao, “Channel Characteristics of Ultra
0.8
outage probability Po

Wideband Systems with Single Co-Channel Interference,” Wireless

Communications and Mobile Computing, vol. 13, issue 9, pp. 864-873,
Jun. 2013.
0.6 [10] C. C. Chiu, M. H. Ho and S. H. Liao, “MIMO-UWB Smart Antenna
Communication Characteristics for Different Antenna Arrays of
Transmitters,” International Journal of RF and Microwave Computer -
0.4 Aided Engineering, vol. 23, no 2, pp. 378-392, May. 2013.
[11] S. H. Liao, C. C. Chiu, M. H. Ho and C. L. Liu, “Channel Capacity of
8X8-Linear Multiple–Input Multiple–Output Ultra Wide Band Systems with Single
0.2 8X8-Rectangular Co–channel Interference,”, International Journal of Communication
Systems, Vol. 23, Issue 12, pp. 1600–1612, Dec 2010.
8X8-Cube
[12] S. H. Liao, C. C. Chiu, C. H. Chen and M. H. Ho, “Channel
0
0 5 10 15 Characteristics of MIMO-WLAN Communications at 60GHz for
transmission rate R (bit/sec/Hz) Various Corridors,” EURASIP Journal on Wireless Communications
and Networking, 96, Apr. 2013.
Figure 6. outage probability for 8×8 with CCI [13] C. C. Chiu, C. Y. Yu, S. H. Liao and M. K. Wu, “Channel Capacity of
Multiple-Input Multiple-Output Systems for Optimal Antenna Spacing
by Particle Swarm Optimizer,” Wireless Personal Communications, vol.
IV. CONCLUSION 69, issue 4, pp. 1865–1876, Apr. 2013.
Numerical results show that average capacity for linear [14] S. H. Liao, M. H. Ho, C. C. Chiu and C. H. Lin, “Optimal Relay
arrays are larger than that for uniform rectangular and uniform Antenna Location in Indoor Environment Using Particle Swarm
cube arrays with and without interference. In general, the Optimizer and Genetic Algorithm,” Wireless Personal Communications,
Vol.62, No.3, pp. 599–615, Feb. 2012.
linear arrays is a good choice for indoor environments. Full-
[15] W. Q. Malik and D. J. Edwards, “Measured MIMO Capacity and
duplex design using a single circularly polarized antenna. Diversity Gain with Spatial and Polar Arrays in Ultrawideband
Since we do not need to allocate different antennas for Channels,” IEEE Trans. Commun., vol. 55, no. 12, Dec. 2007, pp. 2361-
receiving and transmitting, and will be the subject of a future 2370.
study.

ACKNOWLEDGEMENTS

A SPECIAL THANKS GOES TO INSTITUTE FOR INFORMATION

INDUSTRY IN TAIWAN, FOR ITS GREAT HELP TO THIS STUDY.
 Least Squares SVM Approach for Abnormal Brain
Detection in MRI using Multiresolution Analysis
Deepak Ranjan Nayak, Ratnakar Dash, Banshidhar Majhi
Department of Computer Science and Engineering,
National Institute of Technology, Rourkela, India- 769008
Email: depakranjannayak@gmail.com, ratnakar@nitrkl.ac.in, bmajhi@nitrkl.ac.in
Abstract—Developing automatic and accurate computer-aided
diagnosis (CAD) systems for detecting brain disease in magnetic
resonance imaging (MRI) are of great importance in recent years.
These systems help the radiologists in accurate interpretation of
brain MR images and also substantially reduce the time needed
for it. In this paper, a new system for abnormal brain detection
is presented. The proposed method employs a multiresolution
approach (discrete wavelet transform) to extract features from
the MR images. Kernel principal component analysis (KPCA)
is harnessed to reduce the dimension of the features, with the
goal of obtaining the discriminant features. Subsequently, a new
version of support vector machine (SVM) with low computational
cost, called least squares SVM (LS-SVM) is utilized to classify
brain MR images as normal or abnormal. The proposed scheme is
validated on a dataset of 90 images (18 normal and 72 abnormal).
A 6-fold stratified cross-validation procedure is implemented and
the results of the experiments indicate that the proposed scheme
outperforms other competent schemes in terms of classification
accuracy with relatively small number of features.
Keywords—Magnetic resonance imaging (MRI); Discrete
wavelet transform (DWT); Kernel principal component analysis
(KPCA); Least squares support vector machine (LS-SVM)
I. I NTRODUCTION
Brain diseases are growing rapidly among children and
adults throughout the world. According to the National Brain
Tumor Foundation (NBTF) in the United States, it has been
estimated that, in children, brain tumors are the reason for
one-quarter of all cancer deaths [1]. In the year 2104, World
Health Organization (WHO) reported that around 250,000
people globally were diagnosed with primary brain tumors
every year. Therefore, early detection of brain disease is very
important. Magnetic resonance imaging (MRI) has been used
as the most suitable medical imaging technique for an accurate
detection of various brain diseases in recent years [2]. It is
a low-risk, non-invasive method that generates high-quality
images of the anatomical structures of the human brain and
gives rich information about the soft brain tissues anatomy [3],
[4]. MRI provides better contrast for different brain tissues
than all other imaging modalities [5]. These advantages have
delineated MRI as the most well-known method of brain
pathology diagnosis and treatment. However, the high volume
of information leads difficulty in analyzing and interpreting
MR images. Computer-aided diagnosis (CAD) systems are cur-
rently used which examines brain MR images with the help of
image processing techniques. CAD systems help radiologists
in accurate interpretation of brain MR images for detecting
abnormal brain. Therefore, it is necessary to develop a CAD
system to increase the diagnosis capability and to reduce the
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
time required for it. One of the most important steps in this
system is to find out a set of discriminative feature that can
classify the normal brain MR image from the abnormal one.
An assortment of techniques has been studied for this purpose.
Over the last decade, several researches have been carried
out for brain MR image classification. The most widely
used approach for feature extraction is the multiresolution
analysis that decomposes original MR image into several sub-
images. These images preserve information about both low
and high frequencies. Wavelet transform is one of the most
important approaches for the texture analysis of the image.
Various researchers have used wavelet transform to extract
the features from the MR image. Chaplot et al. [3] have
utilized the approximation coefficients of two-dimensional
discrete wavelet transform (2D DWT) of level-2 decomposition
as the features and employed self-organizing map (SOM)
and support vector machine (SVM) classifiers. Maitra and
Chatterjee [6] have introduced Slantlet transform (ST) which
is an improved version of DWT, for feature extraction and
applied back-propagation neural network (BPNN) classifier.
El-Dahshan et al. [7] have used the approximation coefficients
of level-3 decomposition of 2D DWT to represent each im-
age. Principal component analysis (PCA) was employed to
reduce the number of coefficients. They used feed forward
back-propagation artificial neural network (FP-ANN) and k
-nearest neighbor (k-NN) classifiers separately to detect the
normal and pathological brain. In [4], [8]–[10], the researchers
have used the coefficients of level-3 approximation sub-
band of 2D DWT to extract features from images and then
employed PCA for feature reduction. They have suggested
different classifiers with some training parameter optimization
approaches, namely, feed forward neural network (FNN) with
scaled chaotic artificial bee algorithm (SCABC) [9], FNN with
adaptive chaotic particle swarm optimization (ACPSO) [8],
and BPNN with scale conjugate gradient (SCG) [4]. Zhang et
al. [10] have used a kernel SVM (KSVM) classifier with three
kernels, viz., linear (LIN), homogeneous polynomial (HPOL),
inhomogeneous polynomial (IPOL) and Gaussian radial basis
(GRB), to segregate the normal and pathological MR images.
They have achieved high classification accuracy with GRB
kernel. Das et al. [5] have presented an efficient mutiscale
geometric analysis tool, Ripplet transform (RT) for feature
extraction followed by PCA for dimensionality reduction. A
less expensive SVM approach, called least square SVM (LS-
SVM) was applied for classification and they have achieved
suitable results over larger datasets. Saritha et al. [11] sug-
gested the combined wavelet entropy based spider web plots
(SWP) to extract features. The entropy values were calculated
 


"#$ 


 
 %
 
 


& 
 

 


'(()# 
 
 ! 
Fig. 1. Block diagram of the proposed scheme for detection of abnormal brain
for the approximation sub-bands of level-8 decomposition of
Daubechies-4 wavelet. Finally, probabilistic neural network
(PNN) was applied for classification. Zhang et al. [12] have
used Shannon entropy (SE) and Tsallis entropy (TE) to get
features from the discrete wavelet packet transform (DWPT)
coefficients and suggested a generalized eigenvalue proximal
SVM (GEPSVM) classifier. Zhou et al. [13] have achieved
classification accuracy of 92.60% by using wavelet entropy
values as the features for each image. They have applied a
Naive Bayes classifier (NBC) to determine the normal and
abnormal brain. Zhang et al. [14] have obtained 82.69% of
accuracy using SVM classifier. To get the features, they have
utilized wavelet-energy values of all the detail sub-bands of
level-2 decomposition.
The literature review reveals different existing schemes for
abnormal brain detection. Most of the schemes are not able
to get a high classification accuracy. It has been observed
that the dimension of the feature space is relatively high in
many cases which may degrade the performance. When the
extracted features have more complicated structures and can
not be well represented in a linear subspace, then PCA will be
not helpful for dimension reduction in such case. Hence, there
is a need to use a new technique for nonlinear dimensionality
reduction. Moreover, PCA requires a high computational cost
for eigenvalue decomposition when the number of features is
more than the number of images. To address the above issues,
we have utilized the coefficients of approximation sub-band
of 2D DWT for feature extraction. A kernel PCA (KPCA)
approach is employed to handle the nonlinear feature values
and to reduce the computational cost. To make the system
more robust and computationally efficient, LS-SVM is used
to classify the abnormal brain from the normal one. The
proposed method is tested on a dataset of 90 images and the
experimental results indicate that the scheme is superior to its
competent schemes. The remainder of this paper is organized
as follows. Section II deals with the working procedure of
the proposed method. In Section III, the simulation results
and comparisons are portrayed. Finally, Section IV gives the
concluding remarks.
II. P ROPOSED M ETHOD
The proposed method includes three important phases,
namely, feature extraction, feature dimensionality reduction
and classification. The overall block diagram of the proposed
scheme is shown in Fig. 1. All the phases of the scheme are
portrayed below in detail.
A. Feature Extraction using mutltiresolution technique
The proposed scheme uses a popular multiresolution tech-
nique, called DWT to extract features from the brain MR
images. Wavelet transform is proven to be a powerful math-
ematical tool for feature extraction [15]. Compared to other
transformation techniques, wavelet transform provides time-
frequency localization of an image which is very important
for classification.
A 2D DWT is implemented using low pass and high pass
filters and down samplers. In case of images, the DWT is
applied to each dimension individually, which results in four
sub-band images (LL, LH, HL, HH) at each level. Among
them, three sub-band images LH (low-high), HL (high-low)
and HH (high-high) are the detail (high frequency) components
in horizontal, vertical and diagonal directions, respectively. LL
(low-low) sub-band image is the approximation (low pass)
component which is used for next level 2D DWT calculation
[4]. Fig. 2 illustrates the wavelet decomposition of a normal
brain MR image up to three resolution levels. In this study,
we have utilized the coefficients of the approximation sub-
band of level-3 decomposition (LL3 ) of Daubechies-4 wavelet
to extract features. Daubechies-4 provides better resolution for
smoothly varying signals in case if MR images of the brain.
Therefore, we have selected Daubechies-4 wavelet, which
gives better classification accuracy. The coefficients of LL3
sub-band are arranged in row-major order to generate a feature
vector. Then a feature matrix is created by combining the
vectors corresponding to all brain MR images. The extracted
features have been normalized before employing to KPCA.
The feature z is normalized to zn using the following formula.
z−μ
zn = (1)
σ
where, μ and σ are the mean and standard deviation of the
features, respectively. The normalized feature vectors are then
sent to the next phase.
B. Feature Reduction using KPCA
The size of the feature space becomes large if the ap-
proximation coefficients are directly used as the features, and
all the features are not relevant for classification. Hence,
to make the classification task feasible, the dimensionality
of the feature vector needs to be significantly reduced, and
informative features need to be extracted. PCA is often used
for this purpose [16]. However, PCA only allows the linear di-
mensionality reduction and it doesn’t perform well on the high-
dimensional features having complicated structures. Therefore,


 


Fig. 2. A normal brain MR image and its wavelet decomposition at three resolution level

a non-linear form of PCA, called kernel PCA (KPCA) is Here, K is a symmetric and positive semidefinite matrix.
employed in this paper for the dimensionality reduction of the We then normalize the eigenvectors β to ensure that the
features. Additionally, KPCA is computationally efficient than corresponding eigenvectors V are orthonormal. The resulting
conventional PCA when the size of the feature space is greater k th principal component of a test sample x is calculated using
than the number of samples [17], [18].
N 
 k 
Consider a dataset {xk ∈ X} of N observations where yk (x) = V .φ(x) = βi φ(xi ) .φ(x)
k

k = 1, 2, . . . , N and each xk is a d-dimensional feature vector. i=1

(10)
Suppose a nonlinear transformation φ(x) from original feature 
N
space to a high dimensional feature space is given, such that = βik k(xi , x)
φ(x) : d → D , where D >> d. Each data point xk is i=1
projected to a point φ(xk ). At first, we have assumed that the In general, {φ(xk )} may not be zero mean, and the Gram
N
 can be used for this purpose to substitute the kernel
data is centered, i.e. N1 φ(xk ) = 0. matrix K
k=1 matrix K [19]. The Gram matrix is defines as,
The covariance matrix of the projected features is defined  = K − 1N K − K1N + 1N K1N
K (11)
by,
1 
N where 1N is the N × N matrix with all the elements equal
CD = φ(xj )φ(xj )T (2) to 1/N . The main advantage of kernel methods is that we do
N j=1 not need to compute φ(xk ) explicitly. Rather, we can directly
construct the kernel matrix from the input dataset X [18].
Now, we have to find out the eigenvalues λ and eigenvectors Three commonly used kernels are polynomial kernel, Gaussian
V of CD using kernel and sigmoid kernel.
λV = CD V (3)
In this paper, we have used polynomial kernel to construct
Multiplying φ(xk ) in both sides of (3), we get, K. Using the above procedure, we have obtained an reduced
feature matrix from the normalized feature matrix. The reduced
λ(φ(xk ).V ) = (φ(xk ).CD V ) ; k = 1, 2, . . . , N (4)
matrix and a target vector is sent to the classifier to determine
The eigenvectors can be expressed as the linear combinations the abnormal brain.
of projected features and is given by,
C. Classification using LS-SVM

N
V = βi φ(xi ) (5) Standard SVM leads to high computational complexity
i=1 in the case when it deals with a large dimensional dataset.
To overcome the computational overhead, a least squares
where βi , i = 1, 2, . . . , N are the coefficients. Substituting (2) version of SVM (LS-SVM) is exploited as the classifier in
and (5) in (4), we have, this paper. Because of the equality constraints in the LS-SVM
⎛ ⎞
 formulation, a set of linear equations has to be solved, instead
1  ⎝ 
N N N
λ βi (φ(xk )φ(xi )) = βi φ(xk ) φ(xj )T ⎠ of a quadratic programming problem for standard SVM [20].
N N
i=1 i=1 j=1 Given a training set of N data points {xk , dk }k=1 with
(φ(xj )φ(xi )) , ∀k = 1, 2, . . . N input data xk ∈ m and class labels dk ∈ , where dk =
(6) {−1, +1}, LS-SVM can be formulated as the optimization
problem:
and defining an N × N matrix K by,
1 2
N
Kij = (φ(xi ).φ(xj )) 1
(7) min J (w, b, e) = wT w + ζ ek (12)
w,b,e 2 2
k=1
we can write (6) as,
subject to the equality constraint
N λKβ = K 2 β (8) 
dk wT ϕ (xk ) + b = 1 − ek , k = 1, 2, . . . , N (13)
where β is a N -dimensional column vector with entries
β1 , β2 , . . . , βN and can be solved by the eigenvalue problem where w is the weight vector, ϕ (.) the mapping function,
ζ > 0 the regularization factor, b a bias term and ek the error
N λβ = Kβ (9) variables.
 The Lagrangian can be defined as Algorithm 1 Pseudocode of the proposed system
Require: Brain MR images of size 256 × 256
L (w, b, e, α) = J (w, b, e)
N : Total number of images

N
   (14) M : Total number of features
− αk dk wT ϕ (xk ) + b − 1 + ek R: Number of reduced features
k=1 Ensure: Normal or abnormal brain
where αk are Lagrange multipliers. NThe conditions for op- Step 1: Feature extraction using 2D DWT
timality are: ∂w ∂L
= 0 → w = k=1 αk dk ϕ(xk ); ∂b =
∂L
N for i ← 1 to N do
0 → k=1 αk dk = 0; ∂ek = 0 → αk = ζek ; and
∂L
Read the brain MR images
∂αk = 0 → dk w ϕ (xk ) + b − 1 + ek = 0, which can be
∂L T
Compute the LL3 coefficients of 2D DWT
written as the solution to the following set of linear equations
Arrange the coefficients in row-major order and store
⎡ ⎤⎡ ⎤ ⎡ ⎤ in a matrix Q(N × M )
I 0 0 −Z T w 0
⎢ 0 0 0 −D ⎥ ⎢ b ⎥ ⎢ 0 ⎥ T end for
⎣ ⎦⎣ e ⎦ = ⎣ 0 ⎦ (15)
0 0 ζI −I Normalize the matrix Q using (1) and obtain a new
Z D I 0 α 1 matrix Qn (N × M )
 Step 2: Feature reduction using KPCA
where Z = ϕ(x1 )T d1 ; . . . ; ϕ(xN )T dN , D = [d1 ; . . . ; dN ],
1 = [1; . . . ; 1] , e = [e1 ; . . . ; eN ] , α = [α1 ; . . . ; αN ]. We can Choose a kernel function and find out the kernel
also obtain the solution as matrix K(N × N ) using Qn
     Calculate the normalized kernel matrix of the data
0 −DT b 0 
i.e. K
= (16) 
D Ω + ζ −1 I α 1 Solve the eigenvalue problem, λβ = Kβ
Select R principal components corresponding to the
where Ω = ZZ T and according to Mercer’s condition [20], R largest eigenvalues
Ωkl = dk dl ϕ(xk )T ϕ(xl ) = dk dl K(xk , xl ) (17) Generate a reduced feature matrix X(N × R)
Step 3: Classification using LS-SVM
where, K(., .) is the kernel function. The LS-SVM classifier N
Create a training dataset {xk , dk }k=1 with input data
is obtained by xk ∈ X and class label dk ∈ D = {−1, +1}
N  Apply n-fold cross validation procedure to find train-

f (x) = sign αk dk K(x, xk ) + b (18) ing and testing samples
k=1 Choose a kernel function K and train the LS-SVM
classifier
The kernels that we have used for training the LS-SVM Classify test images as normal or abnormal
classifier are listed in Table I. The parameter θ indicates the
degree of the polynomial and σ is a free parameter which
controls the shape of the kernel.
To make the classifier more reliable and more generalize
TABLE I. D IFFERENT KERNEL FUNCTIONS USED IN LS-SVM to independent datasets, 6-fold stratified cross-validation (CV)
procedure is employed. The setting of the training and the
Kernel Definition
validation images of the dataset is shown in Table II.
Linear K(x, xk ) = xT
kx
K(x, xk ) = (xT
k x + 1)
θ
Polynomial
  TABLE II. S ETTING OF 6- FOLD STRATIFIED CV PROCEDURE
Radial Basis Function (RBF) K(x, xk ) = exp −||x − xk ||2 /2σ2
Total number of images (90) Training images (75) Validation images (15)
Normal Abnormal Normal Abnormal Normal Abnormal
18 72 15 60 3 12

III. E XPERIMENTAL R ESULTS AND C OMPARISONS

The experiments were carried out on a PC with 3.40 GHz
Core-i7 processor and 4 GB of RAM, running under Windows
8 operating system. The proposed algorithm is simulated using
MATLAB toolbox. The pseudocode of the proposed CAD
system is presented in Algorithm 1.
1) Experimental setup: The dataset includes 90 images, 18
from each of the five category normal, brain tumor, stroke,
degenerative disease and infectious disease. So there are total
18 normal and 18 × 4 = 72 abnormal images. The dataset
consists of T2-weighted MR brain images in the axial plane
and 256×256 in-plane resolutions which are downloaded from
the Harvard Medical School website [21]. A sample from each
of the category is shown in Fig. 3.
In this work, we consider the abnormal and normal class as
the positive and negative class, respectively. Table III lists the
measures which are used to calculate the performance of the
proposed scheme and its competent schemes. TPR (Sensitivity)
is the probability that a diagnostic test is positive, given that
the person has the disease, whereas TNR (Specificity) is the
probability that a diagnostic test is negative, given that the
person does not have the disease. ACC is the probability that
a diagnostic test is correctly performed. The classifier’s perfor-
mance can also be evaluated using an important index value,
called the area under the curve (AUC) which is calculated by
the receiver operating characteristic (ROC) curve. For an ideal
classifier, the value of AUC is 1.
 (a) (b) (c) (d) (e)

Fig. 3. Sample of brain MR images with (a) Normal brain, (b) brain tumor, (c) stroke, (d) degenerative disease, (e) infectious disease

TABLE III. D IFFERENT CLASSIFICATION PERFORMANCE MEASURES TABLE VI. C ONFUSION MATRIX FOR ‘LS-SVM + RBF’ CLASSIFIER

Measures Definition Output (predicted) class

True Positive Rate (TPR) TP/(TP+FN) Abnormal (positive) Normal (negative)
True Negative Rate (TNR) TN/(TN+FP) Target class Abnormal (positive) 72 0
Accuracy (ACC) (TP+TN)/(TP+TN+FP+FN) Normal (negative) 0 18

TP (True Positive): correctly classified positive cases, TN (True Negative): correctly clas-
sified negative cases, FN(False Negative): incorrectly classified positive cases, FP(False
Positive): incorrectly classified negative cases
is 93.06%, 98.61% and 100%, respectively. We have also
compared the ROC curves obtained by LS-SVM classifier with
2) Results and Discussion: The proposed method utilizes the three kernels and are shown in Fig. 4. Table VIII presents
the coefficients of LL3 sub-band as the primary features of the classification performance comparison of our proposed
each MR image. However, the size of the feature space is 1444 method with the existing schemes. It is observed that the
for Daubechies-4 wavelet, which is quite large for computa- suggested scheme is superior to its competent schemes while
tion. Thus, KPCA approach is used to reduce the dimensions it requires relatively less number of features.
of features to only 7. These reduced features are the first
7 principal components(PCs) which are only 0.48% of the TABLE VII. P ERFORMANCE METRICS FOR THREE CLASSIFIERS
primary features. The polynomial kernel has been selected to Classifier Sensitivity (%) Specificity (%) ACC(%) AUC
calculate the kernel matrix in KPCA method. The performance LS-SVM+Linear 93.06 100 94.44 0.965
of the method is tested with different number of principal LS-SVM+Polynomial 98.61 100 98.89 0.986
components to find out the required number of features. It has
LS-SVM+RBF 100 100 100 1
been observed that the proposed system works efficiently with
7 PCs on the given dataset. The LS-SVM classifier has been
trained with the three kernels (linear, polynomial, and RBF).
To estimate the optimal value of the parameters, viz., θ, σ and
ζ, various pairs of the (ζ, θ) and (ζ, σ) are tested and finally 1

the pair with the low error rate is chosen to train the classifier. 0.9

The confusion matrix for linear, polynomial, and RBF kernel 0.8
is illustrated in Table IV, V, and VI, respectively.
0.7

TABLE IV. C ONFUSION MATRIX FOR ‘LS-SVM + L INEAR ’

0.6
CLASSIFIER
Sensitivity

0.5
Output (predicted) class
0.4
Abnormal (positive) Normal (negative)
Target class Abnormal (positive) 67 5 0.3
threshold
LS−SVM+Linear
Normal (negative) 0 18 0.2
LS−SVM+Polynomial
LS−SVM+RBF
0.1

0
TABLE V. C ONFUSION MATRIX FOR ‘LS-SVM + POLYNOMIAL’ 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

CLASSIFIER 1−Specificity

Output (predicted) class

Abnormal (positive)
Target class Abnormal (positive) 71
Normal (negative) 0 18
The performance measures of the three classifiers are
shown in Table VII. From the table, it is evident that the
classification accuracy and AUC value for RBF kernel is higher
than other two kernels. However, all the three kernels achieve
a specificity of 100%. The sensitivity of the three kernels
Fig. 4. ROC curves for LS-SVM classifier with three kernels
Normal (negative)
1
IV. C ONCLUSION
This paper presents an efficient scheme to detect the brain
MR images as abnormal. The scheme uses DWT to extract
features from the images. The features have been normalized
to enhance the efficiency. A computationally less expensive
approach, KPCA is employed to select the most significant fea-
tures from the high-dimensional normalized features. Finally,
LS-SVM classifier has been used to build an automatic and
 TABLE VIII. P ERFORMANCE COMPARISON BETWEEN THE PROPOSED METHOD AND EXISTING SCHEMES

References Total number of images Feature extraction Feature reduction Classifier Number of features ACC (%)
Chaplot et al., 2006 [3] 52 DWT — SVM with RBF 4761 98
El-Dahshan et al., 2010 [7] 70 DWT PCA FP-ANN 7 97
El-Dahshan et al., 2010 [7] 70 DWT PCA k-NN 7 98.6
Zhang et al., 2011 [9] 66 DWT PCA FNN with SCABC 19 100
Zhang et al., 2011 [4] 66 DWT PCA FNN with SCG 19 100
Zhou et al., 2015 [13] 64 Wavelet entropy — Naive Bayes 7 92.60
Zhang et al., 2015 [14] 66 Wavelet energy — SVM 6 82.69
Proposed method 90 DWT KPCA LS-SVM+Linear 7 94.44
LS-SVM+Polynomial 7 98.89
LS-SVM+RBF 7 100

accurate CAD system for brain MR image classification. The
classification accuracies of LS-SVM with respect to the linear,
polynomial, and RBF kernel are 94.44%, 98.89% and 100%,
respectively. The results show the efficacy of the suggested
scheme with considerably less number of features as compared
to other schemes. Though the feature reduction technique
and the classifier used in this paper are less expensive than
the schemes proposed in the literature, however, the feature
extraction step is more time consuming. The proposed work
can be experimented with larger datasets, and the performance
of the feature extraction stage can be enhanced using other
advanced transformation techniques.
R EFERENCES
[1] E. A. El-Dahshan, H. M. Mohsen, K. Revett, and A. B. M. Salem,
“Computer-aided diagnosis of human brain tumor through MRI: A
survey and a new algorithm,” Expert Systems with Applications, vol. 41,
no. 11, pp. 5526–5545, 2014.
[2] C. Westbrook, Handbook of MRI technique. John Wiley & Sons, 2014.
[3] S. Chaplot, L. M. Patnaik, and N. R. Jagannathan, “Classification of
magnetic resonance brain images using wavelets as input to support
vector machine and neural network,” Biomedical Signal Processing and
Control, vol. 1, no. 1, pp. 86–92, 2006.
[4] Y. Zhang, Z. Dong, L. Wu, and S. Wang, “A hybrid method for MRI
brain image classification,” Expert Systems with Applications, vol. 38,
no. 8, pp. 10 049–10 053, 2011.
[5] S. Das, M. Chowdhury, and K. Kundu, “Brain MR image classification
using multiscale geometric analysis of ripplet,” Progress In Electromag-
netics Research, vol. 137, pp. 1–17, 2013.
[6] M. Maitra and A. Chatterjee, “A Slantlet transform based intelligent
system for magnetic resonance brain image classification,” Biomedical
Signal Processing and Control, vol. 1, no. 4, pp. 299–306, 2006.
[7] E.-S. A. El-Dahshan, T. Honsy, and A.-B. M. Salem, “Hybrid intel-
ligent techniques for MRI brain images classification,” Digital Signal
Processing, vol. 20, no. 2, pp. 433–441, 2010.
[8] Y. Zhang, S. Wang, and L. Wu, “A novel method for magnetic resonance
brain image classification based on adaptive chaotic PSO,” Progress In
Electromagnetics Research, vol. 109, pp. 325–343, 2010.
[9] Y. Zhang, L. Wu, and S. Wang, “Magnetic resonance brain image
classification by an improved artificial bee colony algorithm,” Progress
In Electromagnetics Research, vol. 116, pp. 65–79, 2011.
[10] Y. Zhang and L. Wu, “An MR brain images classifier via principal
component analysis and kernel support vector machine,” Progress In
Electromagnetics Research, vol. 130, pp. 369–388, 2012.
[11] M. Saritha, K. P. Joseph, and A. T. Mathew, “Classification of MRI
brain images using combined wavelet entropy based spider web plots
and probabilistic neural network,” Pattern Recognition Letters, vol. 34,
no. 16, pp. 2151–2156, 2013.
[12] Y. Zhang, Z. Dong, S. Wang, G. Ji, and J. Yang, “Preclinical diagnosis
of MR brain images via discrete wavelet packet transform with tsallis
entropy and generalized eigenvalue proximal support vector machine
(GEPSVM),” Entropy, vol. 17, no. 4, pp. 1795–1813, 2015.
[13] X. Zhou, S. Wang, W. Xu, G. Ji, P. Phillips, P. Sun, and Y. Zhang,
“Detection of pathological brain in MRI scanning based on wavelet-
entropy and naive bayes classifier,” in Bioinformatics and Biomedical
Engineering, 2015, pp. 201–209.
[14] G. Zhang, Q. Wang, C. Feng, E. Lee, G. Ji, S. Wang, Y. Zhang, and
J. Yan, “Automated classification of brain MR images using wavelet-
energy and support vector machines,” in 2015 International Confer-
ence on Mechatronics, Electronic, Industrial and Control Engineering
(MEIC-15), 2015.
[15] S. Mallat, A wavelet tour of signal processing. Academic Press, 1999.
[16] S. Theodoridis and K. Koutroumbas, Pattern recgnition. Academic
Press, 2008.
[17] B. Schölkopf, A. Smola, and K.-R. Müller, “Kernel principal component
analysis,” in Artificial Neural NetworksICANN’97, 1997, pp. 583–588.
[18] ——, “Nonlinear component analysis as a kernel eigenvalue problem,”
Neural computation, vol. 10, no. 5, pp. 1299–1319, 1998.
[19] C. M. Bishop, Pattern recognition and machine learning. Springer,
2006.
[20] J. A. K. Suykens and J. Vandewalle, “Least squares support vector
machine classifiers,” Neural processing letters, vol. 9, no. 3, pp. 293–
300, 1999.
[21] “Harvard medical school data,” http://www.med.harvard.edu/AANLIB/.
 Introducing FERPS: A Framework for Enterprise
Resource Planning Selection
ON Singh
School of Management, Information Systems and
Governance
University of Kwa Zulu-Natal
Durban, South Africa
omesh.narain.singh@gmail.com
Abstract—Enterprise Resource Planning (ERP) systems, over
the years, have evolved into integral business systems in both
medium and large organizations. The strength of an ERP is
that it is a computerized transactional information system,
with a centralized data repository. This allows for significant
data availability and seamless collaboration between business
functions. Realizing associated benefits through an ERP
implementation is challenging, as experienced by a South
African water utility. On the establishment of the utility the
failure to implement an ERP system successfully, forced the
organization to critically review the position it found itself in.
This study identifies some of the critical shortfalls in the initial
ERP implementation, and introduces the process of the creation
of an ERP evaluation framework, named FERPS, that assists in
ERP evaluation and selection, prior to acquisition. Ultimately,
FERPS assists in providing an organization with evaluation
results of ERP systems to support the implementation of an
optimal ERP system, based on organizational fit.
Keywords—evaluation framework; enterprise resource
planning (ERP) systems; ERP selection; ERP modules; ERP
procurement
I. INTRODUCTION
With the implementation of ERP systems,
companies are able to better manage their resources,
standardise business processes across the organisation and
assist management with real time financial and production
information [9]. There are many commercial off-the-shelf
(COTS) software packages available in the market that
organisations can configure to their specific needs with
inherent benefits such as outsourced system maintenance,
support and system improvements. Some of these COTS
software include Oracle, SAP, BAAN, MS Dynamics etc.
Organisational reasons for implementing ERP systems may
vary from replacing legacy systems to transforming business
operations [15].
When considering the public sector, citizens in
many countries expect their administration to provide
quality services that are adapted to the technological
environment at the lowest cost. This expectation is in
response to the business practices in the private sector where
evolved business methods, techniques and practices appear
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
UG Singh
School of Management, Information Systems and
Governance
University of Kwa Zulu-Natal
Durban, South Africa
singhup@ukzn.ac.za
promising, more so in the area of information systems and
information technology [16]. Researchers argue that IT
management in the public sector is ―more of the same‖
when compared to the private sector whilst others view the
public sector as ―a whole new ball game‖ [17].
The water utility under study was established in 2004,
and the following business units were created - strategy and
leadership, customer services, engineering, operations,
environmental services, human resources, corporate services,
finance and technology. In order to operate a sound business,
with strong governance principles, it was imperative for this
water utility to implement financial systems to account for
business transactions. Thus, the utility implemented an ERP
system, where the acquisition process was out-sourced to a
consulting company. At the time the utility lacked the
resources and technically skilled staff to champion this
project.
The background of this particular water utility is that it is
responsible for the provisioning of water, waste water
management and other environmental services. These
services include the supply of water, from the source to
consumer, from sewer connection to waste water treatment
and finally, disposal into the environment. These water
services are provided to a range of consumers including
community households, industries, businesses, and
government institutions. The highest governance structure of
a water utility is usually composed of a board of directors.
The management of this particular water utility is divided
among two committees; the leadership committee (Leadco) -
who is responsible for executing the overall business strategy
set by the board of directors, and a management committee
(Manco) - whose responsibility it is to manage the
operational activities of the utility.
However, the ERP acquisition by the consulting
company, resulted in an ERP implementation failure, as the
system was unsuccessful in satisfying the utility‘s
operational and strategic objectives. Post implementation
review of the ERP implementation identified two major
shortcomings:
 1. Manco team members were unable to manage their
budgets and operational activities as the ERP was
unable to provide them with reliable reporting.
2. The finance team reported that they were unable to
reconcile creditors, debtors, and bank statements.
Further, they complained of poor training and more
importantly, the inability to process transactions
correctly with the selected ERP system.
Subsequent analysis by the Leadco team, in
consultation with the Manco team, at the utility,
detailed seven major shortcomings, in the process of
selecting and acquiring the current ERP system. These
shortcomings are summarized in Table 1, and discussed
briefly below.
Table 1: Weaknesses identified in the ERP implementation process
Shortfalls
1 Poor project planning
2 Insufficient user training
3 Poor testing procedures
4 No user sign-offs
5 Lack of management involvement
6 Lack of vendor technical insight
7 Lack of utility technical skill
Shortfalls Description Result
Poor project planning existed since basic project
planning techniques were not instituted at any phase of the
process of implementing the ERP system at the utility.
Furthermore, project deadlines were not met and not all
relevant modules were procured. ERP implementation
methodology typically consists of defined steps that can be
managed throughout the project in order the guide the
process to success. Reference [1] describe these steps as:
project definition, requirements, solution design, build,
testing, production (go-live) and rollout.
Insufficient user training was provided, and where
informal training did occur, this was in an environment that
lacked proper test data. Hence, this resulted in transactions
being captured incorrectly when the ERP system was
implemented. The object of training is to familiarize the
users with the standard functionality, basic look and feel and
the methods of navigation of the ERP. This should be an
integral part of the project planning phase. According to [2]
appropriate and timely training is essential to
implementation success as training assists in improving the
quality of project results and to meet user expectations.
Poor testing procedures were followed. Insufficient
testing scenarios and procedures due to a lack of testing
data, led to poor testing procedures being implemented. This
resulted in time wasted as transactions that were captured
incorrectly due to incompetence, had to be re-captured
correctly. The testing phase of ERP implementation
attempts to emulate how the software would run in a
production environment. The testing of COTS software
generally consists of unit testing - the testing of individual
transactions, cycle or scenario testing - that will test a chain
of transactions that constitute a business process, and
integration testing - which includes interfaces and final
testing [3].
No user sign-offs occurred upon completion of modules,
hence there was a lack of accountability and ownership by
stakeholders. According to [4] and [5] the project sign-off
steps should be carried out after the successful completion
of all implementation phases as required by the
implementation methodology.
Lack of management involvement was distinctly
noticeable. Senior management had minimal oversight of
the overall implementation. This allowed vendors to make
decisions in isolation, to their benefit, and not necessarily in
the best interests of the utility. The role of top management
is critical to ERP implementation success and the lack
thereof lends to deviation from the project goals. The
organization suffers from a gap between what the ERP
promises and the actual realized benefits on completion of
the implementation. The role of the ERP sponsor in critical
in driving a successful system implementation [4].
Vendor lacked technical insight, as they were from
an engineering background, hence they were lacking in
terms of business and financial processes, specific to the
water utility. Therefore, post implementation the utility
faced many financial reporting errors and difficulties, such
as:
 inability to reconcile general ledger to sub-ledger
entries with respect to debtors and creditors.
 the cash book could not be reconciled.
 data was corrupted as the data input via the
backend, not complying to data standards.
 poor financial reporting , especially since financial
year end statements could not be compiled.
Experienced external consultants generally possess
specialized skill and knowledge which can be applied to
rapidly solve technical problems, shorten implementation
time and help realise higher quality results. Appointing
experienced consultants results in a successful ERP
implementation effort [6].
Utility lacked technical skill, especially because it
was newly established. All Finance functions of the utility
during that period were not carried out internally – many of
these functions were outsourced. Further, some finance staff
required reskilling as they failed to execute the processes as
required in the ERP system.
A skilled team lends to dynamic collaboration, employees at
different levels share knowledge and best practices adding
value to the ERP implementation process [1].
 It is for the above reasons that the utility was unable to
operate effectively. Collectively, the overall result of this
ERP implementation was counter-productive, especially
since the ERP implementation resulted in the utility‘s
inability to reconcile the general ledger and sub-ledgers;
report accurate financial information; ascertain actual
expenditure to budget. This resulted in the utility being
unable to produce accurate annual financial statements.
According to [11] the top five benefits that should result
from such an implementation, in the first year, is listed as:
integration of applications, organizational learning, internal
information sharing, accuracy and timeliness of information
and data/transaction processing time. The utility failed to
achieve these benefits.
Following this unsuccessful implementation of an ERP
system at this water utility, the researcher set out to identify
the criteria that are essential to evaluate an ERP system,
prior to implementation, to ensure a ‗best-fit‘ in the
adoption process. The developers and suppliers of ERP
systems have a perspective that a plain vanilla approach that
results in a ―one size fits all‖. This perception arises from
the argument that systems operate on the most widely
accepted best practice [10] [13]. This proved incorrect in the
case of the utility.
This research reports on the process of identification, as
well as the criteria identified, which have been incorporated
into FERPS (Framework for Evaluating an Enterprise
Resource Planning System)
II. LITERATURE REVIEW
Enterprise Resource Planning Systems are computer
based information systems that have evolved over the years
to process organisational transactions and facilitate
integrated real-time production planning, customer response
and reporting [7].
Over the years the largest investment in information
systems software has been in ERP systems [8]. With the
implementation of ERP systems, companies are able to better
manage their resources, standardise business processes
across the organisation and assist management with real time
financial and production information [9].
Essentially an Enterprise Resource Planning (ERP)
system is a computer-based software that consists of various
modules that relate to specific business functions and
transactional information that is stored in a central database.
This allows for significant data availability and collaboration
between business functions.
Although ERP systems bring inherent competitive
advantages, there also exists a high risk of failure. About
70% and ERP implementations fail to deliver the expected
benefits and three quarters of ERP implementations are
unsuccessful. On average these projects where 178% over
budget and took 2.5 times longer than initially projected and
delivered a mere 30% of the promised benefit [7].The
implication of these statistics implies that ERP
implementations projects are very complex, costly and
require considerable planning.
No literature was found that comprehensively identified the
criteria that specifically need to be evaluated prior to
investing in the purchase of an ERP system. As part of the
process of selecting and evaluating an ERP solution, this
South African water utility, designed a framework to include
a comprehensive set of categories and criteria that can assist
organisations to determine the best ERP system for
individual organisational fit.
Hence through this study, the researchers will address this
gap, and add to the body of literature on the acquisition of
ERP systems.
III. RESEARCH DESIGN AND METHODOLOGY
This research followed an exploratory approach to
identifying the criteria that are essential for inclusion in a
framework that can be used to evaluate ERP system
suitability, prior to their purchase.
Since the current ERP implementation process, was
regarded as a ‗failure‘, as the current system was not fully in
use in the utility, senior management at the utility initiated
this ‗criteria determination process‘, based on commercial
off-the-shelf (COTS) ERP modules.
The newly appointed IT manager was appointed as the
project champion and was tasked to project manage the data
collection process. The key categories that were identified
by the IT manager for data collection included the processes
of ERP evaluation framework development, ERP
evaluation, ERP selection and ERP implementation.
Data was collected using non-probability purposive
sampling. Key role players were identified based on their
field of expertise and functional responsibilities in the
utility. These participants included the chief financial officer
(CFO), income accountant, expenditure account,
procurement officer, asset officer and budget officer of the
utility.
Six semi-structured interviews were conducted to
determine what modules and scope were essential to the
water utility.
Thereafter a second round of interviews was conducted.
These purposive in-depth interviews were conducted with
finance departmental heads and key staff. The purpose was
to collect data on the actual business processes of the utility.
Interviewees were asked specific questions regarding the
modules applicable to their job functions and processes.
 The information from both these sets of interview was Request for quotation (RFP)
qualitative in nature. Hence narrative analysis was Purchase order
conducted and responses were grouped into themes. Emergency orders
Essentially these themes formed the categories of FERPS. Receiving
Goods returns note
Vendor master file
Subsequent narrative analysis of the same data, once Inventory
again through thematic representation, identified the Inventory master file
individual criteria essential in each category. Inventory reorder
Dispatch
These categories and criteria formed the basis of the Stock takes
modular framework FERPS. Reporting requirements

IV. FINDINGS Contract Management New contracts

Based on the 2 rounds of interviews that were conducted, Existing contracts
Deletion of contracts
eleven essential categories were identified - Supply Chain
Call-offs
Management, Contract Management, Accounts Payable, Reports
Accounts Receivable, Fixed Assets, Project Finance, Cash Escalation on contracts
Management, Interfaces, Budgeting, Management Accounts Payable Invoice matching
Information and Information Technology. These categories Debit invoicing (no purchase order)
were identified as essential modules in any ERP system, Debit & credit notes
specifically motivated by the experiences of the participants Accruals journals
with the ‗non-operational‘ ERP system. Prepayments
Reporting requirements
Leadco in consultation with the IT manager agreed that Views
this approach was fair since majority of ERP solutions are Accounts Receivable Debtors master file
Invoicing
designed and coded in a modular functional architecture. Lab debtors
Water debtors
The Scope of each of these eleven modules in FERPS is Effluent debtors
presented in Table 2. Sundry
Staff
Any ERP system comprises of modules – each module Invoice print out
referring to a specific business function - for which Receipts
Late payments
applications are developed to support the business processes
Statements
[12], [1]. For example, Supply Chain Management (SCM) is Credit notes
a business function for which the processes are encapsulated Journals
in a configurable software application. Due to the Key reports used
configuration capability of ERP systems, these modules can Fixed Assets Asset Masterfile
be setup in several ways, depending on the organizational Depreciation
requirement, thus allowing for customisation, based on the Disposal
way in which business processes are carried out [14]. Repairs and maintenance
WIP and forecasting
The scope of the module refers to the functional business Revaluation
processes within the module. For example, within the SCM Impairment
Wear and tear
module one of the business processes would be
Journals
procurement. Procurement is the business process that Reporting requirements
facilitates the acquisitioning of goods and services. Thus the General
scope of the SCM module, would include Procurement, as Project Finance General requirements
one its functional business processes. Project details capture
Project management - approval
The Modules and Scope identified in the round of routing
interviews are presented in Table 2. Budget capture
Cashflow
Table 2: Modules and Scope required in an ERP system for a water Purchase requisitions and orders
utility Invoicing
Payment - internally funded projects
Modules Scope Externally funded projects
Supply Chain Management Procurement Credit notes
Purchase requisition Reporting requirements
 Summary views
Cash Management General requirements
Borrowings - loan schedule &
interest accrual
Interest accrual schedule
Reporting requirements
Investments
Interfaces GIS, Payroll, HR
Budgeting Operational costs
Income
Forecasting
Dashboards
General
Management Information General ledger codes
Journals
Accounting standards
Financial calendar
Period closure procedures e.g.
Month-end requirements
Data take-on
Cost and management accounting
Financial statements
General reporting requirements
Information Technology User profile management
Login procedures
Environments
Archiving
Paper types
Menu allocation
General
Technical
Once the modules and the scope were identified, it was
necessary to get specific detail on the actual of business
rules and procedures that are followed in each module.
These would constitute the actual evaluation criteria. For
example, service providers to the utility must be managed
and maintained in a master file in the module scope. This
would form a criteria in the of supply chain management,
procurement module.
Thus, to complete FERPS, the second round of
interviews conducted provided the comprehensive overview
of the ERP modules, scope and criteria that would constitute
the key attributes of FERPS.
FERPS has been designed as a comprehensive
framework, hence the number of criteria across all
categories exceeds 100. The summary of the criteria in each
module is presented in Table 3, below, to provide the reader
with the overview of what needs to be evaluated in each
module, prior to the ERP purchase decision.
Table 3: Criteria of requirements for an ERP system for a water
utility
A SUPPLY CHAIN MANAGEMENT
A1 PROCUREMENT
A1.1 General Requirements
The general requirements for the Procurement function of
the SCM identified 10 criteria. These criteria covered
aspects of web-based requisition creation, routing,
delegation of authorities and requisition querying.
A2 PURCHASE REQUISITION (PR)
A2.1 Non-Stock Items (Directs)
Requisitioning, 17 criteria, for non-stock capital project
items. These criteria all for the requisitioning process,
capturing thereof, approvals and notifications.
A2.2 Request for Quotation (RFQ)
12 Criteria for the creation of request for quotations from
purchase requisitions, capturing thereof, checking budgets
and approvals.
A3.1 Purchase Order
Creating of purchase orders from request for quotation. 11
criteria from generating purchase orders, approving and
communicating via fax/e-mail.
A3.2 Emergency Orders
5 criteria for the creation of emergency orders for vendor
not on vendor master file.
A4.1 Receiving
Receiving as a result of purchase order process. 13 criteria
on the receiving process including full and partial orders.
Processing the purchase order and the associated financial
postings.
A4.2 Goods Returns Note
4 Criteria for the process of returning ordered goods and
associated financial postings.
A5 Vendor Master File
Management of vendor master information. 15 criteria with
rules and required information for vendor maintenance.
A6 INVENTORY
A6.1 Inventory Master File
Management of inventory master information. 10 criteria
with rules and required information for inventory
maintenance.
A6.2 Inventory Reorder
3 criteria for the reorder (requisitioning) for inventory
items that have reached minimum stock levels.
A6.3 Dispatch
9 criteria for the requisitioning of goods from stock items.
A6.4 Stock Takes
4 criteria for the management of stock takes, scheduling
and stock information.
A7 Reporting Requirements
22 criteria on the reporting requirements.
B CONTRACT MANAGEMENT
B1 New Contracts
30 Criteria for the management and capturing of new
contracts
B2 Existing contracts
2 criteria for the viewing and queries of existing contracts.
B3 Deletion of contracts
3 criteria for the deletion of contracts.
B4 Call-offs
4 Criteria for the termination of contracts.
B5 Reports
6 criteria on reporting requirements.
B6 Escalation on contracts
2 criteria for the change in price of contracts.
B7 Contract extension
1 criterion for the extension of contracts.
C ACCOUNTS PAYABLE
C1.1 Invoice Matching
 Matching invoices to the goods received note. 20 criteria to 5 criteria for the generation of credit notes and associated
manage in process and associated financial postings. financial postings.
C1.2 Debit Invoicing (No Purchase order) D10 Journals
4 criteria for the processing of invoices with no goods 5 criteria for the creation and posting of journals.
received note. D11 Reporting Requirements
C2 Debit & Credit Notes 12 criteria on the reporting requirements.
3 criteria for processing of debit and credit notes and E FIXED ASSETS
associated financial postings. E1 General
C3 Accruals Journals 10 general criteria for the management of fixed assets.
4 criteria the performing accruals on outstanding orders, E2 Asset Masterfile
outstanding goods received note, unposted invoices and Management of asset master information. 16 criteria with
invoices without orders. rules and required information for asset maintenance.
C4 Prepayments E3 Depreciation
2 criteria for the payment of outstanding invoices. 5 criteria for the rules on the calculation of depreciation.
C5 Payments E4 Disposal
16 criteria for the payment of invoices, approvals, bank 3 criteria for the disposal of assets.
uploads and associated financial postings. E5 Repairs & Maintenance
C6 Reporting Requirements 1 criterion for the updating of assets based on repairs and
17 criteria on the reporting requirements. maintenance.
C7 Views E6 WIP & Forecasting
7 criteria on application views of transactions, purchase 5 criteria for the updating of asset information based on
orders, general ledger codes etc. work-in-progress.
D ACCOUNTS RECIEVABLE E7 Revaluation
D1 Debtors Master File Accommodate for revaluation of assets as and when
Management of debtor master information. 9 criteria with required.
rules and required information for debtor maintenance. E8 Impairment
D2 Tariffs Computation Cater for impairments of assets.
13 criteria to facilitate computation of various tariffs E9 Wear And Tear
charged to customers based on the budgeted figures i.e. Cater for wear and tear.
operational costs and budgeted volumes. E10 Journals
D3 Tariffs Maintenance Table 6 criteria to create journals on additions, depreciation,
6 criteria to maintain applicable tariffs. disposals and transfers of assets.
D4 Water And Waste Water Volumes Table E11 Reporting Requirements
10 criteria for the capture and maintenance of water 11 criteria on the reporting requirements.
volumes, licenses and budgeted volumes. F PROJECT FINANCE
D5 Invoicing F1 General
D5.1 General 3 criteria for general project finance requirements.
11 criteria for the invoicing/billing of debtors. F2 Project Details Capture
D5.2 (1) Lab Debtors 7 criteria detailing the capture of project details.
7 criteria for the invoicing of laboratory debtors. F3 Project Management - Approval Routing
D5.3 2) Water User to define approval routings, authorisers & delegated
16 criteria for the invoicing of water debtors. authority values on the project.
D5.4 (3) Effluent F4 Budget Capture
7 criteria for the invoicing of effluent debtors. 8 criteria on the capturing of project budgets.
D5.5 (4) Sundry F5 Cash flow
2 criteria for the invoicing of sundry debtors. Generation of straight-line cash flow for each set of budget
D5.6 (5) Staff values.
4 criteria for the invoicing of staff debtors. F6 Purchase Requisitions & Orders
D5.7 6)Invoice Print Out 10 criteria on the capture purchase order against project
3 criteria for the generation of a pre-numbered tax invoice allocated budget.
with all the details of applicable debtor. F7 Invoicing
D6 Receipts 15 criteria on the project invoicing.
5 criteria for the capturing of receipts against the respective F8 Payment - Internally Funded Projects
debtors account into the system and associated financial Payment process as accounts payable.
postings. F9 Payment For - Externally Funded Projects
D7 Late Payments 3 criteria for payments of externally funded projects.
5 criteria for the automatic generation of reminder letters F10 Credit Notes
for overdue payments. Ability to process credit notes
D8 Statements Reporting Requirements (For Both Internally And
3 criteria for the process of generating consumer F11 Externally Funded Projects)
statements. 15 criteria for reports printed according to pre-defined
D9 Credit Notes
 parameters. I1 User Profile Management
F12 Reporting Requirements 16 criteria on the management of user profiles in terms of
25 criteria on the reporting requirements. security access and functionality.
F13 Summary Views K2 Login Procedures
Views on Budgets vs. PO's, PO's vs. Invoices and Invoices Criteria for architecture related login.
vs. Payments K3 Environments
G CASH MANAGEMENT Criteria for demo, test and live environments.
G1 General K4 Archiving
7 criteria for the processing of cash receipts, payments, Capability to archive event logs and audit trails.
EFT transfers, standing orders, the integration with bank K5 Paper Types
account and associated financial postings. Criteria for printing and paper types.
G2 Borrowings - Loan Schedule & Interest Accrual K6 Menu Allocation
Criteria for a loans maintenance function. Criteria for configuration of menu allocation.
G3 Interest Accrual Schedule K7 General
System should calculate on a monthly basis the interest on Audit trail on changes to all master file information.
loans and associated financial postings. K8 Technical
G4 Reporting Requirements Criteria detailing organisational IT infrastructure
2 criteria on the reporting requirements. architecture.
G5 Investments
The system should automatically allocate net interest to the As presented in Table 3, Supply Chain Management forms
various schemes. part of the procurement process. This module allows the
H INTERFACES user to create process, manage and report on requisitions,
H1 General
purchase request processing, purchase order processing,
A list of systems that the ERP is required to interface with.
receiving and inventory control. Contract Management
I BUDGETING
I1 Operational Costs
manages all types of contracts related to supply and delivery
6 criteria for operational budgeting. of operations, installments, rentals and leasing. Accounts
I2 Income payable records and manages vendor accounting data while
Interface with other systems and allow for information to Accounts receivable records and manages customer
be downloaded and uploaded from the ERP system. accounting data. The Fixed Assets module allows for the
I3 Forecasting tracking, analyze, manipulate and report on fixed assets. The
Facility to allow preparation of business forecast/plan up to Project Finance module is used to manage project
5 years. definitions, costing, accounting, portfolio management,
I4 Dashboards resource management and billing. Cash Management is the
Allows management to view budgets based on real time bank control system that is used to reflect all transactions
information from the financial system. that affect the cashbook that is linked to the organisations
I5 General bank accounts. This module is used for the generation of
8 criteria listing general budgeting requirements.
payments that are generated by sub-modules. The Interfaces
J MANAGEMENT INFORMATION
module refers to standard interfaces to send and accept data
J1 General Ledger Codes
5 criteria on the management and setup of the chart of
from a number of external systems. The Budgeting module
accounts. is a process within the ERP that allows for the capture,
J2 Journals amend, track revisions and approval of budgets. The
Criteria for accounting journals. Management information module provides comprehensive
J3 Accounting Standards financial information that would include financial and
The system must be in compliance with GRAP, PFMA, management reporting, financial planning, budgeting,
MFMA and Water Services Act. forecasting, profitability management, overhead costs,
J4 Financial Calendar working capital and cash flow management. Finally, the
Criteria for management of financial calender. Information Technology module relates to the associated
J5 Period Closure Procedures technology requirements in terms of setup and management.
Criteria for month end financial closures.
J6 Data Take-On V. CONCLUSION
Data take on should be for the prior and current year, The creation of FERPS has been advantageous to
actuals, as well as the current year budget.
the water utility for the following reasons;1. Senior
J7 Cost And Management Accounting
Criteria for the allocation of costs.
management involvement in the process created a sense of
J8 Financial Statements ownership and accountability, compared to their minimal
6 criteria for the generation of financial statements. involvement in the initial failed ERP implementation. 2.
J9 General Reporting Requirements Since there was a wider range of user participation in the
16 criteria on the reporting requirements. process, a greater buy-in was created with all stake holders,
K INFORMATION TECHNOLOGY including organizational leadership, executive management
and finance staff. 3. The process allowed for technical
operational input from finance staff (chief financial officer,
income accountant, expenditure accounting, asset officer,
budget officer, finance clerical staff and management).
The essence of ERP technology is to support business
processes and organisational strategies. The benefits realised
from ERP systems are directly related to the result of
effective preparation and implementation [6]. The
introduction of FERPS forms an integral aspect in the
planning and preparation process of ERP implementation.
VI. FUTURE RESEARCH
FERPS can be converted into an electronic evaluation
tool that can serve a dual purpose. One, being an evaluation
tool to determine ‗best-fit‘ by the utility, and the other
providing the implementation specification for the
implementing vendor. This would provide the vendor with
the list of requirements for each module setup or a
functional checklist.
FERPS can also be adopted as a testing procedure for
sign-off, when the end-user tests the functionality of the
modules against the framework specification.
Finally, since FERPS was based on the requirements of
the utility and was also guided by the municipal finance
management act (MFMA - The MFMA is piece of South
African legislation that aims to put in place finance
governance regulations for local government by clarifying
and separating the roles and responsibilities of all
stakeholders) future research could apply the framework to
other municipal entities in South Africa, that may be
interested in implementing an ERP system, and further
develop FERPS making it more comprehensive.
[5] Strauss, E., 2012. Project management: guidelines for the
implementation methodology of an enterprise resource
management system.
[6] Steyn, E.A., 2008. A framework for assessing enterprise
resource planning systems.
[7] Amid, A., Moalagh, M. & Ravasan, A.Z., 2012. Identification
and classification of ERP critical failure factors in Iranian
industries. Information Systems, 37(3), pp.227–237.
[8] De Carvalho, R.A. & Johansson, B., 2009. Enterprise Resource
Planning Systems for Small and Medium-Sized Enterprises.
Handbook of Research on Software Engineering and
Productivity Technologies: Implications of Globalization:
Implications of Globalization, p.373.
[9] Nah, F.-H. & Delgado, S., 2006. Critical success factors for
enterprise resource planning implementation and upgrade.
Journal of Computer Information Systems, 46(5), p.99.
[10] Edwards, C.A., 2013. Business process integrity and enterprise
resource planning systems: an analysis of workaround practices
in a large public sector organisation.
[11] Hart, C.A., 2010. The organisational performance impact of
ERP systems on selected manufacturing companies within
South Africa.
[12] Khan, O.U., 2012. Evaluation of an enterprise resource planning
system as a performance enhancing tool for non-profit
organisations.
[13] Mukwasi, C.M., 2013. Enterprise resource planning business
case considerations: an analysis of Small to Medium-Sized
Enterprises in developing countries.
[14] Seo, G., 2013. Challenges in implementing enterprise resource
planning (ERP) system in large organizations: similarities and
differences between corporate and university environment.
[15] Tazyeen, F., 2012. Modeling government ERP acquisition
methods using system dynamics.
[16] Uwizeyemungu, S. & others, 2005. Motivations for erp
adoption in the public sector: an analysis from ―Success
Stories‖.
[17] Zmud, R., Carte, T. & Te‘eni, D., 2004. Information systems in
nonprofits and governments: do we need different theories?
ICIS 2004 Proceedings, p.86.

Acknowledgments
The water utility - consisting of the board, leadership
committee and management committee, chief financial
officer and finance staff for allowing the data collection.

References
[1] Ganesh, K. et al., 2014. Enterprise Resource Planning:
Fundamentals of Design and Implementation, Springer.
[2] Turton, J.W., 2010. A manager‘s view of critical success factors
necessary for the successful implementation of ERP.
[3] Thomas, G. & Jajodia, S., 2004. Commercial-off-the-shelf
enterprise resource planning software implementations in the
public sector: practical approaches for improving project
success. Journal of Government Financial Management, 53(2),
pp.12–19.
[4] Al Rashid, W., 2013. Managing stakeholders in enterprise
resource planning (ERP) context-a proposed model of effective
implementation.
 Performance Evaluation of Matrix- and EXIF-Based
Video Rotation Methods in MJPEG-Based Live
Video Encoder
D. Skvorc, M. Ostrek, M. Horvat
University of Zagreb
Faculty of Electrical Engineering and Computing
Zagreb, Croatia
dejan.skvorc@fer.hr, mirela.ostrek@fer.hr, matija.horvat@fer.hr
Abstract—Contemporary camera-equipped wearable devices,
such as smartphones and tablets, are now powerful enough to be
used as ubiquitous remote sources of live video streams.
However, as these devices may operate in various positions,
including upside down or rotated, encoding the captured video
into a data format suitable for live streaming also includes the
rotation of video frames to compensate the device rotation effect.
In this paper, we analyze the impact of the video rotation
procedures to the frame rate of the produced live video. MJPEG
is used as a video encoding format. We compare the performance
of two video rotation methods: classical pixel-level rotation based
on rotation matrix and JPEG tagging using EXIF metadata. The
results show that EXIF-based tagging outperforms matrix-based
calculations significantly, and gets superior as video resolution
increases. However, since EXIF metadata are not yet widely
supported in modern web browsers, if compatibility with open
user community is required, matrix-based rotation is the option
to go with.
Keywords—live video streaming; video rotation; MJPEG;
rotation matrix; EXIF metadata; Android
I. INTRODUCTION
Many people today are using smartphones, tablets, and
similar forms of mobile computers in every aspect of their
lives as a standard and undetachable personal outfit [1, 2, 3].
Besides the significant amount of computing power, storage
capacity, and communication capabilities, these devices are
equipped with a set of embedded sensors. This makes them
not only sufficient for being reliable on-site assistants for their
owners, but also powerful enough to deliver information
collected from the environment to the remote users through
the Internet. Having an open community of smartphone users
involved as sources of environmentally collected data reduces
the need for a dedicated infrastructure preinstalled throughout
the environment, yet providing a satisfactory level of territory
coverage [4, 5, 6, 7, 8].
In this paper, we focus on smartphone and tablet devices
operating as mobile producers of live video streams. This
means that video frames captured by embedded device camera
are delivered in real time to the remote users through the
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Internet. Depending on a particular system configuration,
mobile device can either operate as an independent video
streaming server where remote clients are directly connected
to or may deliver the video stream to a proxy server which is
further responsible to handle client requests. Regardless of
these variations in system deployment, two general
requirements were taken into the consideration while seeking
for an appropriate solution for implementation of live video
streaming application that runs on a smartphone:
 lightweight implementation of a live video encoder
 reuse of a regular web browser with no additional plug-
ins as a live video player
Since live video encoding and streaming is a heavy-duty task
even for powerful desktop machines, the former is important
to achieve a satisfactory level of system performance while
running on a resource constrained mobile device. The latter
enables the consumption of the produced live video from
various types of access devices, such as desktop or mobile
computers, running any general-purpose operating system,
without the need for installation of application-specific
components, except a regular web browser.
To support as many browsers as possible in their up-to-
date but also older versions, we were seeking for a video
encoding format that is suitable for live streaming over the
standard HTTP (HyperText Transfer Protocol) protocol [9,
10]. MJPEG (Motion JPEG) [11, 12, 13] is a video
compression format that is nowadays widely used by various
video-capture devices, such as digital cameras, IP cameras in
video surveillance settings, and webcams. MJPEG is also
streamable over the HTTP [14, 15] and natively supported by
Safari, Google Chrome, and Mozilla Firefox web browsers, as
well as several standalone video players.
However, having a live video capturing software running
on a mobile device carried around by ordinary people exhibits
one particular aspect that makes it substantially different from
capturing video in usual settings. While video-capture devices
are usually mounted on a fixed spot or handled manually with
their upside up, smartphone users are accustomed to use their
devices in all possible orientations. To make the video
reproducible in the remote player as expected, smartphone-
running live video encoder needs an additional component that
rotates the video frames to compensate the device rotation
effect. While encoding a live video in real time on a mobile
device is a resource-demanding task by itself, an additional
step needed for video rotation further exhausts the device CPU
and reduces the frame rate of the produced video.
For the purpose of our research, we developed an MJPEG-
based live video encoder for Android operating system that
supports two methods of video rotation. The first one is a
classical pixel-level rotation based on rotation matrix that
produces a video playable in many existing video players. The
second one is based on EXIF (Exchangeable Image Format)
tagging which is supposed to be less resource demanding, but
compatible with only a restricted set of video players.
The rest of the paper is organized as follows. In Section 2
we describe the system architecture and emphasize the
importance of video rotation when smartphone-like mobile
devices are used for live video capturing. Section 3 briefly
describes the MJPEG video compression format and how is it
delivered to the web browsers through the HTTP protocol.
Two video rotation methods used in our experiments are Fig. 2. Impact of device orientation on video encoding
presented in Section 4, while their implementation for
Android-based devices are described in Section 5. left or right. Based on their previous experience with majority
Performance comparison of the two video rotation methods is of other mobile applications, they expect the device is going to
given in Section 6. Section 7 concludes the paper. take care of the proper function of the application, regardless
of how it is physically handled. Therefore, to meet user
II. VIDEO ROTATION IN MOBILE LIVE VIDEO STREAMING expectations, live video encoder should be extended with an
additional component that compensates the device rotation
Fig. 1 shows a simplified architecture of the live video effect.
streaming system based on smartphones and tablets acting as
mobile live video streaming devices. Cameras embedded on Fig. 2 shows how different device orientations impact the
mobile devices are used to capture raw video frames, which video encoding process. When user holds the camera in its
are then encoded into MJPEG format and delivered to end normal upside up position, the video is displayed in a player as
users either directly or through a streaming proxy server. End expected, without the need for any modification. However, if
users access the video streams using their web browsers via the user rotates the device, the camera is rotated together with
HTTP protocol. the device and its top edge is no longer aligned with the top
edge of the scene being captured. Without performing the
When capturing video with their mobile devices, users video rotation during the encoding process, the video gets
take the comfort to hold the devices in their hands in all unusually rotated on client side, as shown in the middle row of
possible orientations: normal, upside down, or rotated to the the Fig. 2. Strange effects of camera rotation are compensated
by embedding a video rotation procedure into a video encoder.
Compensated video that meets client expectations is shown in
the bottom row of the Fig. 2.
To rotate the video, the encoder needs information about
the device rotation angle, which is supplied by the device
orientation sensor [16]. As a first step, we normalize the
device rotation angle by transforming it from a continuous [0,
360] degrees range to a discrete space, using (1)

  

where D is the device rotation angle read from the orientation

Fig. 1. Architecture of the live video streaming system based on
smartphone and tablet computers as video capture devices
sensor with the values in range from 0 to 360 degrees, while
N is a normalized angle in discrete space with four possible
values: 0, 90, 180, and 270 degrees. In the second step, we
Mobile device Client browser
calculate the angle that the video should be rotated for to
compensate the device rotation effect. The video rotation
angle is calculated using (2) for the back-facing, and (3) for HTTP request for live video stream
the front-facing camera.
HTTP response
Content Type: multipart/x-mixed-replace
  
JPEG image (frame 1)
<boundary>
  
JPEG image (frame 2)
<boundary>
where  is the video rotation angle used by video encoder,

...
while  is the angular offset of the camera from the normal
orientation of the device. For example, if device with portrait JPEG image (frame N)
<boundary>
New frame
arrived
from camera Fig. 4. Delivery of MJPEG-based live video stream over HTTP

orientation has a camera mounted in portrait, the value of the 

Retrieval of device
rotation angle parameter is 0 degrees. If device with portrait orientation has a
front-facing camera mounted in landscape, the value of the 
Device
parameter is 270 degrees, etc. The value of the parameter is a
rotation angle hardware characteristic of a given combination of the device
Uncompressed
[0, 360] and the camera and can be retrieved from the camera settings
frame data using device camera API. This value does not change during
(YUV, RGB, …) the device life cycle, but varies from device to device.
Normalization of
device rotation
angle
III. MJPEG AND HTTP
Normalized In our experiments, we use MJPEG (Motion JPEG) [11,
device rotation angle 12, 13] as a video encoding format. MJPEG was chosen for
{0, 90, 180, 270}
two reasons: its straightforward implementation on Android
platform, which is a test platform in our experiments, and its
broad client-side support. Most major web browsers and
90, 180, players, including Safari, Google Chrome, and Mozilla
Rotation or 270 Rotation EXIF Firefox, provide native support, while plug-ins are available
angle? type? for the rest.
0 Matrix MJPEG is a video compression format that uses intraframe
method of compression. This means that each video frame of a
Matrix-based digital video sequence is compressed separately and
rotation
independently of all other frames. In MJPEG, each frame is
compressed as a separate JPEG image. The video sequence is
Compression then composed of a series of JPEG images which are
to JPEG
Compression displayed interchangeably in video player at a required frame
to JPEG
rate. MJPEG is simple to implement because it uses a mature
Compressed JPEG compression standard with well-developed libraries and
frame data
in JPEG format
EXIF-based requires minimal hardware because it is not computationally
rotation intensive. As such, MJPEG is suitable for implementation of a
+
orientation live video encoder for execution on resource-constrained
mobile devices.
Delivery to Fig. 3 shows the procedure performed by MJPEG-based
the network
live video encoder for each video frame captured by a device
camera. Since we are using Android-powered devices in our
experiments, raw frame data arriving from the camera is
End of frame processing formatted as YUV image [17, 18]. Since the user may change
Fig. 3. Device-orientation-aware MJPEG-based live video encoder with
the orientation of the device anytime, the encoder retrieves the
matrix and EXIF-based video rotation support current device rotation angle from the orientation sensor and
normalizes it using (1). If the rotation is necessary, it then uses
(2) or (3), depending on the camera being used, to rotate the
image and compress it into a JPEG format. The order in which
rotation and compression are applied depends on the preferred
type of rotation. Finally, rotated JPEG is delivered as a video
frame to the client through a computer network.
To make the video playable in a web browser, we deliver
the MJPEG-encoded video stream through the HTTP protocol,
as shown in Fig. 4. Delivery of continuous data stream over
request-response oriented HTTP protocol is enabled through a
special multipart/x-mixed-replace value of the
Content Type HTTP header. This information, contained in
HTTP response, tells a web browser to replace previously
received content of a web page with a newly arrived one. That
way, the sequence of JPEG images sent at a required frame
rate looks like a live video sequence in client‟s web browser. Fig. 5. Structure of the JPEG image with EXIF metadata in image
header

IV. VIDEO ROTATION TECHNIQUES USED IN EXPERIMENTS

Since MJPEG video is composed of a sequence of
independent JPEG images, valid video rotation techniques are
all those that are applicable for rotation of still JPEG images.
Two JPEG rotation methods are most common in practice
today: matrix-based rotation which makes pixel-level
alterations over the image data, and EXIF tagging which
modifies image headers but leaves the actual image data
intact. In our experiments, we compare and analyze the
performance of these two rotation methods while used in
MJPEG video encoder.
A. Matrix-Based Rotation
Matrix-based rotation [19] consists of two steps. First step
is creation of the rotation matrix which varies depending on
the required video rotation angle. Rotation matrix shown in (4)
is created using the value of the video rotation angle 
calculated previously using (2) or (3).
which tells the image viewing software how to shift pixels to
display the image correctly. Hence, to provide the information
about the video rotation, MJPEG video encoder sets the value
of the EXIF orientation attribute according to the Table 1 and
stores it in the image header.
Since setting the value of the image header does not affect
pixels, the complexity of EXIF-based rotation is O(1). This
indicates that EXIF-based rotation is not affected by the
resolution growth, rather it has a constant speed. Once the
orientation value has been saved in EXIF header, it depends
on the browser whether the image will be displayed correctly
or not. The browsers that support the EXIF orientation tag will
rotate the image and display it as expected, while others will
ignore the EXIF silently and display the image unrotated.
TABLE I. EXIF ORIENTATION ATTRIBUTE VALUES
Type of Rotation EXIF Value
Undefined 0

Normal 1
   Flip horizontal 2

Using the created rotation matrix, second step is mainly Rotate for 180˚ 3
pixel multiplication. One by one, pixels are multiplied by the Flip vertical 4
rotation matrix, thus creating a rotated image. Consequently,
complexity of matrix-based rotation is O(width*height), which Transpose 5
is approximately O(n2) complexity class. It is valid to assume Rotate for 90˚ 6
that this algorithm will cease its performance with video
resolution growth. However, alterations made on pixels are Transverse 7
permanent and due to that, this kind of rotation is supported by Rotate for 270˚ 8
all relevant browsers and image viewers.

B. EXIF Tagging
EXIF (Exchangeable Image Format) metadata [20] is the
extra information about the image, such as date and time when
the image was taken, model and manufacturer of the camera
being used, image resolution, etc. EXIF metadata is appended
to the image data as a file header, as shown in Fig. 5.
Through the orientation attribute, EXIF metadata
enables to store the information about image orientation,
V. ANDROID-POWERED IMPLEMENTATION OF MJPEG-BASED
LIVE VIDEO ENCODER
To compare the performance of the two video rotation
methods on mobile devices, we developed a simple MJPEG
video encoder based on Android operating system. The
encoder performs the algorithm described in Fig. 3. The goal
of the experiment was to determine the extent the rotation
procedures reduce the frame rate of the produced live video.
 public void onPreviewFrame(byte[] yuvData, which takes frames in the range of 10-30 video frames per
Camera camera) { second (fps). Video rotation techniques have been analyzed
int rotationAngle = calculateRotationAngle();
ByteArrayOutputStream jpeg = and compared with regard to rotation angle and lighting
new ByteArrayOutputStream(); conditions. Results show video rotation speed is not in
if (rotationType == MATRIX) { correlation with the rotation angle for both techniques.
Bitmap rotBmp = rotateMatrix(yuvData, Therefore, further discussion is focused on lighting conditions.
rotationAngle);
rotBmp.compress(Bitmap.CompressFormat.JPEG,
MJPEG_QUALITY, jpeg); A. Performance Limits of the Camera
}
else { Android cameras use adaptive frame rate that depends on
YuvImage yuv = new YuvImage(yuvData, the lighting conditions and is not under direct control of an
ImageFormat.NV21, application developer. The best the developer can do is to
W, H, null); specify the preferred frame rate range among the supported
yuv.compressToJpeg(new Rect(0, 0, W, H),
MJPEG_QUALITY, jpeg);
ones. Therefore, in our first experiment, we examine the
rotateExif(jpeg, rotationAngle); performance of the camera without the encoder being
} attached. The goal is to determine the maximum speed at
sendToClient(jpeg); which the camera is able to deliver raw frames to the encoder
} under different lighting conditions. Results are shown in Fig.
--------------------------------------------------
private Bitmap rotateMatrix(byte[] yuvData, 7a.
int rotationAngle) { We confirmed there is a difference in frequency in which
byte[] rgbData = convertYuvToRgb(yuvData);
Bitmap bmp = Bitmap.createBitmap(rgbData, camera captures frames in regard to the amount of light
0, 0, W, H, null); entering camera lens. Camera speeds under all three lighting
Matrix rotationMatrix = new Matrix(); conditions remain steady with the resolution growth because
rotationMatrix.postRotate(rotationAngle); camera speed is a fixed hardware characteristic which has
Bitmap rotBmp = Bitmap.createBitmap(bmp, nothing to do with image size. However, the speed in daily
0, 0, W, H,
rotationMatrix, false); light is almost twice as big as interior and street lighting
return rotBmp; speed. This proves that camera adjusts its speed in a way
} which does not affect a human eye drastically. In the poor
--------------------------------------------------
private void rotateExif(ByteArrOutputStream jpeg,
int rotationAngle) {
// Android’s EXIF library manages EXIF
// metadata through a file system
writeToTempFile(jpeg, EXIF_PATH);
ExifInterface exif =
new ExifInterface(EXIF_PATH);
exif.setAttribute(
ExifInterface.TAG_ORIENTATION,
String.valueOf(rotationAngle));
exif.saveAttributes();
readFromTempFile(EXIF_PATH, jpeg);
}
Fig. 6. Implementation of the MJPEG live video encoder using the
Android camera API and ExifInterface library
a) Performance of the camera without running the encoder
MJPEG encoder is implemented using the Android camera
API [21] and Android EXIF library [22], as shown in Fig. 6.
Raw frame data are delivered to the encoder using the
onPreviewFrame() callback method of the
Camera.PreviewCallback class. This method is called
autonomously by the Android framework whenever a camera
has a fresh frame data to be delivered to the application. The
programmer then has the ability to manipulate the frame data
by implementing the internals of the callback method. Within
the callback, we differentiate between the two video rotation
methods and invoke the appropriate rotation routine and image
processing libraries, respectively.

VI. PERFORMANCE EVALUATION OF VIDEO ROTATION b) MJPEG encoder performance during JPEG compression
TECHNIQUES
Fig. 7. Performance of the camera and MJPEG encoder when no video
Performance evaluation of video rotation techniques has rotation is required
been done on Android-powered device with 5Mpx camera

a) Daily light
b) Interior lighting
c) Street lighting
Fig. 8. Comparison of MJPEG encoder performance for two different
video rotation methods under different lighting conditions
lighting conditions, human eye cannot recognize as many
details as it does during a sunny day. This is the reason why
camera slows down when lighting conditions are not so great.
B. Impact of the JPEG Compression
To determine the point at which the video encoding starts
to impact the camera performance, we examined the encoder
speed when it performs JPEG compression, but skips the
rotation step. As JPEG compression speed depends on the
frame size, the encoder performance degrades with the
resolution growth, as shown in Fig. 7b. However, for small
resolutions, the encoder is able to perform the compression
fast enough that the performance loss is not visible to the user.
This means that the encoder finishes with the frame processing
before new frame arrives from the camera. For larger
resolutions, we notice a proportional decline in performance
since the compression becomes dominant resource-consuming
process and starts to slow down the encoder.
Since the camera delivers frames faster as lighting
conditions get better, the turning point at which the encoder
starts to reduce the frame rate appears earlier on a daily light
than on interior and street lighting. Also, once the turning
point has been reached, we can notice a slight variations in
encoder performance for the same resolution, but different
lighting conditions. Street lighting speed is the fastest one and
daily light speed is the slowest one. The reason for this
occurrence is a different possibility for pixel arrangement
coding in different lighting conditions which makes image
taken during a day much less compressible than the others
because they have more details and shades. In other words,
they contain a lot more different pixels than darker images.
C. Impact of the Rotation
When JPEG compression is followed by a video rotation,
the effects of performance loss caused by these two processes
are added to each other. The cumulative performance loss in
video frame rate for three types of lighting is shown in Fig. 8.
While both video rotation techniques reduce the frame rate to
some extent when compared to plain JPEG compression, each
graph clearly shows the advantage in speed of EXIF-based
rotation over the matrix-based one.
Since EXIF metadata manipulation does not depend on the
image size, EXIF-based rotation decreases the performance of
the encoder marginally when compared to the JPEG
compression. Also, the EXIF manipulation library from
Android SDK that was used in experiments is intended to
work with still images and is not optimized for on-the-fly
EXIF manipulation. It requires the image to be saved to a file
prior to metadata manipulation can begin and, likewise, to
import modified image with metadata embedded from the file
back into memory to be available for further processing. File
system access causes the performance loss more noteworthy
than expected.
On the other hand, performance of the matrix-based
rotation steadily decreases with the resolution growth. The
encoder that uses matrix-based rotation is two to more than
three times slower than its EXIF-based counterpart on a
daylight. As lighting conditions get worse, EXIF-based
encoder is even more superior and gets almost four times
faster than matrix-based one at highest resolution. The
advantage of EXIF-based rotation gets more significant as the
resolution grows. For example, using an EXIF-based encoder
on a daily light doubles the frame rate on lowest resolution if
compared to the matrix-based one, but even triples it on
highest resolutions.
 VII. CONCLUSION [8] J. Chamberlain, U. Kruschwitz, and M. Poesio, “Motivations for
participation in socially networked collective intelligence systems”, in
Although superior in performance, EXIF-based rotation of Proceedings of CI2012, MIT, Cambridge, MA, USA, 2012.
MJPEG video is not yet widely supported by web browsers [9] D. Gourley et al., HTTP: The Definitive Guide, Sebastopol, CA:
and other video players. Among major web browsers, Mozilla O‟Reilly Media, Inc., 2002.
Firefox, Google Chrome, and Safari natively support the [10] R. Fielding et al., Hypertext Transfer Protocol – HTTP/1.1,
MJPEG, but still ignore the EXIF orientation tag without http://tools.ietf.org/ search/rfc2616
additional plug-ins installed. Internet Explorer does not even [11] S. Kamath and J.R. Jackson, “Low-bit rate motion JPEG using
support the MJPEG without a third-party plug-in. Therefore, differential encoding,” in Conference Record of the Thirty-Eighth
Asilomar Conference on Signals, Systems and Computers, November
for any practical application, matrix-based rotation is the only 2004, pp. 1723-1726.
option to choose when MJPEG-based live video need to be [12] A. Cozzolino et al., “Evaluating the effects of MJPEG compression on
displayed in a regular web browser. However, the potential of motion tracking in metro railway surveillance,” Advanced Concepts for
the EXIF orientation metadata deserves to pay the attention to, Intelligent Vision Systems, Springer Berlin Heidelberg, pp. 142-154,
especially with the mass production of multimedia content 2012.
produced by mobile devices. [13] L. Chen, N. Shashidhar, and Q. Liu, “Scalable secure MJPEG video
streaming,” in Proceedings of the 26th International Conference on
Advanced Information Networking and Applications Workshops
References (WAINA), Fukuoka, Japan, March 2012, pp. 111-115.
[14] Pelco Developer Network, “MJPEG HTTP push”,
[1] A. Rahmati and L. Zhong, “Studying smartphone usage: lessons from a http://pdn.pelco.com/content/mjpeg-http-push
four-month field study,” IEEE Transactions on Mobile Computing, vol.
12, no. 7, pp. 1417-1427, July 2013. [15] en.code-bude.net, “Motion JPEG”, http://en.code-bude.net/tag/how-
does-mjpeg-work/
[2] H. Falaki et. al., “Diversity in smartphone usage,” in Proceedings of the
8th International Conference on Mobile Systems, Applications, and [16] Android API Guides: Position Sensors, http://developer.android.com/
Services (MobiSys „10), San Francisco, CA, USA, 2010, pp. 179-194. guide/topics/sensors/sensors_position.html
[3] K. Church and B. Smyth, “Understanding mobile information needs,” in [17] DiscoveryBiz.net, “YUV, YCbCr, YPbPr color spaces”,
Proceedings of the 10th International Conference on Human Computer http://discoverybiz.net/enu0/faq/faq_YUV_YCbCr_YPbPr.html
Interaction with Mobile Devices and Services (MobileHCI „08), [18] D.A. Rumball, “Method and apparatus for converting digital YUV video
Amsterdam, The Netherlands, 2008, pp. 493-494. signals to RGB video signals,” US Patent 5124688 A, May 1990.
[4] A. Lih, The Wikipedia Revolution: How a Bunch of Nobodies Created [19] D. Salomon, Computer Graphics and Geometric Modeling, Volume 1,
the World‟s Greatest Encyclopedia, 1st ed, New York: Hyperion, 2009. Springer-Verlag, New York, 1999.
[5] BOINC, http://boinc.berkeley.edu/ [20] “Exchangeable image file format for digital still cameras: Exif version
[6] E. Korpela, D. Werthimer, D. Anderson, J. Cobb, and M. Leboisky, 2.2,” in JEITA CP-3451: Standard of Japan Electronics and Information
“SETI@home – massively distributed computing for SETI”, Computing Technology Industries Association, 2002.
in Science & Engineering, vol. 3, no. 1, pp. 78-83, January-February [21] Android API Guides: Camera, http://developer.android.com/guide/
2001. topics/media/camera.html
[7] F. Smadja, “Mixing financial, social and fun incentives for social [22] Android Reference: ExifInterface, http://developer.android.com/
voting”, in Proceedings of WEBCENTIVES‟09, Madrid, Spain, 2009. reference/android/media/ExifInterface.html
 Integrating Learning Analytics for higher-order
thinking e-assessments
S Gunness
Centre for Innovative and Lifelong Learning
University of Mauritius
Ebene, Mauritius
sandhya.gunness@uom.ac.mu
Abstract—This paper attempts to understand the broader
implications for integrating learning analytics in e-learning
systems. This is also the commencement of a collaborative effort
for developing research bases around learning analytics between
the University of Mauritius and the University of KwaZulu-
Natal. The focus on e-assessments is deliberate: the tendancy in
both our universities to appraise student learning as summative
and quantitative does not reflect contemporary workplace and
scholarly needs where learning processes are more valuable than
outcomes. The necessity for epistemological rethinking and
developing systems based on contextual requirements is
important, whether it is related to curriculum development or
workshop organisation or even conducting a class. After a survey
of the literature on the potentials and hypes around learning
analytics, we look towards the literature linking learning
analytics, data mining and e-assessment for educators to
interface more effectively and efficiently with the LMS platform.
We conclude this concept-paper with a workshop outline which
can help educational technologists organise practical, sustainable
and meaningful hands-on sessions with educators.
Keywords—learning analytics, big data, data mining, e-
assessments, higher order thinking skills (HOTS)
INTRODUCTION
The University of Mauritius (UoM) falls under the Applying
stage of technology integration. At this stage, teaching still
takes the priority in the classroom. However, computer-based
learning, through Learning Management Systems (LMS), are
integrated into regular teaching, and thus the users (lecturers
and students) become more confident in adopting specialized
technology-based tools to teach and learn their subject fields.
The Centre for Innovative and Lifelong Learning (CILL),
responsible for academic induction of e-learning integration at
the UoM has adopted the Moodle LMS as its e-learning
platform. It has been observed that most of the academics at
the UoM adopt e-learning services for communication and
provision of supplementary reading material. Essentially
Moodle is used by academics as the typical placeholder for
lecture material. Academics are clearly not incorporating any
of the interactive features available in Moodle. In addition to
the Quizzes tool which Moodle provides for electronic
assessment (e-assessment), the latest Moodle 2.8 version
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
UG Singh
School of Management, Information Systems and
Governance
University of Kwa Zulu-Natal
Durban, South Africa
singhup@ukzn.ac.za
provides tools for assessing the performance of students which
is usually grounded in the data analysis gathered during
interactive online sessions. This paper aims at investigating
how to introduce educators to adopt these tools to enhance
their teaching practices, in particular, to use learning analytics
for appraising metacognitive skills in e-assessments. After a
survey of the literature on the potentials and hypes around
learning analytics, we look towards the literature linking
learning analytics, data mining and e-assessment for educators
to interface more effectively and efficiently with the LMS
platform. We conclude this discussion-paper with a workshop
outline which can help educational technologists organise
practical, sustainable and meaningful hands-on sessions with
educators. The aim of this workshop is to familiarise
academics with learning analytics for e-assessments, that go
beyond the simple recall of knowledge, and that are geared
towards equipping their learners with a disposition towards
lifelong learning.
BIG DATA FOR EDUCATION
Traditionally, teaching pedagogies provide limited immediate
feedback to students. Thus educators often have to spend long
hours marking regular assignments. As a result they lack
enthusiasm to guide students on how to improve their
comprehension skills. Furthermore, many educators fail to
experiment with, and provide support, by adopting
technology-based resources to provide variety in the learning
process. Adopting data-driven approaches to teaching
facilitates better analysis of students‟ actual learning, as well
as provide systematic feedback to both students and educators.
The potential of “big data” provides opportunities to mine the
learning patterns of students, to get a better insight into both
their student performance, and learning approaches [1].
Data mining is defined as the „broad field of discovering
innovative and potentially valuable information, from large
amounts of data‟ [2].
Educational Data Mining specifically refers to „an emerging
discipline, focusing on developing methods for exploring
unique data that is extracted from educational settings‟. These
approaches allow educators to better understand their students,
as well as the context of this learning.
A model for effective data mining is presented by [3]. This
model presents the order of activities that are followed to
extract knowledge from data, as shown in Figure 1.
Fig. 1. Sequence for extracting knowledge from data [4]
Contextualizing the above figure to the LMS, Moodle, when
an e-assessment is taken by students, using the Quizzes tool,
the „Raw data‟ collected includes the students‟ marks. These
would be associated with the student registration details in the
„Selection‟ process. The „Pre-processing‟ activity would
arrange these individual scores per question into logical
formats to facilitate the extraction of data. „Data mining‟
would allow the user to specify the criteria for extraction
purposes, e.g. if the user wanted to view all the student
responses for a particular question. This is then mined from
the „test results‟ and extracted for the user to view in a format
that is useful for him/her. „Patterns‟ can easily be identified by
the representations provided e.g. colour coding for full, part
and zero marks; graphs/charts/average values. The educator
can then identify student learning trends and thus „knowledge‟
is produced about student learning of that particular concept.
Educational Data Mining can provide multi-context research
data about student learning, including „individual learning
from educational software, computer supported collaborative
learning, computer-adaptive testing, e-assessment, as well as
the factors that are associated with student failure or non-
retention in courses‟ [2], [19].
Differing algorithms and techniques, adopted for knowledge
discovery in databases include „Classification, Clustering,
Regression, Artificial Intelligence, Neural Networks,
Association Rules, Decision Trees, Genetic Algorithm, and
Nearest Neighbour‟ method [4].
For example, the decision tree model is often used to analyze
students learning behaviour, in order to predict student results.
This can warn educators of poor-performing students -, prior
to their final exams, so that additional support can be provided
to these students [5].
Educational data mining methods, are often unique, when
compared to regular data mining methods, due to the need for
unambiguous accountability required in educational data [5].
Reference [6] categorizes the „processing‟ activity, as
illustrated in Figure 1, in educational data mining, into the
following categories:
 Statistics and visualization
 Web mining
 Clustering, classification, and outlier detection
 Association rule mining and sequential pattern mining
 Text mining.
The key areas of application of educational data mining
include:
 improved student models, of student‟s characteristics
e.g. student‟s current knowledge, motivation, meta-
cognition, and attitudes[7];
 improved models of a domain‟s knowledge structure
[9];
 studying pedagogical support to understand which
method was most effective for different groups of
students, or in different situations, or even for all
students overall [7];
 finding empirical evidence to improve educational
theories [9]; and
 understanding the key factors that impact learning, to
design better learning systems [5];
The advantages of adopting educational data mining include:
 the ability of the educator to analyse student learning in
far more effective ways. With the ability to mine into
educational data in a real time manner, means that the
data analytics provided help educators to understand what
students know, at that point of time, and what techniques
are most effective for each student, rather than them
relying on periodic test performance for this feedback.
 Effective feedback techniques usually encompass
information on:
o what is regarded as good performance,
o implementing self-assessment in the learning
process,
o delivering better quality feedback to students,
o promoting peer collaboration around learning,
o positively motivating students,
o assisting students to close the gap between their
current and desired performance, and
o informing educators on effective feedback
mechanisms [7], [9].
e-Assessment facilitates the provision of immediate feedback
about academic performance, to both students and educators
Adopting e-assessment practices allows educators to design
the assessment to automatically adjust test items, based on
students‟ performance on earlier questions. Thus assisting
„weaker students‟ by providing them with more questions, on
the concepts they have not yet mastered.
Furthermore, the flexibility to experiment with alternatives,
and observe how students respond becomes feasible, without
putting an extra workload on the educator.
E-Assessment also enhances the usage of „branching‟ items,
which allow educators to probe more deeply into
individualized information of student performance. For
example, if students answer all questions correctly (or
incorrectly), a better understanding can be established,
regarding their knowledge levels, thus directing the
assessment to proceed to either higher or lower skill mastery
levels. This will assist in understanding where students
required more help.
Additionally, e-assessment can be adopted to evaluate
problem-solving approaches and provide feedback to students
along the instructional path. Messages can be provided to the
student, to guide them if an incorrect approach is adopted.
Hints may also be provided to students who request it this
additional assistance.
The data analytics provide educators with a detailed analysis
of how the students solved the problem.
Extending e-assessment, the inclusion of embedded
assessment “engages and satisfies students at a higher level
than do average courses and presents a more realistic and
integrated view of business decision making.” By following
this student solution process, educational data analytics tools:
 provide guidance on students preferred learning styles and
helps the educator tailor the instruction process to that
preference.
 present educators with underlying patterns to assist them
in predicting student outcomes like:
o potential drop out students,
o students in need of additional help, and
o students who are capable of more challenging
tasks [7].
o evaluation of a variety of student actions, which may
include:
o the time students devote to readings,
o where students access electronic resources from,
and
o how fast they master key concepts [5],[7],[8].
Data mining also allows educators to probe student
understanding of „core concepts, mark expectations, and
specific skills, as well as, quickly diagnose students‟ strengths
and instructional needs‟ [3]. Further, data mining is effective
in identifying pedagogic approaches that are successful with
students. This is facilitated through interactive „web of
inquiry‟, which is a common technique to teach scientific
analysis. This method enables students to:
 „test hypotheses,
 learn the language of science,
 design experiments,
 code variables,
 collect data,
 analyse data, and
 develop tables‟ [3],[7].
Thus, equipped with statistical information provided by
educational data analytics, many educational institutions adopt
dashboard software and data warehouses to monitor learning,
performance, and behavioural issues for students. Dashboards
compile key metrics in a simple and easy to interpret interface,
so that educators can quickly and visually see how the
institution is doing. These dashboards are updated
automatically, based on data stored in Student Information
Systems.
Higher education dashboards offer a wide array of material,
thus the adoption of data analytics assist in providing instant
information about assessment tasks. They also provide more
effective systems for mining information. Some examples of
the application of these data analytics are presented briefly. A
study conducted by [4] indicated, that through data mining, a
high correlation between factors like „mother‟s education and
student‟s family income, with student academic performance
was found. Other data analytics in another study [8] showed
that females with a „high socio-economic status had relatively
higher academic achievement in science streams,‟ compared
to males. However, males with a low socio-economic status
had relatively higher academic achievement in general.
The limitations associated with educational data mining in
Higher Education institutions is that consistent data formats
are required across the dashboard, - to facilitate better result
comparisons. Inconsistent terminology and coding, or
incomplete data captured with regard to student dropouts or
graduation, can be problematic. For effective data analytics,
the information captured into these dashboards must be simple
to understand and coded in equivalent ways. Adopting
common semantics and measure will facilitate the aggregation
of material and better the analysis of the data analytics
information provided [2],[3],[6].
Most importantly, Higher Education Institutions must
understand the value of a data-driven approach to education,
to contributing towards more informed decision-making.
Adopting this approach, institutions will be able to identify
trends, pinpoint problem areas, and direct resources in an
efficient manner. Furthermore, these analytics will guide
educators on overall student performance, as well as ways to
improve the learning process [3],[4].
The challenges of creating data sharing networks necessitates
the balancing of student privacy on the one hand with access
to data, for research purposes, on the other. Maintaining the
confidentiality of student records is essential, but these records
also need to be accessible for researchers to mine data, for
understanding vital trends, and implementing helpful
interventions [3],[7].
To summarise, this section has dealt with Educational Data
Mining (EDM) as a means to identify and, bouncing on the
geological metaphor, unearth knowledge from educational
databases, particularly in relation to at-risk learners. The
different techniques for educational data mining have been
listed along with the advantages of EDM. Caution, however
should be exerted before investing massively into a university-
wide datamining system, as issues of student privacy, levels of
data access, the way data is formatted for comparing results
across systems must be properly understood and accounted
for.
The next section deals with pedagogical and usability
considerations for designing learner-centred dashboards
LEARNING ANALYTICS AND DASHBOARD DESIGNS
Educators are well placed to understand the importance of data.
Statistics based on marks, attendance, student demographics
are consulted regularly to explain performance. In an online
environment, accessing and analysing these data are done
much more efficiently and effectively through the data logging
systems that a Learning management system provides. While
many data-mining software exist commercially and in open-
source formats, their origins in business and non-educational
fields make that these platforms are sometimes too complex for
teachers and educators to use. Therefore the cognitive load
associated with operating the program itself is too great to
make them practical, as they reduce the users' cognitive
resources to actually perform their task.
Visualisation and customisable interfaces.
Visualisation contributes positively to the intuitiveness of the
tool and is considered an important factor for adoption and
understanding of the learning analytics by the end users.
Originally, the dashboard was a business application which
allowed business analysts to have a snapshot of Key
Performance Indicators arranged in small data boxes on a
single webpage. Google Analytics, for instance, can
accommodate up to 12 indicators depending on each
individual organisations‟ needs.
Fig. 2. Google Analytics data boxes
Data pertaining to Visits and Site Usage which are
determinants of visitor loyalty, data on increased accessed
during particular times – which can be then useful for targeted
advertising, or depending on the amount of time spent on a
particular page will determine the value of the content to
particular users. Similarly, learning dashboards provide
support to users to graphically represent their course activity.
Customisation and visual representation of relevant data is
important for dashboard applications. In this way, learners and
educators can personalise their dashboards to represent key
data as indicators for performance. Data that students find key
and which they are given access to are different from
accessible data for educators.
Learner analytics is different from data mining, in that
analytics support data-driven decision-making and provides
information for educators, policy makers and administrators to
improve learning processes. Reference [10] adds that learning
analytics is about sense-making and interpretation, whereas
educational data mining is more about “developing methods
and models for existing data in educational settings so that
bigger questions about learning could be answered”. Data
mining tries to organize and reduce educational data whereas
data analytics views the entire data systems to better
understand learner behaviours. Table 1attempts to differentiate
Educational data mining and Learning Analytics.
Table 1: Distinguishing betwen EDM and LA
Educational Data Learning Analytics
mining
Control Systems Educators, Learning
leveraged to administrators, designers, Learners
Educational
Technologists
Nature Technical / Artifical Epistemic and
intelligence, number epistemological
 crunching activity streams, and tabular overviews provide the basis for
Analysis Individual Complex systems as reflecting on these data and assessing their usefulness-
components and a whole
relationships
Techniques Classification, Social network
and models clustering, Bayesian analysis, sentiment
modelling, analysis, influence
relationship mining, analytics, discourse
discovery models, analysis, learner
visualization success prediction,
concept analysis
Example of Students‟ preferred Understand
information style of learning. Students‟ learning
that can be Recommendations and optimising the
collected based on similar learning process
interests. (Amazon based on human
link to suggested interpretation of Fig. 3. Learning Analytics processes
readings) EDM results.
Deriving models and
methods based on Having thought out the various implications from the data, the
data acquired. learner and/or educator will try to identify conceptual
Visualisation Tabular/ relationship Graphical, user weaknesses and find answers to these. For instance at the
diagrams friendly awareness stage, teachers can use the dashboard to get an
overview of the course activity – number of students having
accessed the resources that were provided, number of times
Learning Analyticadoption and ethical considerations these resources were accessed, time spent on accessing (and
presumably reading) these documents which would assist the
Studies by [11] demonstrate that adoption of a specific tool for educators to reflect on their teaching practices at stage 2. At
practice is based on the pedagogical and organisational utility stage 3, they could identify particular patterns for instance to
of the learning analytics. If the lecturer feels that it is worth find at risk students or isolated lurkers, providing deeper sense-
their time and effort to investigate a particular tool, only then making opportunities. Reference [14] mentions that at stage 4,
will they make use of the tool. The kinds of analytics that the which was to measure impact, very few dashboards provided
tool offers determine its helpfulness along with the feedback this feature. Course Signals [15] was highlighted as one such
that the tools provide. For instance, LOCO-Analyst was dashboard that explicitly demonstrated impact of Dashboards
perceived [12] to provide suitable feedback regarding weak- on Learning. Course Signals uses a student success algorithm
points of students‟ online interactions in terms of visualising that predicts students‟ risk status based on four factors namely:
the connection strengths that each student had with their (1) performance (% marks earned in the course), (2) effort
peers.[13] propose a toolkit (Exploratory Learning Analytics (interaction logs on Blackboard Vista LMS), (3) prior
Toolkit-eLAT) to support educators reflection-on and academic history (high school GPA, standardised test scores)
improvement of online teaching methods. They realized that and (4) student demographics. However, this focus on
educators need easy-to-use tools that “help them observe the students-at risk is not really pointing to actual learning
consequences of their actions and evaluate their teaching processes.[16], [17] mention the risks of overemphasising
interventions. They need to have appropriate access to data to analytics for student retention and predicting drop-outs over the
assess changing behaviours and performances of their students more important use of learning analytics on student learning.
to estimate the level of improvement that has been achieved in They also highlight the need for learners to exercise more
the learning environment”. [14] compared 15 dashboard control over what is being done to their data. The issues of
applications with an evaluation focus on usability, usefulness, Informed consent, institutional surveillance and online privacy
efficiency and effectiveness. The learning analytics pertained have to be discussed openly with the learners. Ultimately, it is
to time spent on a particular reading or activity, Social the learner‟s choice about control and what is being done to
interactions between peers and tutors, Documents and tool their data and who will be accessing it for educational or other
use, artefacts produced and exercises results and quizzes. uses. Reference [18] provides a framework for ethical
Based on their learning analytics process model, [14] posit considerations for institutions to provide guarantees for
that “when data can be related to goals and progress toward protecting their students against misuse of personal data.
these can be tracked, meaningful feedback loops can be Reiterating that learning analytics should be focussing on their
created that can sustain desired behaviour”. The figure 4 impact on learning, metacognitive skills such as information
below describes the different processes that learners and retrieval and processing via the internet seems to be a complex
educators experience during a learning activity - Data from issue for students, and should also be part of the assessments if
we want to encourage higher-order skills such as metacognitive
abilities and autonomous learning. Getting the right answers
are not as important as the learner‟s enquiry process [1].
Reference[19]suggests the creation of the “Intelligent
Curriculum”- as disrupting the traditional view of courses.
Instead of assessing summative end-of-course activity, learners
should be evaluated based on the interactions that the learner
makes with his network. “The knowledge, attitudes, and skills
required in any domain can be rendered as a network of
relations. The semantic web and linked data are partial
instantiations of this concept. Knowledge domains can be
mapped, and learner activity can be evaluated in relation to
those maps.” This is reiterated in [1] who put forward that
assessment is should only be related to knowledge of facts, but
rather knowledge-how, demonstrating skills in information
processing, synthesis and metacognitive abilities. This clearly
calls to placing analytics in the hands of the students. Instead of
being evaluated on understanding a particular learning
resource, learners would be provided with guidance to better
formulate their search queries, “identify and collect appropriate
evidence, present results systematically, analyse and interpret
results, formulate conclusions” [20].
To summarise this section, learning analytic dashboards should
be designed such that their users can easily monitor their
progress, but more so should develop a culture of using data for
making instructional decisions. Data should be processed to
show success, but also provide learning paths to enhance the
learners‟ ability to think about context, reliability, validity,
certainty and connectedness of the knowledge they have
acquired. In the next session we will look at e-assessments that
integrate metacognitive skills. This will help us to design
learning objectives, outline and assessment criteria for a
workshop targeted at educators who would like to integrate and
make use of learning analytics in their teachings.
of high level knowledge – it is assessment for learning.
Fig. 4. The Epistemology-Assessment-Pedagogy triad
Thus epistemology is important to learning analytics. The
ways that we adopt resources for assessment tasks, the actual
tasks we set for learners and the kinds of learning we believe
to take place (and aim for) are linked to our understanding of
epistemology. Reference [1] presents interesting implications
for conceptualising the relationship between LA and learning
design. Learning in contemporary times depends not only on
what knowledge is, but also how is it learnt and valid.
Learning analytics for more factual information would be
provided through test scores and access logs, while
constructivist pedagogical approaches that require guided
discovery or problem-based learning would look at learning
paths of the leaner as s/he spent time on particular resources
and responded to particular questions and quality of his/her
discussions. Possible metrics for analysing quality
considerations in online discussions could be related to levels
of Paraphrasing, Value addition and Critical thinking. These
could be possible constructs when applying “subjectivist
approaches” where personal satisfaction with the contents
being provided, are measured [14]. Use of feedback forms,
blog-tagging and sentiment analysis are presented as analytic
indicators. Table 2 presents the various Learning analytic tools
which can be associated with higher-order learning skills.

Table 2: Higher order thinking skills and LA tools

E-ASSESSMENTS THAT INVOLVE HIGHER-ORDER SKILLS.
Students, traditionally, focus on assessments. When asked to Higher Order thinking skills Learning analytics
design the interface of a learning analytic dashboard, students and activities
tend to list personal rankings in the classroom, reminders of Analysing trends and Network size, mapping
assessment due dates and exams, percentage assessments directions of scholarly connections and quality of
completed as well as grade that they need to obtain to receive arguments interactions
different levels of grades [21]. As explained by [1], Group annotation exercises Annotated bibliography
epistemology could be a driver of assessments that are aimed Online Role plays Classifying expert and
at uncovering student knowledge, However, epistemology can novice users
also drive pedagogy to build high quality knowledge. As Information retrieval and Argumentation in
outlined in figure5below, assessment is targeted at the learning interpretations sensemaking deliberations
Collaborative Wiki Editing of Sociocultural discourse
an article analysis, argumentation
theory
Metacognition and Personal Sentiment analysis of
Learning journal writing communication data.

While the actual working of these LA tools are beyond the

scope of this paper, snapshots of the type of analysis that can
be performed and the interface design are relevant for our next
discussion around choice and potentials of Learning analytics.
Network size and quality
Figure 7 below is an example of analysing couplings between
papers. Each node is a paper published in a particular
conference and joining lines represent the papers that they
refer to in common. Analysing this particular case, it was
inferred that the most influential leaders of the field had the
most connections and were the bigger nodes. By creating a
network on the basis of shared references, [22] propose a way
to understand whether an article bridges among present
strands of literature and infer the size of its research
community and its embeddedness.Outliers on the figure which
are not linked represented workshops and panels which would
not contain as many references as conference papers.
Fig. 5. Bibliographic Network Coupling
Causal Diagrams
In this tool, learners are provided with quiz questions which
are represented in a causal map. Coherence analytics imply
applying information from the hypertext resources to edit a
map and regulating the system to represent correct causal
links. Reference [23] indicates that learners with higher levels
of action coherence possess stronger metacognitive knowledge
and task understanding.
Fig. 6. Coherence Analysis
Social Discourse Analytics
Through this tool, linkages between the various participants
and the concepts that they are discussing can be analysed.
Each node would represent the discussant or the concept and
the strength of the linkages could determine the frequency of
discussion around a particular topic.
Fig. 7. Social Discourse networks
Conclusion and Workshop planning
This paper has provided an overview of learning analytics for
e-assessments. We focussed on metacognitive and higher-
order learning while realising that visualisations and
personalisation of learning interfaces were key to user buy-in.
The insights gained through the literature allowed us to gauge
important features of learning analytics and key considerations
when planning the integration of LA in an institution. While
learning analytics is booming in the educational industry,
blindly following the masses will not be beneficial to
universities where technology integration is still at applying
stage. Based on our review of the current literature, we can
recapitulate our findings as follows:
o Meaningful data-mining can only occur if we make use of
common semantics and metrics to facilitate the
aggregation of material and better the analysis of the data
analytics information provided
o Possible metrics for analysing quality considerations in
online discussions could be related to levels of
Paraphrasing, Value addition and Critical thinking.
o Learning analytic dashboards should be designed such that
their users can easily monitor their progress, but more so
should develop a culture of using data for making
instructional decisions
o These analytics should be integrated as situated learning
activities during the workshop so that participants can
experience the learning outcomes at first hand.
Our recommendations for an initial workshop therefore would
be to set the learning objectives of the workshop as follows:
a) Introduce the main epistemological groundings for
learning analytics.
b) Provide an overview of learning dashboards and
their potentials for gathering learning evidences and
impact beyond simply identifying at-risk students
c) Focus on learner-centred activity designs which
provide higher-order learning and assessments.
d) Provide hands-on real-time data analytical exercises
for educators to experience.
e) Gather feedback on pedagogical and cultural
aspects.
This paper has been a joint effort at understanding the state of
play of learning analytics. Our future collaborative research
between the University of Mauritius and University of
KwaZuluNatal would be to firstly integrate higher order
thinking learning analytics tools within the Learning
Management Systems (LMS). We could then investigate and
compare educator responses to learning analytic tools
integrated in our LMSes at our respective universities. It
would be interesting to also investigate levels of online
interactions between peers and tutors and correlate the
findings to impacts on learning as has been highlighted.
REFERENCES
[1] Knight, Simon; Buckingham Shum, Simon and Littleton, Karen (2014).
Epistemology, assessment, pedagogy: where learning meets analytics in
the middle space. Journal of Learning Analytics, 1(2) pp. 23–47.
[2] West, D. M. (2012). Big data for education: Data mining, data analytics,
and web dashboards. Governance Studies at Brookings, 1-10.
[3] Arnold, K. E. (2010). Signals: Applying Academic Analytics. Educause
Quarterly, 33(1), n1.
[4] Baradwaj, B. K., & Pal, S. (2012). Mining educational data to analyze
students' performance. arXiv preprint arXiv:1201.3417.
[5] Baker, R. S., &Yacef, K. (2009). The state of educational data mining in
2009: A review and future visions. JEDM-Journal of Educational Data
Mining, 1(1), 3-17.
[6] Romero, C, &Ventura,S. (2010) "Educational data mining: a review of
the state of the art," IEEE Transactions on Systems, Man, and
Cybernetics, Part C: Applications and Reviews, vol.40(6), 601-618.
[7] Beck, J. E., &Mostow, J. (2008, January). How who should practice:
Using learning decomposition to evaluate the efficacy of different types
of practice for different types of students. In Intelligent Tutoring Systems
(pp. 353-362). Springer Berlin Heidelberg.
[8] Shaffer, D.W., Hatfield, D., Svarovsky, G.N., Nash, P., Nulty, A.,
Bagley, E., Franke, K., Rupp, A.A., and Mislevy, R. (2009) Epistemic
Network Analysis: A prototype for 21st Century assessment of learning.
The International Journal of Learning and Media 1: 33-53.
[9] Nicol, D. J., & Macfarlane‐Dick, D. (2006). Formative assessment and
self‐regulated learning: A model and seven principles of good feedback
practice. Studies in higher education, 31(2), 199-218.
[10] Siemens, G. (2013). Learning analytics: The emergence of a discipline.
American Behavioral Scientist, 57(10), 1380–1400.
[11] Ali, L., Asadi, M., Gašević, D. Jovanović, J. and Hatala, M.(2013).
Factors influencing beliefs for adoption of a learning analytics tool: An
empirical study. Computer Eduation.62 130-148.
[12] Jovanović, J., Gašević, D., Brooks, C., Devedžić, Y.,Hatala, M.(2007).
LOCO-analyst: a tool for raising teachers' awareness in online learning
environments, Proceedings of the Second European Conference on
Technology Enhanced Learning: creating new learning experiences on a
global scale, September 17-20, 2007, Crete, Greece.
[13] Dyckhoff, A. L., Zielke, D., Bültmann, M., Chatti, M. A., & Schroeder,
U. (2012).Design and Implementation of a Learning Analytics Toolkit
for Teachers. Educational Technology &Society, 15 (3), 58–76.
[14] Verbert, K., Duval, E., Klerkx, J., Govaerts, S., & Santos, J. L. (2013).
Learning Analytics Dashboard Applications.American Behavioral
Scientist.
[15] Arnold, K.,E.,& Pistilli, M.,D. (2012). Course signals at Purdue: using
learning analytics to increase student success. In Proceedings of the 2nd
International Conference on Learning Analytics and Knowledge (LAK
'12), Simon Buckingham Shum, Dragan Gasevic, and Rebecca Ferguson
(Eds.). ACM, New York, NY, USA, 267-270. D
[16] Kruse, A., & Pongsajapan, R. (2012). Student-centered learning
analytics. Retrieved from
https://cndls.georgetown.edu/m/documents/thoughtpaper-
krusepongsajapan.pdf
[17] Prinsloo, P., & Slade, S., (2013). An evaluation of policy frameworks
for addressing ethical considerations in learning analytics. In
Proceedings of the Third International Conference on Learning
Analytics and Knowledge (LAK '13), Dan Suthers, Katrien Verbert, Erik
Duval, and Xavier Ochoa (Eds.). ACM, New York, NY, USA, 240-244.
[18] Slade, S., & Prinsloo, P. (2013). Learning analytics: ethical issues and
dilemmas. American Behavioral Scientist, 57(10) pp. 1509–1528.
[19] Long, P., & Siemens, G. (2011). Penetrating the fog: Analytics in
learning and education. EDUCAUSE Review, 46(5), 30-32.
[20] Lee,V., et al. (2004), “What is Inquiry-Guided Learning?”
in Teaching and Learning Through Inquiry, ed. Virginia
Lee Sterling, VA: Stylus, 9.
[21] Duval, E., Klerkx, J., Verbert, K., Nagel, T., Govaerts, S., Parra Chico,
G.A., Santos, J.L., & Vandeputte, B.(2012). Learning dashboards and
learnscapes. Educational Interfaces, Software, and Technology, May, 1-
5.
[22] Biscaro C, Giupponi C (2014) Co-Authorship and Bibliographic
Coupling Network Effects on Citations. PLoS ONE 9(6): e99502.
doi:10.1371/journal.pone.0099502
[23] Segedy, J.R., Kinnebrew, J.S., & Biswas, G. (2015). Using Coherence
Analysis to Characterize Self-Regulated Learning Behaviours in Open-
Ended Learning Environments. Journal of Learning Analytics, 2(1), 13-
48.
 A Survey of Context-Aware Recommender System
and Services
Ebunoluwa Ashley-Dejo, Seleman Ngwira, Tranos Zuva
Department of Computer Systems Engineering
Tshwane University of Technology
South Africa
ebunashley@gmail.com

Abstract—The advancement of the Internet, mobile and
wireless technologies have produced a rapid change of
information in terms of volume and accessibility. The enormous
volume of information can be devastating especially to mobile
users exceeding human ability to differentiate information which
is relevant and that which is irrelevant. For many yzears now,
recommender system have become well-known, and have been
studied in various domains such as e-learning, online shopping,
tourism to help overcome information overload.
Recommendations are produced based on users who have
interests in a particular thing or item. This recommendation
process was further enhanced by incorporating context such as
time, weather, and location to make recommendations more
accurate and efficient. However, these systems have introduced
context-aware recommender systems. This paper presents a
survey of Context-aware recommender systems, the background
and algorithms of Context-aware Recommender System, and also
discusses the open issues of context-aware recommender systems.
Keywords—Recommender System; Context; Context-aware
Recommender systems
I. INTRODUCTION
In the past, many researchers have studied Context-aware
computing and have also built several applications to prove
the importance of it [1] context-aware computing is
considered to automatically use context data so as to run the
services that are suitable for a specific time, places or events.
It was incorporated to improve the traditional user request
response pattern that requires the user to express the desire for
recommendation.[1]
The traditional recommendation is focused in suggesting
the most important items/objects to users without considering
any additional contextual information, like location, weather
and time.[2] though, in several applications such as travel
package, e-learning etc., it might not be adequate to consider
items and users only, it is essential to integrate the context
information into the process of recommender systems so as to
recommend services/items to users under some circumstances.
[2] Context-awareness is a potential technology for mobile
devices by means of facilitating the use of device in
challenging condition by dynamically adapting the device
behaviour.[3]
The main contribution of this paper is to give a survey of
Context-aware Recommender system and services. The
remainder of this paper is as follows: Sections II discusses the
background of Context-aware Recommender Systems, Section
III explains Context-aware systems and its Components,
Section IV discusses Context-aware Recommender System
CARS and the algorithm, Section V presents the application,
Section VI talks about the issues and challenges, Section VII
concludes the paper
II. BACKGROUND
A. Definition of Context
Different definition of context exits, according to [4]
Schilit and Theirmer introduced the term Context-awareness,
they defined context as location, objects and identities of
individual nearby. The general definition of context is defined
by Dey and Abowd, in [5-7]they defined “context” is any
information that can be used to characterize the situation of an
entity. An entity is a place or object that can be considered
relevant with the interaction between an application and a
user. The context information of a user consists of different
attributes, like physical location, emotional state such as
“calm, distraught or anger” physiological state such as “body
temperature heart rate or” behavioural patterns, personal
history etc.[8]
B. Categories of Context
Various researchers categorized context in different ways.
This will be of advantage for application designers to reveal
the most possible pieces of context that will be of benefit
when designing their application. [4]According to [5, 9]
Context is divided into two categories: Static and dynamic
Static Context often changes during system operation, such as
contact list, calendar, user profile, to-do-list, address book etc.
Dynamic Context is an information that is highly variable, like
temperature, time, location of the user, mood of the user
etc.[5]. And also work done by Schilit in [2] divided context
into three categories, they are: User Context, Physical Context,
Computing Context shown in Figure 1.

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

 Sensor
Context

Middleware

User context Physical Computing

context Context
E.g. user profile, E.g. Temperature, E.g. work
Location weather stations, printers
Figure 1: shows the different categories of Context and their examples.
 User Context consists of social situation, people nearby,
user’s profile and location.
 Physical context is made up of weather, temperature,
traffic conditions, lighting and noise level.
 Computing Context consists of the following which are;
resources close by such as work stations, printers and
displays, communication bandwidth and also network
connectivity [2].
The study researched by [6] characterised context into three
which are: Social context, Location Context, User Context.
Social context includes, related profiles, common interest etc.
Location Context consists of geographical information,
nationality etc. the User Context consist of the current activity
of the user.
End-user
Application
Figure 2: Distributed Context-aware systems
The sensor layer and end-user application layer are
compulsory, as they are context information consumers and
producers , the middleware is in between the sensor and end
application, it addresses the interaction or communication
issues between distributed components [8]. However, in
simple context-aware systems, the end-user application
connects directly with the sensor, for example a mobile phone
that switches off GPS automatically when the battery is below
a certain capacity. But [8] stated that this happened in early
systems, they specified that the incorporation of infrastructural
components are required so as to improve maintainability, to
decrease the complexity of using context-aware application
and support reuse [8].
B. The Components of Context-aware system
Context-aware system components was enumerated in [8] they
are shown in the table below:
Table 1: Context-aware system components

III. CONTEXT-AWARE SYTEMS Layer 0 Context Context Sensors and

Sensors and Actuators: this provides the
Context-aware systems provides new opportunities Actuators interface with the
entirely for end users and application developers by bringing environment either by
together data of the context and adapting the behaviour of the adjusting it actuators or by
system respectively. Mostly in the grouping of mobile devices, capturing it sensors.
the value of this mechanism are high and they are used to Layer 1 Components Components: they help with
improve usability [9]. Context-aware systems sense and adjust processing the outputs of the
their actions based on the changing of the contexts [5]. Current sensor to produce context
applications adopts a context-aware perspective to the information that are used by
interaction between the user and the system, or interaction applications, and map update
among users and systems [10] Context-aware system can be operations on the high order
defined as end-user applications that uses the context information back down to
information provided by sensors [8]. actions on actuators.
Layer 2 Context Context repositories:
A. Distributed Context-aware systems repositories provides persistent storage of
context information.
Context aware systems is distributed in three layers which are
Sensor, Middleware and End-user application. Shown in Layer 3 Decision Decision support: this tool
Figure 2 below. support assist applications to select
suitable activities and
adaptation based on the
 contextual information that is
available.
Layer 4 Application Application components:
components they are incorporated in client
applications using
programming toolkits.
C. Types of Context-aware system.
Context-aware systems can be categorized into two large sets
which are local and distributed. Distributed context-aware
system is then divided into two types which are; Collaborative
and Non-collaborative [8].
Context-aware Systems
Local Distributed
Collaborative Non Collaborative
Figure 3: Types of Context-aware System.
 Local systems are the systems in which applications
and sensors are joined firmly through a direct
physical connection, for instance a mobile phone
application that’s on silent mode, while the owner is
walking/running is a local system, as the
accelerometer that provides the Walking/running
context is attached directly to the mobile phone and
also the application that uses context to activate the
silent mode.
 Distributed systems does not have a direct physical
connection between application and sensor. As a
result the coupling is being loosed, it is likely to have
multiple applications getting information from the
same sensor, and it is also likely that multiple
dispersed sensors provide information to be
consumed by one application. For instance a mobile
phone might broadcast to a group of people that its
owner is currently running/walking to reduce the
number of incoming calls. This means the
accelerometer is not joined firmly to the receiver of
its information.[8].
The distributed systems is divided into two types: the
Collaborative and non-collaborative. [8] The
distributed collaborative systems are the type of
systems that helps multiple dispersed human achieve
a common goal and on the contrary non-collaborative
supports only individual goal.
IV. CONTEXT-AWARE RECOMMENDER SYSTEM
(CARS)
Traditional recommender system, Content-based,
Collaborative, hybrid and Knowledge-based recommender
system deal with types of entities which are users and items
[14]. This user/item recommender system was extended in
order to add more dimensions capturing the Context in which
recommendations are made. Context awareness majorly
covers understanding the user’s objective with services that
users might find interesting by knowing the broad range of
contextual attributes [15] this contextual attributes can be
acquired in some ways [14].
A. Attributes of Context-aware Recommender Systems
 Explicit: the context is captured when users input
data manually, modules of registration are frequently
used to capture information of users and rating
modules are being used to acquire interests and
preferences of the user [14]. This modules could
involve RFID tagging, through scanning [3].
 Implicit: these methods automatically captures
contextual information from the environment, for
example by obtaining the current location [14]. The
location Context is frequently discovered via implicit
actions accompanied by Location Based Service for
example (Wi-Fi) Wireless Fidelity location sensors,
Global Positioning System (GPS). They are used in
knowing and discovering the user’s physical location
at that time. Such detection are used as vital source of
information to accommodate information made
available through the system and can have effect in
the group of a recommendation. The Context of a
location is necessary in places such as galleries, Art
centres [3].
 Contextual Information: these method can be
concluded by analysing user interactions with
resources and tools for example, by estimating the
current task of a user. These information are used to
find or request a certain resources or repository, an
example is searching for restaurants [14].
B. Algorithms of Context-aware Recommender System
The three general algorithm of Context-aware
recommender systems are as follows: Contextual Pre-filtering,
Contextual Post-filtering and Contextual Modelling [11, 12].
 Context pre-filtering: In Context-aware pre-filtering,
the information about the present context is used to
select the set of data that are relevant and ratings are
predicted using 2D (users items ) traditional
recommender system [11]. The data rating which is
not related to the context is filtered out before
applying the recommendation algorithm [12]. The
major benefit of this algorithm is that it allows the
 positioning of any recommender system technique
[11].
 Context-aware post-filtering: In Context post-
algorithm, the result list of the recommendation is
being filtered taking the context values into
consideration [12] the ratings are predicted using 2D
recommender system on the whole data which is
(users items ) .[2]
 Context modelling algorithm: These systems make
use of the context information straight into the
recommendation function as a predictor for users
rating an item explicitly. The context information is
being added to the 2D traditional recommender
system which is the user and item data. Therefore the
approach formulates a multidimensional
recommendation which mainly signifies predictive
models which are made using regression, decision
tress, probabilistic models or heuristic calculations
that enhances the context information to item and the
user. Which is;
R : user  items  context  Rating [2].
Adomavicius et al., pointed out that in addition to the
existing two dimension (2D) recommendation
methods to multidimensional methods, there has been
some new methods developed specially for context-
aware recommender systems based on the context
modelling approach [11]. Numerous context
recommender systems have been evolved using this
algorithm in several application field [13].
V. APPLICATIONS OF CONTEXT-AWARE
RECOMMENDER SYSTEMS
Context-aware recommender system is a popular topic in
the research of recommender systems, there are various
applications in Context-aware Recommender System and they
can be categorized into six parts according to [14] which are;
smart space that provides smart environment for users, A tour
guide which directs tourists, communication system that
provides m-commerce, web service social community,
information system. However, new applications are constantly
suggested [11]. Most especially in the aspect of tourism [7], in
the banking environment [15] etc. For instance Gallego and
Huecas [15] presented a model and architecture of context-
aware recommender system in a banking environment, taking
into consideration the social, mobility and ubiquitous
requisites to provide recommendations to customers.
Meehan et al. [7] proposed a mobile application Virtual
Intelligent System for Informing Tourists (VISIT) to prove the
concept of context-awareness. The application used a hybrid
recommendation approach which is made up of collaborative
filtering, content- based filtering and demographic filtering.
Intelligent reasoning was performed as part of the hybrid
system to determine the importance/weight of each different
context type.
Madhusudanan et al.[4] presented a Mobile Assistance
System (MAS), he explained that MAS is classified into three
modules which are; Environment module, which tells the
present status of the environment such as noise temperature,
lighting, etc. the second is Time Module, which is used to
know the current country, time, date and the current season
and Location Module which is been used to recover the
suitable information that the user would need based on his/her
location.
Research paper by Champiri et al. [16] elaborates the role
of context for digital libraries, the article stresses out the need
for qualitative study on the subject so as to find out the
elements that develops context from the user’s point of view,
and the way context-aware recommender system could be
designed in an effective way.
Wu and Shih [17] proposed a new framework for context-
aware recommender system to ease the issue of cold start and
making use of rich resources for user generated content, the
frame work collects ratings and extracts related contextual
information from a social media.
Beer et al. [18] presented a general software architecture
and proposed a general framework for the implementation of
context-aware recommendation, the frame work of this system
suggests high flexibility according to the configuration and
definition of new impact functions, which influences the
recommendation of items for given users.
Work done by Molina et al. [19] proposed a content
model that is made on an evaluation and measurement
framework named Information Need, Concept Model,
Attribute, Metric and indicator. (INCAMI) to allow
recommendations in the measurement and also the evaluation
of project. The representation of the designed contextual
information can also be used in any field, and can be
considered as a domain-independent method.
Hariri et al.[20] presented a Context-aware recommender
system for music that gathers context information which is
based on the latest sequence of songs that the user likes. The
system mines popular tags for songs from the last FM website,
the data tag that was used by the topic modelling module
which is Latent Drichlet Allocation model (LDA) was used to
infer the topic probability distributions for songs.
Rodriguez et al.[21] proposed a Context-aware system
using multi-dimensional ontology model to represent mobile
user contexts, web services and application domains, the
model includes application domain, semantic and the user
contextual information, web services interface modelling. The
main idea was to support the development of more complex
and intelligent mobile context-aware applications.
Pessemier et al. [22] proved that Bayesian Classifier
permits the addition of the consumption context such as time,
location, user’s mood to exploit the flexibility and also to
address the context dependent likes of user.
Braunhofer et al.[23] proposed a novel context-aware
mobile recommender system for places of interest (POI)
which they named South Tyrol Suggests (STS) that learns
users preferences from their past ratings and their personality
as well, the users personality is acquired through a brief five
item questionnaire that is subsequently used to actively elicit
ratings for POI that were assessed to be experienced by users.
VI. DISCUSSION AND OPEN ISSUES
Context information is been used as the main basis for
efficiency and the accuracy of predictions in recommender
systems [16] various research have been revealed that using
the context information of a user can improve the performance
of recommender systems. [16] however, the major concern in
designing a successful context-aware recommender system is
the finding of context factors that are of value when predicting
and making suggestions/recommendations, this issue needs
framing of informed estimations on the effect of some factors
that are certain before gathering data in real environment.[2].
Another aspect which is also important is privacy and security.
Contextual information takes into consideration the user
profile information and other sensitive information. Also,
Proactivity can be integrated in Context-aware recommender
system, this system help users to get recommendation of items
and places automatically without submitting a query or
request. [1] Especially in the aspect of tourism, a tourist would
like to have a mobile system with an application that would
suggest places to them right on the spot.[24]. There is need for
ideas to express strategies and rules and also to define
ownership of context information.[25].
VII. CONCLUSION
Recommender systems has been a fast evolving area of
software systems, it has been a very important tools for
distinguishing information that are relevant from irrelevant
information for users in various areas or fields such as tourist
field, restaurant field, education, and also in the entertainment
aspect. However, recent development in recommender system
incorporates the use of contextual information to make
recommendations in certain situations such as ubiquity and
mobility. This paper made use of existing literature to search
for this aspect of recommender system. Context-aware
recommender systems has received great attention in recent
years, it recommends items/services to users based on the
contextual information, and it was integrated to increase the
effectiveness of recommender systems. Therefore there are
some limitations enumerated in the discussion part of the
paper that should be treated in order for the system to be more
efficient and trustworthy.
ACKNOWLEDGEMENT
The author acknowledge the financial support and resources
made available by Tshwane University of Technology.
REFERENCES
[1] W. Woerndi, J. Huebner, R. Bader, and D. G. Vico, "A Model for
Proactivity in Mobile, Context-aware Recommender Systems,"
presented at the RecSys, USA, 2011.
[2] N. Y. Asabere, "Towards a Viewpoint of Context-Aware
Recommender Systems (CARS) and Services," International
Journal of Computer Science and Telecommunications, vol. Vol 4,
pp. 19-29, 2013.
[3] K. Shah, A. Raundal, G. Bhandari, S. Rathi, P. Raikar, and P.
Mahalle, "Proposed Profile Translation in Context Aware
Proactive System" Indian Journal of Computer Science and
Engineering (IJCSE), vol. vol 3 pp. 763-767, 2012.
[4] J. Madhusudanan, A. Selvakumar, and R.Sudha, "Frame work for
Context aware Applications," presented at the Second international
conference name on Computing, Communication and Networking
Technologies, 2010.
[5] K. F. Yeung and Y. Yang, "A Proactive Personalized News
Recommender System," presented at the Development in E-
systems Engineering (DESE)London, 2010.
[6] D. Gallego, E. Barra, P. Rodriguez, and G. Huecas, "Incorporating
Proactivity to Context-Aware Recommender Systems for E-
Learning " presented at the IEEE, 2013.
[7] K. Meehan, T. Lunney, K. Curran, and A. McCaughey, "Context-
Aware Intelligent Recommendation System for Tourism "
presented at the Work in Progress session at PerCom San Diego,
2013.
[8] P. Alves and P. Ferreira, "Distributed Context Aware Systems,"
2011.
[9] J. Kwon, "A Study on Effective Proactive Service in Pervasive
Computing Environment," International Journal of Computer
Science and Network Secirity (IJCSNS), vol. Vol 12, 2012.
[10] C. Bolchini, G. Orsi, E. Quintarelli, F. A. Schreiber, and L. Tanca,
"Context Modelling and Context Awareness: steps forward in the
Context-ADDICT project.," presented at the IEEE Computer
Society Technical Commitee on Data Engineering, 2011.
[11] G. Adomavicius, B. Mobasher, F. Ricci, and A. Tuzhilin. (2011)
Context-Aware Recommender System. Association for the
Advancement of Artificial Intelligence. [Magazine].
[12] A. Smirnov, A. Kashevnik, A. Ponomarev, N. Shilov, M.
Schekotov, and N. Teslya, "Recommendation System for Tourist
Attraction Information Service " presented at the PROCEEDING
OF THE 14TH CONFERENCE OF FRUCT ASSOCIATION,
2013.
[13] K. Verbert, N. Manouselis, X. Ochoa, M. Wolpers, H. Drachsler, I.
Bosnic, et al., "Context-aware Recommender Systems for
Learning: a Survey and Future Challenges," JOURNAL OF LTEAX
CLASS FILES, vol. Vol 6, pp. 1-20, 2007.
[14] J. Hong, E.-H. Suh, J. Kim, and S. Kim, "Content-Aware System
for Proactive Personalized Service based on Context history,"
Expert Systems with Application, pp. 7448-7457, 2009.
[15] D. Gallego and G. Huecas, "An Empirical Case of a Context-aware
Mobile Recommender System in a Banking Environment,"
Journal of convergence ,Future Technology Research Association
International, vol. Volume 3, pp. 49-56, December 2012 2012.
[16] Z. D. Champiri, S. S. B. Salim, and S. R. Shahamiri, "The Role of
Context for Recommendations in Digital Libraries," International
Journal of Social Science and Humanity, vol. Vol 5, pp. 948-954,
Nov 2015.
[17] C.-C. Wu and M.-J. Shih, "A Context-Aware Recommender
System Based on Social Media " presented at the Int'l Conference
on Computer Science, Data Mining & Mechanical Engg.
(ICCDMME’2015), Bangkok (Thailand), 2015.
[18] W. Beer, W. Hargassner, C. Derwein, and S. Herramhof, "General
Framework for Context-Aware Recommendation of Social
Events," presented at the The Second International Conference on
Intelligent Systems and Applications 2013.
[19] H. Molina, L. Olsina, and G. Rossi, "Context-Based
Recommendation Approach for Measurement and Evaluation
Projects " J. Software Engineering & Applications, vol. Vol 3, pp.
1089-1106, 2010.
[20] N. Hariri, B. Mobasher, and R. Burke, "Context-Aware Music
Recommendation Based on Latent Topic Sequential Patterns,"
presented at the RecSys' 12, Dublin, Ireland. , 2012.
[21] J. Rodríguez, M. Bravo, and R. Guzmán, "Multi-dimensional
Ontology Model to Support Context-aware Systems: A Mobile
Application to Search and Invoke Semantic Web Services "
presented at the The Seventh International Conference on Internet
and Web Applications and Services (ICIW), 2012.
[22] T. D. Pessemier, T. Deryckere, and L. Martens, "Extending the
Bayesian Classifier to a Context-Aware Recommender System for
Mobile Devices," presented at the 2010 Fifth International
Conference on Internet and Web Application and Services 2010.
[23] M. Braunhofer, M. Elahi, and F. Ricci, "A Context-Aware Mobile
Recommender System for Places of Interest," in Proceedings of
UMAP 2014 pp. 75-80.
[24] H. A. Tair, M. J. Zemerly, M. AL-Qutayri, and M. Leida,
"Architecture for Context-Aware Pro-Active Recommender
system," International Journal Multimedia and Image Processing
(IJMIP), vol. 2, pp. 125-133, March/June 2012 2012.
[25] M. Baldauf, S. Dustdar, and F. Rosenberg, "A survey on context-
aware systems " Int. J. Ad Hoc and Ubiquitous Computing, vol.
Vol. 2, pp. 263-277, 2007.
 A Standard Compliant Blood Bank Management
System with Enforcing Mechanism
Amarjeet Singh Cheema, Siddharth Srivastava, P K Srivastava, Dr. B K Murthy
Centre for Development of Advanced Computing,
Noida, India
{ascheema, siddharthsrivastava, pksrivastava, bkm}@cdac.in
Abstract—Blood is a non-replenishable entity, the only source
of which are humans. Timely availability of quality blood is a
crucial requirement for sustaining the healthcare services.
Therefore, maintaining quality of blood and identification of
Professional Donors is a major responsibility of blood banks.
NACO (National AIDS Control Organization) and NABH
(National Accreditation Board for hospitals and Healthcare
Providers) have provided guidelines for ensuring the quality of
blood and identifying Professional Donors. Moreover, manually
monitoring standards and identifying professional donors is a
challenging job. In this work, we develop a standard compliant
Blood Bank Management System with a novel rule based
enforcing mechanism. The developed system is an end-to-end
solution for not only managing but implementing enforcing
strategies and providing decision support to the users. The
proposed Blood Bank Management System has been
implemented across 28 blood banks and a major hospital. It has
been found extremely effective in streamlining the workflow of
blood banks.
Keywords—Blood Bank Management System; Blood Stock;
NACO; NABH; Professional Donor; Donor Repository
I. INTRODUCTION
The major concern of blood banks is to ensure efficient and
effective collection and maintenance of quality blood stock as
well as identification of Professional Donors (Section II-E).
This becomes crucial since the span of time, especially in
emergency situations, between requirement, arrangement and
delivery of blood is very narrow. Moreover, blood banks
across the state, districts are not able to utilize the available
blood stock appropriately due to lack of connectivity and time
taken to propagate information via conventional channels.
National AIDS Control Organization (NACO) and National
Accreditation Board for hospitals and Healthcare Providers
(NABH) have provided guidelines to ensure the quality of
blood. But there is absence of effective enforcement strategy
to ensure the adherence to these guidelines. In view of this, we
propose a comprehensive IT solution i.e. a Blood Bank
Management System (BBMS) attempting to address this
problem by providing means to connect, digitize and
streamline the workflow of blood banks.
The need for automating blood banks have been there for a
long time. In early days of digitization, the primary purpose of
an IT solution for blood banks was inventory management
[6][13]. With time, the processes involved in management of
services of blood bank as well as Blood Transfusion System
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
[5][4] have become more complex. The main issue, which
plagues BBMS in the country is enforcing the standards and
identification of the Professional Donors. Therefore, in the
modern world the purpose of BBMS is not only to passively
act as inventory management system but to actively enforce
standard operating procedures along with providing decision
support. Recent Blood Bank Management Systems tend to
focus on adapting the system to local practices instead of
enforcing standard practices. The developed solution
augments the functionality of the contemporary systems.
Authors in [15] developed a blood bank management system
which adhered to the requirements of a single hospital. The
system developed in [3] caters to a National level transfusion
service, but limits the scope to providing citizen centric
services and inventory management. Authors in [8], attempt to
address the issue of safe transfusion by developing an end to
end solution. A real time system for blood bank has been
proposed in [1].
A recent study [12] has observed that the existing
workflow of blood banks needs to be strengthened, both in
terms of planning and monitoring. Also, there are many gaps
in the management of blood supply [16] [9]. Our work
addresses the following gaps as compared to similar systems.
Firstly, the existing systems have been designed to take care of
routine functioning of blood bank and care not able to enforce
the guidelines and standards based on rules. Secondly,
identification of professional donors, if available, is usually
based on bio-metric devices only. Such identification
mechanisms mostly require an additional level of integration
effort and scrutiny by the blood bank staff. Thirdly, quality
checks are based on manual entry processes. In addition to
these fundamental issues, there are many challenges that are
encountered when such a system is implemented across a large
number of locations. These challenges are discussed in Section
II-A. In view of the above, the contributions of this paper are:
 Development of a standard compliant Blood Bank
Management System including a stringent rule based
enforcing mechanism.
 Architecture for a fault-tolerant deployment
especially for rural and areas with sparse
connectivity.
 Identifying key requirements and learnings from im-
plementation of a Blood Bank Management System
for large scale deployments
 The rest of the paper is organized as follows. The next
section describes the proposed system. Section III presents the
details of implementation. The impact of the system as
compared to pre-deployment scenario is discussed in Section
IV. Finally, Section V concludes this paper.
II. DESCRIPTION
A. Challenges
The major challenges faced by Blood Banks, in general,
due to lack of a comprehensive IT solution are:
 Absence of centralized Information and Knowledge
Management
 It is not possible to identify Professional Donors. The
donors who repeatedly donate blood usually in exchange
of money or other remunerations are termed as
Professional Donors. Mostly, such donors donate blood
multiple times even within the minimum donation
eligibility period. This results in reduced quality of blood,
making the receiver vulnerable to infections etc.
 In emergency cases, the process of searching records in
registers and contacting suitable donors was highly
inefficient.
 Notifying about blood donation camps to registered
donors is time consuming.
 There is no provision to view real-time stock reports.
Manual search through registers was required for
obtaining details about a donor or blood type.
 Continuous monitoring of Standards and quality checks is
not possible for each donation in the blood banks.
 Due to large number of samples being processed
simultaneously, the chances of typographical errors and
incorrect entries are high.
 The manual system of blood collection and requisitions
has long waiting time in queues, which in turn leads to
delay in issue of blood. In addition, the manual record
keeping results in inefficient utilization of man-power at
the blood banks.
 The process of manual record keeping limits the
effectiveness of monitoring mechanisms.
 Lack of efficient tracking facility for soon to expire blood
stock. Moreover, transfer of such bags to other blood
banks is also not possible due to lack of information about
requirement at other places.
 The maintenance of records during non-peak hours is a
major challenge as there was shortage of staff that is
responsible for maintaining elaborate manual registers.
B. Proposed System
The proposed solution strictly enforces NACO (National
AIDS Control Organization) and NABH (National
Accreditation Board for hospitals and Healthcare Providers)
guide-lines ensuring safer collection and donation, effective
management and monitoring the quality of the donated blood.
In order to realize this, the BBMS has been developed on
configurable rule based architecture allowing customizations
to easily incorporate specific requirements from the
stakeholders. The BBMS has six major components for
management of the blood donation life cycle. Firstly, the
biometric Donor Management System for identifying, tracking
and blocking donors based on donors health, donation history
etc. Secondly, it provides features such as cross matching ,
antibody screening, blood grouping, component preparation,
issue of compatible blood and blood components, TTI
screening etc. as per the guidelines. Thirdly, a centralized
Blood Inventory Management System for keeping track of the
blood stock across numerous blood banks. In addition to this,
it also manages inventory for refreshments and other
consumable items. Fourthly, Bio-Medical Waste Management
System for disposal of unsuitable blood and other waste
generated during this process. Fifthly, an Alert and
Notification System having two modes. An internal mode,
where an alert is generated when the system observes any
anomaly from the defined processes and rules. An external
mode, where the eligible donors registered with the system are
notified of donation requests and scheduled camp information
via SMS/Email. The eligibility of the donors is identified
based on the proximity between the geographical location of
the request and the donor, previous donation, health and
donation history as available in the system. Lastly, a BBMS
API for ubiquitous integration with multiple platforms and
services. This results in easy integration of BBMS into other
services such as existing health portals, allowing for instant
availability of the service though existing modes of delivery to
the citizens and organizations. All components of BBMS work
in synchronization, hence transforming the traditional isolated
structure of blood banks to centralized structure. It helps
bringing transparency, enhancing quality of service and
effectively reducing the time for delivery of the service to the
citizens. The citizens of the country benefit the most from this
initiative since during medical emergency, one can directly
access the blood stock details and request for the same to
multiple blood banks through a single point of contact i.e.
BBMS.
C. Workflow
Figure 1 shows key features of various modules of the Blood
Bank Management System, i.e. Core Services, Support
Services and Back Office. The system constitutes of
registering the donor at the blood bank premises or at the
camps arranged by the blood bank itself. Every donor
undergoes a physical examination test and the donors
examination results are entered into the system. Depending on
the examination result, the system decides whether to accept
or reject the donors blood. The blood of the accepted donor is
collected in a bag. The system assigns a unique number to the
bag, which can also be used to identify the donor. The
standard procedure requires performing certain tests on the
collected blood. Initially serology tests are performed and the
results are entered into BBMS through the Result Entry
Process. The results are then validated by an authorized staff.
Next step is to perform blood grouping whose results are again
entered into the system. A senior person validates this blood
Fig. 1. Modules of Blood Bank Management System
D. Enforcing Mechanism: Ensuring the NACO and NABH
Guidelines
The enforcing mechanism is a set of rules that are enforced by
the system. These rules can be configured using masters which
are easy to use graphical user interfaces designed especially
for this purpose. The rules are used for validating the data
when it is processed by the BBMS (Figure 2). Whenever the
data is saved into BBMS, it first passes through the validation
engine where integrity of the data as per the defined rules is
verified. If the data is valid it is processed further else the
system enforces the defined strategy to reject or update the
status of the data. In Addition to the NACO and NABH
guidelines the validation engine can even be configured for
incorporating blood bank specific rules in certain modules.
Fig. 2. Workflow of Enforcing Mechanism
E. Biometric Authentication: To Restrict Professional Donors
The Professional Donor refers to a person who donates blood
in lieu of some valuable consideration on behalf of the
recipient-patient. This may include paid as well as commercial
donors. To identify such donors biometric authentication
process has been integrated with the Donor Registration
Process. The system allows blood bank to set a customized
fingerprint capturing policy which may involve verification of
more than one fingerprint. On registration, the system searches
for a previous record based on the fingerprint. If a match is
found, the donor details, donation history etc. are returned.
The system then automatically calculates the locking periods
of donation as per the last donation type and date. If a donor
has already donated blood within the locking period, the
system will not allow further processing and suggests the next
group detail on the system. After this, various blood
components are separated. For every requisition, blood is
issued after cross-matching. The stock inventory is updated
immediately and the issue details are stored in the system.
visit date. In addition to this, the system has a secondary
verification process for the identification of professional
donors. If biometric system doesn’t get any matching details,
then the system provides a list of existing donors which match
the information provided by the donor during registration as
per the matching criteria configured in the system.
E. Health Standard and Guideline Compliance
The BBMS is compliant with the following health standards
and guidelines.
 Standard Code Sets
o ICD 10: International Classification of Diseases
[11]
o SNOMED: Systemized Nomenclature of
Medicine [2]
o LOINC: Logical Observation Identifiers
Names and Codes [7]
 Guidelines
o NACO: National AIDS Control Organization
[10]
o NABH: National Accreditation Board For
Hospitals and Health Care Providers
F. Interfacing with BBMS
The Blood Bank Management System provides interfacing
capability with various channels. The developed solution is
Compatible with the following interfaces, using which the
Functionality of the system can be extended.
 Bar Code Interface
 Biometrics
 Temperature Tag
 Smart Card
 SMS and e-Mail Server
 Generic Medical Equipment Interface [14]
 III. IMPLEMENTATION AND RESULTS
The Blood Bank Management System described above has
been implemented across 28 blood banks across a state and a
major hospital. The blood donations received are shown in
Table I. The system is deployed over a cloud infrastructure
adhering to stringent security and privacy policies. The system
is accessible to the staff of the blood banks via a dedicated
private network ensuring security and privacy of the
information while the citizen services can be accessed through
a web portal.
TABLE I. BLOOD DONATIONS USING BBMS
Statistics
Total Blood Donations 2,94,571
Average Donations/Blood Bank 10,520
Average Donations/Day 807
The key features of the system are briefed below:
 Web Based Application.
 Aadhar Number Integration
 Configurable for Multiple Blood Banks.
 Enforcing Mechanism for monitoring standard practices.
 Biometric Interfacing for restricting professional donor.
 Bilingual Questionnaire and Consent Printing.
 Bar Coded Bag No/Sample Number Generation.
 Generation of Alerts, MIS Reports and Trends.
 Dynamic Dashboard.
 User Management: Roles and Rule.
 Compliance to various Standards.
 Provision of Smart Cards for Master Donors
 Web Services for Heterogeneous Integration
 Mobile Apps
A. Deployment Architecture
The system has been developed using Java-J2EE Architecture,
Struts framework and Oracle 11g/PostgreSQL as backend.
The solution architecture is designed to operate between two
sites viz. a centralized Data centre and Intranet of systems
setup in the beneficiary blood banks. The system architecture
is shown in Figure 3.
The application is further divided into layers based on a
responsibility layering strategy that associates each layer with
a particular responsibility. This strategy has been chosen as it
isolates various system responsibilities from one another,
hence improving efficiency while developing the system as
well as its maintenance.
Fig. 3. System Architecture
B. Infrastructure
1) Centralized Data Center: The application has been
deployed in a data-centre, which is running 24x7 and is
making the services available over multiple Blood Banks
using the backbone network of MPLS/VPN cloud. This
backbone cloud is set up with more than 99.9% availability to
ensure that the services are available in the blood banks with
minimal downtime. The system in Data centre consists of (a)
SAN Storage (b) DB Servers (c) Application Servers (d)
Backup and SAN Management Server (e) Anti-virus Server (f)
DNS Server (g) NMS Server (h) Mail Server and (i) The File
and Print Server. These are interlinked over a Gigabit LAN.
The dedicated hardware for the project comprises of two
Application Servers based on RAID-5 and two Database
Servers based on RAID-1 technology. The Application
Servers are connected to a Load Balancer (CISCO Content
Manager) that regulates external traffic from the MPLS-VPN
Network via router. The model is based on double redundancy
in order to minimize system failure and hence increase the
operational efficiency.
2) Connectivity between Blood Banks and Centralized Data
Center: The connectivity between the central Data Centre and
the Blood Banks is one of the major components of the cloud
architecture. A Virtual Private network (VPN) is established
between the Blood Banks and the data centre using physical
network links. This is achieved by sourcing the services from
an Internet Service Provider (ISP). For the purpose of
redundancy, services from multiple ISPs have been sourced.
3) Infrastructure at the Blood Banks: The infrastructure at the
blood banks consists of Gigabit LAN connection from (two
distinct providers, client PCs and additional hardware such as
printers etc. as per the requirement. In case of the failure of
connectivity between the beneficiary blood bank and the data
center, the system automatically switches over to the other
provider. In case, both the networks are down, local servers
have been installed at the blood banks to continue the
operations till the connectivity is resumed.
IV. DISCUSSION
Prior to the implementation of the solution, the entire
workflow of the concerned blood banks was manual. When a
donor arrived, his/her entry was made into multiple registers.
Once the blood was collected in blood bags, various tests were
performed on the collected samples. The results were again
written into registers. In case of requirement of blood, a citizen
had to enquire to each blood bank separately either through
phone or in person. Moreover, there was no effective
monitoring mechanism at the blood banks to ensure the quality
of the collected blood. In such a scenario, chances of the
stocked blood being expired without being used were high. The
reasons were two fold. Firstly, there was no timely intimation
to the concerned official about the stock going to expire. This
prevented appropriate measures to be taken by authorities to
transfer the stock to places where there may be requirement.
Secondly, since each blood bank operated in isolation,
determining such requirements was tedious and had to be
initiated at the end of the blood banks. In addition to this, there
was no provision at all to identify donors including professional
donors. The professional donors were not only troublesome to
the blood banks since their blood did not satisfy the quality
parameters for donation wasting time and effort but also
harmful to the recipients. For administrative purposes, the
authorities received statutory reports of donor and stock
separately from each blood bank on monthly basis. This helped
the concerned authorities to provide controlling guidelines to
blood banks. The controlling guidelines may include
scheduling blood donation camps, reducing demand supply gap
by encouraging voluntary donations, enhancing infrastructure at
blood banks based on the stock etc.
After the deployment of the solution, the complete process of
donating blood and its management has not only become
simpler but effective as well. The system has helped the
authorities to enforce standard practices and at the same time
makes it convenient for the staff, saving their precious time and
efforts. Due to real time availability of stock, the system has
allowed citizens to enquire and arrange blood a lot quicker,
hence saving many invaluable lives. The statutory reports can
be viewed on the most recent data, allowing authorities to issue
controlling guidelines quickly and extract statistical
information easily. Most importantly, the solution has brought
transparency and accountability to the system, putting the
power at the hands of citizens. In Table II, we show specific
improvements on numerous parameters brought in by BBMS as
compared to the pre-deployment scenario.
A. Key Learnings
A successful and effective system unfolds and solves many
issues. In the following, describe the key learnings from our
work.
1) Key Learning for Improving Citizen Centric Services:
 Automation Should Reduce Effort: Any such application
should aim at prioritizing the reduction in the workload of
the end user rather than only automating the workflow. In
order to achieve this, the system should be designed to
minimize the inputs at multiple stages including
intermixing of offline and online data entries; efforts
should be made to capture the information directly into the
system with proper validation instead of first capturing on
registers and then entering them into the system. Moreover,
specially designing the Graphical User Interface (GUI) as
per the existing processes not only minimizes the training
time but also enhances the productivity of the end user.
 Secondary Notifications: In a domain such as health and
especially services dealing with emergency cases such as
blood banks cannot solely rely on passive generation of
reports or alerts confined within the system. Critical alerts
such as blood expiry notifications, unattended emergency
blood requests etc. should be intimated to concerned
authorities via SMS/Push Notifications.
 Importance of Mobile Applications: A key requirement
was to make the camp details, reports, stock details etc
available to the public through web portal. During the
operation of the project, it was observed that availability of
such information via Mobile Apps would help reaching a
wider range of people. It was also observed that push
notifications to the users via Mobile Apps and their
consent/feedback through the same would result in
reaching out a higher number of donors.
 Online Training Videos/Manuals: Transfer of staff,
change in work profile etc. are a few common issues in
such organizations. In such cases, continuous retraining is
not possible. Therefore, online training material including
interactive tutorials and videos should be available to the
staff. In addition to this, a comprehensive help including
tips to use the system should be provided for boosting the
efficient use of the system.
2) Key Learning for Project Implementation:
 Site Readiness: Site preparation should be parallel activity
with development and it should be initiated immediately
after the initiation of project. Roles and responsibility of
stakeholders should be cleared and communicated
properly. Regular meeting should be scheduled to discuss
the status of site. Responsibility for site preparation should
be well defined in Project Signing Document.
 Stakeholder Participation: Stakeholder engagement is
the process of involving people who either influence or are
affected by the decisions taken by an organization. It plays
a vital role in successful implementation of the project.
Their inclusion and recognition in the BBMS decision-
making process is the key to the successful implementation
and use of the system. The module wise nodal officer and
team for requirement gathering and workflow finalization
should be identified. End user should be involved from the
requirement gathering phase. Requirement gathering team
should also interact with end users, analyze their working,
examine the register and reports which are being used so
that complete requirement is gathered.
 Service Provider for Backbone Connectivity: Reliable
network connectivity critically influences the performance
of an application. The requisite quality is achieved by
opting for two service providers. First with fiber based
lease line connectivity and the second with RF based last
mile connectivity. This ensured lower network downtime
while keeping the system operational in a transparent
manner.
 Initial Teething Troubles: Since, the staff at the blood
banks was not acquainted to use such system; a Project
 Management Unit (PMU) was required to be stationed at enforcing mechanism. The paper discussed an architecture for
various blood banks for handling technical issues. Due to a fault-tolerant deployment especially for rural and areas with
the critical nature of services at blood banks, an additional sparse connectivity. The challenges faced in such large scale
responsibility of the PMU was to educate and enable the deployments of citizen-centric services were identified along
blood bank staff to handle the technical issues as well. with key learnings from such implementation.
V. CONCLUSION
The paper discussed development of a standard compliant
Blood Bank Management System with a novel rule based
TABLE II. COMPARISON OF PRE-DEPLOYMENT AND POST-DEPLOYMENT SCENARIOS
S.No. Activity Pre-Deployment Post-Deployment
1. Stock Availability a) Telephonic/in-person visit by Patients, Hospital staff to a) Online availability of stock.
each blood bank. b) Information from multiple Blood Banks is shown in
b) Chances that the operator either does not know the exact consolidated manner.
stock or hides the stock details
2. Donor Repository a) No centralized repository of donors. a) Centralized donor repository.
b) Tedious and time consuming process to search registers b) One click search of registered volunteer donors across
for volunteer donors having a particular blood group. multiple blood banks.
3. Enforcing and a) NACO and NABH guidelines were manually monitored. a) The NACO and NABH guidelines are strictly enforced.
Authenticity Mechanism b) Chances of typographical errors and lack of consistency b) Bar coded labels available on Bags synchronized with
due to manual entry. the database. This helps in reducing manual errors due to
misplaced tags or entries.
4. Statutory Reports a) Manual preparation of stock reports. a) Real time availability of statutory reports.
b) Delay in dispatch or generation of reports, b) Easy to extract statistical information.
c) No statistical analysis
5. Duplicity a) Repeated entry in registers, a) Once the information is captured, it is available to be
b) Result from Medical Equipments were entered manually used in other forms and reports without re-entry.
in registers b) Generic Medical Equipment Interface for automatically
mapping results from machines to blood samples and
inserted into BBMS.
6. Efforts a) The staff was occupied in preparation of statutory reports a) Online Reports reduce this effort.
and register entries. b) Automatic and immediate availability of statistical data
b) Stock count, donor count were done manually based on various parameters.
7. Transparency a) No provision to display the real time stock in Blood Bank. a) Comprehensive User Management and Access Policies,
b) In case shortage of blood in blood banks, no mechanism b) Transaction Log availability,
to get other blood bank stock status c) Real Time information available through public portals.
8. Accuracy a) Donors Physical fitness was directly dependent upon a) On the basis of keyed values system suggests donor
Doctors Examination. fitness. This help in cases where a doctor is not available
b) Bag Grouping and whether the bag is reactive for disease or a technician conducts the tests.
depended upon technician entry in registers. b) With validations, cross checks, bar codes in place at
c) Chances of typographical errors were more. various stages, the chances of errors have been reduced.
9. Response Time a) Long queues due to manual and multiple register entries, a) Waiting time has been reduced,
b) Patient's relatives needed to travel to the blood banks in b) With the availability of portal and capability to
search of Blood. integrate KIOSK, registration time can further be reduced.
c) Stock enquiry was time consuming c) Users can get stock information through the
Portal.
10. Temperature Monitoring a) No monitoring system available during the transportation a) With the Temperature Tag, the chain of temperature is
during blood of bags from camps to the blood bank and from blood bank noted during transportation. In case the bag is exposed to
transportation to transfusion centers. The bags which should have been an unsafe temperature, the bag can be rejected at the
rejected due to exposure to unsafe temperature (if any), were recipient’s end.
accepted and used.

REFERENCES
[1] N. Adarsh, J. Arpitha, M. D. Ali, N. M. Charan, and P. G.
Mahendrakar. Effective blood bank management based on rfid in real
time systems. In Embedded Systems (ICES), 2014 International
Conference on, pages 287–290. IEEE, 2014.
[2] L. Bos and K. Donnelly. Snomed-ct: The advanced terminology and
coding system for ehealth. Stud Health Technol Inform, 121:279–290,
2006.
[3] E. Ekanayaka and C. Wimaladharma. Blood bank management system.
Technical Session-Computer Science and Technology & Industrial
Information Technology, page 7, 2015.
[4] L. T. Goodnough, M. E. Brecher, M. H. Kanter, and J. P.
AuBuchon.Transfusion medicineblood transfusion. New England
Journal of Medicine, 340(6):438–447, 1999.
[5] D. M. Harmening. Modern blood banking and transfusion practices.
FA Davis, 2012
[6] H. Lowalekar and N. Ravichandran. Blood bank inventory
management in India. OPSEARCH, 51(3):376–399, 2014.
[7] C. J. McDonald, S. M. Huff, J. G. Suico, G. Hill, D. Leavelle,
R. Aller, A. Forrey, K. Mercer, G. DeMoor, J. Hook, et al. Loinc, a
universal standard for identifying laboratory observations: a 5-year
update. Clinical chemistry, 49(4):624–633, 2003.
[8] M. F. Murphy, E. Fraser, D. Miles, S. Noel, J. Staves, B. Cripps, and J.
Kay. How do we monitor hospital transfusion practice using an
end-to-end electronic transfusion management system? Transfusion,
52(12):2502–2512, 2012.
[9] M. F. Murphy and M. H. Yazer. Measuring and monitoring blood
utilization. Transfusion, 53(12):3025–3028, 2013.
[10] N. A. C. Organization. Standards for blood banks and transfusion
services. 2007.
[11] W. H. Organization et al. International classification of diseases
translator: ninth and tenth revisions: user’s guide to electronic tables
[computer file]. 1997.
[12] K. Ramani, D. V. Mavalankar, and D. Govil. Study of blood-
transfusion services in maharashtra and gujarat states, india. Journal of
health, population, and nutrition, 27(2):259, 2009.
[13] C. Sapountzis. Allocating blood to hospitals from a central blood bank.
European Journal of Operational Research, 16(2):157–162, 1984.
[14] S. Srivastava, R. Gupta, A. Rai, and A. Cheema. Electronic health
records and cloud based generic medical equipment interface. National
Conference on Medical Informatics (NCMI), AIIMS, New Delhi, 2014.
[15] S. Sulaiman, A. A. K. A. Hamid, and N. A. N. Yusri. Development of
a blood bank management system. Procedia-Social and Behavioral
Sciences, 195:2008–2013, 2015.
[16] L. M. Williamson and D. V. Devine. Challenges in the management of
the blood supply. The Lancet, 381(9880):1866–1875, 2013.
 A Survey of Recommender System Feedback
Techniques, Comparison and Evaluation Metrics
Masupha Lerato, Omobayo A. Esan, Ashley-Dejo
Ebunoluwa
Dept. of Computer Systems Engineering
Tshwane University of Technology, TUT
Soshanguve, South Africa Country
{masuphaLE, esanoa}@tut.ac.za,ebunashley@gmail.com
Abstract—Recommender system is an application that
filtered personalized information based on user preference
from large information obtained. However, in order to provide
effective and efficient recommendation it is necessary to
accurately construct user’s interest and taste. In constructing
the user interest and taste there is need to obtain accurate
feedback from user on the recommendation. In order to
achieve this, various Recommender Systems techniques can be
employed. This paper intends to highlight various
Recommender Systems, their merits and demerits; and
overview of recommender System is discussed.
Keywords—Recommender Systems; personalized
information; interest; taste
I. INTRODUCTION
Users often encountered difficulties in finding the
content they want quickly and in easy way due to the large
amount of information available on the internet [1, 2]. Most
of the time user tries to seek assistance from others who had
previously had the same needs for those particular items or
alternatively choose item that are closer to what they are
looking for and this sometimes resulted in information
overload problem. Employing information retrieval
technique such as Recommender Systems (RS) can assist in
addressing this challenge as this can help user to select items
from internet based on user’s interest and taste.
Also, the issue of inaccurate estimation of user
taste and interest is another challenge affecting RS [2]. This
is due to the fact that for RS to work effectively and
efficiently, it always request user’s model interest [3]. In
order to have smooth and effective RS, there is need to
highlight various feedback techniques and determine which
is most suitable to address issue of user satisfaction accurate
and also highlight various performance evaluation scheme
that can be used for feedback Recommender Systems
techniques.
The major contributions of this paper are as
follows:
 To highlight the classification of various
Recommender Systems (RS) feedback techniques,
their merit and demerits.
 Comparison between Implicit, Explicit and Hybrid
feedback techniques using last.m database.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Ngwira SM, Tranos Zuva
Dept. of Computer Systems Engineering
Tshwane University of Technology, TUT
Soshanguve, South Africa
{ngwiraSM, zuvaT}@tut.ac.za
 Overview of Recommender Systems (RS)
evaluation and performance technique.
The rest of the paper is divided into: Section II
overview of various Feedback Techniques; Section III is
comparison study of Implicit, Explicit and Hybrid; Section
IV is Evaluation and Performance Metric of Recommender
Systems (RS) Techniques and Section V Concludes the
paper.
II FEEDBACK TECHNIQUES
The information feedback is basic process of any
Recommendation Systems, as it gives the information that
RS needs in order to give appropriate recommendations to
the users based on their preferences. With this, classification
of information feedback technique is necessary in RS
implementation in order to know the appropriate feedback
techniques to use for particular RS which will be based on
user preferences.
In this sense the feedback techniques are classified
into three types [4]:
 Implicit Feedback Technique(IFT)
 Explicit Feedback Techniques(EFT) and
 Hybrid Feedback Technique(HFT)
A. Explicit Feedback Technique(EFT)
Explicit Feedback Technique is a process that involves
users assigning either numerical or scores rating to evaluate
the items or goods retrieved to the user by the system
(survey).
The frequent case of Explicit Feedback ratings are
made on an agreed discrete scale (e.g star ratings for
restaurants, mark out ten for films etc). Ratings made on
these scales allow these judgements to be processed
statistically to provide average, ranges, distributions etc. The
EFT helps users in expressing their interest and taste on
particular object [4, 5].
There are three main approaches to get explicit relevance
feedback [3, 6]:
 Like/dislike – items or goods are classified as
relevant/ irrelevant using binary rating scale.
 Ratings – judgement of items or goods are made
using numerical scale.
  Text comments – in this approach, the comments
about an item are obtained and showed to users so
as to help them in process of making decision
about the particular items or goods. For instance
the user’s feedback comment feedback on
particular item can assist users to know whether the
community appreciate that particular item or not.
However, there are merits and demerits of EFT [6, 7].
Merits
1. EFT is simple to use [7].
2. The rating can be done either be positive or negative
i.e helps user to specify what they like and what they
don’t like [7].
3. EFT tends to concentrate on either side of the rating
scale, as user as more likely to express their
preferences if they feel strongly for or against an item
[7].
4. The accuracy of EFT is seems to be higher than IFT.
Demerits
1. EFT is absolute. For example, a user listening to music
Fig. 1: Observation behavior of IFT
many times, the users can still express is interest if the
user listens to the music once without listening to it
several times [7].
2. The problem of intrusiveness is one of the challenges
affecting EFT [7].
3. It wastes time, as most of the time can be wasted in
only checking the items rather than doing the items
rating.
4. Using numerical scale can be confusing as user might
not be consistent in giving their rating [7].
5. User’s rating might not show true opinion of users.
6. It provides unambiguous labels as the truthfulness of
such labelling might be questionable.
7. EFT is susceptible to noise [7].
8. It is sensitive to user context.
B. Implicit Feedback Technique(IFT)
IFT is refers to as unobtrusive approach of rating the
state assessment about usefulness of any individual
particular item. An IFT captures and interpret user’s
feedback by using application domain-dependent tools and
some methodologies [3-7].
However, IFT operates and makes decision based on
the use’s actions. This approach measures user’s taste and
interest without seeking for user’s consent. This simply
mean information is captured without user’s awareness but
obtained based on the user’s action during the process.
This type of IFT can be found in various applications
such as web purchase history, browsing history, search
pattern or even mouse movements. This is shown in Fig. 1.
RS function in assisting user’s in finding the
desired items or goods of their interest and taste. This works
by integrating inference and prediction, and Fig. 2 and 3
show alternative strategies used for this.
Fig. 2 depicts a modular strategy in which the
inference stage helps in giving rating that is almost the same
as those that user would have assigned by himself, and those
estimated rating is then used for predict future ratings [7].
Observation
Inference
Estimated Ratings
Prediction
Predicted Ratings
Fig. 2. Rating Estimated Strategy [7]
 2. HFT is combination of both Implicit and explicit
techniques.
Observations 3. HFT uses both numerical rating and human expression
for prediction.
Prediction
Demerits
1. HFT is not cheap
Predicted Observations
2. It is computationally intensive.
Inference
III. COMPARATIVE STUDY OF VARIOUS FEEDBACK
Predicted Ratings TECHNIQUES
We do comparison between the IFT, EFT and HFT RS
. using the data obtained from Last.fm. The Table I show the
comparison result [5, 6].

TABLE I. COMPARATIVE STUDY OF FEEDBACK TECHNIQUES

Fig. 3. Predicted Observation Strategy [7]
Features
Implicit Explicit Hybrid
Here, in Fig.3 the predicted observation strategy
Abundance Positive Negative Positive
shows how the past observations are used to predict items or
goods for user’s based on user observation. Context-
There are some advantages and disadvantages of sensitive Yes Yes Yes

IFT [6-9].
Expressivity of
Positive &
user preference Positive
Negative
Positive
Merits
1. IFT can be collected at much lower cost, user time in
examining and rating Accuracy Low High Very High
2. IFT is effortless, it does not put burden on the user of
Noise
retrieval system. Yes Yes Yes
3. It can be continuously collected from user-system Measurement
interaction and can be used for the users profile update. Reference Relative Absolute Absolute
4. It is less accurate when compared with EFT, but large
quantities of information can be obtained at lower cost
from users.
IV. METRIC EVALUATION OF RECOMMENDER SYSTEMS
5. The affective labelling is unobtrusive and difficult to
The evaluation metric of recommender System can be
cheat by users.
used to measure the performance of the system in order to
obtain the error the system can encounter during the
Demerits
implementations.
1. IFT is susceptible to noise
The following metric can be used to predict the rate a
2. The IFT are sensitive to user’s context, albert not the user will give to items. This is shown in (1-6) respectively
same extent [10, 11].
3. IFT is less accurate compared to EFT Mean Absolute Error (MAE): This measures the
4. Implicit feedback is always positive. For instance, if a average of the absolute deviance between the predicted
user did not watch a track that does not mean the user rating and the actual rating given by the users in the system
does not like the track [1, 2] as in (1).
5. IFT is difficult to interpret.

N
pri  ari
 MAE  i 1
 
C. Hybrid Feedback Technique(HFT) N
HFT is the combination of both implicit and EFT. This
approach utilizes both combination of numerical rating Mean Square Error (MSE): This is used in order to
scores and human behavior in predicting items of interest give more importance to cases with larger deviance from the
and taste to the users. actual rating. It is used instead of MAE as in (2).

Merits
1. HFT helps to improve the prediction rating accuracy.
  V.
N
( pri  ari ) 2 CONCLUSION
MSE  i 1
In this paper, techniques that are used to construct user’s
N interest and taste on items in RS Feedback have been
(2) highlighted. The merits and demerits of IFT, EFT and HFT
on RS have been discussed.
Root Means Square Error (RMSE): This is the The performance evaluation of RS has been discussed.
variant of MSE is the Root Mean Squared Error (RMSE), In a summary, accurate feedback in RS can be of assistance
which was the error metric, used in the Netflix. The RMSE to users in various applications such as in helping user’s to
between the predicted and actual ratings is given as in (3). find items of their interest from huge items within a limited
time.

  pr  ar 
N 2 In future, IFT, EFT and HFT we be implemented in real-
life experiment to address some of the challenges
RMSE  i 1 i i
(3)
N highlighted in this paper such as accuracy, speed and
efficiency.
In application where a list of recommendations are
provided for user to evaluate relevant and irrelevant items. ACKNOWLEDGMENT
Precision, Recall and DGG are used for information
retrieval in such scenarios. The authors acknowledged the contribution of Tshwane
Precision: This is the fraction of relevant item University of Technology.
recommended to the items in the recommendation list as in
(4). REFERENCES

[1] L.H Li, R-W Hsu and Fu-Ming Lee. Review on “The Technologies
relevant items recommended
Pr ecision  and The Assessment Methods of Recommender Systems” Pg.1-46.
items in the recommendation list [2] Mild, A. & Natter, M., “Collaborative Filtering or Regression Models
for Internet Recommendation Systems” J. of Targeting, Measurement
(4) and Analysis for Marketing, Vol.10, No. 4, pp. 304-313, 2002.
[3] Gawesh Jawaheer, Martin Szomsor and Patty Kostkova Comparison
The number of the list can be huge depending on the type of of Implicit and Explicit Feedback from an Online Musinc
Recommendation Service, Her ’10, Sptemebr 2010.
recommendation technique and the size of the database
[4] F Ricci, L. Rokach, B. S Paul. KantorRecommender Systems
used. Handbook”, pg.77
Recall: This is defined as the fraction of item [5] E R Núñez-Valdéz , J.M Cueva Lovelle , O.S Martínez , V García-
recommended to be relevant by user to relevant items as in Díaz, P Ordoñez de Pablos , C.E Montenegro Marín “Implicit
(5). feedback techniques on recommender systems applied to electronic
books Computers in Human Behavior” Vol.,28 2012, 1186–
1193Elissa, “Title of paper if known,” unpublished.
relevant items recommended [6] Pasquale Lops, Marco de Gemmis and Giovanni Semeraro, Content-
Re call  based Recommender Systems: State of the Art and Trends,
relevant items Recommender Systems Handbiik pg 73-85
(5) [7] Amatrian, X; Pujol, J (2009)’I like it…I like it Not: Evaluating User
Rating Noise in Recommender Systems.
[8] Tong Queue Lee, Young Park, Yong-Tae Park “A time-based
Discounted Cumulative Gain (DCG): This is used approach to effective recommender systems using implicit feedback”
for measuring the effectiveness of the recommendation Expert Systems with Applications, Vol., 34, 2008, Pg., 3055–3062.
method in locating the most relevant item at the top and the [9] Douglas W. Oard and Jinmook Kim ”Implicit Feedback for
fewer items at the bottom of the recommendation list as in Recommender Systems”Digital Library Group, 2000.. Young, The
Technical Writer’s Handbook. Mill Valley, CA: University Science,
(6). 1989.
[10] Ioannis Arapakis, Joemon M. Jose and Philip D.Gray Affective
p
2reli  1
DCG  
Feedback: An Investigation into the Role of Emotions in the
information Seeking Process” SIGIR’08, Pg.395-402.
i log 2 (1  i) [11] Jonathan l. herlocker, Joseph a. Konstan, Loren g. Terveen and John t.
Riedl “Evaluating Collaborative Filtering Recommender Systems”
(6) ACM Transactions on Information Systems, Vol. 22, No. 1, January
2004, Pages 5–53.
 Implementing Protocol to Access White-Space
Databases on Smart Set-top Box
Elesa Ntuli
Department Of Computer Systems Department Of Computer Systems
Engineering, TUT
Pretoria, South Africa
Ntulife@tut.ac.za
Abstract—in television white space networks, secondary
users are mandatory to query an accredited geo-location
spectrum database (GSDB) in order to determine the vacant
channels or white spaces. The recent developments of the
Protocol to Access White Spaces (PAWS) by the Internet
Engineering Task Force (IETF) have proposed to standardize
communication between the GSDB and white space devices. To
enable efficient sharing of spectrum, white spaces devices
(WSDs) and their related restrictions, the mechanism or
channel selection remain open issues in research. In this paper,
we proposed use of PAWS and JavaScript Object Notation
(JSON) to address the communication between the GSDB and
WSDs to enable optimal channel selection decisions in the set-
top box making it a smart set-top box.
Keywords—TV whitespace; set-top box; digital television;
cognitive radio; base station; geo-location spectrum database; Protocol to
Access White Spaces; White space device.
I. INTRODUCTION
[1] and [2] in their prior work on TV spectrum
measurements in both urban and rural areas in Southern
Africa, they found that TVWS availability ranges between
100 to 300 MHz The key element in the worldwide
deployment of digital television (DTV) is the digital set-top
box receiver and related interactive services. Set-top boxes
with First-generation were quite simple with support only
for channel-tuning and audiovideo data decoding.
Increasingly, however, these set-top boxes can be used as a
gateway to the Internet and as a hub in home networking
and entertainment. Digital-quality broadcast television,
personal video recording, and high-speed Internet access are
some of the features for extensive range of consumer
broadband multimedia services. These services with such
capabilities are paving the way for the set-top box to
become not only a residential entertainment center but also a
smart central piece of home computing equipment capable
of handling various advanced applications such as browsing
the internet. Sharing of TV spectrum successfully between
broadcast and broadband services depends on the approval
of dynamic spectrum access (DSA) regulatory approach. In
DSA based broadband networks and cognitive radios.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Seleman Ngwira Tranos Zuva
Department Of Computer Systems
Engineering, TUT Engineering, TUT
Pretoria, South Africa Pretoria, South Africa
NgwiraSM@tut.ac.za ZuvaT@tut.ac.za
(CRs) or white space devices (WSDs) operate as secondary
users (SUs) without creating interference access the white
spaces to the certified or primary users (PUs). There are
mainly two techniques which are commonly considered for
discovering the white spaces namely the spectrum sensing
and geo-location spectrum database (GSDB) [3].
Experimental TVWS broadband networks using GSDBs
have been piloted in many parts of the world, Worldwide
there is a coordinated move for Digital Switch Over (DSO)
by terminating analogue television transmission. This
Digital Dividend (DD) has created new opportunities on
spectrum for many new wireless technologies and research
including South Africa [7].
We review the portable devices when switching over to
digital transmission using the Television White Space
(TVWS), Base Station (BS), Geo-location and the
possibility of turning the set–top box (STB) into a smart set-
top box (SSTB) by introducing the possibility of browsing
the internet which is suitable for rural areas in South Africa.
Spectrum is becoming a limited natural resource, although
the evidence shows instances continually where the
spectrum is not fully exploited by the allocated facilities
making it inefficiently utilized [1].
spectrum of about 1720 MHz will be required according to
ITU predictions many new wireless services/ applications
cannot be rolled out due to nonavailability of spectrum by
the year 2020 [17], which demands dynamic allocation of
spectrum instead of static [1]. The technology of CR would
assist in meeting the constantly growing demand of radio
spectrum and aid in dealing with resource in more
methodical and in more efficient way. The basic impression
of CR is to reuse the spectrum whenever it is vacant by the
primary/licensed users (PUs). The frequent spectrum
sensing for detecting the TV switchover to full digital
broadcast service (DD) is required to be performed by the
Secondary/unlicensed Users (SUs) which creates new
spectrum opportunities due to higher spectral efficiency
compared to analogue services. This TVWS technology
opens the door for CR technology. In various countries
Spectrum regulatory bodies are studying the pros and con of
CR devices. Other countries have made provisions for CR
already and also FCC for the use of CR device in TV bands
[4].
II. GEO-LOCATION COMBINED WITH DATABASE
The Global Positioning System (GPS) and its database will
determine the channels that can be utilized by using the
Geo-location technique; it uses location technology at a
certain location and connects parameters for an unlicensed
user which has maximum transmission powers. The Geo-
location process could be described as follows:
A. Conceptual diagram of TVWS using the PAWS Protocol.
Fig. 1. TVWS Device Using PAWS Protocol to access the internet.
Based on the central information of the geographic locations
for primary users and locations for unlicensed users which is
being held by the Geo-location database, that database can
calculate the protected service then outline each station
which can be determined by the GPS, thus determining the
available frequencies list in a specific location, the strengths
of the possible signal can be calculated based on the
matching occupants interference, determining the possibility
of maximum transmission power for a unlicensed user, one
can require WS access with the condition of not causing too
much interference for the occupants for the television set, to
access these frequencies there had to be a device/ a set-top
box to translate and change these frequencies from analog to
digital then yet again the same device/ set-top box can be
used to access the internet using white space the vital trick is
the device/ set-top box accurate location that must be known
by the database, additional connectivity is needed in a
devices of a different band in order to access the database
before any transmission in Digital Television bands. Using
JSON and PAWS in addressing the communication factor
between the GSDB and WSDs requires the process of
Hypertext Transfer Protocol (HTTP) to be in session
enquiring for the device to browse the internet.
The device will access the database checking available
channel in a different band before getting access to available
channels for TV utilization or internet. Since the challenge
is in the specification design and implementation of
switching from analog to digital and/or browsing the
internet, sensing the Geo-location technique is the method
that is viewed as the most significant instrument for
spectrum detection as recommended by Ofcom [2].
B. Wide are coverage in Rural Areas (IEEE 802.22)
IEEE 802.22 wide regional area network (WRAN) aims to
provide fixed wireless access with a typical cell radius of 33
km with operative Isotropic Radiated Power (OIRP) of 4W
and a maximum of 100 km in rural and remote areas using
CR technology in TVWS [12]. [15] Mentions that the
spectral efficiency of 802.22 systems ranges from 0.5 to 5
bit/s/Hz. with 3 bit/s/Hz average value, having 6 MHz
channel capacity to reach 18 Mbit/s. Broadband to rural
areas might therefore prove to be very cheap.
III. BUSINESS FOR RURAL SOUTH AFRICA
The South African government agreed to subsidies families
especially in the rural areas with set-top box when the
migration to digital is underway, with that set-top box we
are proposing that we use JSON and PAWS in addressing
the communication between the GSDB and WSDs.
There are few reasons for TVWS technology to bring
outstanding possibilities for the rural areas and they are:
 Service providers can serve customers they cannot
reach with existing wireless technology.
 Rural industrialists for provisioning a broadband
Internet access will have a framework of a small
fee design from a smart mesh network.
 Businesses and governments will expand the
capability in opportunities for rural communities
that demands high-speed Internet and also
extending communications for private networks.
IV. PROTOCOL FUNCTIONALITIES
According to [6] they mention the components for
PAWS and some regulatory domains or Database
implementations that have a mandate for the use of a
component, even when it’s not in use but it’s mandated by
PAWS. Protocol Parameters contains a more detailed
discussion of the limitations that make up the PAWS
request and response messages (encoding process of
messages). The Binding of HTTPS describes the use of
HTTPS over TLS for transferring PAWS messages and
optional device verification. Protocol Parameters are being
used for reference and contain the data type and name of
each parameter and checks if the parameter exist and
available for the protocol transaction in question. Data types
must be defined in Protocol Parameters or in one of the
following basic or organized types [6]:
 string: is JSON classified to the UTF-8 encoding.
 int: is JSON classified, without a slight or
proponent part.
 float: is JSON classified.
 Boolean: is JSON classified.
  list: represents a list of elements and is a structured C. Proposed Spectrum Decision Hierarchy Structure
type, is also JSON classified as an array type.

V. ACCESS TO GSDB
The TVWS-BS is connected to one of the approved GSDBs
the TVWSBS establishes an HTTP session with the national
GSDB in order to register and initiate a query for available
channels. The messages between the TVWS-BS and the
GSDB are standardized by the PAWS [8], while messages
between the WSD and the GSDB is use dependent, i.e. not
standardized by the PAWS. Once selected a channel for
itself, the TVWS-BS then uses each WSD’s geo-locations
(which are readily known to the TVWS-BS) to query the
GSDB for more available channels using the AVAIL
SPRECTRUM BATCH REQ request.
Fig. 2. Channel Selection hierarchy structure
The WSDs are also called slaves because they do not have
direct access to the GSDB [5]. Depending on the amount of The figure above illustrations the hierarchy structure where
traffic demand or class of service (i.e. real-time or best the top level presents the main goal on the structure, which
effort) by the WSDs and the distance from the BS, the best select the best available channel. The selection of the
suitable number of channels will be selected (also channels is based on the user’s preference which requires a
depending on the availability of white spaces). Finally, by specific quality of service (QoS) for each class of service.
sending a SPECTRUM USE NOTIFY, the TVWS-BS Considering two classes of service (CoS): Real Time (RT)
notifies the GSDB of the selected channels used by the and Best Effort (BE) which are on the second level of the
entire network. This will assist the GSDB not to provide the hierarchy. On the third level of the hierarchy are the three
same list of available channel to other secondary networks, independent criteria to be compared when selecting the
thereby preventing any co-channel interference [6]. channels. At the bottom of the hierarchy are at least four
alternative channels which are compared in order to select
VI. AVAILABLE SPECTRUM RESPONSE MESSAGE the best channel for the user.
According to PAWS protocol, attributes are extracted from D. Overview of openPAWS protocol
the AVAIL SPECTRUM RESP message which is sent to
TVWS-BS by the GSDB [5]. This message contains several
parameters which include timestamp, spectrumSchedules,
maxTotalBwHz, maxContiguousBwHz, etc. The most
important parameters used in the model are maxTotalBwHz
and spectrumSchedules which are explained below.

1) The maxTotalBwHz Parameter Provides: the maximum

total bandwidth (Hz) it can be adjacent or not [5].
2) The spectrumSchedules Parameter Provides: a
combination of EventTime element and Spectrum elements
which specifies the start (startTime) and stop (stopTime)
times of an event. The TVWS-BS will then calculate these
elements to discover the availability of the time span of each
white space channel [16].

The spectrum element in the spectrumSchedules parameter

consists of two parameters which characterizes a list of
frequency ranges and permissible power levels for each
range. The frequencyRanges lists the maximum permissible
power levels within a frequency range [5]. Fig. 3. Messages exchanged between PAWS slave device, PAWS
master device, and the database. [4]

Fig. 4. PAWS master device state diagram, obtains list of available
channels. [4]
Fig. 5. PAWS slave device state diagram, once the slave device is
connected to the Master, the rest of the PAWS messages are exchanged in a
manner similar to the Master-database communication. [4]
VII. CONCLUSION
In this paper, a protocol to access TVWS database was
proposed. The Geo-location database based system using a
PAWS and JSON in establishing the communication
between the GSDB and the set-top box was proposed. GPS
receiver determines the location of WSD and the
accessibility of a database at a particular location.
PAWS address the communication between the GSDB
and WSDs. The advantage of using JSON strings is that it
has multiple parameters encoded in them. The disadvantage
is that without the Geo-location database JSON would not
give accurate results when searching for available channels.
The need to bridge the digital divide in under-served remote
and rural areas cannot be over emphasized. In [10], it is
evident that low spectrum utilization, particularly in rural
areas is inseparable from, and directly relevant, to the digital
divide. Therefore, spectrum regulators from developing
countries in particular, should compete with the policy
initiatives taken by the FCC to bridge the digital divide by
boosting telecommunications operators to provide low cost,
long-range wireless broadband connectivity in rural areas.
To achieve this you need to allow innovative opportunistic
access (using DSA-based technologies) in specified portions
of the RF spectrum such as the digital dividend and TVWS
in the VHF and UHF bands on a license-exempt basis
ACKNOWLEDGMENT
We would like to thank TUT for supporting this paper and
appreciate Sibusiso Colby Ntuli for all the support given.
REFERENCES
[1] M. T. Masonta, D. Johnson, and M. Mzyece, The White Space
Opportunity in Southern Africa: Measurements with Meraka
Cognitive Radio Platform, R. Popescu-Zeletin, et al., Ed.
Springer, vol. 92, pp. 64-73, Feb. 2012.
[2] A. Lysko, M. Masonta, D. Johnson, and H. Venter, ―Fsl based
estimation of white space availability in UHF TV bands in
Bergvliet, South Africa,‖ in SATNAC, George, South Africa,
Sep. 2-5 2012.
[3] M. Masonta, Y. Haddad, L. De Nardis, A. Kliks, and O.
Holland, ―Energy efficiency in future wireless networks:
Cognitive radio standardization requirements,‖ in IEEE
CAMAD, Barcelona, Spain, Sep. 17-19 2012.
[4] Ghosh, S.; Naik, G.; Kumar, A.; Karandikar, A., "OpenPAWS:
An open source PAWS and UHF TV White Space database
implementation for India," in Communications (NCC), 2015.
[5] TENET, ―The Cape Town TV white spaces trial,‖ Available
from: http://www.tenet.ac.za/tvws, 2015, [Accessed:
14/03/2015].
[6] V. Chen, Ed., S. Das, L. Zhu, J. Malyar, and P. McCann,
―Protocol to access spectrum database,‖ Jun. 2013, draft-IETF-
PAWS-Protocol-06, 2013.
[7] C. Ghosh, S. Roy, and D. Cavalcanti, ―Coexistence challenges
for heterogeneous cognitive wireless networks in TV white
spaces,‖ IEEE Wireless Communications, vol. 18, no. 4, pp. 22–
31, Aug. 2011.
[8] FCC Spectrum Policy Task Force. http://www.fcc.org. ET
Docket No. 04-186, 2008.
[9] T. Dhope, D. Simunic, and M. Djurek. Application of DOA
estimation algorithms inSmart antenna systems. Studies in
Informatics and Control, 19(4), 445–452, 2010.
[10] K-C. Chen and R. Prasad. Cognitive Radio Networks. John
Wiley & Sons, 2009.
[11] Ofcom: Digital Dividend: Clearing the 800 MHz Band.

[12] http://www.ofcom.org.uk/consult/condocs/cognitive/,
[Accessed: 14/03/2015].
[13] http://ieee802.org/22/, [Accessed: 14/03/2015].

[14] Masonta, M.T, Olwal, T, Mekuria, F and Mzyece, M. 2014.

Standard compliant channel selection scheme for TV white
space networks. In: Southern Africa Telecommunication
Networks and Applications, Port Elizabeth, South Africa, 31
August - 3 September 2014
[15] Masonta, M.T, Mzyece, M, Ntlatlapa, N, "Spectrum Decision in
Cognitive Radio Networks: A Survey," Communications
Surveys & Tutorials‖, IEEE , vol.15, no.3, pp.1088,1107, Third
Quarter 2013
[16] W. Ho, ―Integrated analytic hierarchy process and its
applications a literature review,‖ European J. of Op. Res., vol.
186, p. 211228, 2008.
[17] L. Mfupe, M. Masonta, T. Olwal, and M. Mzyece, ―Dynamic
spectrum access: regulations, standards and green radio policies
consideration,‖ in SATNAC, George, South Africa, Sep. 2-5
2012.
[18] ITU, ―Provisional Final Acts‖, World Radio communications
Conference (WRC-12), Geneva, 23 Jan-17 Feb, 2012.
 Personal Preferences Analysis of User Interaction based on
Social Networks
Cheng-Hung Tsai, Han-Wen Liu, Tsun Ku, Wu-Fan Chien
Institute for Information Industry,
Innovative DigiTech-Enabled Applications & Service Institute,
Taipei, Taiwan, R.O.C.
{jasontsai, tonyhwliu, cujing, juliettechien}@iii.org.tw
Abstract—Under the current situation the booming social
networks, users interact between people with the way of social
networks platforms (such as: press like, join fans pages and
groups), and for these interactive information on social platform
can fully represent that oneself is interested in the content of
information sources in different social group. Therefore, how to
collect user behavior patterns (Users Behavior) generated by
social interaction, and then analyze and understand user
preferences by these interactive data would be the purpose to
discuss and to do the research of the paper.
Furthermore, for many brand enterprises, it is important to
know how to understand individual preferences, because when
you know the individual preference information, it can carry out
personal preference for advertising, product recommendation,
article recommended…and other diversified personal social
service, which can increase the click rate and exposure of the
products, better close to the needs of the user's life. Therefore,
with the above through social science and technology
development trend arising from current social phenomenon,
research of this paper, mainly expectations for analysis by the
data of user’s personal interaction on the social network, such as:
user clicked fan page, user press like article, user share data etc.
three kinds of personal information for personal preference
analysis, and from this huge amount of personal data to find out
corresponding diverse group for personal preference category.
We can by personal preference information for diversify personal
advertising, product recommendation and other services. The
paper at last through the actual business verification, the
research can improve website browsing pages growth 8%, site
bounce rate dropped 11%, commodities click through rate
growth 36%, more fully represents the results of this research fit
the use’s preference.
Keywords—Social Networks; Social Personal Analysis; Cost
Per Click; Personal of Interest analysis
I. INTRODUCTION
Recently, the flourish of information technology caused
people to think how they can promote the chances of
interaction with others by information technology. With the
progress of information technology and the extension of the
concept of interaction, virtual communities (e.g. Facebook,
Weibo, Blogspot, PTT, forums, news…) had begun forming
and growing. Users have gotten used to interacting and
communicating with their families, colleagues and friends
over all kinds of social networking sites. Even they will
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
exchange information with others having similar interests or
demands. Hence, under the influence of social network, there
is a considerable amount of social activities shifting from
reality to virtual online platforms. This brings a direct impact
to the consumer market, which is too immense to be ignored.
Social networks for business operators, the value of the
information can be one of the important sources for
understand customer's opinion and behavior analysis which
retained by the virtual social group. And the "human" play a
very important link in the social network, and also because of
the interactive link between people, the social platform has
become a very huge information exchange center, spread and
scattered that people for any event and personal preferences
are derived from the words and social behavior. By collecting
and analyzing personal interaction behavior and the content of
writing comments on the social platform, we can more
understand that the preference for personal interest has more
significant results. Therefore, how to effectively analyze the
personal preference of the social platform has become a
problem to be solved in this paper.
The paper mainly put forward the research of
solving social personal preference analysis, through the three
modules planning and implementation that the paper
mentioned about and to collect and analyze data, according to
the algorithm of each module for processing, the effective
personal preference information analysis for social network
platform, through modules planning in this paper, after the
system actual verification, web browsing pages growth 8%,
site bounce rate dropped 11％, product click through rate
growth 36%, the results can provide extended of follow-up
research, the method follow up will be introduced in order.
The framework of this paper is as follows: the first
section introduces the motivation and background of this paper.
The second section, through the previous research put forward
that the relevant literature is discussed and analyzed. The third
section introduces the used of the method and technology in
the paper. The fourth section puts forward the implementation
results and analysis. The fifth section is the conclusion and the
future outlook.
 II. RELATED WORK
A. Social Network Development Trends
According to the estimation of the international market
research authority eMarketer, the number of global users of
social network has reached 1.5 billion. The number of active
user of Facebook, Twitter, Weibo and Tencent were one billion,
0.5 billion and 0.3 billion at 2012, respectively. Furthermore,
the penetration rate of social network services among the
global users of all ages has reached 79%. Even those once
considered the main composition of digital divide has a rise of
penetration rate up to 9.3%. These facts show that the use of
social network is very popular. In domestic, the number of
user of social network had a breakthrough of ten million at
2011, according to the research of Institute for Information
Industry. [3]Among those under 30 and between from 50 to 59,
the proportions of using social network were 96% and 70%,
respectively, which show an extremely high penetration rate.
As for the period of using social network, a research shows
that there were more than 70% of domestic people using it
over a year. If we take those using over six months but less
than a year into account, the proportion is approximately 90%.
Learning from the high penetration rate, the application of
social network has deeply rooted in our daily life and also
changed our lifestyle.。
III. CONSTRUCTING SOCIAL PERSONAL ANALYSIS
Fig. 1 social personal analysis organization chart
Social Personal Analysis of this research with the
Facebook personal information as an example to develop and
design, and can according to different nature of social
platforms to carried out diversity of Social Personal Analysis
as its sources of information. The research design and
planning as shown in Fig. 1, two modules (Facebook personal
interactive data and Facebook fans pages observation data) to
carry out different data analysis algorithms. Through our
previous research of distributed Web Crawler [3] to collect
social personal information, and carry out the Social Personal
Analysis after the social big data collection.
3.1 Distributed Data Collection Module
With this research we studied previously distributed
Web Crawler [3] to carry out personal social networking sites
(e.g. Facebook) information collect service. Since this
research should be carried out personalization information
collect service, so for the sake of the personal information
protection law of different countries, so we need to obtain
personal authorization in this research. Therefore, this research
practical cooperation with brand owners, and legally acquired
the vendor of Customer Relationship Management
Information, through user consent and legally obtain
personalization authorized to carry out analysis of this
research.
The paper in this research mainly added two data scopes,
1.legally acquired 10,000 users authorized to Facebook
personal information 2.Observing Taiwan’s 10,000 Facebook
fans pages, as shown in Table 1. The research by means of the
user's personal ID Mapping to understand which fans pages
that the user join, and through the method of user's personal
authorization, for personal users in the Facebook on press like
article and share article and other information, to carry out
personalized preference analysis via these two data. After
analyzing the results and provide it to the brand owners to
carry out follow-up recommended services.
Table 1、Analysis of personal information Sources
Social
Data_1 Data_2 Data_3
sources
Press Share Join
like information fans
Facebook
article list pages
list list
3.2 Social Personal Analysis
This research, the main source of personal analysis data
were analyzed through Table 1 user activity behavior on
Facebook, we through the fans pages data that the user have
joined and social interaction behavior data of oneself on
Facebook to conduct personal preferences classification, as
shown in Fig.2.
Fig. 2 personal preference extraction and analysis
3.2.1 User operation complexity analysis

In this research, we think about the operating behavior

property of people on the social platform (Facebook). For
example: user A is assumed to be interested in photography,
user A will concern about related photography fans pages on
Facebook even "join photography fans pages ", and this
behavior action can receive the latest photography information
and satisfy the needs of user A about photography information.
Another situation, user A is interested in photography-related
articles content, and then will "press a Like" or "share" to
close friends, and these kinds of behavior patterns, all with the
behavior pattern on behalf of user A is an interest for
photography.
But there will likely be part of the exceptional
circumstances, for example: user A to part of the article is
represented "agree" or "interesting" on Facebook, user A will
conduct the behavior actions "press a Like" and "share" to the
articles.
Through the description of the related issues above, we
are thinking how to through the user behavior patterns on the
social platform (Facebook), analysis of the user's exact
preferences, therefore, the research on the social personal
behavior pattern analysis, mainly according to the user to
analyze operational complexity on Facebook, because when
the operation is complex, users are willing to spend more time
to obtain the relevant information services, which may
represent the behavior of the most representative preferences.
Therefore, the research based on the difficulty of operation
setting preference weights, as shown in Table 2.
Fig. 3 Personal preference category list
Through Fig.3, we correspond to 320 Categories items
on Facebook for a person's preferences in the social group.
With various classification, we assume a personal favor
graphy (G), U, V nodes are set to the starting point and
end-point, and in accordance with the user on personal social
behavior patterns corresponding to Facebook classification set
each classified as a node. Our V denote the set of all the
vertices of G, the edges of each graph are ordered elements
formed by the two vertices. (u, v) represents from vertex u to v
has connected path. Our E denote the set of all the edges of G,
and the weights of the edges defined by the weight function w:
E → [0, ∞]. Therefore, w (u, v) is a non-negative weight right
from vertex u to vertex v. The weight of the edges can be
imagined as the distance between two vertices. The weight of
the path between two points is the sum of the weights of all
the edges of the path, as shown in Fig.4.

Table 2、The complexity of the user operation

Press Press
Operational Join fans Share
like fans like
complexity pages Article
pages article
Weight
0.35 0.3 0.25 0.1
allocation

For the Table 2 weight parameter setting, this research is Fig. 4 Personal favor graphy
based on the field experiments process of the fourth section.
Through the Fig. 4 personal favor graphy, we can define
The research by obtain 10,000 personal access information to
the results of personal preference based on the data of each
conduct the data analysis, which by A / B test and mining
node preference, through Dijkstra's Algorithm based on the
process of adjustment to obtain optimization, and detailed
shortest path, as shown in Fig. 5, to find out the personal
experimental procedures, refer to the fourth section.
preference on the social group of the higher score data, can
define the results of personal preference.
3.2.2 Personal preference analysis

Through the research to define the user operational

complexity of weights, we will define category list of personal
preferences through the behavior of the user on social group
(join fans pages, press like fans pages, share article, press like
article), such as shown in Fig.4.
 Fig. 6 Field Empirical Chart

Fig. 5 Dijkstra's Algorithm The system on this paper collect 10,000 personal
preference data with the user’s approval through the Personal
To find a vertex t from the U-V set, so that the D [t] is the preference recommendation interface, the WebCrawler of data
minimum, and put the T into the S set, until the U-V is empty processing platform module will carry out the personal data
set. Adjust the values in the D array according to the formula collection. In addition, the research simultaneously observed
below: Taiwan's 10,000 Facebook fans pages data to understand the
fans pages that the user have joined and to collect personal
MinW ( P)  W (u, v)
( u ,v )P
social interaction data. Then, the Personal preference analysis
calculation module will analyze the personal preference of
（3-1）
these data, and finally through the results of personal
D = A [F, I] ( I =1, N ) （3-2） preference analysis to convey the product or the article that
u = {F} （3-3） users interest in.
v = {1, 2, … , N} （3-4） 4.2 The system flow
D [I] = min ( D [I], D [t] +A [t, I; ] ((I , Start

t )E) （3-5）
Users browse the web pages
D: An array of N positions for storing the shortest distance
from a vertex to another vertex. FB APP_Users Login

I: refers to the adjacent vertices of T.

FB Access Token No End
F: denote that starting from a starting point, A [F, I] means the
YES
distance from the point F to point I.
Web Crawler Module
V: set of all node in the network.
U: set of vertices. Persona Raw Data

IV. THE REALIZATION OF PERSONAL PREFERENCE Persona Data Analysis

ANALYSIS OF USER INTERACTION 1.user interest
2.Monitor 10000 fanspages
3.user’s add fanspages mapping

4.1 The system architecture

Persona Data Analysis
The main implementation is set up the system on the 1.user interest API
2.User fanspages interest API
Amazon Web Services platform, the system is divided into
three modules: Personal Preference Recommendation
Interface (PPRI), Data Processing Platform Module (DPPM), Recommendation Engine
Preference Analysis Arithmetic Module (PAAM), to conduct
Personal data collection, analysis and temporary storage
service, and its architecture planning as shown in Fig. 6. Fig. 7 System flow chart
 In this section, we will introduce the details of the process
of system flow, as shown in the Fig. 7 system flow chart. First,
we set up Facebook Login APP on the web page, when the
user login with Facebook Login APP, it will show the user's
personal authorization content. And through the user click
authorized personal content items, Facebook APP will deliver
the user access token to the data processing platform module.
Data processing platform module will conduct two data
collection by the user access token (user_profile, user_friends,
user_post). 1. FB personal page interactive data and 2.
Observation of personal joined fans pages data. With complete
collection of personal data and observation of personal joined
fans pages data (Raw data), we will carry out two Personal
preference analysis calculation module, 1.user interest
2.user’s add fans pages mapping, the technical details as Fig. 8 Recommended service verification field
described in the third section. Finally, with the results of our Our research actually verified on the web site, the
Personal analysis, through two different kinds of Personal verification period is a month, between 2015/01/01 ~
preference analysis API to fetch with the front-end data, the 2015/01/31, the six classifications totally have of 103 popular
front-end web site can carry out the recommendation of articles, which carried out the user recommended verification,
personal commodities and articles by the result of Personal as shown in Table 3.
preference analysis.
4.3.2 System statistics
The research analyzes the user behavior interaction data on
social group, further with the way to initiative recommend The research experimentally verified on the web site during
commodities and articles that the user preferences by personal a month, based on the back-end data which provide by the
preference, and to stimulate the user / the consumer browsing co-operative enterprise to verify the result of this service, such
and purchasing the commodities, enhance the owners on click as: Page View (PV), site bounce rate, the commodities click
through rate and adhesive capacity of the commodities. through rate. The results of verification are shown as follows:
4.3 System Verification Table 4、Verification of Personal Preference recommend
empirical data analysis services
4.3.1 Verify field setting
Verification of Personal Preference
The research is to verify the effectiveness of this system. recommend analysis services
The effectiveness will be actual verification to the the official Click
website that built by the cooperation owners of this research, PV Bounce rate through
as shown in the Fig. 8. On the website, we will plan seven rate
recommended fields (Test_A: article below / you might like, Test_A 174,467 7.23% 1.87%
Test_B: Blog below / you might like, Test_C: video below / Test_B 27,051 9.66% 1.91%
you might like, Test_D: Health wall, Test_E: recipes below / Test_C 8,991 7.89% 2.06%
you might like) to carried out two recommended service Test_D 4,823 5.63% 0.56%
verification (1. Personal Preference Recommend Analysis Test_E 781 8.82% 1.88%
Service 2. Rule of thumb recommendation services)
The verification of this research verified through the data of
Table 3、The resource of recommended articles the five recommended field which collected through the
Personal preference recommend analysis services by the
Finance
Recommended Latest in co-operative enterprise. The results of data as shown in Table
and Hot issue
classification today 4, the total browsing rate of PV reading data is 216,113, the
economics
Article total average of page views is 43,222.6, the average of bounce
30 10 15 rate is 7.85％, and the average of click through rate is 1.66％.
quantity
Recommended Business Fashion Then, we verified and analyzed through the rule of thumb
Magazine recommended method (ROT) at follow up.
classification Management life
Article
13 18 17
quantity
 After experimentally verify on the web site, we compared
Personal ROT
and analyzed through the Personal preference recommend
200000 analysis services and the rule of thumb recommended services,
150000
as shown in Fig. 9,we found that in the process of the user's
interaction of social behavior, we can get the data of social
100000 interaction to further understand the user's preferences. And
the preference recommend method is more superior than the
50000 traditional rule of thumb recommended method, to achieve its
overall effectiveness: website page views growth 8%, enhance
0
Test_A Test_B Test_C Test_D Test_E website bounce rate dropped by 11％, commodities click
through rate growth 36%. Through the actual verification to
Fig. 9-1 PV empirical data verify the system of Personal preference recommend analysis,
it has reached certain accuracy.
Personal ROT
12.00% V. CONCLUSION
10.00%
This research is a project of Research on Intelligence
8.00%
Techniques and Service Modes of Social Media conducted by
6.00%
Institute for Information Industry. By the design and
4.00%
implementation of our research: Personal preference analysis
2.00% technology based on the social networks, with 10,000 user
0.00% authorization acquired the Facebook personal data, further
Test_A Test_B Test_C Test_D Test_E
analyze 10,000 personal user preference data to recommend
services to achieve the overall efficiency as: website browsing
Fig. 9-2 Bounce rate empirical data
pages growth 8%, enhance site bounce rate dropped 11%,
commodities click through rate growth 36%. Through the
Personal ROT
2.50% analysis structure of the system that planning and developing
by our research, after experimentally verified and analyzed,
2.00%
the effectiveness of the system already has certain accuracy,
1.50% and it can effectively achieve precise recommendation to
1.00% enhance the contribution of the future research issue for social
0.50%
analysis. This is certainly one of our goals.

0.00%
Test_A Test_B Test_C Test_D Test_E
The purpose of this paper is how to quickly understand the
personal preferences on social behavior of the user, and then
carry out all the recommendation and other innovative
Fig. 9-3 Click through rate empirical data application services. In the future, due to the rapid
development of social media, through the analysis of the
Table 5、The rule of thumb recommended services data
interest of personal preference, to reach a variety of innovative
The rule of thumb recommended services applications services, will be one of the next big issues of the
(ROT) social analysis. Therefore, personal preference analysis data
Bounce Click conducted by the research can be used as reference of
PV follow-up research, expected the follow-up research will be
rate through rate
Test_A 168,157 8.12% 1.21% put forward further industrial application, population analysis,
Test_B 21,113 10.21% 1.34% social policy ... etc, this will be the future of our concerned.
Test_C 6,571 8.97% 1.52%
Test_D 3,543 6.77% 0.33% VI. ACKNOWLEDGEMENT
Test_E 523 9.93% 1.67%
The verification of this research verified through the data of This study is conducted under the "Social Intelligence
the five recommended field which collected through the rule Analysis Service Platform (3/4)" of the Institute for
of thumb recommended services by the co-operative enterprise. Information Industry which is subsidized by the Ministry of
The results of data as shown in Table 5, the total browsing rate Economy Affairs of the Republic of China .
of PV reading data is 199,907, total average of page views is
REFERENCES
39,981.4, the average of bounce rate is 8.8％, and the average
of click through rate is 1.21％. [1] Zhang, C. ; Fang, Y.A “Trust-Based Privacy-Preserving Friend
 Recommendation Scheme for Online Social Networks” IEEE
Transactions on Dependable and Secure Computing, 2015, 12(4),
413-427.
[2] Louta, M.D. ; Varlamis, I. “A Trust-Aware System for Personalized User
Recommendations in Social Networks”, IEEE Transactions on Systems,
Man, and Cybernetics, 2014, 44(4), 409-421.
[3] Cheng-Hung Tsai, Tsun Ku, and Wu-Fan Chien, “Object Architected
Design and Efficient Dynamic Adjustment Mechanism of Distributed
Web Crawlers”, International Journal of Interdisciplinary
Telecommunications and Networking, 2015, 7(2), 57-71.
[4] Rashid, A. “Who Am I? Analyzing Digital Personas in Cybercrime
Investigations” IEEE Computer Magazines 46(4).
[5] Sheth, A., Thomas, C., & Mehra, P. “Continuous semantics to analyze
real-time data.” IEEE Internet Computing, 2010,14(6), 0084-89.
[6] Cheng-Hung, Tsai, Tsun Kua, Liang-Pu Chena, Ping-che, Yang,
“Constructing Concept Space from Social Collaborative Editing,”
Proceedings of AHFE 2015.
[7] Hui Zheng ; Ji-kun Yan ; Ye Jin “Persona Analysis with Text Topic
Modelling” Proceedings of IEEE ACAI 2012.
[8] Shiga, A. , “A Support System for Making Persona Using Bayesian
Network Analysis” Proceedings of IEEE ICBAKE 2013.
[9] Clark, J.W. , “Correlating a Persona to a Person,” Proceedings of IEEE
SocialCom 2012.
[10] Cunha, B. ; Pereira, J.P. ; Gomes, S. ; Pereira, I. ; Santos, J.M. ; Abraham,
A., “Using personas for supporting user modeling on scheduling systems”,
Proceedings of IEEE HIS 2014.
[11] Letier, E. ; Sasse, M.A., “Building a National E-Service using Sentire
experience report on the use of Sentire: A volere-based requirements
framework driven by calibrated personas and simulated user feedback”
Proceedings of IEEE RE 2014
[12] Bo Wu ; Qun Jin ; Nishimura, S. ; Julong Pan ; Wenbin Zheng ; Jianhua
Ma, “Social Stream Organization Based on User Role Analysis for
Participatory Information Recommendation” Proceedings of IEEE
UMEDIA 2014
[13] Chengjie Mao ; Zhenxiong Yang ; HanJiang Lai, “A new team
recommendation model with applications in social network”, Proceedings
of IEEE CSCWDM 2014.
[14] Piedra, N. ; Chicaiza, J. ; Tovar, E., “Recommendation of OERs shared in
social media based-on social networks analysis approach”, Proceedings of
IEEE FIE 2014.
[15] Djauhari, M.A., “A recommendation on PLUS highway development: A
social network analysis approach”, Proceedings of IEEE IEEM 2014.
[16] Mishra, S., “An analysis of positivity and negativity attributes of users in
twitter”, Proceedings of IEEE ASONAM 2014.
[17] Lu Yang, “A collaborative filtering recommendation based on user profile
and user behavior in online social networks”, Proceedings of IEEE
ICSEC 2014.
[18] Huaimin Wang ; Gang Yin ; Ling, C.X. “Who Should Review this
Pull-Request: Reviewer Recommendation to Expedite Crowd
Collaboration” Proceedings of IEEE APSEC 2014.
[19] Castiglione, A. ; De Santis, A. “Friendship Recommendations in Online
Social Networks” Proceedings of IEEE INCoS 2014.
[20] Shaffer, C.A. ; Weiguo Fan ; Fox, E.A., “Recommendation based on
Deduced Social Networks in an educational digital library”, Proceedings
of IEEE JCDL 2014
[21] Gerogiannis, V. ; Anthopoulos, L., “Project team selection based on social
networks”, Proceedings of IEEE ITMC 2014.
[22] Diaby, M. ; Cataldi, M. ; Viennet, E. ; Aufaure, M.-A., “Field selection
for job categorization and recommendation to social network users”,
Proceedings of IEEE ASONAM 2014.
[23] Zegarra Rodriguez, D. ; Bressan, G. “Music recommendation system
based on user's sentiments extracted from social networks”, Proceedings
of IEEE ICCE 2014.
[24] Sun Yanshen ; Yu Xiaomei, “Personalized recommendation based on link
prediction in dynamic super-networks”, Proceedings of IEEE ICCCNT
2014.
[25] Xueming Qian ; Dan Lu ; Xingsong Hou ; Liejun Wang, “Personalized
tag recommendation for Flickr users”, Proceedings of IEEE ICME 2014.
 Multimodal Biometric Scheme for E-Assessment
S.O. Ojo T. Zuva
Computer Systems Engineering Computer Systems Engineering
Tshwane University of Technology Tshwane University of Technology
Tshwane, South Africa Tshwane, South Africa
Abstract—Problems of impersonation and other malpractices of
various degrees have continued to mar the electronic-assessment (e-
assessment) of students at various levels. The use of traditional
methods of password and token passing for identity management
and control has failed to provide acceptable solution to these
problems, due to their susceptibility to theft and various forms of
violations. The failure of these and other methods in identity
management has propelled the emergence of different biometric-
based systems as solutions. The biometric systems use naturally
endowed traits which include fingerprint, face, iris, voice, etc. for
more reliable human identification and authentication. This paper
proposes a multimodal biometric scheme for e-assessment aimed at
reducing the time spent when performing biometric authentication.
Results obtained indicate that by combining the fingerprint and
facial biometric modes in a multimodal biometric scheme, the
amount of time required to authenticate a learner in an e-
assessment can be reduced.
Keywords— E-assessment; biometrics; fusion
I. INTRODUCTION
Traditionally, the assessment process has been considered a
high stake process and the inclusion of one or more proctors
throughout the assessment process has been one of the means
of ensuring that the assessment is conducted in a credible
manner. Proctors, also known as invigilators, are present to
address security concerns during an assessment.
The C.I.A (Confidentiality, Integrity, and Availability)
security goals [1] which are meant to prevent the compromise
of assets in computer network systems are applicable to the
summative e-assessment [2]. Apart from the hardware and
software used for the purpose of an e-assessment, the student
undergoing the e-assessment too is an asset and certain threats
are posed by this asset. These threats include, an illegal or
unauthorized candidate taking an e-assessment, the
dissimulation of identity detail, and the abuse, or misuse of
authenticity detail [2].
Due to the aforementioned threats posed by a student, the
C.I.A. goals though sufficient to protect hardware and data
used for the e-assessment, will not adequately address the
threats posed by the student; and thus three other goals for the
summative e-assessment process can be deduced [2]:
 Presence – Physical and online presence, which
defines the state of an entity being in a particular
environment.
 Identity – The ability to distinguish one entity from
another
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
S.M. Ngwira
Computer Systems Engineering
Tshwane University of Technology
Tshwane, South Africa
 Authentication – Preventing the misuse of identity
detail.
In an effort to guard against the aforementioned threats and
achieve the goals of Presence, Identity, and Authentication,
several biometric based schemes have been proposed with the
use of biometrics for the authentication of learners in an e-
assessment.
The main purpose of this paper is to present a non-intrusive
multimodal biometric authentication scheme for e-assessment.
This paper is arranged as follows: Section 2 a general overview
of biometrics and some popular biometric modes, Section 3
Proposed method, Section 4 Results, Section 5 Conclusion and
recommendation.
II. LITERATURE REVIEW
A. Biometrics
Historically, human identification of persons has been done
through various means, including visual identification, gait
recognition, fingerprint matching, presentation of access cards,
usernames and passwords, analysis of bioelectrical signals, etc.
Biometrics is the ability to identify and authenticate an
individual using one or more of their behavioral, physical or
chemical characteristics [3]. A biometric recognition system
relies on who a person is or what a person does instead of what
they know or possess. Physical biometric characteristics
include face, fingerprint, retina, iris, etc., characteristics.
Behavioral biometric characteristics include gait recognition,
keystroke dynamics, mouse dynamics, etc. Chemical
biometrics characteristics involve the use of human DNA to
identify a person. A biometric system can be described as a
pattern-recognition system. The biometric system aims to
recognize a person based on a feature vector deduced from
physiological and/or behavioral characteristic that a person
possesses.
The use of a particular biometric mode is largely dependent
on the application domain. Some popular biometric modes
used for identification include gait recognition, fingerprint
recognition, facial recognition, bioelectrical signals, keystroke
dynamics, and mouse dynamics.
The gait is a popular biometric due to the plausibility of
obtaining ones gait unobtrusively as an individual walks about.
The gait has been defined as ―the coordinated, cyclic
combination of movements that result in human locomotion‖
[4]. Research has been conducted into the identification of
humans through gait recognition [5, 6], and the gait has been
found to be reliable even in law enforcement as Closed Circuit
Television (CCTV) camera footage can be used in identifying
people by their gait even in a public and crowded place. Within
the domain of e-assessments however, gait recognition would
not be suitable as learners would ideally not be allowed to walk
about during the e-assessments. More often than not, an e-
assessment is carried out by a learner in a seated position.
Fingerprint biometrics has also been found to be reliable in
identifying individuals. A fingerprint is made up of ridges,
furrows, and minutiae. These three contribute to the uniqueness
of the fingerprint [7]. An array of fingerprint sensors are in use
today and are capable of capturing the fingerprint biometric
and authenticating an individual based on the fingerprint
biometric data obtained. The mouse which is a standard input
device can be equipped with a fingerprint sensor which can
authenticate computer users as they interact with the mouse. In
an e-assessment scenario, a learner’s fingerprint biometric can
be obtained with the aid of a mouse equipped with a fingerprint
scanner.
Facial recognition aims to achieve what humans have been
doing since the beginning of time. The identification of an
individual through facial features is not a new concept. The
facial biometrics relies on features such as the curves of the eye
socket, curves of the nose, curves of the chin, among other
features, to identify an individual. The facial biometric is
obtained with the use of a camera, and in the case of an e-
assessment, a web camera mounted to the top of a monitor
would suffice to obtain the facial biometric.
The use of bioelectrical signals as a biometric has also
drawn significant interest. Bioelectrical signals are low
frequency electrical signals that are generated from humans [8].
The activities of the heart (Electrocardiogram), muscles
(Electromyograph), skin (Galvanic skin response), brain
(Electroencephalograph), and eyes (Electrooculogram) can be
measured and monitored. These activities can be used as a
biometric for human identification [9, 10].
Keystroke dynamics is a behavioral biometric that aims to
identify a person based on metrics that include typing speed,
keystroke seek time, flight time, characteristic errors, and
characteristic sequences. As a computer user types on the
keyboard, the keystroke biometric is obtained and a user can be
identified based on the aforementioned metrics. In an e-
assessment, learners can be authenticated as they make use of
the keyboard. Keystroke dynamics however suffers from low
permanence as is typical of most behavioral biometrics. In an
e-assessment, a learner’s keystroke dynamics can be affected
by nervousness. Also, a large number of keystrokes is required
in order to perform a single biometric authentication [11].
Mouse dynamics, like keystroke dynamics, is a behavioral
biometric and is captured as a computer user makes use of the
mouse. Some factors that are important when obtaining the
mouse dynamics biometric include mouse movement speed,
movement direction, action type (double click, single click,
etc.), travel distance, and elapsed time. The use of the mouse
for authentication has the advantage of unobtrusively obtaining
biometric data and authenticating an individual as (s)he makes
use of the mouse [12]. Mouse dynamics does however suffer
from low permanence as is the case with the keystroke
biometric.
B. E-assessment applications
Given the nature of an e-assessment, it is paramount that
the processes responsible for the authentication of learners
operate in a non-intrusive manner. Also, the authentication of
learners needs to be a continuous process. The term for this is
transparent authentication which can be defined as ―the ability
to non-intrusively and continuously authenticate a user utilizing
(largely biometric-based) credentials obtained from users
whilst they normally interact with the electronic device or
system‖ [13].
Some biometric modes such as keystroke and mouse
dynamics aim to accomplish transparent authentication of
learners as they perform their normal interaction with the
computer system. A lot of research has been conducted
particularly around the use of mouse and keystroke dynamics
biometrics, among others, for the transparent authentication of
learners as an e-assessment is conducted.
[12] propose a model for continuous authentication of
learners in a virtual classroom environment. Two behavioral
biometric modes are used for the authentication process,
keystroke dynamics and mouse dynamics; along with the use
of one physical biometric mode which is facial biometrics.
Authentication for each biometric mode is designed as a
subsystem which independently performs authentication of a
learner. Fusion of the results of the individual subsystems is
done by applying a logical AND to all three results.
[14] employ the use of fingerprint biometrics and mouse
dynamics biometrics in proposing a multimodal, nonintrusive
biometric approach to enrolling and authenticating learners
during an e-assessment. Five factors are taken into account
which makes up the mouse dynamics: Mouse movement speed,
movement direction, action type, travel distance, and elapsed
time. The mouse is equipped with a fingerprint sensor used for
obtaining the fingerprint biometric.
[15] propose a bimodal biometric model for e-assessments;
particularly for non-proctored environments. The model
involves integrating with various systems including the e-
assessment system. Fingerprint biometrics, keystroke dynamics
as well as video monitoring are used for the authentication
process. The authentication process provides for transparent
authentication of learners. In measuring the keystroke
biometric, five metrics were considered: typing speed,
keystroke seek time, flight time, characteristic errors, and
characteristic sequences.
According to [16], two samples of the same biometric
characteristic from the same person—e.g. two samples of the
same fingerprint may not be exactly the same due to adverse
imaging conditions, changes in the user’s physiological or
behavioral characteristics, ambient conditions, and the user’s
interaction with the biometric device. Therefore, a biometric
matching system’s typically makes use of a matching score
(usually a single number) to quantify the similarity between the
input and the database template representations. The higher the
score, the more certain the system is that the two biometric
measurements come from the same person. Usually, a
threshold is used for system decision.
C. Biometric systems errors and accuracy
In measuring the accuracy of a biometric system, some of
the metrics used include:
 FAR (False Acceptance Rate) – The number of
authentications granted erroneously over time
 FRR (False Rejection Rate) – The number of valid
biometric data rejected over time
Given 𝑅 as the number of persons wrongly accepted and 𝑇
the total number of false matches, FAR can be determined as
shown in equation 1 [17]:
𝑅 possible identities represent a waned list of learners from which
𝐹𝐴𝑅 = (1)
𝑇 the fingerprint biometric can match its biometric data against.

Given 𝑃 as the number of persons wrongly rejected and 𝑆
as the total number of correct matches made, FRR can be
determined as shown in equation 2 [17]:
𝑃
𝐹𝑅𝑅 = (2)
𝑆
through a mouse equipped with a fingerprint scanner. The
facial biometric is obtained with the use of a web camera
mounted to the top of the computer screen. This biometric data
is fed into the feature extraction module and passed onto the
matching module. At the matching module the biometric data
is compared with templates from the biometric database
containing templates for both the facial and fingerprint
biometric, similarity scores are obtained for the facial biometric
mode and the LLR (likelihood ratio) is computed which is
passed onto fingerprint matching module. Because the facial
biometric is the less accurate biometric compared with
fingerprint biometric, it is given preference in obtaining an
initial set of possible identities which meet the threshold. These
Being that the e-assessment process is time bound, the
initial sample provided when the learner initially interacts with
the computer system will be used as the valid template against
which subsequent biometric samples will be matched against.

The accuracy of the biometric authentication is directly

affected by the values of FAR and FRR. A reduction in the
FAR and/or FRR value will improve the accuracy of the
biometric authentication. The percentage accuracy of the
biometric authentication can be calculated as shown in
equation 3 [17]:

𝐹𝐴𝑅 + 𝐹𝑅𝑅
𝐴𝑐𝑐% = 100 − % (3)
2

where:
𝐴𝑐𝑐% = 𝑃𝑒𝑟𝑐𝑒𝑛𝑡𝑎𝑔𝑒 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦.
Fig 1 Multimodal Biometric Scheme with fusion at matching module.
As the size of a biometric database increases, the possibility Adapted from [19]
of a false rejection or false acceptance also increases [18]. The
accuracy metrics are commonly used in evaluating a biometric The biometric database contains templates of both the
system, for any given application domain, such as e- learners’ facial biometric and the corresponding fingerprint
assessment. biometric in a one to one mapping. Fig 2 shows a sample
database with five learners enrolled.
III. PROPOSED APPROACH
This study proposes the use of fingerprint and facial
biometrics in a multimodal biometrics authentication scheme
for e-assessment. The facial biometric has been found to be less
precise than the fingerprint biometric when identifying
individuals and will thus be used as the first mode of
authentication to obtain a waned list of possible samples which
have met the threshold. From this waned list, the fingerprint
biometric can then be used to examine the smaller subset of
possible identities of the learners. In so doing, we are able to
reduce the overall amount of time spent matching biometric
samples against a large biometric database.
Fusion will be done at the matching module. Fig 1 shows
the flow chart of the proposed scheme with fusion done at the
matching module. The fingerprint biometric is obtained
 𝑁
𝜇 = 𝑥𝑖 (4)
𝑖 =1

Where:
𝑋 = 𝑥1 , 𝑥2 , … . , 𝑥𝑖 , … . , 𝑥𝑁 and represents a 𝑛 × 𝑁 data
matrix
N = number of facial images being examined
𝑥𝑖 = a vector with dimension n made up of an 𝑝 × 𝑞 image
𝑛 = 𝑝 × 𝑞.

The LLR is then calculated as shown in equation 5 [21].

𝑝 𝑋| 𝐼 = 0
𝐿𝐿𝑅 = (5)
𝑝 𝑋| 𝐼 = 1

where:
𝑋 = 𝑠1 , 𝑠2 , 𝑞1 , 𝑞2
𝑠1 = 𝑚𝑎𝑡𝑐𝑕𝑖𝑛𝑔 𝑠𝑐𝑜𝑟𝑒 𝑓𝑟𝑜𝑚 𝑓𝑖𝑛𝑔𝑒𝑟𝑝𝑟𝑖𝑛𝑡 𝑏𝑖𝑜𝑚𝑒𝑡𝑟𝑖𝑐
𝑚𝑎𝑡𝑐𝑕𝑖𝑛𝑔 𝑚𝑜𝑑𝑢𝑙𝑒
𝑠2 = 𝑚𝑎𝑡𝑐𝑕𝑖𝑛𝑔 𝑠𝑐𝑜𝑟𝑒 𝑓𝑟𝑜𝑚 𝑓𝑎𝑐𝑖𝑎𝑙 𝑏𝑖𝑜𝑚𝑒𝑡𝑟𝑖𝑐 𝑚𝑎𝑡𝑐𝑕𝑖𝑛𝑔
𝑚𝑜𝑑𝑢𝑙𝑒
𝑞1 = 𝑞𝑢𝑎𝑙𝑖𝑡𝑦 𝑜𝑓 𝑓𝑖𝑛𝑔𝑒𝑟𝑝𝑟𝑖𝑛𝑡 𝑏𝑖𝑜𝑚𝑒𝑡𝑟𝑖𝑐 𝑠𝑎𝑚𝑝𝑙𝑒
𝑞2 = 𝑞𝑢𝑎𝑙𝑖𝑡𝑦 𝑜𝑓 𝑓𝑎𝑐𝑖𝑎𝑙 𝑏𝑖𝑜𝑚𝑒𝑡𝑟𝑖𝑐 𝑠𝑎𝑚𝑝𝑙𝑒
Fig 2 One to one maping of facial biometric templates with fingerprint
biometric template
𝐼 ∈ 0, 1
𝑝 𝑋| 𝐼 = 0 = 𝑡𝑕𝑒 𝑔𝑒𝑛𝑢𝑖𝑛𝑒 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 𝑜𝑓 𝑠𝑒𝑡 𝑋
From the set of five templates shown in Fig 2, the facial
biometric will return three templates which have achieved the 𝑝 𝑋| 𝐼 = 1 = 𝑡𝑕𝑒 𝑖𝑚𝑝𝑜𝑠𝑡𝑒𝑟 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 𝑜𝑓 𝑠𝑒𝑡 𝑋
set threshold. These three templates are ranked in order of a
similarity score with the first template in the set being the most In comparing fingerprint biometric data with template data,
likely correct identity and the last template being the least Minutiae Matching is used as shown in equation 6 [21].
likely identity. The three templates are passed on to the
fingerprint biometric module. The fingerprint module then 2 2
compares its biometric data against the three provided and also 𝑟𝑜𝑤𝑘𝑇 − 𝑟𝑜𝑤𝑟𝑒𝑓
𝑇
+ 𝑐𝑜𝑙𝑘𝑇 − 𝑐𝑜𝑙𝑟𝑒𝑓
𝑇

ranks them according to a similarity score. 𝑟𝑘𝑇

𝑇
∅𝑇𝑘 = 𝑟𝑜𝑤𝑘𝑇 − 𝑟𝑜𝑤𝑟𝑒𝑓 (6)
tan−1
If the similarity ranking result of the fingerprint module is 𝜃𝑘𝑇 𝑐𝑜𝑙𝑘𝑇 − 𝑐𝑜𝑙𝑟𝑒𝑓
𝑇
identical to the similarity ranking done by the facial biometric 𝑇
𝜃𝑘𝑇 − 𝜃𝑟𝑒𝑓
module, the authentication would have been done successfully
and the decision module can accept or reject the learner based
on the first (highest ranking) template in the waned list. In
cases where the similarity ranking of the fingerprint module where, for a template image:
differs from that of the facial biometric module, this will 𝑟𝑘𝑇 = 𝑟𝑎𝑑𝑖𝑎𝑙 𝑑𝑖𝑠𝑡𝑎𝑛𝑐𝑒 𝑜𝑓 𝑘 𝑡𝑕 𝑚𝑖𝑛𝑢𝑡𝑖𝑎𝑒
require human intervention to ascertain whether or not the
∅𝑇𝑘 = 𝑟𝑎𝑑𝑖𝑎𝑙 𝑎𝑛𝑔𝑙𝑒 𝑜𝑓 𝑘 𝑡𝑕 𝑚𝑖𝑛𝑢𝑡𝑖𝑎𝑒
learner is who (s)he claims to be.
𝜃𝑘𝑇 = 𝑜𝑟𝑖𝑒𝑛𝑡𝑎𝑡𝑖𝑜𝑛 𝑎𝑛𝑔𝑙𝑒 𝑜𝑓 𝑘 𝑡𝑕 𝑚𝑖𝑛𝑢𝑡𝑖𝑎𝑒
𝑇 𝑇
By alleviating the need for the fingerprint biometric to 𝑟𝑜𝑤𝑟𝑒𝑓 , 𝑐𝑜𝑙𝑟𝑒𝑓 = 𝑟𝑜𝑤, 𝑐𝑜𝑙𝑢𝑚𝑛 𝑖𝑛𝑑𝑒𝑥 𝑜𝑓 𝑟𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒 𝑝𝑜𝑖𝑛𝑡𝑠
compare its data against the entire biometric database the 𝑏𝑒𝑖𝑛𝑔 𝑐𝑜𝑛𝑠𝑖𝑑𝑒𝑟𝑒𝑑.
amount of time spent for a single biometric authentication is
reduced. For a biometric database of 𝑁 templates, the facial
biometric will match its biometric data against the entire IV. RESULTS
database of 𝑁 templates and select the top 𝑥 ranked by a
similarity score. The fingerprint biometric will thus match its Experiments conducted show that for a set of 100 templates
data against 𝑥 templates as opposed to 𝑁 templates. the time spent for matching biometric data was reduced due to
fewer templates being used by the fingerprint module. Fig 3
In obtaining the facial biometric, the Kanade-Lucas-Tomasi shows the graph of template number against percentage time
(KLT) algorithm is used. Images from the web camera are spent. The values on the x-axis indicate the number of
represented as vectors and the mean vector of the images being templates which the fingerprint biometric needs to match
examined is given by equation 4 [20]:
against for a single authentication once the facial biometric has
obtained a set of likely templates ordered by a similarity score.
Fig 3 Performance of the system in relationship to the number recommended
to the fingerprint module from the facial module
As a result, the overall time for authentication of the learner
was reduced. Fig 4 shows the graph of number of biometric
templates passed onto the fingerprint module against the
percentage time saved. No observations were made with
respect to the FAR and FRR. The significant observations
made were with respect to the database lookup time which was
reduced.
Figure 4 Performkance of the system in relationship to the time saved with
regard to the number recommended to the fingerprint module from the facial
module.
V. CONCLUSION AND REOCOMMENDATION
This paper proposed a multimodal biometric authentication
scheme with a reduced amount of time for biometric
authentication. The scheme makes use of facial and fingerprint
biometrics for the authentication of learners in an e-assessment.
Results obtained from experimentation show that a multimodal
biometric scheme can have a reduced authentication time when
smaller sets of biometric data are used for each subsequent
biometric mode. Further research needs to be conducted into
protecting a biometric system from circumvention particularly
in high stake environments such as an e-assessment.
ACKNOWLEDGMENT
Special thanks to Dr Libor Spacek for making the faces96
database available for use [22]. Thanks to the University of
Bologna for their fingerprint databases DB2 [23].
REFERENCES
[1] A. Simmonds, P. Sandilands, and L. van Ekert, "An ontology for network
security attacks," in Applied Computing, ed: Springer, 2004, pp. 317-323.
[2] K. M. Apampa, G. Wills, and D. Argles, "Towards security goals in
summative e-assessment security," in Internet Technology and Secured
Transactions, 2009. ICITST 2009. International Conference for, 2009,
pp. 1-5.
[3] K. Rabuzin, M. Baca, and M. Sajko, "E-learning: Biometrics as a
Security Factor," in Computing in the Global Information Technology,
2006. ICCGI '06. International Multi-Conference on, 2006, pp. 64-64.
[4] J. E. Boyd and J. J. Little, "Biometric gait recognition," in Advanced
Studies in Biometrics, ed: Springer, 2005, pp. 19-42.
[5] J. P. Gupta, P. Dixit, N. Singh, and V. B. Semwal, "Analysis of Gait
Pattern to Recognize the Human Activities," arXiv preprint
arXiv:1407.4867, 2014.
[6] V. B. Semwal, M. Raj, and G. Nandi, "Biometric gait identification based
on a multilayer perceptron," Robotics and Autonomous Systems, vol. 65,
pp. 65-75, 2015.
[7] D. Kumar and Y. Ryu, "A brief introduction of biometrics and fingerprint
payment technology," in Future Generation Communication and
Networking Symposia, 2008. FGCNS'08. Second International
Conference on, 2008, pp. 185-192.
[8] Y. N. Singh, S. K. Singh, and A. K. Ray, "Bioelectrical signals as
emerging biometrics: Issues and challenges," ISRN Signal Processing,
vol. 2012, 2012.
[9] R. Palaniappan and D. P. Mandic, "EEG based biometric framework for
automatic identity verification," The Journal of VLSI Signal Processing
Systems for Signal, Image, and Video Technology, vol. 49, pp. 243-250,
2007.
[10] K.-S. Song, S. M. Lee, and S. Nam, "Cognitive Biometrics Application
for e-Learning Security Enhancement," International Journal of Bio-
Science and Bio-Technology, vol. 5, pp. 143-152, 2013.
[11] P. S. Teh, A. B. J. Teoh, and S. Yue, "A survey of keystroke dynamics
biometrics," The Scientific World Journal, vol. 2013, 2013.
[12] M. K. Dehnavi and N. P. Fard, "Presenting a multimodal biometric model
for tracking the students in virtual classes," Procedia-Social and
Behavioral Sciences, vol. 15, pp. 3456-3462, 2011.
[13] N. L. Clarke, P. Dowland, and S. M. Furnell, "e-Invigilator: A biometric-
based supervision system for e-Assessments," in Information Society (i-
Society), 2013 International Conference on, 2013, pp. 238-242.
[14] S. Asha and C. Chellappan, "Authentication of e-learners using
multimodal biometric technology," in Biometrics and Security
Technologies, 2008. ISBAST 2008. International Symposium on, 2008,
pp. 1-6.
[15] Y. W. Sabbah, I. A. Saroit, and A. M. Kotb, "A Smart Approach for
Bimodal Biometric Authentication in Home-Exams (SABBAH Model),"
Biometrics and Bioinformatics, vol. 4, pp. 32-45, 2012.
[16] S. Prabhakar, S. Pankanti, and A. K. Jain, "Biometric recognition:
security and privacy concerns," Security & Privacy, IEEE, vol. 1, pp. 33-
42, 2003.
[17] J. Malik, D. Girdhar, R. Dahiya, and G. Sainarayanan, "Reference
Threshold Calculation for Biometric Authentication," International
Journal of Image, Graphics and Signal Processing (IJIGSP), vol. 6, p.
46, 2014.
[18] T. Dunstone and N. Yager, Biometric system and data analysis: Design,
evaluation, and data mining: Springer Science & Business Media, 2008.
[19] K. Sasidhar, V. L. Kakulapati, K. Ramakrishna, and K. KailasaRao,
"Multimodal biometric systems-Study to improve accuracy and
performance," arXiv preprint arXiv:1011.6220, 2010.
[20] X. Lu, "Image analysis for face recognition," personal notes, May, vol. 5,
2003.
[21] R. N. Rodrigues, L. L. Ling, and V. Govindaraju, "Robustness of
multimodal biometric fusion methods against spoof attacks," Journal of
Visual Languages & Computing, vol. 20, pp. 169-179, 2009.
[22] L. Spacek, 'Face Recognition Data', Csessex.ac.uk,
2015.[Online].Available:http://cswww.essex.ac.uk/mv/allfaces/.
[Accessed: 02- Aug- 2015]
[23] U. Bologna, 'FVC2000', Bias.csr.unibo.it, 2015.
[Online].Available:http://bias.csr.unibo.it/fvc2000/databases.asp.
[Accessed: 02- Aug- 2015]
 Survey of Biometric Authentication for E-assessment
T. Zuva
S.O. Ojo Computer Systems Engineering
Computer Systems Engineering Tshwane University of Technology
Tshwane University of Technology Tshwane, South Africa
Tshwane, South Africa
Abstract—Problems of impersonation and other malpractices
of various degrees have continued to mar the electronic-
assessment (e-assessment) of students at various levels. The use of
traditional methods of password and token passing for identity
management and control has failed to provide acceptable solution
to these problems, due to their susceptibility to theft and various
forms of violations. The failure of these and other methods in
identity management has propelled the emergence of different
biometric-based systems as solutions. The biometric systems use
naturally endowed traits which include fingerprint, face, iris,
voice, etc. for more reliable human identification and
authentication. In this paper, various biometric modes are
discussed along with their application within the domain of e-
assessment. The results in this paper show that the need for
multimodal biometrics and transparent authentication in an e-
assessment is still an open issue.
Keywords—E-assessment; Biometrics; Multimodal Biometrics
I. INTRODUCTION
Traditionally, the assessment process has been considered a
high stake process and the inclusion of one or more proctors
throughout the assessment process has been one of the means
of ensuring that the assessment is conducted in a credible
manner. Proctors, also known as invigilators, are present to
address security concerns during an assessment. The C.I.A
(Confidentiality, Integrity, and Availability) security goals [1]
which are meant to prevent the compromise of assets in
computer network systems are applicable to the summative e-
assessment. Apart from the hardware and software used for the
purpose of an e-assessment, the student undergoing the e-
assessment too is an asset and certain threats are posed by this
asset. These threats include, an illegal or unauthorized
candidate taking an e-assessment, the dissimulation of identity
detail, and the abuse, or misuse of authenticity detail [2].
Due to the aforementioned threats posed by a student, the
C.I.A. goals though sufficient to protect hardware and data
used for the e-assessment, will not adequately address the
threats posed by the student; and thus three other goals for the
summative e-assessment process can be deduced [2]:
 Presence – Physical and online presence, which
defines the state of an entity being in a particular
environment.
 Identity – The ability to distinguish one entity from
another
 Authentication – Preventing the misuse of identity
detail.
S.M. Ngwira
Computer Systems Engineering
Tshwane University of Technology
Tshwane, South Africa
In an effort to guard against the aforementioned threats and
achieve the goals of Presence, Identity, and Authentication,
several techniques have been proposed with the use of
biometrics for the authentication of learners in an e-assessment.
The main purpose of this paper is to discuss various
biometric modes and their application for e-assessment. This
paper is arranged as follows; Section 2 a general overview of
biometrics, Section 3 biometrics performance measures,
Section 4 multimodal biometrics, Section 5 applications of
biometrics for e-assessment, Section 6 Challenges and open
issues.
II. BIOMETRICS
Historically, human identification of persons has been done
through various means, including visual identification, gait
recognition, fingerprint matching, presentation of access cards,
usernames and passwords, etc. Biometrics is the ability to
identify and authenticate an individual using one or more of
their behavioral or physical or chemical characteristics [3]. A
biometric recognition system relies on who you are or what
you do [4]. Physical biometric characteristics include face,
fingerprint, retina, iris, etc., characteristics. Behavioral
biometric characteristics includes gait recognition, keystroke
dynamics, mouse dynamics, etc. Chemical biometrics
characteristics involves the use of human DNA to identify a
person. A biometric system is a pattern-recognition system.
The biometric system aims to recognize a person based on a
feature vector deduced from physiological or behavioral
characteristic that the person possesses.
Biometrics recognition entails enrollment, authentication or
identification tasks. Enrollment associates an identity with a
biometric characteristic. In verification, an enrolled individual
can proffer an identity and the system will be responsible for
verifying the authenticity of the proffered identity based on his/
her biometric feature. An identification system identifies an
enrolled individual based on their biometric characteristics
without the individual having to claim an identity. Biometric
recognition has become a principal method of human
identification and verification.
Any human traits can serve as a biometric characteristic
provided it satisfies the measurement requirements of:
universality - each person should have the characteristic,
distinctiveness - any two persons should be different in terms
of the characteristic, permanence - the characteristic should not
change drastically over a short period of time, collectability -
the characteristic should be quantitatively measurable, and for
practical purposes, in addition, performance - meet accuracy,
speed, and resource requirements, acceptability - accepted by Next we identify some behavioral and physical biometrics
the intended population, and circumvention – not easily and discuss how they are used.
hoodwinked [4, 5]. These seven measurement parameters are
used to determine the usefulness of a biometric. For the e- A. Keystroke dynamics
assessment domain, the most important five of these are The keystroke dynamics biometric aims to obtain biometric
universality, distinctiveness, permanence, collectability, and data as a learner types on the computer keyboard. In order to
circumvention. authentication a learner using the keystroke biometrics the
following metrics are considered: typing speed, keystroke seek
Tables 1 and 2 show some physiological biometrics and time, flight time, characteristic errors, and characteristic
behavioral biometrics respectively and how they measured sequences [7]. The time between keystrokes is known as a
against these five parameters, on the scale of LOW (L) – diagraph and is a vital part of the keystroke biometric data. In
MEDIUM (M) – HIGH (H). matching the keystroke data obtained with templates in a
database, correlation is used. The correlation r can be
Table 1 Physiological biometric modes rating based on five characteristics [5] determined by equation 1 [8]:

𝑛
𝑖=1 (𝑘𝑖 ∗ 𝑡𝑖 )
𝑟=
Distinctiveness

Circumvention 𝑛 2 𝑛 2
(1)
Collectability

𝑖=1 (𝑘𝑡 ) ∗ 𝑖=1(𝑡𝑡 )

Universality

Permanence

Where
𝑘 = 𝑣𝑒𝑐𝑡𝑜𝑟 𝑜𝑓 𝑙𝑒𝑛𝑔𝑡𝑕 𝑛 𝑠𝑡𝑜𝑟𝑖𝑛𝑔 𝑡𝑕𝑒 𝑓𝑙𝑖𝑔𝑕𝑡 𝑡𝑖𝑚𝑒 𝑜𝑓 𝑡𝑕𝑒
Biometric 𝑠𝑡𝑜𝑟𝑒𝑑 𝑡𝑒𝑚𝑝𝑙𝑎𝑡𝑒
Fingerprint H H M M H 𝑡 = 𝑣𝑒𝑐𝑡𝑜𝑟 𝑜𝑓 𝑙𝑒𝑛𝑔𝑡𝑕 𝑛 𝑠𝑡𝑜𝑟𝑖𝑛𝑔 𝑡𝑕𝑒 𝑓𝑙𝑖𝑔𝑕𝑡 𝑡𝑖𝑚𝑒 𝑏𝑒𝑡𝑤𝑒𝑒𝑛
𝑘𝑒𝑦𝑠𝑡𝑟𝑜𝑘𝑒𝑠 𝑜𝑓 𝑎 𝑏𝑖𝑜𝑚𝑒𝑡𝑟𝑖𝑐 𝑠𝑎𝑚𝑝𝑙𝑒
Palmprint H H M M M
Retina H H H L L B. Mouse dynamics
Iris H H H M L Mouse dynamics, like keystroke dynamics, is a behavioral
biometric and is also captured non-intrusively as a learner
Face H M M H L interact with the computer (uses the mouse). Any mouse action
can be classified into one of four categories [9]:
Table 2 Behavioural biometric modes rating based on five characteristics [5]
 Mouse-Move – Any general motion of the mouse
 Drag-and-Drop – A mouse down action followed
by a mouse up action
Distinctiveness

Circumvention
Collectability
Universality

Permanence

 Point-and-Click – A mouse motion followed by a

single or double click
 Silence – No motion
Biometric
Mouse actions are measured within 45⁰ angles of a circle.
Signature H M L H M This implies a maximum of 8 directions in which a mouse
Gait L M L H L movement can occur.

Device Use L M L H L Complexities associated with the use of non-standard

computer equipment for authentication are eliminated with the
Typing/Keystroke H M L H M use of keystroke and mouse dynamics.
Voice H M M H M C. Fingerprint biometrics
The fingerprint is the most widely used biometric [10]. The
Behavioral biometrics generally suffer from low fingerprint biometric owes its popularity largely to its
permanence [6]. A learner’s keystroke dynamics can change uniqueness. A fingerprint is made up of ridges, furrows and
drastically within a short period of time as the learners gets minutiae [11]. These three contribute to the uniqueness of the
accustomed to the keyboard being used. Also anxiety or fingerprint.
nervousness can have an effect on the keystroke dynamics An array of fingerprint sensors are in use today that are
during an e-assessment. capable of capturing the fingerprint biometric and
These measurement requirements for biometric authenticating a person based on analysis of the ridges and
characteristics are necessarily to be considered in determining furrows which make the fingerprint distinct. These devices are
their suitability for a given application domain. used for enrolment and template matching [12]. Electro-optical

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

sensors are used to capture an image of the fingerprint and can
be fitted with light emitting polymers which can in turn be
embedded into devices such as a mouse [13].
D. Facial biometrics
Previously, facial biometrics relied on 2D images which
meant the individual being authenticated had to face a camera
directly to be able to capture the biometric. Technology has
since evolved enabling cameras to accurately identify
individuals in a crowded area. [14] describes 3D facial
recognition as a more accurate and therefore reliable approach
to facial biometrics. [15] identify some features of the face that
is measured by a 3D facial recognition system which are curves
of the eye socket, curves of the nose, and curves of the chin.
The authentication of a person by facial biometrics is a
nonintrusive process and can be done transparently. The
enrollee may not even be aware that the authentication process
is ongoing [16].
Algorithms have been developed which aim to detect a face
and track features of the face for the purpose of identification
and authentication. The Kanade-Lucas-Tomasi (KLT)
algorithm is one such algorithm. Images of the face can be
obtained by a camera and are represented as vectors.
The facial biometric can be computed by finding the mean
Eigenvector. The Kanade-Lucas-Tomasi (KLT) algorithm is
one such algorithm. Images of the face can be obtained by a
camera and are represented as vectors.
The mean vector of the images being examined is given
equation 2 [17]:
𝑁 IV. MULTIMODAL BIOMETRICS
𝜇 = 𝑥𝑖 (2)
𝑖 =1
Where:
𝑋 = 𝑥1 , 𝑥2 , … . , 𝑥𝑖 , … . , 𝑥𝑁 represents a 𝑛 × 𝑁 data matrix
N = number of facial images being examined
𝑥𝑖 = a vector with dimension n made up of an 𝑝 × 𝑞 image
𝑛 = 𝑝 × 𝑞.
III. PERFORMANCE MEASUREMENTS
According to [4], Two samples of the same biometric from
a single individual—e.g. two samples of the same fingerprint,
may not be exactly the same due to adverse imaging
conditions, changes in the user’s physiological or behavioral
characteristics, ambient conditions, and the user’s interaction
with the biometric device [4]. Therefore, a biometric matching
system’s typically makes use of a matching score. Usually, a
threshold is used for system decision.
In measuring the accuracy of a biometric system, some of
the metrics used include:
 FAR (False Acceptance Rate) – The number of
authentications granted erroneously over time
 FRR (False Rejection Rate) – The number of valid
biometric data rejected over time.
Given 𝑅 as the number of persons wrongly accepted and 𝑇
the total number of false matches, FAR can be determined by
[18]:
𝑅
𝐹𝐴𝑅 = (3)
𝑇

Given 𝑃 as the number of persons wrongly rejected and 𝑆
as the total number of correct matches made, FRR can be
determined by [18]:
𝑃
𝐹𝑅𝑅 = (4)
𝑇
The accuracy of the biometric authentication is directly
affected by the values of FAR and FRR. A reduction the FAR
and/or FRR value will improve the accuracy of the biometric
authentication. The percentage accuracy of the biometric
authentication can be calculated by [18]:
𝐹𝐴𝑅 + 𝐹𝑅𝑅
𝐴𝑐𝑐 = 100 − % (5)
2
Where: 𝐴𝑐𝑐 = 𝑃𝑒𝑟𝑐𝑒𝑛𝑡𝑎𝑔𝑒 𝑎𝑐𝑐𝑢𝑟𝑎𝑐𝑦.
As the size of a biometric database increases, the possibility
of a false rejection or false acceptance also increases [19]. The
accuracy metrics are commonly used in evaluating a biometric
system, for any given application domain, such as e-
assessment.
In order to implement near absolute security within an
information system using biometric features, it is necessary to
combine several different biometric traits, either those
belonging to one type only or those already presenting a
combination of both types mentioned [9].
Biometric systems can be either unimodal or multimodal.
Multimodal biometric system is the use of two or more
biometric modes for recognition, while a unimodal biometric
system only employs one biometric mode for verification
and/or identification. Multimodal biometric systems are more
noise resistant and less susceptible to spoofing [20].
Minimizing noise and the risk of spoofing is of great
importance given the high stake environment of an e-
assessment. Having multiple biometric modes raises the
question of how authentication can be done based on the inputs
from the multiple biometric sensors, that is, how the fusion of
results from the multiple biometric modes can be done in an
effort to authenticate a person. Biometric systems generally
consist of four modules, which are, the sensor module that
obtains the biometric data, the feature extraction module that
obtains features sets from the raw biometric data, the matching
module that generates matching scores based on comparisons
between the features sets ant the stored templates, and lastly the
decision module that authenticates a person based on the
matching scores. Fusion can take place in the level of any of
these four modules [12].
 V. APPLICATION OF BIOMETRICS FOR E-ASSESSMENT
[21] propose a model for continuous authentication of
learners in a virtual classroom environment. Two behavioral
biometric modes are used for the authentication process,
keystroke dynamics and mouse dynamics; along with the use
of one physical biometric mode which is facial biometrics.
Authentication for each mode is designed as a subsystem
which independently performs authentication of a learner.
Fusion of the results of the individual subsystems is done by
applying a logical AND to all three results.
[9] employ the use of fingerprint biometrics and mouse
dynamics biometrics in proposing a multimodal, nonintrusive
biometric approach to enrolling and authenticating learners
during an e-assessment. Five factors are taken into account
which makes up the mouse dynamics: Mouse movement speed,
movement direction, action type, travel distance, and elapsed
time. The mouse is equipped with ta fingerprint sensor used for
obtaining the fingerprint biometric.
[7] propose a bimodal biometric model for e-assessments;
particularly for non-proctored environments. The model
involves integrating with various systems including the e-
assessment system. Fingerprint biometrics, keystroke dynamics
as well as video monitoring are used for the authentication
process. The authentication process provides for transparent
authentication of learners. In measuring the keystroke
biometric, five metrics were considered: typing speed,
keystroke seek time, flight time, characteristic errors, and
characteristic sequences.
VI. CONCLUSION AND RECOMMENDATION
In this paper, various biometric modes were discussed and
their application within the domain of e-assessments. Given the
continued adoption of e-assessments globally, more research
needs to be conducted to increase the accuracy of biometric
authentication and the protection of a biometric system from
circumvention through multimodal biometrics. The transparent
authentication of a learner using multimodal biometrics is vital
in an e-assessment and thus requires more research.
REFERENCES
[1] A. Simmonds, P. Sandilands, and L. van Ekert, "An ontology for
network security attacks," in Applied Computing, ed: Springer, 2004, pp.
317-323.
[2] K. M. Apampa, G. Wills, and D. Argles, "Towards security goals in
summative e-assessment security," in Internet Technology and Secured
Transactions, 2009. ICITST 2009. International Conference for, 2009,
pp. 1-5.
[3] K. Rabuzin, M. Baca, and M. Sajko, "E-learning: Biometrics as a
Security Factor," in Computing in the Global Information Technology,
2006. ICCGI '06. International Multi-Conference on, 2006, pp. 64-64.
[4] S. Prabhakar, S. Pankanti, and A. K. Jain, "Biometric recognition:
security and privacy concerns," Security & Privacy, IEEE, vol. 1, pp.
33-42, 2003.
[5] H. A. Crawford, "A framework for continuous, transparent
authentication on mobile devices," University of Glasgow, 2012.
[6] P. S. Teh, A. B. J. Teoh, and S. Yue, "A survey of keystroke dynamics
biometrics," The Scientific World Journal, vol. 2013, 2013.
[7] Y. W. Sabbah, I. A. Saroit, and A. M. Kotb, "A Smart Approach for
Bimodal Biometric Authentication in Home-Exams (SABBAH Model),"
Biometrics and Bioinformatics, vol. 4, pp. 32-45, 2012.
[8] E. Flior and K. Kowalski, "Continuous biometric user authentication in
online examinations," in Information Technology: New Generations
(ITNG), 2010 Seventh International Conference on Information
Technology, 2010, pp. 488-492.
[9] S. Asha and C. Chellappan, "Authentication of e-learners using
multimodal biometric technology," in Biometrics and Security
Technologies, 2008. ISBAST 2008. International Symposium on, 2008,
pp. 1-6.
[10] R. Saini and N. Rana, "Comparison of Various Biometric Methods,"
International Journal of Advances in Science and Technology (IJAST),
vol. 2, 2014.
[11] D. Kumar and Y. Ryu, "A brief introduction of biometrics and
fingerprint payment technology," International Journal of advanced
science and Technology, vol. 4, pp. 25-38, 2009.
[12] A. Ross and A. K. Jain, "Multimodal biometrics: An overview," in
Signal Processing Conference, 2004 12th European, 2004, pp. 1221-
1224.
[13] X. Xia and L. O'Gorman, "Innovations in fingerprint capture devices,"
Pattern Recognition, vol. 36, pp. 361-369, 2003.
[14] S. Thorat, S. Nayak, and J. P. Dandale, "Facial recognition technology:
An analysis with scope in India," arXiv preprint arXiv:1005.4263, 2010.
[15] H. Garg and S. Singh, "A Review Paper on Facial Recognition,"
International Journal of Enhanced Research in Science Technology &
Engineering, vol. 3, pp. 80-85, 2014.
[16] T. O. Majekodunmi and F. E. Idachaba, "A Review of the Fingerprint,
Speaker Recognition, Face Recognition and Iris Recognition Based
Biometric Identification Technologies," 2011.
[17] X. Lu, "Image analysis for face recognition," personal notes, May, vol.
5, 2003.
[18] J. Malik, D. Girdhar, R. Dahiya, and G. Sainarayanan, "Reference
Threshold Calculation for Biometric Authentication," International
Journal of Image, Graphics and Signal Processing (IJIGSP), vol. 6, p.
46, 2014.
[19] T. Dunstone and N. Yager, Biometric system and data analysis: Design,
evaluation, and data mining: Springer Science & Business Media, 2008.
[20] A. Mishra, "Multimodal biometrics it is: need for future systems,"
International journal of computer applications, vol. 3, pp. 28-33, 2010.
[21] M. K. Dehnavi and N. P. Fard, "Presenting a multimodal biometric
model for tracking the students in virtual classes," Procedia-Social and
Behavioral Sciences, vol. 15, pp. 3456-3462, 2011.
 A Framework for the Support of Mobile Knowledge
Workers in South Africa
Michelle Robberts Klarissa Engelbrecht
School of Computing Department of Auditing
University of South Africa University of South Africa
Pretoria, South Africa Pretoria, South Africa
Abstract—Over a relatively short time span, mobile
communication progressed from an expensive and unwieldy
commodity to the brink of delivering a truly “agile” workplace.
It has become of importance for organisations to invest in
initiatives aimed at providing workers with a degree of mobility.
The impact of mobility on organisational environments seems to
be generally well understood. The researchers identified a gap
with regard to a holistic analysis of the requirements of mobile
environments. The question as to what framework can be
defined to guide organisations when designing environments to
support mobile knowledge workers was identified.
Having analysed data gathered via an online questionnaire with
the current literature knowledge base, a layered approach to
mobile environments, namely the SAND framework, was
designed. Addressing the aspects captured in the layers of the
framework can assist organisations in designing successful and
optimal mobile environments.
Keywords—Mobile workforce; Mobile knowledge workers;
Mobile worker requirements; Challenges facing mobile worker;
Mobile worker perceptions; Framework
I. INTRODUCTION
Mobile work has introduced a wide range of opportunities
into the technological and economic landscape. With the
continuing increase in mobile device capabilities,
globalisation, and the dramatically fast, “need now, want
now” attitude from consumers and clients, organisations can
no longer ignore the mobility trend, lest they suffer the
consequences of missing business due to being unable to react
as fast as their competitors.
Despite literature addressing best practices and guidelines for
mobile knowledge worker support, a framework that brings
these together into a single holistic whole remains amiss.
Organisations consequently need to spend much time
researching and gathering information to guide them when
implementing their own mobile initiatives [37]. All these lead
to increased cost while competitors gain traction in the market
[1].
The question of “what framework may be defined to address
the needs of mobile workers” therefore gained importance as
the trend for mobile work gains momentum, aimed at
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
John Andrew van der Poll
Graduate School of Business Leadership
(SBL)
University of South Africa,
Midrand, South Africa
realising a number of benefits for a company. However,
failing to elicit the necessary requirements or hastily
implementing mobility initiatives may become costly
exercises [2].
The layout of the paper follows: Section 2 addresses the
research questions that inspired this research, while existing
models and frameworks are discussed in Section 3. The
research methodology followed to gather and analyse the data
is outlined in Section 4. Section 5 describes the layers of the
SAND framework. Sections 6 and 7 present the conclusion
and directions for future work respectively.
The research questions are defined hereunder.
II. RESEARCH QUESTIONS
Main RQ: What framework/model can be defined to address
the requirements of mobile knowledge workers?
Sub Research Questions
Our main question may be deconstructed into the following
sub RQs:
RQ1: What activities and tasks are performed by mobile
knowledge workers?
RQ2: Which technologies are being utilised by mobile
knowledge workers?
RQ3: What constraints are being experienced by mobile
knowledge workers?
The answers to the above sub questions led to the researchers
devising a framework, coined the SAND Framework that
brings together these factors.
Such framework may be used by organisations as a guide
during the implementation of mobile environments, thereby
providing them with a set of guidelines to act faster, hence
reducing the risk of financial losses.
In building the framework, specific focus was placed on
identifying the different aspects of mobility, including the
challenges and constraints that are faced by mobile
knowledge workers when performing tasks in a mobile
environment.
Amongst the findings of this research, it became clear that the
constraints and perceptions of mobile workers differ from
those of their office-bound counterparts. When the traditional
office environment starts to evolve into a mobile
environment, organisations need to pay attention to a wide
array of new and challenging requirements to keep these
mobile workers satisfied and optimally productive.
Apart from the technological challenges that might be
expected from a mobile environment, such as securing the
corporate network and providing employees with the most
efficient means to connect to the network, several “soft”
issues also need to be addressed in order to be successful and
provide optimal support to the users of the environment.
Soft issues that need to be addressed can take many forms.
Examples of these are the prevention of professional
isolation, disconnection and the loss of their organisational
identity, the prevention of an imbalance between work and
personal life, and coping with the digital divide, especially in
cases where workers who find technology challenging may
on some level be forced to perform mobile work.
Furthermore, when the personalities of workers are brought
into the mix (e.g. extrovert vs. introvert), implementing a
successful mobile initiative becomes challenging.
III. EXISTING MODELS AND FRAMEWORKS
The literature deals mainly with creating an environment to
enable mobile work in business markets, but as is also hinted
at by [3] and [4], a clear and comprehensive analysis of the
requirements of the workers that are to function in these type
of environments remains mostly amiss. A reason for this state
of affairs may be that designing mobile environments are
more complex than simply installing hardware and software
not requiring a comprehensive understanding of the needs of
mobile workers.
Brodt et al. [5] point out that a flexible approach is required
to support mobile workers, allowing for the following
attributes:
 No fixed working space
Access to an organisation’s network should be available
wherever a worker happens to be. The physical “main”
office space is simply another place from where the
worker an access the network.
 Internet-based processes
Processes and systems should be designed to be useful
and accessible by both mobile workers and office-bound
workers, and should be available to workers whenever
they need access to the organisational facilities.
 Mobile technology
Technology ought to enable workers to seamlessly
connect to the network in order to enable
anyplace/anytime work. The mobile workforce generally
employ a wide array of different mobile devices to
perform their tasks on. The systems should allow for this
with minimal integration problems.
 Management of mobility and mobile working culture
An organisation should recognise that mobile workers
have different requirements to their office-bound
counterparts, and should ensure that their managers are
well-trained to motivate and manage mobile teams.
 Privacy
Issues of privacy and accessibility should be appreciated
by an organisation, and protocols and measures ought to
be put in place to protect workers and the data they
access and generate.
Neyem et al. [6] agree that the development of mobile shared
workspaces is a complex process. The authors focussed
mainly on the limitations and challenges of mobile
technology and experiences as discussed in the literature. The
requirements they identified are presented in Table 1.
Table 1: The general requirements of mobile shared workspaces [6].
Requirement Brief discussion
Autonomy Users of a mobile system must be able to work
autonomously, storing their work in an individual
workspace.
Flexibility Mobile systems must support frequent changes in
group size and structure. By virtue of being mobile,
users may become disconnected from a group, hence
applications ought to allow users to switch, on
demand between being online or offline.
Consistency and Frequent disconnections and autonomous work tend
Availability to cause inconsistencies and unavailability of
resources. The mobile system must be able to handle
these cases, via replication or synchronisation
processes.
Connectivity Mobile workers do not always have access to
wireless communication support; therefore, the
mobile system should be able to handle connectivity
issues transparently.
Requirement Brief discussion
Heterogeneity Mobile workers may utilise various heterogeneous
and devices, such as laptops, PDAs and smart-phones,
Interoperability each with their own hardware features and
computing capabilities. Thus, the mobile system
should be able to allow the user to seamlessly switch
between different hardware platforms.
Communication As mobile workers tend to be on the move to carry
out their activities, some of them may not be reached
for certain periods. Therefore, the mobile system
should provide mechanisms for automatic
connection/disconnection,
synchronous/asynchronous communication and
attended/unattended message delivery.
Awareness Mobile workers should be able to be aware of
changes in the environment, be it document changes,
or new information becoming available. The mobile
system should, therefore, be able to provide
information on both offline (changes/new
information) and online (list of connection users and
their locations) changes to the mobile environment.
Protection It is important that the mobile workers’ work is safe
and secure. The system should also protect the
privacy of the mobile worker by permitting only
authorised users to participate in the group and share
information.
Brodt et al. [7] further state that the transition between
traditional ways of working and mobile or multi-locational
ways of working are affected by many factors, such as
training, leadership and autonomy, which need to be kept in
mind to facilitate the success of a project.
The notion of dividing mobile environments into layers is
discussed by Neyem et al. [6]. They indicate that mobile
environments ought to be designed on two layers, namely:
Infrastructure – The underlying infrastructure should cover
both wireless and wired communication services in order to
provide a stable platform on a mass scale. The infrastructure
will need to be technically heterogeneous, geographically
dispersed, and institutionally complex without a centralised
coordination mechanism.
Services – While the infrastructure should be heterogeneous
and stable, the services will require personalisation; be able to
be dynamically configured and modified; and combined to
meet the personal needs of the worker.
Most authors (see [8], [9] and [5]) seem to agree on several
issues, commonalities and the interrelation of various
concepts pertaining to the implementation of mobile
environments. These models were not always shown to
address the requirements from the mobile worker’s
perspective, but rather from an organisational standpoint,
focussing on aspects such as data protection and governance.
The differences between the said models could result from the
use of different research methodologies, strategies and
viewpoints by the various authors. Having compared the
results of our survey with the above-mentioned studies, it is
apparent that not all issues garner the same level of
importance, and that the viewpoints of the mobile workers
often did not form part of the said studies.
IV. RESEARCH METHODOLOGY
The literature generally deals with the implementation of a
mobile environment from the organisational standpoint, but
lack input from the actual or future users of these
environments. When the needs of the users are not included
in the design of a mobile environment, they very often
become estranged and unable to successfully and optimally
function in the said environment. In order to gather input
from the users pertaining to their experiences and their
requirements for a successful mobile environment, it is
important to gain greater insight into the daily work and
challenges of mobile workers.
In order to achieve such insight, our research was conducted
in four phases, with each phase commencing only after the
previous phase has been satisfactorily addressed. Where it
was determined that a completed phase required further
attention, the researchers returned to such phase accordingly.
The four phases of this study are:
Phase 1: Literature Review
Phase 2: Data collection
Phase 3: Data mining and analysis
Phase 4: Framework design
The first two phases of the study entailed gaining an
improved understanding of the activities and requirements of
mobile knowledge workers. This was accomplished by
administering a questionnaire, designed to extract as much
relevant information as feasible, in order to perform relevant
data analyses.
The questionnaire was built around the constructs inspired by
our research questions as well as groups of issues related to
mobile environments as identified during the literature review
phase of the study. The literature groups (constructs) are:
Services
Questions in this construct deal with the respondents’
experience regarding the services that were made available
for mobile working. The respondents were also afforded the
opportunity to comment on services they would like to have
access to.
Flexibility and Autonomy
Flexibility and Autonomy questions determine the level of
flexibility and autonomy the respondent are afforded when
performing mobile work.
Technology
Technology-related questions gauged the respondents’ access
to the technology that enables them to perform mobile work.
Apart from general technology aspects, questions pertaining
to connectivity and availability were also posed.
Awareness and Collaboration
With these questions, the goal was to:
 Determine how aware the respondents were of their team
members, and whether this was in fact a necessary
requirement; and
 Ascertain whether the respondents collaborated with their
team members, and if so, how such collaboration was
done and how often it occurred.
Privacy and Security
Questions related to privacy and security dealt the
respondents’ perceived need for privacy and security of their
data and online persona.
During the third phase, the researchers analysed the data
collected, and determined the relevant patterns that emerged
from the information provided. In the final phase, the
researchers focussed on the requirements that emerged from
the perspective of the mobile knowledge worker, and devised
the SAND framework that can be utilised by organisations
when entering into their own mobility initiatives.
The analyses of the data collected allowed the researchers
deeper insight into the daily work and challenges of mobile
workers, and highlighted the pertinent points that should be
included in the design of a framework to optimally meet the
needs of the users. Several recommendations and best
practices stemmed from this.
V. BUILDING THE SAND FRAMEWORK
Analyses of the results indicated that various challenges are
more prominent when they are addressed without adequate
support for the underlying aspects related to such challenges.
One such aspect emerging from the analyses is that mobile
environments may usefully be designed as a series of layers,
with each layer building upon previous layers.
The approach, therefore, taken by the SAND framework
allows organisations to divide issues pertaining to mobility-
enabling environments into various layers, thereby providing
them with a set of recommendations that focuses on the
different areas of mobile environments.
By dividing focus into different layers, the SAND framework
has the potential of providing an understanding of the needs
of mobile workers in a linear and logical manner, and to
highlight the variables of a successful mobile environment.
The SAND layers are underpinned by related security issues.
It is however important to recognise that the layers are
interdependent. By implication, although possible to function
modularly, the layers that make up a mobile environment
function best when addressed as a holistic whole.
The SAND layers are discussed below.
A. Devices and Hardware
One of the main challenges that mobile workers face pertains
to the physical devices and hardware that are used by them to
interact with the mobile environment. By grouping these
challenges together, the first layer of the SAND framework
came to the fore.
Naturally, the devices and hardware layer is of significant
importance, as failing to equip mobile workers with adequate
hardware and devices can have a noticeable effect on the
efficiency of interactions with the environment.
Two main issues are highlighted in this layer:
 Physical requirements of mobile devices
 BOYD (Bring your own device) strategies
The above two aspects are closely related, especially when
employees require guidance with regard to the procurement
of devices.
For mobile workers to perform their daily tasks efficiently, it
was found that four physical requirements are of importance:
 Processing power of the device
 Size of the device’s screen
 Capacity of power sources
 Portability of the device
These findings are consistent with those of [10], which claim
that these requirements are also a main concern for
manufacturers.
According to [11], perhaps the most distinguishable
characteristic of mobile devices is portability, and in order to
achieve maximum portability, manufacturers generally strive
to find a balance between device size and user expectations.
The second aspect addressed by the Device and Hardware
layer of the SAND framework pertains to BYOD (Bring Your
Own Device) strategies. Generally, BYOD does not constitute
a generic technology, but rather, is driven by a multitude of
technological trends and constitutes a business policy of
allowing employees to use personally-owned devices to
access organisational resources, such as email, intranets and
networks [12].
The results of our survey indicate that some kind of BYOD
strategy is increasingly being expected by mobile employees.
However, BYOD introduces a range of concerns to
companies, such as privacy and the safeguarding of company
resources and information [13].
In order to address concerns related to BYOD strategies, the
SAND framework proposes that focus should be placed on at
least the following aspects:
 Where feasible, allow mobile workers to use their own
devices;
 Investigate the feasibility and related risks of
implementing BYOD initiatives;
 Policies and procedures regulating acceptable use and
standards should be incorporated;
 Solutions that offer remote management of devices ought
to be considered;
 Time should be spent analysing the security of the
network, especially at the network edges, in order to
provide maximum security of network resources;
 Employees should be provided with education and
awareness pertaining to the consumption of network
resources.
B. Network connectivity
Irrespective of how functionally and technologically superior
a device or piece of hardware might be, without proper
remote connectivity technologies and platforms, users may
still struggle to connect remotely and access company
resources via extranets and/or intranets. Though the quality of
the connectivity technology might largely be out of the hands
of the organisation, several measures can be put into place to
alleviate some the concerns raised by connection quality.
The very nature of mobility increases the volatility of remote
connections [14]. The efficiency of remote connectivity
technologies is a major concern in mobile environments and
has a decidable influence on how successful a mobile worker
is at interacting with company resources.
Despite connectivity challenges, mobile workers are still
expecting an efficient mobile endeavour, high quality of
service and ubiquitous coverage [15].
Based on the issues reported in this study, connectivity issues
can be grouped into four categories:
 No connectivity: Users are unable to connect.
 Slow connectivity: Users can connect, but the data
transfer speed is experienced as relatively slow.
 Unstable/Unreliable connectivity: Users can connect,
but the connection is unstable.
 Successful connectivity: Users can connect and
authenticate, and are able to successfully interact with
resources.
The priority and strategies for addressing these issues will, in
part, determine how architects design the software for mobile
workers as well as the measures that are put in place to reduce
or mask these issues.
Network connectivity and related issues are subsequently
grouped together into the second layer of the SAND
framework.
The SAND framework ascribes to the notion that the more
autonomous a mobile system is designed to be, the better it is
able to tolerate various connectivity issues, and the less
sensitive it becomes to bandwidth fluctuations and unstable
connections.
In order to achieve a measure of success when addressing
network connectivity issues, developers should consider the
offline usage capabilities of the application, as well as data
synchronisation strategies [16].
C. Applications and Services
Addressing network connectivity issues leads to the next
layer of the SAND framework, namely Applications and
Services.
Much has been written about mobile application development
and the best practices that govern such [16], [17] and [18].
The development of applications for mobile devices
experienced an exponential growth in recent years, especially
since the opening of the iPhone AppStore in 2008 [19].
With the increasing popularity and complexity of devices, the
number of scenarios that need to be considered is also
increasing [20]. Applications capable of integrating and
interacting with organisational resources are becoming
particularly challenging, as aspects such as platform
heterogeneity, data security and privacy need to be
considered.
Modern consumers have an increasing number of devices and
platforms available to choose from, hence developing cross-
platform applications is becoming difficult.
Without an adequate cross-platform development strategy in
place, companies may face the need to develop applications
for each target platform [21], which adds to the costs of
developing and maintaining applications.
The ultimate goal of cross-platform application development
is therefore to create applications that are able to run on as
many platforms as possible while spending the least amount
of time on development and achieving optimal performance.
Another complexity related to the development of
applications and services stems from the fact that the
experience of working on a mobile device is quite different to
working on a desktop computer.
Applications for mobile devices are generally more
constrained than their computer-based counterparts with
regard to various aspects, such as smaller displays, different
ways of interacting with the application, and data
transmission rates.
Most company networks host a variety of resources that are
required by mobile workers in order to accomplish their tasks
[22]. How applications interact with these resources are an
important factor to consider when designing and choosing
applications that will be deployed to mobile employees.
Applications ought to be intelligently integrated into existing
applications and architecture.
At least two issues can be seen as critical to the success of a
mobile application, namely the availability of the application
and the ease with which mobile workers are able to interact
with and access company resources.
The recommendations suggested by the SAND framework are
that companies need to think not only in terms of platform
heterogeneity but also in terms of the network conditions and
environment that the applications will be used in. Hence, the
goal is not only to procure or develop applications that are
appropriate for the tasks to be performed, but which will also
provide the users with a satisfying user-experience. These
include the availability of the application and the ability to
easily and effectively access company resources.
D. Support
The research data has shown that without addressing the
“soft” issues surrounding mobile initiatives, the risk of
alienating employees is increased. When employees are
estranged, they tend to resist change and generally decline
buy into new strategies. Factors such as workplace inclusion,
organisational identity, employee awareness programs and
training thus present hurdles that should not be ignored [23].
Consequently, a third layer of the SAND framework is
introduced. This layer deals with the support and “soft”
issues that need to be addressed to provide an optimal
environment for mobile workers. It also addresses, to some
degree, the management of mobile workers.
Baumeister and Leary [24] argued that the need to belong is a
fundamental human motivation. The implication is that all
humans have an inherent need to belong to, and be accepted
by, a group. In an organisational context, these groups can
include their peers, their department, or their employers.
Mobile workers are not physically present in the office as
their peers are, and this may cause the workers to experience
a feeling of being “left out”. Interpersonal communication is
essential to the motivation and psyche of mobile workers
[25]. Therefore, finding effective methods of making mobile
workers feel like “part of the group” and not alienating them
from the company is an important consideration in mobile
initiatives.
By comparing the results of this study to those of [24] it may
be argued that organisational identity has ties to the
employees’ commitment to their employers. It is therefore
significant for companies to foster good relationships with
their employees and to provide ways in which employees can
build organisational identity.
Naturally, employees can only form an identity with respect
to aspects they are aware of. Consequently, the effectiveness
of communication channels between employees, their peers
and leadership directly impacts identity building [26]. The
SAND framework supports this notion, namely, companies
should set open communication policies, aimed at facilitating
two-way communication.
It can be argued that a lack of face-to-face communication is
an inherent challenge of mobile initiatives. It has furthermore
been shown by [27] that poor communication skills have an
impact on how the communicator and his/her message is
perceived. The implication is that different strategies need to
be investigated to resolve communication-related issues
amongst the mobile workforce, as proper communication
channels are crucial in co-ordinating between members [28].
A further “soft” issue surrounding mobile initiatives revolves
around the education and training of mobile workers. While
mobile workers generally experience much more freedom
than their office-bound counterparts, this perceived freedom
may have disruptive side-effects, such as anti-social
behaviour, distraction and a blurring of work/life boundaries.
Training on how to cope with the challenges presented by
mobile work was noted as beneficial to mobile workers [29].
Mobile workers generally remove intellectual resources from
company premises and it is important that these workers
appreciate the risks involved, and ultimately, be able to accept
responsibility for the company resources they access and
utilise off-premise.
It is, however, not only the mobile workers who need
educational support. Managers of teams that include mobile
workers need to learn the skills necessary to manage these
types of teams.
E. Security
Irrespective of how the previously mentioned layers are
implemented, a final, and arguably the most all-encompassing
set of issues that need to addressed deals with the security
concerns affecting the mobile environment.
Although various security strategies have been proposed [30]
and [31], these artefacts have been shown to have several
attributes in common, thus allowing the researchers to The layers as discussed above can be synthesised as in Fig. 2.
categorise the various aspects. Once these aspects have been
categorised, they can be shown to be related to the categories
(or groupings) as identified in sections 5.1 to 5.4:

Diagrammatically we have:

Devices and Network

hardware Connectivity

Applications
Support
and Services

Fig. 1: Security Layers (researchers’ synthesis)

Synthesising the foregoing discussions and Fig. 1 gives the

following:

 Devices and Hardware: Securing devices and hardware,

including the physical objects and the data stored on
them.
 Network connectivity: Securing the company network
and managing the remote connections made to it.
 Applications and Services: Ensuring that applications,
services and related resources are secure.
 Support: Enforcing security measures by governance
though policies and regulations.
In a heterogeneous mobile environment that is being accessed
by various devices and end-points, it is difficult to adopt a
one-size-fits-all approach to accommodate all the possible
scenarios [32]. Security measures employed in traditional
client-server networks have been shown to not fully address
the challenges introduced by mobile devices [33].
Furthermore, the complexity of securing a mobile
environment significantly lags behind security, for example,
stationery or always-connected desktop computers [34].
In an attempt to address the above complexity, the NIST
(American National Institute of Standards) recommends
various measures for securing a mobile environment [35].
 Implement mobile work-related policies.
 Prepare a security plan covering the deployment of items
to mobile devices and the on-going operation of such.
 Perform risk-assessments and plan for on-going risk-
management.
 Conduct regular education and awareness programs.
 Incorporate central device management and configuration
protocols.
Fig. 2: The SAND framework
VI. CONCLUSION
This research aimed to provide answers to several questions
related to the main research question, “What
framework/model can be defined to address the requirements
of mobile knowledge workers”.
This question remains particularly important to address, as the
trend for mobile work is increasing in momentum, and can
provide companies with many benefits.
However, lacking in understanding of the requirements or
hastily implementing mobility initiatives may lead to waste of
an expensive investment.
In order to address the main research question, the researchers
focussed on three sub-questions, as discussed below.
RQ1: What are the activities and tasks performed by
mobile knowledge workers?
Our research indicated that mobile workers engage with wide
variety of tasks on a daily basis. Such tasks include the
following:
 Collaboration with peers.
 Communicating with employers, peers, clients and
suppliers.
 Solving problems by interacting and researching
available online material.
 Time management, such as scheduling meetings.
 Accessing and interacting with company resources over
an extranet.
 Creating and updating documents and reports.
 Completing work and tasks that were not completed
during the day.
Arguably, the tasks that mobile workers engage in are to a
certain degree the same as what their office-bound
counterparts do, with the exception that mobile workers are
much more reliant on efficient communication because they
are not as readily present at the company premises.
RQ2: Which technologies are being utilised by mobile
knowledge workers?
Two technologies are present in almost all mobile
interactions, namely device technologies, and remote
connectivity technologies. Of the remote connectivity
technologies, ADSL and Broadband were shown to be used
most frequently.
RQ 3: What constraints are being experienced by mobile
knowledge workers?
The challenges experienced by mobile workers mostly
revolve around limitations in connectivity technologies,
hardware issues such as slow responses of devices and limited
power sources; and communication issues, such as
collaboration constraints and loneliness.
The answers to the above sub questions yielded an answer to
the main RQ:
Main RQ: What framework/model can be defined to
address the requirements of mobile knowledge workers?
From the results of the preceding, the researchers devised the
SAND framework, a layered framework that has the potential
of guiding companies through different aspects specifically
pertaining to the implementation of mobile environments.
The SAND framework consists of the following layers:
 Devices and hardware;
 Network connectivity;
 Applications and services;
 Support; and
 Security.
In order to successfully implement mobile initiatives, diligent
planning and understanding of the elements of a mobile
environment are required. Companies should not hastily
embark on mobility efforts. The goal is to find a balance
between the requirements of the users and the requirements of
the company, and to establish an environment that is not only
capable of providing optimal support, but also one that is
safe, secure, modular and scalable, and able to seamlessly
integrate with existing infrastructure.
The framework ought to provide a company with a holistic
view of a mobile environment. The layers are interdependent
and underpinned by a framework-wide security layer. By
understanding the requirements and elements depicted in each
layer, companies are able to gain a better understanding of the
requirements of an environment that will support mobile
workers and their activities.
The SAND framework is arguably flexible enough to be
applicable as a guide to a wide range of companies. The
layered approach gives companies the ability to implement
complimentary strategies that has the potential of providing
an optimal supporting environment for their mobile workers.
VII. FUTURE WORK
Several aspects pertaining to the support of mobile workers
have been addressed in this paper and directions for future
work may branch into a number of areas:
 Research into the impact of mobile work on family
relations.
 Study into the accepted responsibility of mobile workers
when they are issued with company paid connectivity
and devices – To what extent may a mobile worker abuse
these privileges?
 Research into the correlation between personality types
and mobile work – which personality types and traits are
more suited to mobile work?
 The usage of social media as a viable channel for mobile
work.
 The SAND framework was designed with larger
corporate companies in mind; however, validating the
adaptability of the SAND framework to smaller
companies, single owner, or other SMEs will be
worthwhile research.
REFERENCES
[1] Timo Botzenhardt and Key Pousttchi, "Analysing the benefits of mobile
enterprise solutions using the example of dispatching processes," 7th
International Conference on Mobile Business, pp. 260 - 268, 2008.
[2] Geleyn Meijer, Els Van de Kar, and Yan Wang, "Designing Mobile
Solutions for Mobile Workers - Lessons learned from a case study,"
ICEC 05, pp. 582 - 584, 2005.
[3] Volker Derballa and Key Pousttchi, "Extending Knowledge
Management to Mobile Workplaces," ICEC'04 Sixth International
Conference on Electronic Commerce, pp. 583 - 590, 2004.
[4] Bettina Beurer-Zeullig and Miriam Meckel, "Smartphones Enabling
Mobile Collaboration," Proceedings of the 41st Hawaii International
Conference on System Sciences, pp. 1 - 10, 2008.
[5] Torsten Brodt, Liz Carver, Terrence Fernando, Hans Schaffers, and
 Robert Slagter, Mobile Virtual Work, Professor Erik Andiressen and
Matti Professor Vartiainen, Eds.: Springer Berlin Heidelberg, 2006.
[6] Andres Neyem, Sergio Ochoa, and Jose Pino, "Designing mobile shared
workspaces for loosely coupled workgroups," CRIWG, pp. 173 - 190,
2007.
[7] Torsten Brodt, Marc Pallot, Wolfgang Prinz, and Hans Schaffers, The
Future Workspace - Perspective on mobile and collaborative working.:
Telematica Instituut (Enschede), 2006.
[8] Valeria Herskovic, Andres Neyem, Sergio Ochoa, and Jose Pino,
"General Requirements to Design Mobile Shared Workspaces," in
Computer Supported Cooperative work in design, 2008, pp. 582 - 587.
[9] Lasse Berntzen, Tore Engvig, Per Hasvold, and Bente Skattor, "A
framework for mobile services supporting mobile non-office workers,"
Human Computer interaction, vol. 4, pp. 742 - 751, 2007.
[10] Anar Gasimov, Chee Wei Phan, Juliana Sutanto, and Chuan-Hoo Tan,
"Visiting mobile application development: What, how and where,"
Ninth international conference on mobile business, pp. 74 - 78, 2010.
[11] Yufei Yuan and Wuping Zheng, "From stationery work support to
mobile work support: A theoretical framework," Proceedings of the
International conference on mobile business, pp. 1 - 7, 2005.
[12] Louis Jonker, "Brnig your own device: Policies and contracts,"
ITechLaw European Conference, pp. 3 - 20, 2012.
[13] Rupali Gurudatt, Sangita Mohite, Rajnikant Shelke, and Shirkant
Belsare Vikas Solanke, "Mobile Cloud Computing - Bring your own
device," Fourt International Conference on Communication Systems
and Network Technologies, p. 566, 2014.
[14] Chevonne Dancer, Jacqueline Jackson, and Gordon Skelton, "Teaching
software engineering through the use of mobile application
development," Journal of Computing Sciences in Colleges, vol. 28, no.
5, pp. 39 - 40, 2013.
[15] Tolga Ayhan, Leonid Kazovsky, and Shing-wa Wong, "Hybrid optical-
wireless access networks," Proceedings of the IEEE, vol. 100, no. 5, pp.
1197 - 1225, 2012.
[16] Sybase, "Best practices for developing mobile applications: The data
driven approach," 2007.
[17] Motorola, "Best practices for developing enterprise smartphone apps,"
2012.
[18] AQuA, "Best practice guidlines for producing high quality mobile
applications," 2013.
[19] Anthony Wasserman, "Software Engineering Issues for Mobile
Application development," FoSER, pp. 397 - 400, 2010.
[20] Stefan Diewald, Matthias Kranz, Andreas Moller, and Luis Roalter,
"Towards a holistic approach for mobile application development in
intelligent environments," MUM, p. 73, 2011.
[21] Spyros Xanthopoulos and Stelios Xinogalos, "A comparative analysis of
cross-platform development approaches for mobile applications,"
Proceedings of the 6th Balkan conference in informatics, pp. 213 - 219,
2013.
[22] Marco Casole and Yi Cheng, "Secure access to corporate resources in a
multi-access perspective: Needs, problems, and Solutions," The
Institution of Electrical Engineers, pp. 482 - 487, 2003.
[23] Debra Major, Valerie Morganson, and Kurt Oborn, "Comparing
telework locations and traditional work arrangements," Journal of
Managerial Psychology, vol. 25, no. 6, p. 583, 2010.
[24] Roy Baumeister and Mark Leary, "The need to belong: Desire for
interpersonal attachments as a fundamental human motivation,"
Psychological Bulletin, vol. 117, no. 3, pp. 497 - 529, 1995.
[25] Tomonori Hashiyama, Junko Ichino, Yasuhiro Kojima, and Shun'ichi
Tano, "Casual Multi-user Web Collaboration by Lowering
Communication Barriers," , Tokyo, 2010.
[26] Jos Bartels, Menno De Jong, Oscar Peters, and Ad Pruyn, "Horizontal
and vertical communication as determinants of professional and
organisational identification," Personnel Review, vol. 39, no. 2, pp. 210
- 226, 2009.
[27] Jonathan Gardner and Lamar Reinsch, "Do communication abilities
affect promotion decisions? Some data from the C-suite," Journal of
business and technical communication, vol. 28, no. 1, pp. 31 - 57, 2013.
[28] E Haribabu, C Lakshmi, and Avvari Mohan, "Intra-organizational
communication and aspects of organisation culture: A workplace study
among knowledge worker teams of two IT firms in Hyderabad,"
Management of Innovation and Technology, vol. 1, pp. 131 - 135, 2006.
[29] IBM, "The mobile working experience," p. 12, 2005.
[30] Intel, "Enterprise Mobility: Increasing productivity for mobile users,"
2013.
[31] Al Salam and Rahul Singh, "Semantic information assurance for secure
distributed knowledge manaegment: A business process perspective,"
IEEE Transactions on Systems, Man and Cybernetics, vol. 36, no. 3, pp.
472 - 474, 2006.
[32] Vipin Chaudhary, Hanping Lufei, and Weisong Shi, "Adaptive secure
access to remote services in mobile environments," IEEE transactions
on services computing, vol. 1, no. 1, p. 49, 2008.
[33] Lu Ma and Jeffrey Tsai, "Formal modeling and analysis of a secure
mobile-agent system," IEEE Transactions of systems, man and
cybernetics, vol. 38, no. 1, p. 180, 2008.
[34] Max Landman, "Managing smart phone security risks," InfoSec '10, pp.
145 - 154, 2010.
[35] NIST, "Improving critical infrastructure cybersecurity," 2013. [Online].
https://www.huntonprivacyblog.com/wp-
content/files/2013/10/preliminary-cybersecurity-framework.pdf
[36] Sybase, "Best Practices for Developing mobile applications: The data
driven approach," iAnywhere Solutions, 2007.
[37] EVD Kar, SM Muniafu, Y Wang, "Mobile services in unstable
environments: Design requiremetns based on three case studies",
Proceedings of ICEC, pp 302 - 308, 2006.
 Characteristics of Transfer Function of Power Lines
having Tapped Branches in the Propagation Path
Banty Tiru
Department of Physics
Gauhati University
Guwahati-14, India
banty_tiru@rediffmail.com

Abstract— Power lines can be used for communication In practical networks, complex branches like STAR
purpose. However, the channel is highly frequency selective
arising from discontinuities in the network. In this work, the
effect of tapped branches in different topologies (STAR and BUS)
is studied. The frequency selectiveness is measured in terms of the
Coherence Bandwidth. It is found that the Coherence Bandwidth
is more for STAR than a BUS one. In STAR topology, the same
at first decreases and then increases for increasing number of
taps at the nodes. For large number of components, the transfer
function of STAR branch is governed by the characteristics of the
connecting cable while for in BUS topology, the same becomes
very complex. The paper gives an idea of the dependency of the
efficiency of the communication system on branches with
different topologies encountered.
Fig. 1. Different types of branches CD in the propagation path AB. (a)
Keywords—Power Line Communication, Complex Simple branch (b) Branch with a STAR (c) BUS topology.
Trigonometric Hyperbolic Functions, ABCD matrices,
Transmission Line, Star Topology, Bus Topology. and BUS topologies are found. In the former, there are large
numbers of branches in a single node. In the later, the same are
connected in a serial manner. This paper studies the effect of
I. INTRODUCTION different topologies on the communication path. The frequency
selectiveness is measured in terms of the coherence bandwidth
P OWER line communication (PLC) uses the unambiguous
power line (PL) to meet the data requirements of access,
in-house and control [1]. PLC is also used as components of
(CBW) defined as the range of frequencies over which the
channel can approximately be considered as flat. The study
hybrid networks incorporating both cabled and wireless will enable to predict efficiency of PL in a practical network.
channels [2]. PLs are however the worst to deal with The PL channels are assumed to have a linear time invariant
characterized with time varying frequency selective transfer behavior modeled as two port network (2PN). Section II of the
paper describes the use of ABCD matrices in the evaluation of
function (TR). An in-depth knowledge is therefore required to
the TR and its application to PLs. In Section III, effect of
predict the efficiency of the devices.
different types of branches on the TR is studied. A
The frequency selectiveness of PL channels depicts the
comparative study of the CBW is done in Section IV. The
multipath nature arising from the many discontinuities in the paper is concluded with the key results.
communication path. Notches occur when the communication
signal reach the receiver via different paths having different
propagation delays and at opposite phase. The worst cases are II. TRANSMISSION MATRIX OF A TWO PORT NETWORK AND
when the signal encounters open or short circuited branches. APPLICABILITY TO POWER LINE CHANNEL MODELING
The number of notches increases with the length of the
The transmission, chain or ABCD matrix of a 2PN (Fig. 2),
branches and
is defined in terms of the input voltage v1  , input current i1  ,
are periodic in nature. Many researchers have analyzed the output voltage v2  and output current i2  as in (1)
problem of notch formation and the processes are at present
understood clearly. The dependencies on channel variability
like length of branch, loads etc [3]-[10] has been studied v1   A B  v 2 
elaborately. However, most of the papers target simple i       
branches in the propagation path like that shown in Fig.1 (a).  1  C D   i 2 

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

Where
v1 v1 i1 i1
A= B= C= D=
v2 i 20
 i2 v 20
v2 i 20
 i2 v 20

Knowledge of the matrices enables the transfer function to be

obtained as given by (2) [5]-[6].

ZL
 H  f  
Fig. 2. ABCD matrix of a 2 Port Network

AZ L  B  CZ S Z L  DZ S (a) (b)

8
12
10 7
Here, Z s and Z L are the source and load impedances 10
6
8

abs(cothz))

abs(tanhz)
respectively. ABCD matrix of complex networks is equal to 8 5
6
the product of the matrices of the cascaded sections. A PL 6 4
network can be taken as a 2PN with cascaded sections of 4 3
4
transmission line (TL) and branches taps. For a PL of length l, 2
2 2
2
characteristics impedance Z 0 and propagation constant γ, the 0
2 0
0
1
0 10
10 0
ABCD matrix of a simple TL and branched tap is given by (3) 0
y -10 -2 x
y -10 -2 x

cosh l Z 0 sinh l   1 0 Fig. 3. Plot of absolute value of complex hyperbolic (a) cotangent (b) tangent
ABCDTL   sinh l ; ABCD   1  (3) function showing periodicity along the imaginary axis.
 cosh l  Tap
Z 1
 Z0   in 

R  jwL
Where Z in is the input impedance of the branch of length lbr Z0  ;  R  jwLG  jwC    j (7)
G  jwC
and terminated by a load Zbr . Z in is defined by
α and β are the termed as the attenuation constant (nephers/m)
and phase constant (radian/m) of the cables respectively. Both
Zbr  Z 0 tanh(lbr )
Zin  Z 0 (4) α and β are frequency dependent terms and given by (8)[11]
Z 0  Zbr tanh(lbr )
12
  R 2  G 2 
12
RG 
 1  2 2 1  2 2   1  2
w LC (8)
For open and short circuited branches, the input impedance ,  
(and input admittance) is obtained by substituting adequate 2   w L  w C   w LC 

loads limits in (4) yielding (5) and (6)

tanhlbr 
In the simulation, the constants are taken from a real network
Zinopen  Zin  Z 0 cothlbr ; Yinopen  (5) measurement and noted as L=0.69µH/m, G=0.018µmho/m
Z br  Z0 (Seimen/m) and C=38pF/m. For high frequencies, Z0 ~135Ω.
cothlbr 
Zinshort  Zin  Z 0 tanhlbr ; Yinshort  (6)
Z br 0 Z0 III. EFFECT OF DIFFERENT BRANCHES IN THE COMMUNICATION
PATH

The periodicity of the notches occurs due to the periodicity of A communication path from A to B encounters different
the hyperbolic tangent and cotangent functions branches in the propagation path like those shown in Fig. 1. A
(Fig. 3). Notches are found whenever the input admittance PL network can consist of either of the configurations or a
(impedance) becomes a maximum (minimum) so that most of hybrid system one consisting of both.
the signal is shorted out from the load. Z0 and γ depends on A. Simple Branch in the Propagating Path.
the primary line constants of the cable termed as R
(Resistance/length), L(inductance/length), G (conductance In a simple branch CD in the propagating path AB, as the
/length) and C (capacitance /length). At angular frequency w length of the branch is increased, the number of admittance
the same is given by peaks and hence, the number of notches also increases. For
lbr  4.55m , Yopen has a maximum at 10.7MHz and 32.2MHz And the input admittance at node „N2‟ is
and minimum at 21.5MHz and 42.9MHz.
1  Z 0Ytotal  tanh(l ) 
YIN2 _ STAR   
0 0.5  1  Z Y tanh(l ) 
Z0  0 total  (14)

Input Admittance |Yopen|S

Transfer Function (dB)

Yin1  Yin2  Yopen

YIN2 _ STAR 
1  Z 02 Yin1  Yin2 Yopen
(15)

Yopen is the admittance of the branch between (N1,N2) if open.

-50 0 For n number of branches at „N1‟, (15) can be written as
0 25 50 0 25 50
Frequency (MHz) Frequency (MHz)
n

Fig. 4. Magnitude of transfer function of AB with (a) and (c) different branch Y n  Yopen
Where Y1  Yin1 , Y2  Yin2 ..(16)
length and (b) and (d) variation of Z in and Yin for the two lengths YINn _ STAR  n 1

 n 
1  Z 02   Yn Yopen
 n 1 

Analysis. In order to study the effect of STAR topology, a two

branch circuit like that of Fig. 5(a) is considered. For
simplicity Z L1 and Z L 2 are taken to be resistive loads whose
values are very high and the lengths of the branches are taken
to be equal given by l  l br1  l br 2  4.55m . The variation of
YIN2 _ STAR is shown in Fig. 6 and can be analyzed using (17)

1  ZtotalYopen n
YIN2 _ STAR   (17)
Fig. 5. STAR configuration Ztotal  Z 02Yopen d

B. Branch having a STAR Connection at the end

YIN2 _ STAR is highly dependent on Yopen . As in Fig.7, when the
The simplest case of a STAR connection as in Fig. 5(a)
consists of a node „N1‟ having two branches ( lbr1 and lbr 2 ) branches in „N1‟ is increased the number of peaks remains the
same if the lengths of all are equal but increases for increasing
connected to „N2‟ through length l loaded with Z L1 and Z L 2 lengths. Differing lengths increase the peaks further. When the
respectively. The input impedance (admittance) at „N1‟ are lengths are selected using a randomly, the peaks concentrate
near the minimum of Yopen . The analysis shows that for

Z L1  Z 0 tanh(lbr1 ) Z  Z 0 tanh(lbr 2 ) (10) branches with STAR topology, Yin depends more on the
Z in1  Z 0 ; Z in 2  Z 0 L 2
Z 0  Z L1 tanh(lbr1 ) Z 0  Z L 2 tanh(lbr 2 ) lengths of the branches in the node, but for large numbers
whose lengths vary randomly, on the connecting branch to the
1  Z0Ybr1  tanh(lbr1 )  1  Z Y  tanh(lbr 2 )  (11) node. More analysis is done in Section IV.
Yin1   ; Yin 2   0 br 2 
Z0  1  Z0Ybr1 tanh(lbr1 )  Z0  1  Z0Ybr 2 tanh(lbr 2 ) 

Where
1 1
Ybr1  ; Ybr2 
Z br1 Z br2 (12)

The total admittance and impedance in the node „N1‟ is

1
Ytotal  Yin1  Yin2 ; Z total  (13)
Ytotal
 tanhl1 
|

0.5 1
IN-STAR

1 1
Y1  ; y1  ; Yopen1 

|Ytotal|
(e) (18)
(a) 0.5 Z in1 Z in 0 Z0
|Y2

0 0
0 25 50 0 25 50
0.5 4K Yopen1 is the input admittance of the branch of length l1 if it
|Yopen|

|Ztotal|
(f)
(b) were open. The input admittance at „N2‟ is given by
0
0 25 50 0 0 25 50
0.5 10K
|Yin1|

(c) (g)
|d|

0 0 0
0 25 50 25 50
0.5 0.5
|Yin2|

|n/d|

(d) (h)

0 0
0 25 50 0 25 50
Frequency(MHz) Frequency(MHz)

Fig. 6. Analysis of a STAR configuration with two branches using the

impedance of the tapped branch YIN2 _ STAR , K=103. Y axis in Seimen.

0.5 0.5
|

|Y8IN-STAR|
IN-STAR

(a) (b) Fig. 8. BUS connection

|Y3

0 0 Y1  y1  Yopen1
0 50 0 25 50 YIN1 _ BUS  (19)
0.5 0.5 1  Z 02 Y1  y1 Yopen1
|

|Y3IN-STAR|

(c)
IN-STAR

(d)
1
Where the subscript „1‟ in YIN _ BUS denotes that there is only
|Y3

0 0 one node at the branch. If „N2‟ is replaced with another node

0 25 50 0 25 50
with a BUS connection as shown in the Fig. 8(b), then the
0.5 0.5
input admittance at node „N3‟ is given by
|

(e)
|Y3IN-STAR|
IN-STAR

(f)

Y2  y2  Yopen2
|Y3

YIN2 _ BUS  (20)

0
0 25 50
0
0 50 1  Z 02 Y2  y2 Yopen2

Where
Fig. 7. The input impedance YINn _ STAR of a branch having a STAR 1 tanhl2 
y2  YIN1 _ BUS ; Y2  ; Yopen2  (21)
connection having (a) 3 branches of equal lengths l  lbrn  4.55m,n=1,2,3 Zin 2 Z0
(b) 8 branches of equal lengths l  lbrn  4.55m ,n=1to 8 (c) 3 branches of
equal lengths l  lbrn  8.55m ,n=1,2,3 (d) 3 branches l  lbrn , In general,
l=8.55m, lbrn  4.55m, n=1,2,3 (e)3 branches, l  lbr1  lbr2  lbr3 (f) Yi 1  yi 1  Yopen_ i 1
yi  (22)
8 branches, l, lbrn  random (0-7m) , n=1 to 8. Y axis in Seimen. 1  Z 02 Yn 1  yn 1 Yopen_ i 1

C. Branch having BUS Connection

A BUS topology with two branches, as shown in the Fig. 8(a), Yi 1  yi 1  Yopen_ i 1
yi  (23)
1  Z 02 Yn 1  yn 1 Yopen_ i 1
is the same like that of a STAR connection with two branches
given in Fig. 5(a). Here Y1 and y1 are the input admittance of
the two branches loaded with loads Z L1 and Z L0 For n number of such BUS connections as shown in Fig. 8(c),
respectively. the input admittance at node „Nn‟ is given by
 |ZIN-BUS|
5K 10k
Yn 1  yn 1  Yopen_ n 1

|d1|
.Y n
 (24)

2
1  Z02 Yn 1  yn 1 Yopen_ n 1
IN _ bus 0 0
0 25 50 0 25 50
0.5 0.5

|Yopen1|

|y2|
0 0
Analysis. To study the effect of increasing the nodes, the 0 25 50 0 25 50
10K 0.5

|Ytotal2|
simplest form of circuit like Fig. 8(a) is taken and the input

|Zin1|
1
admittance at „N2‟ i.e YIN _ BUS is found out. In the next stage 0
0 25 50
0
0 25 50
0.5 2K

|Ztotal2|
this acts as the input admittance of the branch connection at

|Y1|
„N2‟ and the input admittance at „N3‟ is found out. Likewise 0
0 25 50
0
0 25 50

|
the number of nodes is increased and the admittance at the

total1
1 10K

|d2|
main line found out in each case. To make the analysis simple,

|Y
0 0
the lengths of all the cables taken to be equal 5K
0 25 50
0.5
0 25 50

|Ztotal1|
l1  l2  ...... lbr1  lbr 2  ........ l0  4.55m . As before, the

|y3|
loads are taken to be resistive of very high values. As noted 0 0
0 25 50 0 25 50
Frequency(MHz) Frequency(MHz)
earlier, the length has a Yopen1 maxima at 10.7MHz and
32.2MHz and minimum at 21.5MHz and 42.9MHz. Similarly
Y1 and y1 has maxima and minima at these frequencies. Fig.9 Fig. 10. Analysis of the input impedance at „N3‟ for YIN2 _ BUS .Here K=103 ,

shows YINn _ BUS for various n . In Fig. 10, the formation of Z in1  Z in2  Z in0 , Y1  Y2  y1 , y2  YIN1 _ BUS and

peaks is analyzed for YIN2 _ BUS
APPENDIX A. The Figure shows that the peaks are much
dependent on the admittance of open circuits and loaded
branches. In each addition of more BUS nodes, the admittance
of the open circuit adds to the net impedance forming the load
for the next stage.
0.5
|YIN-BUS|
using the calculations in y3  YIN2 _ BUS .The details are given in APPENDIX A . Y axis in Seimen.
Fig. 9 shows that as the number of BUS connections increased,
the admittance in the node „Nn‟ shows more number of peaks
with decreasing amplitude. Therefore the presence of a BUS
connection in the main propagating path offers a complicated
0.5
transfer function if the number of elements are increased.

|Y2IN-BUS|

IV. COHERENCE BANDWIDTH OF CHANNELS WITH TAPPED

1

0 0 BRIDGES
0 25 50 0 25 50
0.5 0.5
|Y4IN-BUS|

CBW of a channel is defined as the interval of frequencies at

|
IN-BUS

which the normalized autocorrelation function of the TR is

|Y3

0 0
0 25 50 0 25 50
|Y6IN-BUS|

0.5 0.5

TABLE I
|YIN-BUS|

COHERENCE BANDWIDTH OF DIFFERENT TOPOLOGIES

0 0
5

0 25 50 0 25 50 Coherence
0.2 0.2 No. T N Specification Bandwidth
|Y8IN-BUS|

(MHz)
|

l  lbrn  4.55m
IN-BUS

(i) star 2 0.4

(ii) -do- 3 -do- 0.4
|Y7

0 0
0 25 50 0 25 50 (iii) -do- 8 -do- 0.3
Frequency(MHz) Frequency(MHz)
(iv) -do- 10 -do- 0.3
(v) -do- 20 -do- 0.25
Fig. 9. The input admittance at node „Nn‟ for branches in BUS configuration (vi) -do- 50 -do- 0.45
different having different numbers of nodes n. Y axis in Seimen. (vii) -do- 100 -do- 0.85
(viii) -do- 2 l  lbrn  8.55m 0.3
(ix) -do- 3 l  lbrn  8.55m 0.25
(x) -do- 3 l  lbrn 0.40
(xi) -do- 8 l , lbrn  random 0.33
(xii) -do- 40 l , lbrn  random 0.65
(xii) bus 1 4.55m 0.4
(xiii) -do- 2 -do- 0.3
(xiv) -do- 3 -do- 0.2
(xv) -do- 4 -do- 0.2
(xvi) -do- 5 -do- 0.2
(xvii) -do- 6 -do- 0.2
(xviii) -do- 7 -do- 0.2
(xix) -do- 8 -do- 0.2

T: Topology, N: Number of branches in STAR topology/ Nodes

in BUS topology; In BUS topology, all the lengths are 4.55m
 Yopen1
1
Ytotal1 1
higher than a certain value (usually 0.9). The autocorrelation of YIN1 _ BUS  ; Ytotal1  Y1  y1 
1 Z open1
the transfer function is given by (25) [4]  Z 02Yopen1
Ytotal1
Rt , f    H t , f H t , f  f  1  Yopen1Z total1
M
(25) 1 n
_ BUS    1  y2
1
t 0 YIN
Z total1  Z 02Yopen1 Z open1 d1
The CBW of the different channels are given in Table I. It is
seen that for STAR topology, the CBW at first decreases and 1. On addition of another bus as shown in Fig. XX,
then increases as the number of branches at the node is 1 tanh l2 
increased. When the lengths of the branches are increased, then y2  YIN1 _ BUS ; Y2  ; Yopen2 
Z in2 Z0
the CBW decreases due to incorporation of more number of
notches. For BUS connected branch, the CBW decreases as Where
more number of components are included. In general the Z L 2  Z 0 tanh(l2 )
Z in2  Z 0
STAR connection shows a greater CBW than the BUS Z 0  Z L 2 tanh(l2 )
connection.

V. CONCLUSION Then
Y2  y 2  Yopen2
_ bus 
2
In this work, the dependence of transfer functions on the YIN
1  Z 02Yopen2 Y2  y 2 
branched taps in the propagation path for STAR and BUS
topologies are studied. It is seen that the effect on the transfer
function increases when the number of branches in the STAR Yopen2
1
node or number of serial connections in the BUS configuration Ytotal2 1
_ BUS  ; Ytotal2  Y2  y 2 
2
YIN
increases seen from a decrease in the Coherence Bandwidth. 1
 Z 02Yopen2 Z open2
However for STAR topologies, the CBW improves for larger Ytotal2
number of branches. In general, the CBW of STAR topology is
larger than the BUS topology. The results of the work can be 1  Yopen2 Z total2 1 n
used to find out the efficiency of various PLC channels YIN2 _ BUS    2  y3
Z total2  Z 02Yopen2 2
Z IN _ BUS d 2
connecting devices in a complex network. In the future the
procedure will be applied to developing a model for predicting
efficiencies of PLC in more complex topologies. Such a study
REFERENCES
will enhance the use of power line and means can be found out
[1] P. Sutterlin, and W. Downey, “A Power Line Communication Tutorial-
to make the channel more compatible for communication. Challenges and Technologies,” in Proc. ISPLC., Japan, 1998, Available:
http://www.viste.com/LON/tools/PowerLine/pwrlinetutoral.pdf
[2] S. Barmada, M. Raugi, M. Tucci, “Power line communication integrated
in a wireless Power Transfer System : A feasibility study”, in Proc. 18th
APPENDIX
ISPLC, 2014, Glasgow, pp. 116-120.
[3] E. Biglieri, “Coding a Horrible Channel,” IEEE Comm. Mag. vol. 41, no.
1. Taking the first circuit as shown in Fig. 8(a), 5, pp. 92-98, May. 2003.
tanh l1 
[4] P. Mlynek, J. Misurec, M. Koutny, “ Random Channel Generator for
1 1
Y1  ; y1  ; Yopen1  Indoor Power Line Communication”, Meas. Sc. Rev. vol. 13, no. 4, pp.
Z in1 Z in0 Z0 206-212, 2013.
[5] F.Zwane, T.J.O. Afullo, “An Alternative Approach in Power Line
where Communication Channel Modelling”, Prog. Elect. Res. Vol 47, pp 85-
Z L 0  Z 0 tanh(l0 ) 93, 2014.
Z in0  Z 0 [6] B. Tiru, R. Baishya and U. Sarma, “An analysis of indoor power line as a
Z 0  Z L1 tanh(l0 )
communication media using ABCD matrices”, in Advances in
Z L1  Z 0 tanh(l L1 )
Z inl  Z 0 Communication and Computing , Springer, 2015, pp. 171-181.
Z 0  Z L1 tanh(l L1 ) [7] I. Tsokalo, R.Lehnert,”Modelling Approach of broadband in-home PLC
in network simulator 3”, in Proc.19th ISPLC, Austin, 2015, pp. 113-118.
Then
[8] J. Anatory, N. Theethayi, M. Kissaka, N. Mvungi, “Broadband Power
Y1  y1  Yopen1 Line Communications: The factors influencing Wave propagations in
YIN1 _ BUS 
1  Z 02Yopen1 Y1  y1 
Medium Voltage Lines”, in Proc. ISPLC., Pisa, 2007, pp. 127-132.
[9] S. Tsuzuki, S. Yamamoto, T. Takamatsu, Y. Yamada. (2001, Apr.)
Measurement of Japanese Indoor Power Line Channels, Presented at
International Symposium on Power Line Communications. [Online].
Available:
http://www.isplc.org/docsearch/Proceedings/2001/pdf/0687_001.pdf
[10] T.P. Surekha, T. Ananthapadmanabha, C. Puttamadappa, “Analysis of
Effect of Power Line Channel Characteristics Parameters in Broadband
 Power Line Communications (BPLC) systems”, in Proc. IEEE/PES
PSCE., Seattle, 2009, pp. 1-6.
[11] W. Frazer, “Transmission Lines ,” in Telecommunications, 2nd ed., New
Delhi: CBS Publishers and Distributers, 1985, pp. 97.

B. Tiru received her Ph.D. degree

from the Department of
Instrumentation and USIC, Gauhati
University, India in 2011. She is
presently an Associate Professor in the
Department of Physics, Gauhati
University. Her research interest is
hybrid communication systems and
cosmic ray detection. She is guiding
many research students.
 IDPS based Framework for Security in Green Cloud Computing and
Comprehensive Review on Existing Frameworks and Security Issues
Komal Singh Gill, Anju Sharma
CSED, Thapar University, Patiala
komalgillz@gmail.com, anjusharma@thapar.edu
Abstract— IT sector is growing day by day, its data and
energy costs are also increasing, which needs to access the high-
end computing capabilities. To fulfill these requirements, there is
a need to shift from traditional computing practices to new
computing practices that give access to a broader network,
unlimited resources, and enables self-service on demand at a
reasonable cost using pay per use method. All these requirements
can be fulfilled with cloud computing. Cloud Computing has a
number of advantages as well as disadvantages. Security is a
major threat to the cloud computing because the data placed on
cloud servers is always vulnerable as there are lots of attackers
on the internet who tries to compromise the security of the cloud.
To overcome this issue, we use security appliances in the form of
physical machines or by virtualization. A significant amount of
power is consumed by security appliances which not only leads to
the high energy costs but also a major threat to the environment
as carbon-dioxide is released. So there is a need to find optimal
solutions for providing security in cloud computing environment
in an efficient manner. In this paper, a comprehensive survey on
existing cloud security frameworks has been done. Based upon
the limitations on existing frameworks a new framework has
been proposed to provide security in virtual networks that is
based on Intrusion Detection and Prevention System (IDPS) and
its prototype implementation has also been done.
Keywords— security; green cloud computing; IDPS
I. INTRODUCTION
Cloud computing is the most trending technology among IT
sector because of its characteristics like rapid elasticity, broad
network access, measured services, on-demand self-service and
resource pooling [1]. It has given relief to the user by
eliminating the need for the high-cost hardware requirements at
the client side. Many organizations have shifted their data to
the cloud which saves costs and provides ubiquitous and fast
access to the data. As the database increases, it needs more
servers which lead to an increase in the size of data center.
These large data-centers cause two major issues: the first is
maintaining security and the other is energy consumption.
Implementing the security policies which include
confidentiality, integrity and availability are the biggest
challenges in the cloud industry. The recent attacks on Apple
and Sony cloud have created a fear in the mind of customers
about the safety of their data. The security appliances that are
used to prevent attacks and to maintain a safe environment
consumes a significant amount of energy. So to lower the
energy costs and the carbon contents, we need green cloud
environment. Green cloud computing is the necessity of the
environment as global warming is increasing. Green cloud
computing aims to minimize the power consumption costs and
emissions of carbon-dioxide thus making cloud computing
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
more efficient and eco-friendly. So making the cloud green
does not only benefits the cloud providers by saving the costs
but also helps in making environment sustainable for the
coming generations. This work leads to minimize the power
usage by considering virtualization at various levels [10].
Virtualization means running two or more operating systems
on a single physical machine and thus abstract the computer
resources. In this work, an IDPS based framework have been
proposed and its prototype implementation is given. IDPS uses
the concept of autonomic computing to automatically detect a
malicious activity, attempts to stop it and reports the security
Admin.In next section we discussed various papers that are
based on security issues in cloud computing and a survey has
been done on the existing frameworks for cloud security.
II. LITERATURE SURVEY
This section is divided into three parts. Part A discussed the
security issues survey including confidentiality, availability,
integrity etc. Part B discussed the security frameworks for
clouds as well as for green clouds. Part C discussed the
limitations of existing frameworks and a comprehensive
review on existing frameworks has been shown in Table1.
A. Survey on Security issues
S. Subashini et al. [2] surveyed various risks and
security threats existent in cloud computing system at the
service delivery model level (SaaS, IaaS and PaaS). Security
issues in SaaS that are main challenges in the cloud
environment are discussed one by one. Though cloud
computing is very advantageous technology in IT sector but
some issues like SLA agreements, security, privacy, power
efficiency still exists which are a bottleneck for the potential
customers. They suggested that the integrated solution must be
designed by analyzing macro and micro elements and then
deploy them in the cloud.
Dimitrios Zissis et al. [3] discussed the transition of
traditional computing to grid computing and now to cloud
computing. They worked on two challenges: Firstly,
identifying unique security requirements to evaluate cloud
security and secondly, to propose a solution that eliminates
these types of risks. The proposed solution based on
cryptographic techniques assures authentication, integrity and
confidentiality of data communications. The security
requirements, threats and the users of service levels have been
discussed and the categorization of threats has been done. A
discussion has also been done on the trusted third party
describing on whom it can rely, the creation of security
domains and cryptographic separation of data. The assessment
has been done by proposing a new solution based on trust that
operates in a top-down fashion.
Chirag Modi et al. [4] surveyed different intrusions in
cloud resources and services that affect availability, integrity
and confidentiality. A review of IDS/IPS types and techniques
has been done and their characteristics, pros and cons are also
discussed. They emphasized on the positioning of IDS/IPS in
the cloud environment to get the anticipated security level.
Bhavna Makhija et al. [5] discussed trust issues and
proposed a solution by a Third Party Auditor (TPA). TPA
audits the data of client and verifies the integrity of the data,
thus making the task easy for the client. But the trust of the
TPA has not been calculated yet. If TPA changes or eliminates
some important data, it will be a great problem for the user
and the provider. To deal with this, an algorithm based on data
encryption, message digest and RSA are used to ensure data
integrity. A solution has also been proposed to this problem by
using sentinels in which only a small portion of the whole file
is accessed. This method increases the storage overhead on the
server. TPA has its merits and demerits but it does not help in
green cloud computing, hence it increases the overhead.
Maneesha Sharma et al. [6] did a review on
characteristics of service models. This paper mostly focused
on types of clouds, characteristics of service models and the
security threats faced by the service and deployment models.
It described the way to propose and design the solutions for
these threats.
Ramgovind S et al. [7] highlighted the main security
concerns and discussed their presence in cloud delivery and
deployment models. The information security requirements
like authorization, non-repudiation have been discussed. Then
cloud shortfalls are discussed and security has been termed as
the biggest challenge. They proposed the solution to cloud
security by giving strategies like cloud governance, cloud
transparency and cloud computing’s security impact.
Jian Wang et al. [8] proposed a solution to provide
privacy in cloud computing environment. The solution is in
the form of algorithm namely Anonymity-based method which
processes the anonymous data and sends it to service providers
of the cloud. This algorithm is more safe and flexible as
compared to the traditional cryptography approach.
Aman Bakshi et al. [9] proposed an idea of using
Intrusion Detection System (IDS) to secure cloud from DDOS
attacks. They discussed the concept of virtualization and its
benefits. They installed the IDS on the virtual switch and it
examine the packets in real-time according to the pre-set rules.
The IDS determines which type of attack is occurring and
alerts the virtual server that how much risk is involved in this
attack. The proposed system has been implemented by causing
DDOS attack using the Zombie machines. The result shows
that the IDS drops all the packets from that particular IP
address.
B. Survey on existing frameworks
In this section, we have discussed the recently proposed
frameworks for cloud security.
Jianxin Li et al. [10] analyzed the key security issues
in green cloud computing environment and proposed the
solution by virtualization security assurance architecture
named CyberGuarder. CyberGuarder considers three
parameters namely virtual machine (VM) security, virtual
network security and policy-based trust management. The
results for the virtual machine shows that CyberGuarder
VMinsight successfully identifies processes, detects traffic and
monitors CPU usage and file operations thus helps in
preventing malicious activities. In a Virtual network,
CyberGuarder vIDS evaluates the effectiveness of the
dynamic approach. It accumulates maximum packet drop rate
of only 2% and quickly reduces the network workload. An
architecture has been proposed which follow steps to enforce
trust policy. The results show that it makes the system
trustworthy but increase the traffic.
Hassan Takabi et al. [11] proposed a security
framework in which service, trust, security and heterogeneity
management components are considered. These components
maintain trust between the providers and the users.
Ayesha Malik et al. [12] did a review on security
issues and proposed a solution for the security threats,
considered different parameters like fault free API’s,
encryption process, data protection analysis and provider’s
backup.
Mohemed Almorsy et al. [13] introduced a new
security framework which provides better collaboration
between service consumers, cloud providers and service
providers in the management of security. This framework has
been implemented on SaaS model successfully.
Ahmed E. Youssef et al. [14] proposed a framework
for identifying security and privacy challenges (threats,
attacks, risks) in the cloud environment. They considered only
four components namely Verification and validation unit, Data
Protection unit, Attacks Detection unit and Privilege Control
unit.
Prasad Saripalli et al. [15] came up with a framework
to assess the risks in cloud security. The framework has been
named as QUIRC which combines the probability and severity
of a threat to define risk. They identified six security
objectives which could be responsible for the malicious
events.
Ryan K L Ko et al. [16] proposed trust and
accountability framework which uses policy-based and
technical approaches to address accountability. They also
discussed the main challenges faced by the cloud to achieve a
trusted environment.
ENISA [17] proposed a security framework for
Governmental clouds. They used the Plan, Do, Check and Act
(PDCA) security model to deploy the governmental cloud.
Xuan Zhang et al. [18] designed a framework to mitigate
security risks and lower the rate of threat and vulnerabilities in
cloud computing. This framework covers all deployment and
service models of the cloud.
 e

S
F
S
S
F
S
P

E
T
E
E
L
E

R
R
C
R
C

K
O
A
Y
U
N
O
Y
A
N
A
V
A
A
O

G
1.
bl

M
M

SI

W
TI
TI
Ta

RI
XI
Reference Parameters Green Cloud Security Architecture IDPS used Tactics Implemented Types of attacks Advantages Disadvantages
No considered Policy

[10] Green cloud Yes Dynamic De- No Virtualization Yes Attacks related to Supports green Computational cost of
computing, centralized (IDS only) security computing, drop rate IDS, To cope with
security of 2% only dynamic environment
of VMs, Detects not
prevents

[11] Authenticatio No Dynamic De- No Virtualization No Attacks related --- ---

n and Identity centralized with security and
Management, trust
Access
Control,
[13] Security and No Dynamic Centralized No --- Yes --- Fulfills various ---
Policy
hosted standards
Integration,

C. Limitations in existing frameworks

services
Service
[14] Management,
Security and No Dynamic Centralized No --- No Attacks related to --- Very General Model,
Trust
Privacy Security and Implementation is
Management Privacy hard

[15] Threat events No Static Centralized No --- Yes Spoofing, Allow customers to Requires meticulous
compromisin Tampering, comparatively choose collection of data and
g cloud and Repudiation, among different industry SME inputs
internet Information vendors, alleviates
security Disclosure, Fear, Uncertainty and
Denial of Service, Doubt, quantitative
Elevation of and iterative
Privilege convergence approach

[16] Trust and No Dynamic De- No Virtualization No Prevent Found new Detective rather than
accountability centralized diminishing accountability issues preventive approach is
issues controls and lack that are not known proposed, securing the
of privacy before log files is itself a
challenge

[17] Risk, No Static De- No --- Yes Security and Follow PDCA security ---
security, centralized privacy attacks cycle
privacy

[18] Security No Dynamic De- No Virtualization Yes(at SaaS Breaching into Developed in standard ---
centralized model) company’s quality management
network and (PDCA), helpful for
accessing data, risk analysis,
using TELNET to assessment and
access system mitigation, applicable
files using guest on all service and
ID, obtaining deployment models
unauthorized
access
C. Limitations in existing frameworks computation cost and will make the network more
The above discussed frameworks handle security issues but on efficient.
the other hand, they have increased the computational cost. The limitations discussed above section are eliminated by
None of the frameworks have used IDPS. IDPS can IDPS that will not just detect the attack but also attempt to
automatically take actions like dropping the malicious packets, prevent it. The power consumption is reduced by the
resetting the connection and/or blocking the traffic from the virtualization technique as a single hardware system does all
offending IP address etc. Other advantages of IDPS include the tasks. In Fig. 2, switch that is used is virtual in nature, the
automatic correction of Cyclic Redundancy Check (CRC) database is also virtual and even the IDPS itself is running on
errors, cleaning up the unwanted traffic. So IDPS is more a virtual machine. The problem of different security protocols
useful than IDS. There is only one framework that supports is taken into consideration with the concept of nesting of
green cloud computing, quickly reduces the network overload virtual machines. All VMs of same security protocols are
and results in the packet drop rate of only 2% [10]. Some grouped together to make the functioning of IDPS better and
issues do exist in this framework. Firstly, the Intrusion more energy efficient.
Detection System (IDS) only detects the attack but not prevent
it. Sometimes, it is too late to take necessary actions by the
Admin to prevent the attack. Secondly, each virtual machine
has a different configuration and thus require different security
solutions. Thus, rules/signatures in IDS should be relevant to
each VM. Lastly, the IDS itself can be attacked as the attacker
can directly send a malicious packet to the network which
affect the installed IDS. So to overcome above mentioned
issues, an IDPS based framework has been proposed and
prototype implementation has been done as discussed in next
section. A comprehensive survey of above mentioned
frameworks has also been done which consists of various
parameters like architecture, green/non-green, tactics used etc.
and results are shown in Table 1.
III. PROPOSED FRAMEWORK
An IDPS [19] based framework has been proposed and its
prototype implementation is given in this section. It secures Fig.1. Parameters considered for the framework
the network and also takes care of green cloud computing.
Fig.1 shows the layout of proposed framework. The main A. Working of the Framework
tactics used are: The working of the proposed framework as shown in Fig. 2 is
 IDPS: Intrusion Detection and Prevention System given below:
identifies the malicious activity, logs information
about it, attempts to stop it and reports the security
Admin. The IDPS used in this framework is hybrid
comprising of Signature and anomaly based
technique which gives real time protection and the
response time is active. It has minimal impact on the
performance of the overall network [20].
 Virtualization: Virtualization is running two or more
operating systems on a single hardware or it is the
abstraction of computer resources. We have used the
virtualization technology to support green cloud
computing. The virtual switch has been used instead
of the physical switch resulting in energy efficient
environment. The IDPS itself runs on VM and its
data is also stored on a VM.
 Nesting of Virtual Machines (VMs): Nesting of
Virtual Machines means one or more VMs running
under the other VM [21]. By using nesting, we can
combine the VMs of the same configuration together. Fig.2. Framework for security in green cloud computing
It will ease the implementation of security as the  The user got connected to the internet and log into the
security requirements of same types of VMs will be cloud system.
the same. This will also decrease the overall
 Table 2. PROTOTYPE IMPLEMENTATION MODEL

Tool Type/Version Work Input Output Location

Hypervisor Hosted Manages Virtual Instructions from Controls and manages Contains all Virtual
Machines expert or some Virtual Machines machines, virtual
software Switches within itself
Cisco Packet Tracer 6.1 Simulation of network Setup of Real-time view of packet On routers and switches
devices networking devices communicating with
other devices
Switch Virtual Send packets to desired packets packets Entering of packets in
destination the virtual network
after port0
Traffic generator 3.0 Increase/Decrease traffic Page count, Increased/Decreased Between physical switch
timespan number of packets and virtual switch
Wireshark 1.12.6 Sniffs and analyze Packets Analyzed view of packets Between Virtual Switch
packets and Port Mirroring and IDPS
OSSEC --- Detects and prevent packets Detection or prevention On Virtual Machine
security attacks or log file of packets containing IDPS
ORACLE 11g Stores and retrieve Log Query/Log Details of particular log On separate Virtual
files Machine

 The message generated from the user system passes
through the gateway and reaches the router. The
router will check its routing table and forwards the
packet to another router. The process goes on until it
reaches the destination node which is the switch here.
 The switch receives, process and forward the data
and send it to the desired location, that here is port0.
Port0 is a physical network interface and all packets
go through Port0.
 The firewall will monitor and control the incoming
and outcoming traffic. It distinguishes a trustworthy
and secure network from the other outside network.
 The packets through port0 reach the Virtual Switch
which works similarly as physical switch though it is
connected to Virtual Machines rather than Physical
ones. It receives the data packet from port0, process
them and forward them to the desired Virtual
Machine. Here, Linux Bridge is implemented as a
virtual switch.
 For providing security, we have installed IDPS. The
data packet reaches IDPS through VNI-p, which is a
Virtual Network Interface connected with the mirror
port. It duplicates and monitors data packets. IDPS
detects the malicious activities, log information about
those activities and takes action to prevent or block it
and reports it to the Admin. The actions taken by
IDPS include triggering an alarm, blocking the
particular IP, dropping malicious packets and
resetting the network.
 VNI-c is Virtual Network Interface for connecting to
IDPS and port1. Port1 is connected to the physical
network to ensure no disturbance to the whole
system.
 For logging the malicious activities, we have used a
tracking module and DB which is a dedicated virtual
machine that tracks the malicious activities and store
them in the database for further use to prevent the
same type of attacks that had already occurred.
 All the VMs are managed by Hypervisor. Here, we
have used hosted hypervisor which runs on an
operating system like other application software do.
 CloudVisor [21] is used for the nesting of VMs
which helps in running one or more VMs under the
other VM. The VMs having same configurations will
be nested together so that they have same security
protocols.
B. Prototype used for implementation of the framework
The tools used for implementation, different tasks they
perform, the kind of input they take, the output they
provide and where they are used is given in Table 2. Cisco
Packet Tracer is used to route and check the flow of
packets among various routers and switches. Linux bridge
is implemented as a virtual switch by making few changes
in it. It will send the packets to the desired destination. To
check whether the IDPS copes with the sudden increase
and decrease of traffic, we have used a traffic generator
software by which we can provide fluctuations to the flow
rate of the packets. For network protocol analyzing,
Wireshark is used. OSSEC is implemented here as IDPS.
All the log is stored on a separate virtual machine where
ORACLE is installed.
IV. CONCLUSION AND FUTURE SCOPE
Instead of using separate hardware systems, virtualization
technologies are used which leads to better manageable and
efficient cloud computing. But in the multi-tenant
environment it raises the issue of data privacy and security.
The large data centers emit carbon-dioxide which leads to
global warming. So to address both of these issues, we have
proposed a framework which tackles with security and is
energy efficient. Signature and anomaly based hybrid IDPS
detects, logs, prevent and alert the security Admin. The
Nesting of the VMs provide easy management and IDPS can
deal better as the grouping of VMs has been done on the basis
of security protocols. Virtualization gives overall low
computation cost because a single hardware system does all
the tasks (providing security, maintaining log files, managing
virtual machines, implementing switch) and thus consumes
very less power and make our framework energy efficient.
The further research continues to implement this framework in
real time scenarios and based on those results the
computational cost of IDPS could further be reduced. Also, we
aim to improve IDPS for better detection and prevention of
attacks and to lower the rate of false negatives and false
positives.
REFERENCES
[1] NIST, ”The NIST definition of cloud computing”, October, 2011,
<http://faculty.winthrop.edu/domanm/csci411/Handouts/NIST.pdf>
Accessed on 12 Aug, 2015.
[2] S. Subashini, V. Kavitha, “A survey on security issues in service delivery
models of cloud computing”, Journal of Network and Computer
Applications, ELSEVIER, pp. 1-10, 2010.
[3] D. Zissis, D. Lekkas, ”Addressing cloud computing issues”, Journal of
Network and Computer Applications, ELSEVIER, pp. 583-592,2010.
[4] C.Modi, D. Patel, B. Borisaniya, H. Patel, A. Patel, M. Rajarajan, “A survey
of intrusion detection techniques in Cloud”, Journal of Network and
Computer Applications, ELSEVIER, pp. 42-55, 2012.
[5] B. Makhija, V. Gupta, I. Rajput, “Enhanced data security in cloud
computing with third party auditor”, International Journal of Advanced
Research in Computer Science and Software Engineering, Volume 3, Issue
2, February 2013.
[6] M. Sharma, H. Bansal, A. Sharma, ”Cloud computing: different approach
and security challenge”, International Journal of Soft Computing and
Engineering, March 2012.
[7] S. Ramgovind, MM. Eloff, E. Smith, “The management of security in cloud
computing”, International Conference on Information Security for South
Africa, IEEE, 2010.
[8] J. Wang, Y. Zhao, S. Jiang, J. Le, “Providing privacy preserving in cloud
computing”, International Conference on test and Management, IEEE,
2009.
[9] A. Bakshi, B. Yogesh, “Securing cloud from DDOS attacks using intrusion
detection system in virtual machine”, IEEE, 2010.
[10] J. Li, Bo Li, T. Wo, C. Hu, J. Huai, l. Liu, K.P. Lam, “Cyberguarder: a
virtualization security assurance architecture for green cloud computing”,
Journal of Network and Computer Applications, ELSEVIER, pp. 379-388,
2011.
[11] H. Takabi, J.B.D. Joshi, G. Ahn, ”SecureCloud: towards a comprehensive
security framework for cloud computing environments”, Computer
Software and Applications Conference Workshops, IEEE, pp. 393-398,
2010.
[12] A. Malik, M. Nazir, “Security framework for cloud computing
environment: a review”, Journal of Emerging Trends in Computing and
Information Sciences, vol. 3, no. 3, pp. 390-394, 2012.
[13] M. Almorsy, J. Grundy, A. Ibrahim, “Collaboration-Based cloud
computing management framework”, International Conference on Cloud
Computing, IEEE, pp. 364-371, 2011.
[14] A. Youssef, Manal Alageel, ”A framework for secure cloud computing”,
International Journal of Computer Science issues, pp. 487-500,2012.
[15] P. Saripalli, B. Walters, “QUIRC: a quantitative impact and risk
assessment framework for cloud security”, International Conference on
Cloud Computing,IEEE”,2010.
[16] R. Ko,et al., ”TrustCloud: a framework for accountability and trust in
cloud computing”, World Congress on Services, IEEE, 2011.
[17] ENISA, “Security framework for governmental clouds”, February, 2015,
<https://www.enisa.europa.eu/activities/Resilience-and-CIIP/cloud-
computing/governmental-cloud-security/security-framework-for-
govenmental-clouds/security-framework-for-governmental-
clouds/at_download/fullReport> Accessed on Aug 2015.
[18] X. Zhang, N. Wuwong, H. Li, X. Zhang, ”Information security risk
management for the cloud computing environments”, International
Conference on Computer and information Technology,IEEE,pp.1328-
1334, 2010.
[19] NIST, ”Guide to intrusion detection and prevention systems(IDPS)”,
2007,<http://csrc.nist.gov/publications/nistpubs/800-94/SP800-94.pdf>,
Accessed on 4 August,2015.
[20] A. Patel, M. Taghavi, K. Bakhtiyari, J. Celestino, ”An intrusion
detection and prevention system in cloud computing: A systematic
review”, Journal of Network and Computer Applications, ELSEVIER, pp.
25-41, 2012.
[21] F. Zhang, J. Chen, H. Chen, B. Zang, ”CloudVisor: retrofitting protection
of virtual machines in multi-tenant cloud with nested virtualization”,
Symposium on Operating Systems, ACM, 2011.
 Face Recognition Techniques, their Advantages,
Disadvantages and Performance Evaluation
Lerato Masupha, Tranos Zuva, Seleman Ngwira, Omobayo Esan
Tshwane University of Technology
Department of Computer Systems Engineering
Pretoria, South Africa
{masuphaLE,zuvat,ngwirasm,esanoa}@tut.ac.za
Abstract—A human brain can store and remember
thousands of faces in a person’s life time, however it is very
difficult for an automated system to reproduce the same results.
Faces are complex and multidimensional which makes extraction
of facial features to be very challenging, yet it is imperative for
our face recognition systems to be better than our brain’s
capabilities. The face like many physiological biometrics that
include fingerprint, hand geometry, retina, iris and ear uniquely
identifies each individual. In this paper we focus mainly on the
face recognition techniques. This review looks at three types of
recognition approaches namely holistic, feature based
(geometric) and the hybrid approach. We also look at the
challenges that are face by the approaches.
Keywords—Face recognition, biometrics, performance
evaluation
I. INTRODUCTION
Biometrics have emerged the most alternative for
authenticating individuals in this present technological age,
instead of authenticating using conventional passwords, PINs,
smart cards, tokens, keys etc[1, 2]. A Biometric system is an
automated technique of examining an individual using
physiological or biological traits in order to ascertain his/her
identity [1, 3].
The disadvantage of traditional techniques like
vulnerability to loss, theft, misplaced, stolen or forged makes
many application areas such as law enforcement,
banking, time attendance and immigration to migrate towards
using biometric system to improve the security system [1].
The biological traits are signature, gait, speech and keystroke,
these traits change with time[4]. The physiological traits
include face, fingerprint, palm print and iris which could
remain permanent throughout individual’s life time [5, 6].
Fingerprint is the oldest biometric of all times to be used
to for identifying individuals[7]. However face recognition has
an advantage over the other physiological biometrics systems
in that the individual being recognized does not need to
participate or acknowledge the action. Furthermore the facial
images can be obtained easily with a fixed inexpensive camera
as opposed to other biometrics that requires the use of more
expensive equipment like the retina and iris. The face has
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
features that include the eyes, nose, ears, lips, chin, teeth and
chicks. Some of the features are used to recognize individuals.
In this paper we look at the face recognition techniques.
We categorized the techniques under three categories namely
holistic, feature based and hybrid approaches
The major contribution of this paper:
 Highlight various face recognition techniques, their
merit and demerits.
 Challenges and open issues affecting face recognition
technique
 Evaluation metric for face recognition techniques
The rest of the paper is divided into: Section II is the
overview of face recognition techniques; Section III
challenges and Open Issue affecting face recognition
techniques; Section IV Performance Evaluation Techniques
and Section V concludes the paper.
II. FACE RECOGNITION TECHNIQUES
Face recognition techniques can be divided into three
categories”
 Technique that operate on intensity image.
 Technique that deal with video sequences.
 Technique that requires other sensory data such as 3D
information or infra-red imagery.
Obviously, face recognition method has received a great
deal of attention in various applications in the field of image
analysis and computer vision due to several advantages it has
over other biometric methods.
This advantages include low cost equipment for
capturing, face can be done without explicit action on the part
of user, its non-intrusive characteristics.[8, 9].
A. Face recognition based on intensity
Face recognition method based on intensity of images
can be divided into two[8]: (i) Feature-based approach (ii)
Holistic –based approach and (iii) Holistic-based approach as
in Fig. 1.

Face-
recognition
approach
Feature-
Holistic-
based based
Hybrid-
based
Fig. 1: Face Recognition Techniques
1. Feature-based approach
Feature-based approach process the input image to
identify and extract distinctive facial features such as eyes,
mouth, nose etc and then compute the geometric relationship
among those facial points, thus reducing the input facial image
to a vector of geometric feature [8]. Feature based approach is
sub-divided into:
a) Geometric feature based matching
Geometric feature matching techniques are based on the
computation of a set of geometrical feature from the picture of
a face. The entire configuration can be described a vector.
This vector represents the position and the size of main facial
features like the nose, eyes, eyebrows, mouth, chin and the
outline of the face.
The advantage of this technique is that it overcome problem of
occlusion. The disadvantage is does not provide a high degree
of accuracy. Another disadvantage is that it does not require
considerable computational time.
b) Elastic bunch graph
This technique is based on dynamic link structures. A
graph for an individual face is generated using a set of fiducial
points on the face, each fiducial point is a node of a full
connected graph and is labeled with the Gabor filters’
response. Each arch is labeled with the distance between
correspondent fiducial points[10].
A representative set of such graph is combined into a
stack-like structure called face bunch graph. The recognition
of a new face image is done by comparing its image graph to
those of all the known face images and the one with highest
similarities values are selected as closed matching.
1) Advantages
 In feature-based techniques, the feature points precede
the analysis done for matching the image to that of a
known individual.
 The feature-based technique can be made invariant to
size, orientation and lighting.
 It has compactness of representation of the face images
and high speed matching/
2) Disadvantages
 Feature-based techniques lack discrimination ability
 It is difficult to automatic detect feature in this
approach.
2. Holistic-based approach
Holistic based approaches attempt to identify faces using
representation i.e descriptions based on the entire image rather
than on local features of the face. Holistic based approach is
divided into:
 Statistical approach and
 Artificial Intelligent approach.
a) Statistical approach
This is the simplest approach of holistic in which the
image is represented as a 2D array of intensity values and
recognition is done by direct correlation comparisons between
the input face and all other faces in the database. Example of
this includes Principal Component Analysis (PCA), Eigen
faces and fisher face, LDA [11]etc
The Eigenface method is one of the generally used
algorithms for face recognition. The eigenface technique is
based on the Principal Component Analysis (PCA) also
known as Karhunen-Loeve tranform. This method is
successfully used to perform dimensionality reduction.
Principal Component Analysis is used by face recognition and
detection[12].
The advantage of this technique is that it is easy and
efficient as the PCA reduces the dimension size of an image in
a short period of time. Another advantage is that it has high
correlation between the training data and the recognition
data[13]. The disadvantage of this technique is that its
accuracy depends solely on a number of factors amongst them
is the lighting as this decreases the accuracy.
b) Artificial intelligent approach
Artificial Intelligent (AI) approaches utilize tools such as
neural networks and machine learning techniques to recognize
face image. Example of this approach includes: Radial Basic
Fuzzy (RBF), Neural Network (NN), Support Vector Machine
(SVM), Hidden Markov Model (HMM) and Local Binary
Pattern (LBP).
1) Advantages
 They do not destroy any of the information in the
images by concentering on only limited regions or
point of interest.
 The technique produce better recognition results than
the feature-based technique.
2) Disadvantages
 Because the approach does not destroy any images
information this makes the approach to start image
information with the basic assumption that all the pixels
in the image are equally important.
 The approach is computationally expensive and it also
requires high degree of correlation between the test and
training images.
 The approach does not perform effectively under large
variation in pose, scale and illumination etc.
c) Hybrid approach
The hybrid approach is a combination of two or more
approaches which is aimed at yielding more efficient results.
Using more than one approach means that the downfalls of
one approach are improved by advantages of the other
complimenting each other. For example combination of
template matching algorithm and 2DPCA algorithm[14, 15].
B. Face recognition based on video sequences
One of important applications of face recognition is for
surveillance for security purpose. This involves real-time
recognition of faces from images sequence being captured by
a video camera. This modules consists of detecting of face,
tracking of face and recognizing the face.
1. Advantages
 Dynamic does not have advantage over static, the
abundance of data allows the system to select the frame
with the best possible image and discard less satisfactory
ones.
 It provides temporal continuity which improves
recognition performance.
 It allows tracking of face image such that variation in
facial expressions and poses can be addressed, thus
resulted in better recognition result.
C. Face recognition from sensory
Many works have been done on face recognition for
identifying individual using 2-Dimensional (2D) intensity
images. However, no enough work has been done in
recognizing individual by exploiting other sensing modalities
such as 3- Dimensional (3D) or range data and infra-red
imagery.
1. 3D model-based techniques
Using 3D information for face recognition help in
exploiting feature based on the shape and the curvature of the
face such as the shape of forehead, jawline and checks without
being plagued by the variances caused by lighting, orientation
and background clusters that affect the 2D systems[16].
The examples of 3D includes: scanning systems, stereo
vision systems, structured light systems, reverse
rendering/shape from shading etc.
a) Advantage
 It exploits features using shape and curvature of the face.
b) Disadvantage
 The approach is complex and computational cost.
D. Infra-red based techniques
The thermal infra-red imagery is insensitive to variations
in lighting which makes such image to be used for detecting
and recognizing faces. When using infra-red the images vein
and tissues structure which are unique in each individual yield
a good results when applied to these images.
1. Advantage
 It improves face recognition performances.
2. Disadvatages
 It requires substantial cost of thermal sensors.
 It is subjected to a low resolution and high noise in the
images.
 There is lack of widely available data sets of infra-red.
 There is possibility of occlusion in part of face with eye
glasses.
III. CHALLENGES FACING OF FACE
BIOMETRIC SYSTEM
A. Facial form
Most common form (i.e., the frontal view) daces appear to
be roughly alike and the differences between them are quite
subtle. Consequently, frontal face images form a very dense 1. Data quality
cluster in image space which makes it virtually impossible for The quality of raw data obtained during image acquisition
traditional pattern recognition technique to accurately is one the main factors affecting the face biometric system
discriminate among them with a high degree of success. overall performance.

2. Usability
This is refers to extent at which the product can be used
B. Uniqueness by specified users to achieve certain goals with effectiveness,
Human face is not unique; there are many factors that efficiency and satisfaction.
cause the appearance of the face to vary. These are categorized
as extrinsic and intrinsic.
B. Performance metrics
C. Variation The basic fundamental performance metrics for face or
Age, illumination, and pose also affect face recognition any biometric system are:
system. Although most face recognition system works well Failure-to-Enroll Rate (FER): This is the number of users who
under constrained condition, the performance of most of these the system fails to capture.
systems degrades rapidly when they are put to work under Failure-to-Acquire Rate (FTR): This is the number of
conditions where none of the features are regulated. verification or identification attempts that the biometric system
unable to capture.
Failure-Match Rate (FMR): This is the rate of
incorrect positive matches when using matching algorithm for
IV. EVALUATION METRIC FOR FACE
single template comparison.
BIOMETRIC RECOGNITION SYSTEMS
False-Non-Match Rate (FNMR): this is the rate of incorrect
negative matches when using matching algorithm for single
A. Evaluation classification template comparison.
The evaluation of face biometric system can be classified
into three as shown in Fig. 2 below:
C. Verification performance metric
EVALUATION OF False Rejection Rate (FRR): This is the number of
BIOMETRIC geneui user that the system incorrectly denied.
False Acceptance Rate (FAR): this is the number of
impostors that the system accepted.
Receiver Operating Characteristics Curve (ROC):
This is the plot of FMR or FAR (i.e accepted impostor
attempt) on x-ax0s against FNMR or FRR (i.e rejected geneui
Data quality Usability Security
attempts) on the y-axis plotted using a function of decision
threshold as a parameters.
Equal Error Rate (ERR): This is a point at which
FAR and FRR correspond. The more EER is near to 0%, the
better the system performance.

D. Identification system performance metric

Quantitative Qualitative The identification rate is the number of identification
transactions by users enrolled in the system in which the user’s
correct identifier is among those returned.
False-Negative Identification Error Rate (FNIER):
This is the number is identification transaction by user
enrolled in the system that the user’s identifier is not returning.
False-Positive Identification-Error Rate: This is the
number of identification transactions by user not enrolled in
Efficiency Effectiveness User Satisfaction
the system and system identifier is returning.
Cumulative Match Characteristics Curve (CMC):
Fig. 2: Face Recognition Evaluation Techniques This is a graphical representation of the results of
identification task test, plotting rank values on the x-axis and
the probability of correct identification that rank on the y-axis.
 V. CONCLUSION
In this paper, techniques that are used to construct
face recognition system have been highlighted. The challenges
of these techniques have been discussed. The performance
evaluation techniques of recommender systems have also been
looked at. In summary, face recognition systems have added-
value to security and surveillance areas such as airports,
schools, company, corporation etc.

ACKNOWLEDGMENT
The authors acknowledged the financial contributions
and resources made available by Tshwane University of
Technology.

REFERENCES
[1] O. A. Esan, T. Zuva, S. M. Ngwira, and K. Zuva, "Performance
Improvement of Authentication of Fingerprints using Enhancement and
Matching Algorithms," International Journal of Emerging Technology
and Advanced Engineering, vol. Vol. 3, 2013.
[2] O. S. Adeoye, "A survey of emerging biometric technology,"
international journal of computer applications, vol. 9, p. 2, 2010.
[3] M. D. Dhameliya, "A multimodal biometric recognition system based on
fusion of palmprint and fingerprint," international journal of engineering
trends and technology, vol. 4, p. 1908, 2013.
[4] L. masupha, T. Zuva, and S. Ngwira, "A Review of Gait Recognition
Techniques and their Challenges," presented at the Third international
conference on digital information processing, E-Business and cloud
computing, Reduit Mauritius, 2015.
[5] H. A. Aboalsamh, "Vein and Fingerprint Biometrics Authentication-
Future Trends," International Journal of Computer and Communications,
vol. 3, 2009.
[6] O. Esan, S. Ngwira, L. Masupha, and T. Zuva, "Health Care
Infrastructure Security using Bimodal Biometrics System," international
journal of computer and information technology, vol. 03, pp. 299-305,
2014.
[7] A. Senior and R. Bole, "Improved Fingerprint Matching by Distortion
Removal," IEICE Trans. Inf. System, vol. 8, pp. 825-831, 2001.
[8] R. Jafri and H. R. Arabnia, "A Survey of Face Recognition Techniques,"
Journal of Information Processing Systems, vol. 5, pp. 41-63, june 2009
2009.
[9] R. Patel and S. B. Yagnik, "A literature survey on face recognition
techniques," international Journal of computer trends and technology,
vol. 5, pp. 189-194, 2013.
[10] M. Sharif, S. Mohsin, and M. Y. Javed, "Face recognition techniques,"
Research journal of applied science engineering and technology, vol. 4,
pp. 4979-4990, 2012.
[11] B. a. Draper, K. Baek, M. S. Bartlett, and J. R. Beveridge, "Recognizing
faces with PCA and ICA," Computer vision and image understanding,
vol. 91, pp. 115-137, 2003.
[12] V. Vijayakumari, "Face recognition techniques: A survey," world journal
of computer application and technology, vol. 1, 2013.
[13] A. Mir and A. G. Mir, "feature extraction methods(PCA fused with
DCT)," international Journal of advances in engineering and technology,
vol. 6, pp. 2145-2152, 2013.
[14] J. Wang and H. Yang, "Face detection based on template matching and
2DPCA algorithm," presented at the congress on image and signal
processing, 2008.
[15] M. Agarwal, N. Jain, M. Kumar, and H. Agrawal, "Face Recognition
Using Eigen Faces and
Artificial Neural Network," international Journal of computer theory and
engineering, vol. 2, pp. 624-629, 2010.
[16] R. Brunelli and T. Poggio, "Face recognition: Feature versus Templates,"
IEEE transaction on Pattern Analysis and Machine Intelligence, vol. 15,
pp. 1042-1052, 2003.
 Designing Budget Forecasting and Revision System
Using Optimization Methods
M.J. Tarokh
K.N.Toosi University of Technology
Tehran, Iran
mjtarokh@kntu.ac.ir
Abstract—The sales procedures are the most important
factors for keeping companies alive and profitable. So sales and
budget sales are considered as important parameters influencing
all other decision variables in an organization. Therefore, poor
sales forecasting can lead to great losses in an organization
caused by inaccurate and non-comprehensive production and
human resource planning. Hence, in this research a coherent
solution has been proposed for forecasting sales besides refining
and revising it continuously by ANFIS1 model with consideration
of time series relations. Data has been collected from the public
and accessible annual financial reports related to a famous
Iranian company. Moreover, for more accuracy in forecasting,
the solution has been examined by BPN2 and PSO3 as
optimization methods. The comparison between taken prediction
and the real data shows that PSO method can optimize some
parts of prediction in contrast to the rest which is more
coincident to the output of BPN analysis. As a consequence, a
hybrid integrated system including both has been designed. This
system uses them depending on their abilities to optimize each
part, so it can produce more precise results relatively.
Keywords—ANFIS; PSO and BPN methods; hybrid method
I. INTRODUCTION
The importance of sales forecasting for a firm has often
been stressed [1] and is best expressed by what happens when
it is absent. “Without a sales forecast, in the short term,
operations can only respond retroactively, learning to lose
orders, inadequate service and poorly utilized production
resources. In the longer term, financial and market decision
making misallocate resources so that the organizations
continuing existence may be brought into question” [2]. The
forecasts are used for a number of purposes in a firm,
including production planning, budgeting, sales quota setting
and personnel planning ([3],[4]).
The primary objective of most business enterprises is
securing a profit and the accumulation of wealth. Budgeting
aides management in realizing its profit objective by providing
a scientific technique for forecasting business operations and
establishing standards [5]. Managers use the budget as a road
map for allocating the company's resources. The main purpose
of budgeting is aligning the company activities with the
1
Adaptive Neuro Fuzzy Inference System
2
Back Propagation Neural Network
3
Particle Swarm Optimization
K. Koochakpour
Islamic Azad University
Iran, Qazvin
K.Kouchakpour@gmail.com
objectives. Due to the changing conditions and the influences
of many internal and external factors, decision making about
the budget allocation is so important and has too much
complexity. To make this decision making easy and flexible in
these changing conditions of organizations is so essential.
Therefore, it is important to periodically and flexibly adjust
the budget and systematize this process can help to be done
more accurately and reduce human faults.
Efforts to balance and adjust the budget, lead to a better
understanding of income, expenditure and cash flow in a
business. Inappropriate Budgeting in an organization can’t
help the managers and can’t improve performance. A Budget
which is planned inappropriately in an organization will be
ignored because it can’t give staff, reasonable criteria for
comparison with actual performance. That is why it is very
important to review and revise the budget periodically in
accordance with the actual performance. Additionally,
analysis of variance from budget during the budget revises
process; help managers to determine when to adjust their
operations and costs.
The successful combination of some methods, such as
neural networks, fuzzy logic and evolutionary computation,
developed a new method called Soft and Intelligent
Computing, That these soft techniques can be used in
estimation, forecasting and decision making in various
contexts. Neuro fuzzy or Fuzzy Neural system is a hybrid
system that combining the fuzzy logics ability of making
decisions with the neural network’s ability of learning, in high
levels of complexity, to be able to present a modeling /
estimation system. Fuzzy Neural system is a neural network
performing with neuro fuzzy inference system coordinately
[6].
For sales forecasting and budget assessment it is necessary
to identify the relation between variables influencing this
forecast. There are several methods to correlate variables with
sales volume such as multiple linear regressions and
computational intelligence regression which are statistical
methods for studding and modeling the relations between
variables. Artificial neural networks (ANNs) and Adaptive
Network Fuzzy base Inference System (ANFIS) are two
common nonlinear techniques for sale forecasting in recent
years. ANFIS first proposed in [7] which is two combine of
ANNs and Fuzzy Inference System (FIS). In addition, ANFIS
can do training more precisely due to using fuzzy system
 TABLE I. INTERNAL AND EXTERNAL FACTORS
Variable Abbreviation Source Test Data
Operating Income Margin OIM prior financial statement 0.814
Inventory Turnover IT prior financial statement 26.21
Debt Ratio DR prior financial statement 0.252
Return On Assets ROA prior financial statement 0.328
Employee Count EC company data 127 B. Adaptive Neuro Fuzzy Inference Systems
Asset A company data 350000
Industry Share IS company/Industry data 0.023 ANFIS is an adaptive network of nodes and directional
Currency C Industry factors 24846 links with associated learning rules. It is called adaptive
Inflation rate IR Industry factors 182.5
Sale Sale company data 424786
because some, or all, of the nodes have parameters which
which causes to get membership function parameters and influence the output of the node. These networks identify and
optimizes them. There are several methods for optimizations learn relationships between inputs and outputs. The basic
and composing them with ANFIS. Training algorithm used in architecture of ANFIS consist five layers with different
this research for error reduction are PSO and BPN methods. functions that completely described in reference [8].

In this study, we proposed a different way to find a III. LEARNING ALGORITHM

suitable method for forecasting sales budget of a company in
Iran.
II. METHODOLOGIES AND MODEL SETTING
A. Selection of Prediction Variable
Sale forecasting is a complex activity. It requires careful
and systematic analysis of various components of the internal
and external environment. The main factors affecting sales
forecasting are discussed below:
1. Economic conditions: it refers to level of income,
stage of business cycle and standard of living.
2. Condition within the industry: also the aggregate
demand of a product of the industry affects the
sales –level of individual business unit.
3. Condition within organization: conditions within
the organization refer to the internal environment
of the organization. It includes price policy of the
organization, sales promotion schemes,
promotion budget, product design and distribution
and development policies.
4. Level of competition: if level of competition is
high then it will adversely affect the level of sales.
5. Past sales and growth trend in sales. Previous
sales levels and trends and selling expenses are
necessary to generate former sales volumes,
trends of the company's need to borrow (supplier,
trade credit, and bank credit) to support various
levels of inventory and trends in accounts
receivable required to achieve previous sales
volumes.
6. Government policies.
Along with neural network-fuzzy, there are numerous
methods for network training and optimization set of the rules
and output parameters. Each of these methods is trying to
improve the network performance. The following, is a brief
overview of the BPN and PSO technique which has less
complexity and fast convergence.
A. Back Propagation Neural network (BPN)
The ability to create useful features distinguishes BPN
from earlier and simpler methods such as perceptron
convergence procedure. Using this method for Training data,
initially the error must be calculated and finally that is
minimized. To minimize errors, networks must be modified in
order to reduce the weight of communication networks. The
total error in the performance in the network with a particular
set of weights can be computed by comparing the actual and
desired output vectors for every case. The total error is defined
as:
1 p
E  (Oi  t i )2
2 i 1 (1)
Where O i is the desired output of the ith unit and t i is the
actual state of an output unit. Unit i is a typical unit in the
output layer.
B. Particle Swarm Optimization (PSO) Algorithms
PSO method as a powerful way to optimize could be an
internal block structure which is used by artificial neural
networks for Training data. This method was discovered in
1995 by Kennedy [9] and since then is used as a strong
algorithm with a relatively simple structure. This method is an
optimization method based on population and a multitude of
particles.

7. Amount of promotion budget.
8. Fashion and taste of consumers.
Applying business experience extracted from past financial
statements help management and salespeople to make
decisions better. According to these, we could find some
factors that seem to be more effective in sales forecasting than
others. All of these factors come in table 1.
A technique that PSO uses is a vector for moving. The
motion vector which intelligently tries to update itself every
moment. In order To find the best moving one must follow
two types of motions: 1) Move to the previous experience or
x
local best ( x ib ) 2) Move to the pattern or global best ( gb ).
x
Since full motion of particle it is impossible to gb , as well as
x
full motion to x ib particle the best answer that is gb Away,
Therefore, a move that Particle Selects, movement between

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

these two terms and in this way particle will update your
location. The position of each particle is updated using its
velocity vector as shown in Fig.1.
Fig. 1. Depiction of the velocity and position updates in PSO
k
Where in Fig. 1 the current position of the particle is x i
x k 1
and update position is i . Finally the position of the ith
particle is updated according to [10]. Position update is the last
step in each iteration. The new location is given equation (2)
and (3).
V. INDUSTRIAL CASE STUDY
A. Data Collection
This case study provides two data components. The first
component is Company data, and the second one is industry
data. We collected the company data from the databases with
the annual financial reports that the companies mail and
distribute to investors. We had to provide sufficient historical
data as much as possible, to conduct reliable time-series
analysis and neuro fuzzy learning.
Therefore, we use the yearly and seasonal data of the years
2008 to 2013. The sales figures of these data are shown in
Fig.3 (a) and the seasonal pattern of these time series is clearly
recognizable in the Fig.3 (b).

v ik 1  v ik  c1r1 (x ib  x ik )  c 2 r2 (x gb  x ik ) (2)

x ik 1  x ik  cv ik 1
(3)
Where,  is the inertia factor that   (0.4,14) , x ib and
x gb are local best and global best and c1 and c 2 are random
variables that c1 and c 2 are defined as c1  rb
1 1
and c 2  r2b2
Fig. 2. Flowchart of the proposed PSO technique
, with r1 and , and c1 and c 2 are positive acceleration
constant. Kennedy asserted that c1  c 2  4 guarantees the
stability of PSO [10]. Figure 2 shows the flowchart of the
proposed PSO algorithm.
IV. TIME SERIES MODEL
In this contribution an additive model with the following
components to mimic the time series is applied. The purpose
of a time series is a collection statistical data, collected equally
and at regular time intervals. Statistical methods are used to
this kind of statistical data called time-series analysis.
(a)
A technique was used in this article using time-series
methods to predict the future according to defined patterns
liable of the past. This model works as explained. First a
certain number of data end of to assess forecast regardless and
other data are evaluated and Trained. In fact, if x t
represents the amount of data in the moment now, It can be
expressed in equation 3.
x t  f (x t  1 , x t   2 ,..., x t   n ) (4)

Where  i is delayed. In fact equation (4) expresses the (b)

amount of data in the current situation related to the values of Fig. 3. (a). The value of sales Six month of year data (b) Seasonal data
the past. In this paper, time series analysis using ANFIS is
done Based on PSO and BPN optimization algorithm
optimization.
B. Evaluation Workflow
Data
The process of this research can be described in the
proposed framework that is shown in the five steps in Fig. 4.
To design the budget revision system, the steps are: Testing Date
Year[a…(a+M)]
Training Data
Year[1…a]

Step1:
Outputs Inputs

 (Sales) (Internal &

Collecting the main factors which have effect on External Factors)

budget sales.
 Completing the source data with internal and external Seasonal Sales
ANFIS-2
(Time Series
Building Fuzzy
Inference
Year[1…a]
factors and accessible data. Analysis) System(FIS)

Step2: Improve

 Designing and training the first adaptive neuro fuzzy Seasonal coefficient ANFIS-1 Learning Engine
By Using PSO

inference system (ANFIS-1). Sales Prediction Sales Prediction

 Try and test improving learning engine with PSO ANFIS-3
(Time Series
For Second
Semiyearly
For First
Semiyearly
Analysis) Year:a+1
algorithm. Year:a+1

Step3:
 Designing (ANFIS-2), to predict second semiyearly ∑

period by using the time series analysis. (By using Seasonal coefficient
Sales Prediction
ANFIS-2, prediction could be done without any date Prediction For
Total For Total
about effective factors). Year:a+1 Year:a+1

Step4: *
 Calculating the ratios of every season sales to total
sales with seasonal sales data in various years Sales
 Designing ANFIS-3 by using time series of ratios, to
Winter Autumn Summer Spring
Budget Sale Sale Sale Sale
Version1 Prediction Prediction Prediction Prediction
predict the next period. Sales Budget
 Obtaining the sales forecasting of all the seasons as a
- Deversion 1
Real
Spring
result. * Sale For

 One part of prediction system will be realized and

Year:a+1

sales forecasting can be more accurate, after Sales Winter Autumn Summer Spring
Budget
happening each season. Version2
Sale
Prediction
Sale
Prediction
Sale
Prediction
Realized
Sale
Step5:
 Gathering the realized data of last two seasons. *
Real
Summer -
Sales Budget
Deversion 2
 Returning to the first step and going through the Sale For
Year:a+1

previous steps with a new data to calculate sales

Sales Winter Autumn Summer Spring
forecasting and its revision of the next year. Budget Sale Sale Realized Realized
Version3 Prediction Prediction Sale Sale
As seen above, the budget revision is done four times
which can cause budgeting decisions adapted to criteria and Real
Sales Budget
* Autumn -
circumstances with more flexibility. Sale For Deversion 3
Year:a+1

Sales
Budget Winter Autumn Summer Spring
Version4 Sale Realized Realized Realized
Prediction Sale Sale Sale

Real
Winter Sales Budget
Sale For - Deversion 4
Year:a+1

Winter Autumn Summer Spring

Realized Realized Realized Realized
Sale Sale Sale Sale

+ Second Semiyearly
a=a+1
Realized Sale
Year:a+1

Fig. 4. The proposed Frame work

 VI. RESULTS

A. Semiyearly Modeling
In this section the output of ANFIS-1 has been trained with
both BPN and PSO methods. In Fig.5 (a) prepare the
comparison between real data and the trained output of
ANFIS-1 and Fig.5 (b) ones illustrates the percentage error.
These outputs show when we have effective factors as inputs,
forecasting system trained by PSO method has better and more
accurate results than by BPN method.
Three common indices such as Mean Square Error (MSE), (a)
Root Mean Square Error (RMSE), Mean Absolute Percent
Error (MAPE), is used as performance evaluation criteria (see
table II).

(b)
Fig. 6. Seasonal modeling a) Training with BPN & PSO b) Percentage of
error

(a) These outputs in table III show that trained forecasting

system, when using time series data as inputs, is ambiguous
and is not clear.

TABLE III. THE COMPARING RESULTS OF DIFFERENT OPTIMIZATION

METHODS FOR SEASONAL COEFFICIENT OF PREDICTION WITH TIME SERIES
METHOD

Testing Data
Season
Optimization MSE MAPE Seasonal MAPE MSE
of the
year Method (%) Cofficent (%)
(b) Prediction
BPN 3.98e-6 0.897 0.1912 1.24 5.5e-6
Fig. 5. Semiyearly modeling a) Training data with BPN & PSO methods b) Spring PSO 3.99e-6 0.898 0.1913 1.26
Percentage error
Summe BPN 1.48e-5 1.547 0.2375 2.39 3.4e-5
r PSO 1.48e-5 1.547 0.2376 2.38
BPN 1.63e-6 0.391 0.2659 1.50 1.66e-
Fall 5
TABLE II. THE COMPARISON OF TWO DIFFERENT OPTIMIZED RESULTS PSO 1.63e-6 0.391 0.2658 1.48
FOR SALE PREDICTION WITH ANFIS-1
BPN 1.76e-5 1.174 0.3051 2.43 5.08e-
5
Training Data Testing Data Winter
PSO 1.77e-5 1.168 0.3050 2.36
Optimization MSE RMSE MAPE Sale MAPE
Method (%) Prediction (%)

BPN 654,360 808.925 0.5714 163,330 3.7205

PSO 0 0 2.055e-4 164,990
B. Seasonal Modeling by Time Series
In this section the output of ANFIS-2 has been trained with
both BPN and PSO methods. The Fig.6 (a) show the
comparison between real data and the trained output of
ANFIS-2 and In Fig.6 (b) ones illustrate the percentage error.
2.7482 C. Semiyearly Modeling by Time series
In this section the output of ANFIS-3 has been trained with
both BPN and PSO methods. The Fig. 7 prepares the
comparison between real data and the trained output of
ANFIS-3 and percentage of error.
 D. Forecast Testing
For testing the results, Sales forecasting for each season in
2014 is provided. Table V illustrates the forecast and error of
both models.

TABLE VI. SALE PREDICTION FOR EACH SEASON USING BP AND PSO
Optimization Sale prediction for
Season MAPE (%)
Method each season
Spring BPN 72,588 2.1044
PSO 72,177 2.6582
Summer BPN 90,122 5.622
PSO 89,611 6.1573
Fig. 7. Second Semiyearly Modeling Training with BPN & PSO Fall BPN 100, 850 4.7649
PSO 100,290 5.2998
Winter BPN 115,780 0.9935
Outputs in table IV show that, trained forecasting system PSO 115,070 1.5998
by BPN method, when we are using time series data as inputs,
has better and more accurate results than by PSO method. Table VI indicates the revising process based on one
realized data. Realized spring sales are replaced with its
TABLE IV. THE COMPARING RESULTS OF DIFFERENT OPTIMIZATION prediction, and revise all other seasons accordingly and this
METHODS FOR SECOND SIX- MONTH PREDICTION USING TIME SERIES METHOD process repeats for summer as well.
Training Data Testing Data
Second six
E. Hybrid model
Optimizatio MAPE month MAPE The results show that PSO method could improve the
MSE RMSE
n Method (%) prediction (%)
performance of BPN in ANFIS model when it analyze the
BPN 1.8739e5 432.89 0.359 214,530 3.729 effects of the internal and external factors in sale forecasting,
PSO 1.683e6 1297.2 1.034 213,970 3.982 while its effects on the analysis results of time series data is
ambiguous. It seems that PSO does not always perform well
and its behavioral parameters may need tuning.
Considering the fact that we face high test errors although We found that the best fit model for forecasting the
we start from low training errors, one can assume that the discussed process was the Hybrid ANFIS model which can
training set was too small for this specific problem and the produce more accurate predictions. The hybrid model applies
model is over fitted. Especially as data points which are not in a combination of the PSO method and BPN method to train
close proximity to the training set are hard to predict correctly. membership function of fuzzy inference system (FIS).
Table VII illustrates the optimization results of hybrid
method and its percentage improvement from BPN and PSO

TABLE V. SECOND SEMIYEARLY PREDICTION USING HYBRID MODEL

The actual The actual Optimization The amount of sales Forecast for second Forecast seasonal MAPE
amount of amount of method of the first Semiyearly coefficient (%)
spring summer Semiyearly
prediction with real data of prediction with real data of

Spring ----- ------

Summer 0.2377 2.2861
Seasonally Revised

BPN 164,270 260,850 Fall 0.2658 1.4339

spring sales

Winter 0.3051 2.4143

74,148 ----
Spring ------ ------
Summer 0. 2378 2.2654
PSO 163,760 259,110 Fall 0.2659 1.4574
Winter 0.3051 2.4039

Spring ---- ----

spring and summer sales

Summer ---- ----

Seasonally Revised

BPN 169,640 266,560

Fall 0.2665 1.2226
Winter 0.3052 2.3909
74,148 95,491
sales

Spring ---- ----

Summer ---- ----
PSO 169,640 264,300 Fall 0.2665 1.2490
Winter 0.3051 2.3838
methods in first half-year. According to the results, hybrid 1986.
model is better fitting model in comparison of two others due [5] Ch. H. Walker, and C. Marine Corps, “The George Washington
University, Budgetary Administration and Control Facilitates Intelligent
to less MAPE value. This hybrid model has a better Management. 1956.
performance in historical time series data and analyzing the [6] R. Ata, and Y. Kocyigit, “An adaptive neuro-fuzzy inference system
internal and external factors for sales forecasting. approach for prediction of tip speed ratio in wind turbines,” Expert
Systems with Applications, vol. 37 (7), pp. 5454-5460. 2010.
For testing the results of hybrid model, Sales forecasting [7] J. S. R. Jang, “ANFIS: Adaptive-network-based fuzzy inference
for each season in 2014 is provided. The table VIII illustrates system,” IEEE Transactions on Systems, Man, Cybernetics, vol. 23, pp.
the forecast and error of model and its percentage 665–685. 1993.
improvement from BPN and PSO. [8] J.K. Jeon, and M.S. Rahman, “Fuzzy Neural Network Models for
Geotechnical Problems,” Department of Civil Engineering, North
Carolina State University, Raleigh, NC. 2008.
VII. CONCLUSION [9] J. Kennedy, and R.C. Eberhart, “Swarm Intelligence,” Morgan
Kaufmann Publishers, San Francisco. 2001.
A budget sale is the factor which influences all other [10] D. P. Rini, S. M. Shamsuddin, and S. S. Yuhaniz, “Particle Swarm
decision variables in an organization. It is obvious that reliable
forecast can increase the overall performance of an
organization and the effectiveness of all decisions will be
dependent on the validity and accuracy of the forecast. For this
reason, sale budgeting and sales budget revision is the main
strategic activities in decision making process. The purpose of
this paper is expressed as an integrated system to forecast
sales budget of a company by some features about
company/industry characteristics.
To achieve more accuracy, we applied the benefit from
learning capabilities of Neural Network in ANFIS and PSO.
The comparison between optimized prediction by PSO and
BPN and real data indicates that each method can optimize
some parts of prediction better than other parts therefore we
could have improved forecast accuracy more if we combined
both methods according to their abilities of optimization. The
results showed that this hybrid method can improve sale
budgeting and forecasting significantly.

REFERENCES
[1] S.G. Makridakis, S.C. Wheelwright and R.J. Hyndman, “Forecasting:
Methods and Applications,” Third edition, John Wiley and Sons. 1998
[2] R. Fildes, and R. Hastings, “The organization and improvement of
market forecasting,” Journal of the Operational Research Society, vol.
45, pp. 1–16. 1994.
[3] J. Mentzer, and K. Kahn, “Forecasting technique familiarity,
satisfaction. Usage and application,” Journal of Forecasting, vol. 14(5),
pp. 465–476. 1995.
[4] H. R. White, “In Sales Forecasting: Timesaving and Profit-Making
Strategies that Work,” London, UK: Scott, Foresman and Company.

TABLE VII. SECOND SEMIYEARLY PREDICTION USING HYBRID MODEL

Testing Data
Optimization Method Second semiyearly MAPE (%) Improvement Percent compared Improvement Percent compared
prediction with BPN with PSO

Hybrid 216,540 2.8282 %0.92 %1.18

TABLE VIII. SALE PREDICTION FOR EACH SEASON CALCULATE IMPROVEMENT PERCENT USING HYBRID MODEL

Season Optimization Sale prediction for each MAPE Improvement Percent Improvement Percent
Method season (%) compared with BPN compared with PSO
Spring Hybrid 73,754 0.531 1.58% 2.138%

Summer Hybrid 91,569 4.106 1.58% 2.138%

Fall Hybrid 102, 480 3.230 1.59% 2.137%

Winter Hybrid 117,580 0.550 1.53% 2.134%

 Ameliorate Pursuance of TCP in CR-MANET
Utilizing Accessible Resources
A.Senthamarai Selvan,
Research Scholar,
Department of Computer Science and Engg.,
Pondicherry Engineering College,
Puducherry,India.
senselvana@gmail.com
Abstract—Cognitive Radio (CR) networks is well-known for its
expertness in solving the problem of spectrum scarcity that reside
with wireless resources where unlicensed users can
opportunistically perform transmission without impacting the
operations of licensed users. Channel switching is inherently
necessary to make unlicensed and licensed users to appropriate
propagation in the channel. All through the progression,
communication characteristics such as bottleneck bandwidth and
Round-trip time (RTT) have to be modified. However this change
has to adaptively update by TCP in its congestion window
(CWND) to make an efficient use of the available resources. TCP
CRAHN is well-known mechanism which generates spectrum
alertness by retrieving explicit feedback from relay and
destination nodes. In this paper we proposed TCP CR-MANET
and it is evaluated with respect to bottleneck bandwidth and
RTT varying characteristics. This protocol updates its CWND
based upon the available buffer space of the relay node. TCP CR-
MANET system is implemented in NS2 simulator and analyzed
with various characteristics of the network. Experimentally, it
has been proved that our proposed TCP CR-MANET provides
better throughput with respect to TCP CRAHN.
Keywords—Cognitive Radio, Congestion control, Spectrum
Sensing, Transport Protocol, and Mobile ad hoc network.
I. INTRODUCTION
Cognitive radio technologies have gained the rapid
growth of wireless communication popularity now a days, due
to its capability of improving spectrum utilization by
deploying the unused spectrum in dynamically changing
environments. The unlicensed bands, mostly that fall under
900MHz and 2.4GHz, are getting more and more congested
[1]. Various wireless applications are activated under
unlicensed bands. The radio spectrum demand has been
increased dramatically and still few more available spectrums
can be allocated. However, according to Federal
Communications Commission’s (FCC) report [2], the same
spectrum bands are underutilized due to existing amount of
idle spectrum holes at spatial and temporal measurements.
Cognitive Radio Technology has the potential to ameliorate
the scarcity of wireless resources. Cognitive Radio Networks
(CRN) has gained increasing popularity now a days due to its
capability of improving spectrum utilization by manipulating
the vacant spectrum in energetically varying environment.
Secondary Users (SU) or Unlicensed Users are able to
resourcefully function in the permitted spectrum bands, not
Ka.Selvaradjou,
Professor,
Department of Computer Science and Engg.,
Pondicherry Engineering College,
Puducherry, India.
selvaraj@pec.edu
being obstacle to the activity of Primary Users (PU) or
Licensed Users. In this paper, we cautious to pin-point that
Cognitive Radio Mobile Ad Hoc Network (CR-MANET),
internally neither consists of federal party to obtain the
spectrum usage information from the neighborhood nor
external third party provision (spectrum broker) that
empowers the distribution of the offered spectrum resources.
In classical ad hoc networks, the mobility of relay
nodes and the ambiguity residing with wireless channels are
the two key factors that affect the reliable distribution data
from source to destination [3]. TCP CR-MANET works
progressively in considering the temporal spectrum sensing,
channel switching and the cognizance of primary User’s (PU)
activity. Here, we proposed window based TCP for CR-
MANET. CR-MANET constructed from many mobile
secondary users interconnected to each other in a distributed
manner. CR-MANET can be deployed in various aspects of
Intelligent Transport Systems (ITS) applications [4].
The main challenges of transport layer in a classical
wireless ad hoc networks are [5]
1. Congestion.
2. Packet drops based on channel related problems.
3. Packet losses on mobility.
In case -1: The RTT value was increasing based on the
increased queuing delay of relay nodes. When RTT value goes
beyond the given limit, the relay nodes fail to forward packets,
likely this event degrades the performance of TCP.
In case 2. In the network, a packet drops due to channel
related problems or channel induced, likely of fading and
shadowing performance of channel.
In case 3. Relatively Packet losses occurred in the network,
when there was mobility related losses or Permanent losses
[6].
The source node would mistakenly consider the above
mentioned cases as congestion event.
All these losses are taken as inducing factors that are
applicable to CR-MANET. In CR-MANET, we rely an
intermediate nodes, which periodically piggyback the
spectrum information with Acknowledgement (ACK) and also
update’s Primary User’s arrival on time, explicitly informing
the source. Whereas, recently many dimensions of spectrum
sensing algorithms are focused [1]. Handling diverse channel
information deprived from various channels and attaining
insight about performance of these techniques remains an
open challenge, from end to end protocol perspective [7]. Our
protocol ensures prospectively channel switching event by
adhering to the momentous updation in band width of the
interfered link. On account of bottleneck spectrum permitting
low data rate there will be a sudden rise in number of packet
pertaining to the network at unit time period. Thus we propose
a TCP congestion window (CWND) which levers rapidly to
the transformation in the environment. TCP is a broad
recognized region of wireless networks where numerous
hypothetical models has been proposed that delivers
explanation and predicting its performances. Hence the
objective of TCP CR-MANET is to facilitate window based
methodology of the classical TCP, and increased applicability.
II. MOTIVATION
Here we analyze the one of the problems of the
preceding approaches of transport protocols like TCP
NewReno in CRAHN which drive us to have enhanced
performance with the proposed system TCP CR-MANET.
With respect to CRAHN, each node is furnished with a RF
transceiver. The key deeds of the Cognitive Radio Networks
are
1. Spectrum sensing.
2. Impact of primary user activity.
3. Spectrum change.
Primary Users are to be established as Poisson arrivals, by
having an “on” time as and as well as an “off” time as
respectively to all kinds of provided channels of the network.
A. Spectrum sensing
Before transmitting packet on the current channel, the
Secondary User or Cognitive User and the intermediate nodes
has to do a periodic check over the current channel according
to pre-defined sensing time for identifying a presence of
licensed users. Meanwhile the individual node’s data
transmission and the path framed with respect to it are
virtually disconnected on multi-hop network in spectrum
sensing state. When a sensing time ts is zero, the sensing is
disabled on that period, we observed that the congestion
window keep increasing until it reaches the maximum
capacity of the channel. TCP sender side obtains a
multiplicative delay which leads to timeout events during
spectrum sensing. On account of this TCP will be re-tuned to
slow start state by condensing the congestion window to 1
segment. When the sensing delay reach the maximum
retransmission timeout (RTO), RTO event could be triggered
which results the degradation of the end to end performance
[8]. Sensing time plays the vital role in deriving the optimized
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
[9] performance via 1) thorough recognition of primary user
and 2) effective utilization of the channel, which are the two
contradictory goals that need to be achieved by diagnosing a
better blending factor. The transport layer acclimatizes the
current rate of sensing state and decides optimal setting of
sensing time [11]. During this period, it detects the primary
users and utilizing channels, also minimize the interference on
primary users and maintains throughput.
B. Impact of Primary User Activity
A primary user’s activity is periodically detected
during spectrum sensing or data transforming. In case, on
arrival of the primary user, if secondary user’s operation
affects the current channel, the system will be in search of
various unoccupied channel in the spectrum. When the current
channel’s spectrum sensing is of periodic and well defined
interval, two activities will be performed as follows.
1. Available Channel set Discovery at various
spectrum bands
2. Harmonize with the subsequent hop neighbors to
derive mutually adequate channels in the set.
In this set, a risky factor arises i.e. the path towards
destination got disconnected and remains the same until the
system detects the new channel. The source will not be aware
of this event in that period of time, the affected node has to
transmit the explicit feedback notification to the source node.
The transport protocol [11] to differentiate these states, based
on the value of “on” and “off” stage (α and β) of Primary User
activity, that comprises four different patterns as follows.
1. High Activity .
2. Low – Activity
3. Short term Activity
4. Long term Activity
To have an enhanced TCP performance, without
affecting the primary user activity [12], the secondary users
should be allowed to use opportunistically the licensed
spectrum. When primary user frequently arrives on the
channel switches such as “on” and “off”, the congestion
window size will be low, this fact will degrade the
performance of TCP. To overcome this issue, the information
regarding spectrum hand off operations at the lower layers are
believed to be updated in the transport layer which facilitate to
differentiate, the packet losses that would appear due to
congestion and from the one occurring due to primary user’s
interference.
C. Spectrum change state
The effective utilization of the spectrum resources by
Cognitive Radio Technology. Secondary users have
opportunistically transmitting in the license bands on limited
duration. In the simulation, three different bandwidth is using
that 2/3Mbps, 4/3Mbps and 2Mbps. Primary users is modeled
as position arrival with “on” and “off” time respectively. TCP
must be regulating accordingly the new available bandwidth
[8][12]. TCP must be scaled the CWND to meet the channel
condition.
III. TCP CR-MANET
The CR-MANET nodes are establishing a single radio
transceiver, it can be regulated to whichever channel in
licensed spectrum. The channels in different spectrum bands
may have disparate channel bandwidth. Channels are using
spectrum band and data transfer with CSMA/CA and priority
queue at Medium Access Control (MAC) layer. In CR
networks, nodes are maintaining all free channels
(unoccupied) belong to different spectrum bands.
A. Network Modeling
The chain topology network is simulated in CR-
MANET.
( 0 to 200 s)
Node 0 Node 1 Node 2 Node 3 Node 4
Figure-1 Simulation Network
The Chain Topology wireless nodes are constructed and
symbolized in Figure-1. Where, there is primary user and
secondary users and the initial parameters are set. In the
cognitive radio scenario the transmission is initiated among
the CR users when the PU starts transmitting the packets, if
the channel used by CR user is the licensed channel of PU, it
switches to another available channel, allocating the current
channel to PU in order to maintain the performance of primary
user. Total simulation time is set to 200 seconds.
The packet transmission are initiated between the CR
users as follows. Node 0 to 4 configured as CR users and
Node 0 transfer packets towards Node 1, Node 1 transfer
towards Node 2, Node 2 transmit towards Node 3 and Node 3
transfer towards Node 4. The primary user Node 5 starts
transmitting to the Node 3 at specified simulation time 2.58
where the channel switching take place if the channel
allocated to the primary user. The wireless channel type is
used with priority queue where the maximum number of
packets in queue is set to 200. The AODV protocol is used as
a routing protocol.
B. Connection establishment
The three-way handshake protocol are used to
establish the connectivity of CRAHN. The source sends a
Synchronization (SYN) packet to the Sink. A relay nodes in
the routing path appends the following information to the SYN
packet: ID, Timestamp, and the tuple. On the receiver getting
the SYN packet, then it sends a SYN-ACK message to the
source.
C. Spectrum sensing
CR-MANET uses ED (Energy Detection) as the
sensing technology. Each SU performs sensing and data
transmission processes in an asynchronous time-division
manner. Therefore, the TCP sender regulates the timing and
duration in a routing path based on sensing schedules [13].
The node requires different schedules for sensing, it may
happen that a node receives messages from more than one
TCP sender [14]. Sensing time ts of each node is calculated
using following equation (1)
It is calculated with the parameters bandwidth (W),
standard function (Q-1), Probability of missed PU detection
(Pf), SNR (ᵞ), Probability of PU on period (Pon), and
Probability of PU off period (Poff). Transmission range is
assigned as 500Kbps and bandwidth for each channel is
assigned as 2Mbps.
D. Channel Switching
The channel switching is detected periodically in
spectrum band, when PU is arrival in the network then the
secondary users switched to available bandwidth. Whenever
the licensed channel allocated to the PU is used by the
secondary user, the CR user actively sense the arrival of the
PU. The nodes are switching their channels, to evaluate
communication characteristics link bandwidth W, link delay
(LT). The TCP sender aware these characteristics and RTT
then compute easily the bottleneck bandwidth (W′b). In
channel switching, the sender must update CWND and RTT
from equation (2) and (3). L′T is the link delay before channel
switching. A relay node change its channel, if primary user
communication is detected and also measure ssthresh.
RTTnew = RTTold + LT - L′T (2)
CWND = . W ′b RTTnew (3)
 IV. MODULE IMPLEMENTATION
When a relay node changes its channel, if PU
communication is detected, its bandwidth and link delay (L’T
i;i+1) can also be changed. This change is drastic when the
bottleneck bandwidth or RTT changes. Therefore,
appropriately updates its congestion control parameters cwnd
and ssthresh, using the feedback information received from the
relay node.
Fig.2. Congestion window Analysis
The bottleneck node is located in two cases, the first
case is that the bottleneck node is placed on a path from the
TCP sender to bottleneck node, the buffer will be empty by
the end of channel switching when the TCP sender stops
sending packets during channel switching. The second case is
that the bottleneck node is just the switching node and forward
to the TCP sender. Comparison of the CRAHN and CR-
MANET congestion windows representation is shown in Fig-
2, as well as the rectangular box denotes better performance of
CR-MANET congestion window.
Fig.3. Algorithm to update the cwnd
The TCP sender receives the feedback message after
channel switching, then it calculates the RTTnew by using
equation (4) and bottleneck bandwidth W ′ b . Estimation of
the CWND and RTT process shown in Fig 3.
RTTnew = L1,2 T + ..+ L i−1,i T + L′T i, i+1 + …. (4)
Fig.4. Throughput Analysis for CRAHN and CR-MANET
The network consists of the Primary user and the CR
users where the CR users actively participate in the packet
transformation when the PU is idle and handoff the channel
when the primary user enters the packet transferring mode.
The comparison result of CRAHN and CR-MANET is shown
Fig.4, and it shows that CR-MANET performs better
throughput for CRAHN.
V. CONCLUSION
TCP CR-MANET protocol implementation involves
adjacent connection with the fundamental link and network
layers, particularly during on channel switching and mobility.
An analysis result of throughput for TCP CR-MANET is in a
significant improvements. Transport protocols in cognitive
radio network selects a channel from a wide spectrum range.
The transport protocol and the relay node should be
redesigned to make an efficient use of available wireless
resource. TCP sender should consider both the bottleneck
node and buffer resource in the network. The sender in TCP
updates the CWND when either the bottleneck bandwidth or
RTT is changed after channel switching. Experimentally, it
has been proved better throughput with respect to TCP
CRAHN.
REFERENCES
[1] I.F. Akyildiz, W.Y. Lee, and K. Chowdhury, “CRAHNs: Cognitive
Radio Ad Hoc Networks,” Ad Hoc Networks J., vol. 7, no. 2, pp. 810-
836, Elsevier, July 2009.
[2] FCC, Second Memorandum Opinion and Order, ET Docket No. 10-174,
September 2010.
[3] M. D. Felice, K. R. Chowdhury, W. Kim, A. Kassler, L. Bononi, “End-
to-end Protocols for Cognitive Radio Ad Hoc Networks: An Evaluation
Study,” Elsevier Performance Evaluation, Vol. 68, No. 9, pp. 859-875,
2011. 492, 2011
[4] A.O. Bicen and O.B. Akan, “Reliability and Congestion Control in
Cognitive Radio Sensor Networks,” Ad Hoc Networks J., vol. 9, no. 7,
pp. 1154-1164, Elsevier, Sept. 2011.
[5] J. Liu and S. Singh, “ATCP: TCP for Mobile Ad Hoc Net-works,” IEEE
Journal on Sel. Areas of Comm., Vol. 19, No. 7, pp. 1300-1315, 2001.
[6] T. Melodia, D. Pompili, and I.F. Akyildiz, “Handling Mobility in
Wireless Sensor and Actor Networks,” IEEE Trans. Mobile Computing,
vol. 9, no. 2, pp. 160-173, Feb. 2010
[7] K.R. Chowdhury, M. Di Felice, and I.F. Akyildiz, “TP-CRAHN: A
Transport Protocol for Cognitive Radio Ad Hoc Networks,” Proc. IEEE
INFOCOM, pp. 2482-2491, Apr. 2009.
[8] A.M.R. Slingerland, P. Pawelczak, R.V. Prasad, A. Lo, and R. Hekmat,
“Performance of Transport Control Protocol over Dynamic Spectrum
Access Links,” Proc. Second IEEE Int’l Symp. New Frontiers in
Dynamic Spectrum Access Networks (DySPAN), Apr. 2007.
[9] M. Di Felice, K.R. Chowdhury, and L. Bononi, “Modeling and
Performance Evaluation of Transmission Control Protocol over
Cognitive Radio Ad Hoc Networks,” Proc. 12th ACM Int’l Conf.
Modeling, Analysis and Simulation of Wireless and Mobile (MSWIM
’09), pp. 4-12, 2009
[10] D. Sarkar and H. Narayan, “Transport Layer Protocols for Cognitive
Networks,” in Proc. of IEEE INFOCOM Computer Communication
Workshop, 2010, pp. 1-6.
[11] K. Sundaresan, V. Anantharaman, H. Y. Hsieh, and R. Sivakumar,
“ATP: A Reliable Transport Protocol for Ad Hoc Net-works,” IEEE
Trans. On Mobile Computing, Vol. 4, No. 6, pp. 588-603, 2005.
[12] Kaushik R. Chowdhury, Marco Di Felice, and Ian F. Akyildiz, “TCP
CRAHN: A Transport Control Protocol for Cognitive Radio Ad Hoc
Networks” IEEE Transactions On Mobile Computing, Vol. 12, No. 4, pp
790 – 803, APRIL 2011.
[13] N. Parvez, A. Mahanti, and C. Williamson, “An Analytic Throughput
Model for TCP NewReno,” IEEE/ACM Trans. Networking, vol. 18, no.
2, pp. 448-461, Apr. 2010.
[14] A. Al-Ali and K. R. Chowdhury, “TFRC-CR: An Equation-based
Transport Protocol for Cognitive Radio Networks,” Elsevier Ad Hoc
Networks Journal, Vol. 11, No. 6, pp. 1836-1847,2013.
 An Improved Energy Efficient Chain Based Routing
in Wireless Sensor Networks
Zanele Mahlobogwane Prof. SM Ngwira
Dept. of Computer Systems Eng Dept. of Computer Systems Eng
Tshwane University of Technology Tshwane University of Technology
Pretoria, South Africa Pretoria, South Africa
Abstract—Wireless Sensor Networks (WSNs) consists of
sensor nodes that are battery powered, these nodes are often
deployed in remote areas and left unattended for long periods
between battery charges. According to the studies, over sixty
percent of energy is consumed during the transmission and
reception process of the network. Due to the reason stated above,
protocols and algorithms have been designed to manage energy,
to make sure that the life span of the network is prolonged. This
study proposes the improvement of the Energy Efficient Chain
Based Network protocol (ECBSN). This protocol has challenges
in a way it forms the chains using fixed number of nodes per
chain instead of varying the number of nodes in a chain at
different levels. This cause the chain furthest to the Base Station
(BS) to die faster if it’s Cluster Head (CH) is the chosen leader
which has to relay all aggregated data to the sink just like in the
Improved Energy Efficient Chain Based Network (IEBSN). We
then propose the Improved Energy Efficient Chain Based
Routing (IECBR) in WSNs. This study will consider varying the
number of nodes at different levels, that is more nodes will be
assigned to the chain closest to the BS and the further you are to
the BS few nodes will be assigned. It will also consider the level of
the energy and the distance of the node which is the CH to the
BS. The simulation results showed a great improvement in the
lifespan of the nodes in the network as compared to ECBSN.
Keywords—Wireless Sensor Networks; Sensor Nodes; Energy
Efficient , Hierarchical Clustering.
I. INTRODUCTION
Wireless Sensor Networks (WSNs) consist of sensor nodes
which cost less and have small power[1].The structure of the
sensor node consists of four basic components: a sensing unit
formed from a sensor and an analogue to digital converter, a
processing unit formed from a microcontroller with memory, a
communication unit which is a transceiver and a power unit
whose energy source is usually a battery.
Normally a WSN have numerous sensor nodes which are
thickly conveyed in the region of interest, and the sink which
may not be exceptionally a long way from the sensor nodes.
The sensor nodes are sent in a zone to screen changes in
distinctive applications e.g. natural applications, military
applications, human services applications, security and
observation applications, etc.
The sensor will monitor the physical phenomena and generate
an electrical signal based on what it has observed. The signal
will be passed to the analogue to digital converter so that the
analogue signal can be understood by the microcontroller,
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Dr. T Zuva
Dep. Computer Systems Eng
Tshwane University of Technology
Pretoria, South Africa
which will process the signal and take the decision about it.
The communication device will exchange the information
between the nodes. The communication device converts the bit
stream coming from the microcontroller to radio waves [2].
The design of the WSNs is influenced by a number of factors.
According to [3] these factors are fault tolerance, scalability,
production cost, operating environment, sensor network
topology, hardware constraints, transmission media, and
power consumption.
The WSNs must be designed in such a way that even when
one node fails due to power failure or physical damage the
network should be able to deliver reliable messages to the end
user. Looking at the factors listed above, the major factor that
needs to be considered when designing the WSNs is power
consumption. Since the WSNs is battery powered, power from
the nodes can be drained very fast if there is no adequate
power management strategy that is put into place.
To address this challenge a number of protocols and
algorithms have been proposed to manage power in the entire
WSN.
Power consumption in the network can be divided into three
main categories: sensing, communication and data processing.
According to [4] and other researchers over sixty percent of
the energy is consumed in the transmit, receive, idle and sleep
modes. In other words most of the energy is consumed during
the communication process. According to [5], some sensors
can consume more energy than the radio. This means choosing
the right hardware components is also very important when
designing WSN. The type of hardware components you
choose for your design contributes significantly to the total
energy consumed in the entire network.
This paper is arranged in this manner: Section 2 is the
Literature Survey, Section 3 is the Proposed technique,
Section 4 is Conclusion.
II. ROUTING PROTOCOLS
This literature review will focus on energy efficient routing
protocols for wireless sensor networks taking into
consideration the hierarchical chain based protocol. Routing
is very important for energy efficiency in WSNs. [6] shows
that a good routing protocol can avoid challenges faced in
WSNs such as Scalability, fault tolerance, Energy Efficiency,
Quality of service, etc.
A. Routing Protocols in WSNs Where P gives you the desired percentage of how may CH
The routing protocols can be classified into three groups, that you want, r the current round, G is the arrangement of nodes
is Flat routing, Hierarchical routing, and Location Based
that have not been Cluster Heads in the last 1 rounds. Each
routing. P
1) Flat routing
CH telecast a message to different nodes in the network
In this sort all sensor nodes assume the same part, they are
utilized to sense information furthermore as passage. In flat notifying them that it is the cluster head for that round. At that
routing all sensor nodes are dealt with uniformly. When the point every sensor node will decide to which group it need to
node need to send data to the other node, it will calculate the have a place and notifies the CH that it want to be member of
shortest path from itself to the base station by checking its group utilizing Carrier Sense Multiple Access (CSMA).
different paths as shown in the figure below. Directed
Diffusion (DD) [7], Secure Sensor Protocol For Information After the setup stage a steady stage will start. In the steady
via Negotiation (Secure-SPIN) is used stage the nodes will send message to each CH utilizing the
designated Time Division Multiple Access (TDMA) schedule
for every sensor node that belongs to its group. The sensor
nodes will send their information to their closest CHs and the
CH will total its information and after that send the combined
information to the BS. This is the way this protocol lengthens
the lifetime of the network.
Figure1: Example of flat routing
Source Node Advantages of LEACH
Intermediate node  Aggregating its data to the data of the sensor nodes in
Base station a cluster reduces the information that is sent to the
BS.
The disadvantage of the protocol is that if the BS is far from
the source node, it will drain the energy of the nodes faster.  Only few sensor nodes that are selected as CH will be
Direct transmission will also experience the problem of transmitting the data to the BS, by so doing the
redundancy where one information is send many times by number of direct communication to the BS will be
different sensors. These problems are solved on the
reduced as compared to a flat topology.
hierarchical communication protocols.
2) Hierarchical Communication Protocols
In this type the network forms a group or a cluster where a  TDMA allows other sensor to sleep until their time
cluster head (CH) will aggregate or fuse the data coming from slot comes and this saves energy.
different nodes and send the aggregated data to the BS.
In [8] the authors propose Low-Energy Adaptive Clustering  The use of Clusters makes scalability to be possible.
(LEACH) that uses randomization of Cluster heads (CH) with
a specific end goal to disperse the energy load among the Disadvantages of LEACH
whole sensor networks equally, LEACH operation is divided  The cost of overhead to form CH is too high
into two rounds in which the first round is the set up stage and
the second round is the steady stage. In the set up stage that  It does not consider the distance from the CH to the
where CH and clusters are created. The sensor nodes elect BS
themselves as cluster head based on the suggested percentage
P and their previous record of being a CH. This is done in a  The amount of energy left in the battery of the CH is
way that all sensor nodes which were not CH in the previous not considered.
round, will produce an arbitrary number somewhere around 0
and 1, and compare it with the limit or the set threshold and if
the number is less than that limit then the sensor will be a CH.
The formula below is used in LEACH to show how a limit
value is set:-
 P
 if n  G
  1
T ( n )  1  P *  r mod  (1)
  P

0 otherwise
LEACH Topology using only ten nodes following the shortest
route [10] propose an energy efficient chain based routing protocol
in wireless sensor network (EECB). This protocol plays as an
improvement of PEGASIS there by considering the distance
between nodes and the base station. PEGASIS uses the greedy
algorithm which start with the furthest node from the base
station but doesn’t consider the distance between the next
nodes. This causes a long link between other nodes in the
network, as a result EECB checks if the distance between the
communicating nodes is less than the threshold distance and
makes that node the neighbour. It also chooses the leader node
base on the amount of energy left in the battery of that node
and how far is that node to the BS using these two equations:-

Figure 2: LEACH Topology di  xi  xBS 2   yi  yBS 2

(2)
Paper [9] proposes Power Efficient Gathering in Sensor E
Information System (PEGASIS). In PEGASIS the sensor Qi  residuali (3)
nodes utilizes the greedy algorithm when shaping chain di
among the sensor nodes so that every node will get The drawback of this protocol is that even though it performs
information from and send information to a nearby neighbour better that PEGASIS it also uses a long chain that also causes
and these nodes alternate transmitting to the Base Station a bottle neck to the node closest to the BS it also have a
(BS), in this way reducing the measure of energy spent per problem of delay.
round. On the other hand, the BS can figure this chain and EECB RESULTS
telecast it to the sensor nodes. In PEGASIS node one will
communicate with node two and node two will take
information from node one and fuse it with its own
information then send it to the following close node. Each
node in this methodology has the ability of aggregating
information to lessen the measure of information sent in the
middle of sensors and the base station. At the point when a
sensor node deplete all its energy, the chain is formed again
utilizing the same calculation. This PEGASIS utilizes a long
chain; this may over-burden the pioneer sensor node, and
make it die early. This can likewise be enhanced by shaping
cluster chains with one and only leader that can send its
information to the neighbouring Cluster Head.

PEGASIS RESULTS
Figure 4: EECB[10]

[11] proposed an energy efficient chain based network

(ECBSN). This technique proposes a parallel chain formation
that will ensure maximum profitable use of energy which is
considered a higher priority in the sensor networks. Their
method is compared with PEGASIS[9] and is proved to
reduce energy consumption in a network by 15-20%. It also
proved to raise the reliability of the network. Power Efficient
Gathering in Sensor Information Systems (PEGASIS) in[9]
due to its one long chain poses a communication delay, that
also lead to an increase in energy consumption because of the
increase number of packets as the chain approaches the base
station. ECBSN can be further improved by introducing an
extended multiple layer hierarchical chain based protocol that
makes all the CH communicate to one another in a chain form
until the BS. The MAC layer protocols can also be introduced
Figure 3: PEGASIS[10] in the active and sleep cycle.
 ECBSN RESULTS
Figure 5: ECBSN
3) Location Based Routing
This type of network utilizes the position information to relay
the information to the BS. To determine the distance between
the neighbouring nodes it make use of the signal strength.
Some of the location based protocol has been explained in
[12] Location‐Aided Routing (LAR) in mobile ad hoc
networks.
III. PROPOSED
To improve the deficiency explained above, our study will
attempt to improve an Energy Efficient Chain Based Networks
(ECBN) which is part of the hierarchical Protocol that has
shown improvement of energy compared to the other
categories.
A. Assumptions:-
 The base station to be far away for the sensing field
and the sensor nodes are stationery.
 In the beginning sensor nodes are assigned the same
energy.
 It consumes the same energy when transmitting and
when receiving.
B. Radio energy model
The protocol will use the same energy model as that one
described in [2, 8] for communication energy. We chose the
same radio model as in LEACH and PEGASIS for
consistency. The formulas underneath depict the energy used
to transmit and get information.
Where the energy consumed to transmit b bits over a distance
d is:
ETx (k , d )  Eelec  b  Eamp  d k (4)
Energy consumed by a sensor to receive b bits is:
ERX (k )  Eelec  k (5)
This convention will expect free space and direct line without
hesitation where k=2.
Where Eelec stand for energy for the hardware and for our
calculation we use 50nJ/bit, Eamp energy for the amplifier is
100pJ/bit/m2, b being the number of bits, and d as the space
between the nodes and the coefficient k being the signal
strength loss. The value of k will depend on the data
propagation model utilized in the system that is: direct line of
sight and when there is no obstacle k=2 and for environments
with extreme impedance k =4.
This protocol will assume cases where we don’t have
obstacles and direct line on sight where k=2.
Improved Energy Efficient Chain Based Routing Protocol
(IECBR).
This study proposed the Improved Energy Efficient Chain
Based Routing (IECBR) in WSNs to enhance the lifespan of
the system and diminish the utilization of energy during
correspondence (that is transmission and reception). The
operation of the Protocol will be founded on stages as follows
 Clustering Stage
 Chain formation Stage
 Data transmission stage
In this protocol clustering is not performed every round since
that consumed a lot of energy of the sensor nodes.
1) Clustering stage
The clustering stage consist of the following sub stages
 Cluster Head Selection
 Cluster Formation
On the Cluster Head formation we assume all the nodes know
the position of the neighbour nodes and the initial energy of
the nodes is also known that is 0.5J per node. Each node will
broadcast its information to all the nodes that are in its radio
range and the nodes that have the strongest signal will be
considered as the neighbour.
During the first round, the node whose distance to the BS is
smallest will be CH since all the nodes will be having the
same energy. After the energy of the CH has been depleted
then remaining energy in the battery will be calculated then
divided by the distance between the node and the BS, if the
outcome of that for a particular node is more than the other, These protocol defer from ECBSN in a way that it varies the
then the node will be taken as the CH number of nodes per cluster and the way it selects the leader
node per chain. In this protocol there will be longer chains in
2) Chain formation stage the area closest to the BS in order to distribute the energy
The chain development stage is likewise done in two rounds. evenly in the entire network and to avoid nodes closest to the
There will be chain arrangement inside of groups then taken BS dying earlier just like in[13] we will have a leader node in
after by chain arrangement of the CHs. After the formation of each chain which is the node having more energy than the
clusters the CHs are selected, these CHs create a chain other and the head leader will be the node that is closest to the
between the nodes within the clusters, so that each node BS. The leaders from each chain will be nodes closest to the
receives data from its previous neighbour just like the one in BS and with more energy than the other in the chain. The head
[9]. The chain in the cluster is formed starting from the leader will total all the information originating from
furthest node from the CH. Then the node which is close distinctive leaders from diverse chains and send it to the BS.
neighbour of the selected nodes based on the signal strength is
IECBR RESULTS
selected as the next node of the chain. It continues with the
process until the last node is added to the chain. After
construction of the chain the CH will select the TDMA
schedule for each node to avoid collision within the cluster,
this will specify the time in which the node will send the data
to the CH.
Chain formation amongst the CHs start by CH sending their
position and the remaining energy to the BS. Then the BS will
create chain of CHs based on the received information that is
the CHs with shortest distance and more energy will be the
leader CH. The BS will then apply greedy algorithm in
forming the chain within CHs.
3) Data Transmission Stage
After the procedure of chain development, the CH will show
this parameters: MIN limit, Max limit and Change variable to
every one of the nodes in its groups. The nodes will then
utilizing its own particular time space to send the information
to the neighbour node. The neighbour node will pass its
information in addition to the one got from the neighbour to
the following node, if the following node is the leader, it will
then total the information got with its own information then
send to head leader. The procedure will proceed until every Figure 6: IECBSN
one of the CHs from distinctive chains have sent their
information to the head leader, then the head leader will IECBR GRAPH
aggregate all the information originating from CHs and sent it
to the BS. ECBSN AND IECBR
2000
I ECBR
The sensor nodes monitors the environment constantly. On the 1990 ECBSN
off chance that the detected estimation of the hub is not as
1980
much as Min limit, then the CH doesn't total its detected worth
with the information got from its past node in chain. By so 1970

doing the CH will specifically send the information got to the 1960
head leader without collecting it. The Min limit spares the No of Nodes
1950
energy of the CHs by not permitting them to do information
accumulation. 1940

1930
At the point when the CHs get the information from its group
1920
individuals, the transmission of information will begin in the
CHs chain. This will ensure that the CH totals their 1910

information with the past neighbor's information and send it to 1900

0 200 400 600 800 1000 1200 1400 1600 1800 2000
the head pioneer. Finally the head leader send all the No Of Rouds
information to the BS.
Figure 7: ECBSN against IECBR
 IV. RESULTS
The simulation was done using MATLAB programming. In
this simulation an arbitrary system of 100 nodes densely
deployed in a territory of 100 by 100 meters. We expected the
BS to be situated at (100,100). Every one of the nodes in the
beginning they have the same energy of 0.5J and the energy
taken to total the information to be 5nJ/bit/signal. In this
paper we assumed the BS to be at (100,100).
Figure 6 shows the results of the proposed method, from
which it can be noticed that the chains are formed with
varying nodes and the more we approach the base station the
chain have more nodes than the chain furthest to the BS.
Figure 7 shows the MATLAB result of comparing ECBSN
against IECBR. The MATLAB simulation has been done with
100 nodes. IECBR performs better than ECBSN showing that
varying the number of nodes at different levels, taking the
level of energy into thought and the space between the leader
VI. REFERENCES
[1] Gholamzadeh, B., and Nabovati, H.: ‘Concepts for designing
low power wireless sensor networks’2008 pp. Pages
[2] Raghunathan, V., Schurgers, C., Sung, P., and Srivastava, M.B.:
‘Energy-aware wireless microsensor networks’, Signal
Processing Magazine, IEEE, 2002, 19, (2), pp. 40-50
[3] Akyildiz, I.F., Weilian, S., Sankarasubramaniam, Y., and
Cayirci, E.: ‘A survey on sensor networks’, Communications
Magazine, IEEE, 2002, 40, (8), pp. 102-114
[4] Akyildiz, I.F., Su, W., Sankarasubramaniam, Y., and Cayirci,
E.: ‘Wireless sensor networks: a survey’, Computer Networks,
2002, 38, (4), pp. 393-422
[5] Alippi, C., Anastasi, G., Di Francesco, M., and Roveri, M.:
‘Energy management in wireless sensor networks with energy-
hungry sensors’, Instrumentation & Measurement Magazine,
IEEE, 2009, 12, (2), pp. 16-23
[6] Al-Karaki, J.N., and Kamal, A.E.: ‘Routing techniques in
wireless sensor networks: a survey’, Wireless Communications,
IEEE, 2004, 11, (6), pp. 6-28
[7] Intanagonwiwat, C., Govindan, R., Estrin, D., Heidemann, J.,
and Silva, F.: ‘Directed diffusion for wireless sensor
networking’, IEEE/ACM Trans. Netw., 2003, 11, (1), pp. 2-16
[8] Heinzelman, W.R., Chandrakasan, A., and Balakrishnan, H.:
‘Energy-efficient communication protocol for wireless
node and the Cluster Head (CH) enhances the life compass of
the network.
V. CONCLUSION
Our method showed that taking into consideration the varying
of nodes in a chain at different levels has an effect of the
lifespan of nodes in a network. The energy in each node has a
bearing on how the chain are formed and also the space
between the leader node and the CH. The IECBR shows
potential in ensuring a prolonged lifespan of a network. The
protocol can be further enhanced by improving the algorithm
when forming chains to avoid longer link between neighbours.
The network can also be partitioned into three so that every
partition will have its head leader which will send the data to
the BS rather than having only one Head leader. For larger
network the sleep and wake sections may also be applied to
avoid sending one message a number of times.
microsensor networks’, in Editor (Ed.)^(Eds.): ‘Book Energy-
efficient communication protocol for wireless microsensor
networks’ (2000, edn.), pp. 10 pp. vol.12
[9] Lindsey, S., and Raghavendra, C.S.: ‘PEGASIS: Power-
efficient gathering in sensor information systems’, in Editor
(Ed.)^(Eds.): ‘Book PEGASIS: Power-efficient gathering in
sensor information systems’ (2002, edn.), pp. 3-1125-1123-
1130 vol.1123
[10] Yongchang, Y., and Yichang, S.: ‘An Energy-Efficient Chain-
Based routing protocol in Wireless Sensor Network’, in Editor
(Ed.)^(Eds.): ‘Book An Energy-Efficient Chain-Based routing
protocol in Wireless Sensor Network’ (2010, edn.), pp. V11-
486-V411-489
[11] Mahajan, S., and Mahotra, J.: ‘A novel chain based wireless
data sensor network (ECBSN) technique’, International journal
of computer science and telecommunications, 2011, 2, (8), pp.
13
[12] Ko, Y.B., and Vaidya, N.: ‘Location‐Aided Routing (LAR) in
mobile ad hoc networks’, Wireless Networks, 2000, 6, (4), pp.
307-321
[13] S. Mahajan, J.M.a.S.S.: ‘Improved Enhanced Chain Based
Energy Efficient Wireless Sensor Network’, Wireless Sensor
Network, 2014, 5, (4), pp. 6
 An Integrated Model for ICT Governance and
Management
Applied to the Council for Evaluation, Accreditation and Quality Assurance of Higher
Education Institutions in Ecuador (CEAACES)

Carlos W. Montenegro1; Denys A. Flores2

Escuela Politécnica Nacional
Department of Informatics and Computer Science (DICC)
Quito, Ecuador
1
carlos.montenegro@epn.edu.ec
2
denys.flores@epn.edu.ec

Abstract—As an effect of the modernization policy in the and General Norms. Out of these, the Internal Control Norms
administrative processes inside the Ecuadorean Central (ICN) of the General State Comptroller‘s Office [5], the
Government, the usage of Information and Communication Agreement for Information Security (IS) of the National
Technologies (ICTs) has increased during the last years within Secretariat of Public Administration [6] and the Agreement for
more than 300 of the most important and high-ranked public Process-Based Government (PBG) [7] make direct reference to
institutions. Likewise, the citizen access to Internet and e- ICT Management Processes. In Table I, the main
Government services has grown due to the democratization of the characteristics of the legal framework are described.
governmental ICT platform to ensure access to the most
important public services, including those offered by Higher TABLE I. ECUADOREAN LEGAL FRAMEWORK FOR PUBLIC SECTOR‘S
Education Institutions (HEIs). However, the applicable ICT MANAGEMENT
regulatory and legal framework in the public sector has very
little compliance with governance and management Process
Authority
considerations. Our work develops a combined model for ICT Instrument Description
governance and management based on academic models for The processes that should be
General State
governance and strategic alignment, aided by professional Internal Control implemented in public institutions are
Comptroller‘s
practices in the field, fully compliant with the regulatory Norms (ICN) listed and briefly explained from a
Office (CGE)
mechanisms that operate within the Central Government of controlling perspective.
National
Ecuador. Also, our work analyzes the results, limitations and The directives for using ISO 27001 [8]
Secretariat of Agreement for
future work regarding the applicability of this model inside the as National Technical Norm for
Public Information
Council for Evaluation, Accreditation and Quality Assurance of implementing Information Security
Administration Security (IS)
Higher Education Institutions (CEAACES) in order to support Management Systems (ISMS).
(SNAP)
the process of continuous improvement of public higher Agreement for The components for establishing a
Ecuador
education in Ecuador. President‘s
Process-Based process-based government, such as
Government governing processes (subjects) and
Office
Keywords—ICT; IT; governance; management; cobit; model; (PBG) products/services (adjectives)a.
iso; practices; standard a. The adjectives depicted in the agreement are translated as products and services which are provided to
the governing processes that are shown as subjects in this agreement.

I. INTRODUCTION
According to the Art. 226 of the Ecuadorean Constitution
[1], the Executive Function (Central Government) is part of the
public sector, comprised of more than 300 dependencies where
more than a hundred of e-Government applications are used [2]
[3]. Currently, there are more opportunities to use these e-
Government services, as the central Government has promoted
citizen access to ICT platforms. In fact, in between 2010 and
2015, the percentage of Internet users amongst the Ecuadorean
population has increased from 17 to 51 per cent [4].
In case of governmental dependencies, these maintain ICT
platforms to support e-Government services along with their
internal processes. Management of these platforms is based on
a regulatory and legal framework, which includes Organic and
Common Laws, Statutory Instruments, Agreements, Decrees
The Central Government‘s public institutions, created
before 2013, have equivalent processes to those described in
Table I. These processes are known as Governing, Value-
Aggregator and Enabling Processes (consulting and support).
The process known as ―Information and Communication
Technologies (ICTs)‖ is considered inside the Enabling
Processes (adjectives in PBG). ICTs is seen as an ―adjective‖
process, makes it evident that both institutional and ICT
processes must be aligned so that the relevance of the latter can
be assured. Nonetheless, these criteria and alignment processes
are not detailed enough in the Ecuadorean Legal Framework
for ICT Management.
Because of the aforementioned reasons, in our current
work, we propose de design of a model-type artifact in order to
integrate ICT Governance and Management for the Ecuadorean
public sector.
This artifact is developed by consolidating systematic
alternatives, and partial but complimentary solutions, both built
over the basis of the analyzed problem, state-of-the-art research
and the best practices to solve this issue; the artifact strength
relies on keeping internal consistency of its components with
the Ecuadorean Framework for ICT Management in the public
sector. Then, in order to assess the applicability of this artifact-
like model, the Council for Evaluation, Accreditation and
Quality Assurance of Higher Education Institutions
(CEAACES) is considered as a case study so that criteria for its
implementation can be devised from the evaluation results.
This research approach is based on the Design Science
Paradigm proposed by Hevner, Ram, March and Park [9].
From this point forward, this article is organized as follows:
In Section II, considerations regarding governance, ICT
Management and strategic alignment are discussed in depth.
Later, these considerations are used to define the model which
is fully defined in Section III. In Section IV, the evaluation of
the model is described, as it was applied in CEAACES,
Finally, in Section V, a discussion of the obtained results
and further work is presented.
II. GOVERNANCE AND IT MANAGEMENT CONSIDERATIONS
According to Webb, Pollard and Ridley [10], whilst IT
executives and managers (Information Technology
Management - ITM) deploy and supervise business strategies,
other structures handles organizational policies, culture and IT
investment (Information Technology Governance - ITG). In
other words, ITG defines and spreads mechanisms required to
ensure the current and future business-technology alignment
objectives [11] whereas ITM must ensure that the governance
mechanisms are in place in order to fulfil the corresponding
strategies [12].
ITG can be implemented by combining diverse relational
structures, processes and mechanisms; plenty of research has
been done in the last decade about how to implement ITC in a
structured and process-oriented way [13] [14]. Recently, ISO
38500 [15] has become the first international standard that
depicts general directives about implementing ITG inside
organizations, however it does not include mechanisms,
techniques or specific tools, so it lacks of practical
contribution.
Regarding ITM, the Business-Driven Information
Management model (BDIM) is the application of models,
practices, techniques and tools for mapping and quantitatively
evaluate the interdependencies amongst the business
performance and IT solutions in order to improve IT service
quality and the related business outcomes [16] [17] [18].
Currently, ITIL [19] and COBIT [20] are the most developed
frameworks for IT service delivery and governance,
respectively.
For the integrated implementations of ITG and ITM two
models can be found in specialized literature Business-
Objectives driven IT Management (IT-MBO), and the
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Guidelines and Areas-based Model. The first is conceived
around a set of key concepts: objectives, key performance
indicators (KPIs) and organizational perspectives (financial,
customer, processes and learning/growth) [21]. In this case,
COBIT is a clear adaptation of this model.
A combined framework has already been created based on
Guidelines and Areas [22] [23]. After comparing this model
with others, it has been reported to be similar to COBIT 4.1. In
COBIT 5 [20] all the criteria of the previous version are
covered and enhanced, so COBIT 5 can be considered as the
most complete framework for combining ITG and ITM [24].
Returning to a broad approach, in spite of COBIT and ITIL
providing an excellent choice of IT frameworks for governance
and service provision, respectively; they both have limitations.
In one hand, the definition of impact, risk and other
measurements are vague and not necessarily quantitative [21]
[25]. On the other hand, they do not provide implementation
details, which obliges the usage of other guides or standards for
filling the gaps, as it has been reported in previous work [26]
[27].
A. IT Strategic Alignment
IT alignment aids institutions by maximizing IT investment
return, generating competitive advantage through information
systems, and delivering orientation and flexibility to react to
new opportunities [28].
Some models have been identified and described as
alignment research trends [29] [30] [31]. These models adopt
the Strategic Alignment Model (SAM) [32] which has been
widely adopted [31], and has been considered as a design
paradigm [9] on which our proposal is developed. The SAM
model is comprised of four domains which are related in
various ways. Each of them describes one perspective that
displays both links (strategic adjustment and functional
integration) in order to obtain a proper IT alignment inside an
institution (Fig. 1).
Fig. 1. IT Strategic Alignment Domains [45]
For example the second perspective, or technologic
transformation, describes the business strategy implementation
through IT. Then, as business strategies are conductors for
alignment, they provide vision and objectives to meet business
goals by changing and organizing IT infrastructure and its
processes at operative level.
A case in point for strategic alignment can be found in
COBIT 5, which states that the needs of each interested party
must be transformed into a feasible corporative strategy
through its ‗goal cascade process‘ [20]. This process translate
the needs into useful and customized business goals, IT-related
goals and specific catalyst goals. Thus, the whole goal cascade
process in COBIT is fully compliant with the second
perspective in the SAM model.
B. ITG in the Public Sector
A plethora of research have been done in recent years that
address the challenges for ICT management in the public sector
[33] [34] [35]; particularly, regarding centralized management
figures and decision making. For example, according to
Winkler [36], unlike the private sector, which is focused on
generate value for investors, the public sector business core is
based on the services it can provide to the community (public
value). Furthermore, beyond improving governmental
operations, it is required to consider political and social effects
due to the fact that in the public sector, IT governance
objectives are multifaceted and more difficult to measure than
in the private sector. Here, Winkler analyzes the fact that, in
contrast with private economy, public organisms create even
more centralized structures for decision making, as it happens
in the Ecuadorean Central Government, as described in the
instruments depicted in the Ecuadorean Legal Framework for
ICT Management in the public sector (Table I).
C. ITG Implementation
As mentioned before, there are many advantages of
implementing ITG since different governance perspectives
contribute to achieve different results. Hence, it is important to
analyze possible implementation patterns that can be used by
organizations.
A first approach establishes that institutions assign their
‗decision rights‘ to different ‗archetypes‘ [37]. The results
show three important patters for decision making; one of them
depicts the preponderance of business, which is very similar to
the Ecuadorean public sector schemes.
Secondly, a baseline for ITG practices can be taken, as
suggested by De Haes and Van Grembergen [38] whose
application was evaluated in financial service institutions in
Belgium. This research shows that some of these practices are
more effective and easier to implement than the others. The set
of practices perceived as ‗very effective‘ are Management
Committees, management communication lines (CIO reports to
CEO), and Executive Boards in which CEOs are active
members.
Finally, it has also been suggested [39] to select a set of
structures, processes and relationships for ITG out of case
studies analyzed within specialized literature review. From the
global results, characteristics are selected considering a
frequency value which is set depending the number of times
that a characteristic is applied in the case study. These
characteristics can be structures, processes or any other
relevant relationships that can be generalized and applied to
ITG implementations.
III. INTEGRATED MODEL FOR ITG/ITM
Based on all the previous considerations, we propose an
integrated model for IT Governance and Management, based
on the SAM Model proposed by Henderson and Venkatraman
[32]. In Table II, the components of the Integrated Model for
ITG/ITM are summarized.

TABLE II. INTEGRATED MODEL FOR ITG/ITM

ITG ITG/ITM
Implementation
Public Sector IT-MBO Guidelines and Areas
Criteria
- IT Alignment Task
Integration in roles
Governance and
(ITG) responsibilities.
- Internal and - Management
External Committee
Hybrid or
Approach establishment.
Centralized Guidelines:
- Organization - IT organizational
decision COBIT as - IT Objectives
as a Whole Basic structure
Structure structure Governance Domain. - IT and Business Alignment
- Strategic Framework considerations.
organized as a ISO 38500 Areas
Future - CEO as member or
duopolistic - Strategy
- Benefit- Executive Board.
archetype.
Making - CIO reporting to
- Adequate CEO
Investment - Existence of
- Delegation Manager, as a role
for business/IT
relationships.
 ITG ITG/ITM
Implementation
Public Sector IT-MBO Guidelines and Areas
Criteria
- Acquisition. - IT Strategic
- Service Q&A. Planning.
- Information - IT Project
Processes Architecture. Portfolio
Management Management.
- IT Strategic - IT Project Follow-
Planning. Up.
- Mechanisms for
Communication
Improvement in
between IT and - Shared
Organizational understanding of
Sectors. business objectives.
Relationships - Development of - ITG awareness
required skills campaigns.
for decision - Shared management
making and of IT accounts.
critical ITG
processes
execution.
A hierarchical
COBIT: Goal
model like the
Cascade. Guidelines:
Strategic Technology
ITIL: - IT and Business Alignment.
Alignment SAM Model Transformation
Duopolistic Areas:
(BITA) Alignment
Committees (IT + - Strategy
Mechanism.
Business).

Business- Guidelines:
A service- - Business
Driven
oriented Structures/Processes/Mechanisms.
Information
Management model based - IT Processes.
Management
(ITM) on COBIT COBIT: - IT Balanced Score card (BSC).
Model
- Internal and ITIL Management - Best-practices.
(BDIM)
Approach. Domains. - Auditing.
- Departments ITIL: - Improvement.
and Specific - Service Strategy - Innovation.
Individuals. Strategies: (SS). Areas:
- Present - Information - Service Design - Service.
Strategy. Security. ISO 27000 (SD). - Resources.
- Projects and - Business ISO 22301 - Service Transition - Risk.
Operations. Continuity. ISO 31000 (ST). - Development.
- Cost and - Risk ISO 12207 - Service Operation - Architecture.
Quality. Management. CMMI (SO). - Projects.
- Budget - Software Others. - Service - Support and Q&A.
Accountability. Development. Configuration (CM). - Investment.
- Current work. - Other - Outsourcing.
strategies. - Compliance.
- Improvement.
- Innovation.

The strategy for defining the model is explained as
follows:
i. COBIT is adopted as generic practice for Governance
and Management, considering its content and
completeness level.
ii. For ITG, a basic framework is adopted: structures,
processes, relationships and strategic alignment. Also,
the recommendations for implementation are aligned
with the local legal framework for the public sector so
that compliance can be ensured throughout the whole
model.
iii. For ITM, ITIL and COBIT service-related processes
are considered along with the eventual usage of
specific norms and guidelines according to the needs.
The model components or mechanisms are comprised of
recommendations, processes, strategic alignment, guidelines
and norms. These mechanisms must be applied as follows:
1) If strategic alignment employs technology
transformation in the SAM Model (Business Strategy -> IT
Strategy -> IT Infrastructure): The COBIT Goal Cascade
technique has to be applied in order to define the main ITG
COBIT processes to be implemented.
2) According to the organization restrictions and
reality: Some ITG recommendations depicted in the model
should be ignored. For instance Acquisition, IT Project
Portfolio Management and IT Project Follow-Up are
centralized activities in the public sector that may be
overlooked. Whenever it is feasible, recommendations
should be adapted to meet the closest ITG COBIT
processes.
3) Adopting COBIT processes related to ITM:
Alternatively, there are ITIL services that can be considered;
however, in doing so, these services must be integrated with
ITG components. Then, it has to be defined if processes
need to be decomposed, and if the granularity level given to
them requires applying specific guidelines and norms.
Eventually, current process capabilities and their maturity
may be assessed by using a maturity/capacity model, or else
including/discarding processes according to specific
organizational needs.
IV. APPLYING THE ITG/ITM MODEL IN THE COUNCIL FOR
EVALUATION, ACCREDITATION AND QUALITY ASSURANCE
OF HIGHER EDUCATION INSTITUTIONS IN ECUADOR
(CEAACES)
The proposed ITG/ITM Model has been applied to the
Council for Evaluation, Accreditation and Quality Assurance
of Higher Education Institutions in Ecuador (CEAACES).
The model was considered suitable for its application in
CEAACES due to the multifaceted nature of this institution
from the political and public point of view.
From the political context, this institution has its own
organic structure [40] created to ensure transparency, quality
and continuous improvement within the processes for
evaluation, accreditation and academic quality assurance
inside Ecuadorian higher education institutions (HEIs).
Mainly, CEAACES along with the Council of Higher
Education (CES) are in charge of ensuring political and
strategical alignment in between the National Plan for Good-
Living [41], and the academic strategy inside the
Universities‘ current study programs and research projects.
From the public context, CEAACES is a public
institution that is regulated by the Ecuadorean Legal
Framework (Table I). Therefore, as a public institution, it is
controlled by the Central Government in Ecuador. Since
CEAACES regulates academic quality inside Ecuadorean
HEIs, it is clear that its institutional mission is focused in
education as a public service. Furthermore, Ecuador has
promoted e-Government platforms to democratize access to
public information and provide equal opportunities for
citizens, including access to public higher education
programs.
As consequence, the information that CEAACES
manages have direct impact in the public and political
scenario in which Ecuador deploys its national strategy. This
information and its management processes requires
Governance models that can start at Strategy levels and end
up at Operational IT Areas. Thus, in order to apply the
proposed model inside CEAACES, three alignment scenarios
have been identified.
i. Corporative goals alignment in which COBIT
processes are taken into account.
ii. IT goals alignment to ensure that information
infrastructure, its processes and strategies are aligned
with the corporative goals.
iii. Business Processes Alignment to ensure that the
processes are deployed in compliance with the whole
corporative and IT goals. These processes are classified
by their priority according their contribution level: one
IT goal (low priority), two IT goals (medium priority),
three or more IT goals (high priority).
The alignment results are shown in Table III.

TABLE III. IT STRATEGIC ALIGNMENT IN CEAACES

Scenario 1: Corporative Goals Scenario 2: IT Goals Scenario 3: Business Processes

Focused on customer orientation and
internal policy compliance. Emphasis on
benefit making; majorly on risk
management optimization.
Customer orientation and business
process functionality optimization.
Emphasis on risk management and
resource optimization; majorly on
benefit making.
Customer orientation. Emphasis on
benefit making, resource optimization;
majorly on risk management
optimization.

Mission
Exercise management of public politics for quality assurance or higher education in Ecuador, through processes of evaluation, accreditation
and categorization of Higher Education Institutions (HEIs).
Mission and Strategic Objectives
Strategic a) To evaluate and accredit universities and polytechnic schools, their undergraduate and graduate academic programs.
Objectives b) To evaluate and accredit superior institutes and their academic programs.
c) To place CEAACES as a reference in matters of higher education quality, introducing it in the national, regional and international
academic debate.
d) To ensure academic quality of undergraduate and graduate students of HEIs.
Mission of To advice and provide technological support regarding management of evaluation, accreditation, categorization and quality assurance
ICT processes; as well as information technology management, through infrastructure and computer services for management, storage, custody
Management and data, information, and knowledge technical management.
Customer-oriented service culture. Customer-oriented service culture.
Corporative
Business service continuity and Business service continuity and
Goals Business service continuity and availability.
availability. availability.
 Scenario 1: Corporative Goals
Focused on customer orientation and
internal policy compliance. Emphasis on
benefit making; majorly on risk
management optimization.
Scenario 2: IT Goals
Customer orientation and business
process functionality optimization.
Emphasis on risk management and
resource optimization; majorly on
benefit making.
Scenario 3: Business Processes
Customer orientation. Emphasis on
benefit making, resource optimization;
majorly on risk management
optimization.

Business process functionality Information-based strategic decision

Internal policy compliance
optimization. making.
IT and business strategy alignment. IT and business strategy alignment. IT and business strategy alignment.
IT Compliance and Support regarding business, laws and
Managed IT-related business risks.
external regulations.
Managed IT-related business risks. Managed IT-related business risks. Managed IT-related business risks.
IT service delivery according to
IT Goals IT service delivery according to business requirements.
business requirements.
Information security, processing infrastructure and Information security, processing Information security, processing
applications. infrastructure and applications. infrastructure and applications.
Relevant and useful information available for decision Relevant and useful information Relevant and useful information
making. available for decision making. available for decision making.
IT compliance with internal policies.
High
EDM03, APO13, BAI06, DSS03, DSS04,
(+3 IT APO13, BAI06, DSS03, DSS04 APO13
DSS05, MEA01, MEA02
goals)
EDM01, EDM02, APO01, APO02, EDM01, EDM02, EDM03, APO02,
Medium APO08, APO09, APO10, APO012, APO08, APO09, APO10, APO012, EDM03, APO012, BAI01, BAI06,
Priority

COBIT (2 IT goals) BAI01, BAI02, BAI04, BAI10, DSS01, BAI01, BAI02, BAI04, DSS01, DSS03, DSS04, DSS05
Processes DSS02, DS06, MEA03 DSS02, DSS05, DS06, MEA01
EDM01, EDM02, APO01, APO02,
APO03, APO05, APO07, APO08,
Low APO01, APO03, APO05, APO07,
EDM05, APO03, APO05, APO07, BAI03 APO09, APO10, BAI02, BAI04,
(1 IT goal) BAI10, MEA02, MEA03
BAI10, DSS01, DSS02, DSS06,
MEA01, MEA02 , MEA03

In Scenario 1, the high-priority processes encompass the others, so they are considered for further analysis. Next, considerations
are made regarding corporative governance in CEAACES, along with particular processes interests so that definitive mechanisms
can be established (Table IV).
TABLE IV. MECHANISMS IDENTIFIED AFTER APPLYING THE ITG/ITM MODEL IN CEAACES

Mechanisms
- ICT structure with centralized decision, inherent of supporting public processes, according to
organizational structure of CEAACES.
- Creation of Technology Committee at Advisor Lever, led by CEAACES Chairman, or
representative.
Structure
- Integration of alignment ICT tasks in roles and responsibilities attached to the ICT Department
and/or Technology Committee.
- Presidency of ICT Department in the Technology Committee, reporting and accomplishing
Basic administrator roles of ICT business relationships.
Framework - EDM03 Risk Optimizations Assurance
Governance
(ITG) - APO02.05 Define Strategic Plan and Waybill
Processes
- AP009 Service Level Agreement Management
- APO03 Enterprise Architecture Management

- Shared understanding of business objectives through the Committee of Technology.

Relationships
- ITG Awareness Campaigns.

Strategic
Alignment SAM Model - See Table II. Integrated Model for IGT/ITM
(BITA)

Business-Driven Information
Management Model (BDIM)
Management
(ITM)
Specific Strategies and Norms:
- Information Security.
- Security services
- Problem Management.
- Business Continuity.
Mechanisms
- APO13 Security Management
- BAI06 Change Management
- BAI03 Identification and Solution Construction Management.
- DSS03 Problem Management.
- DSS04 Continuity Management.
- DSS05 Service Security Management.
- MEA01 Supervise, Evaluate and Value Performance and Continuity.
- MEA02 Supervise, Evaluate and Value Internal Control System.
- ISO 27001 and ISO 27002 for Information Security and Security Services (APO13 + DSS05)
- ISO 22301 and ISO 27001 for Business Continuity and Problem Management (DSS03 + DSS04)
- ISO 25000 and Agile Methods for Solution Development (BAI03)

A. Model Implementation Restrictions professional frameworks contribute with their practical

Service orientation that constitutes one of the model
elements was not explicitly covered in the case study
because the strategic alignment mechanism employed in
CEAACES did not required this characteristic.
Now, in those cases where model implementation is
required, doubts coexist regarding the difficulties that may
come up. These uncertainties are more evident during
COBIT implementations [38]. On this count, strategic
alignment constitutes a practical ‗filter‘ to establish
important processes to be implemented. In this case study, 12
out of 37 processes were required to be implemented; almost
a third of the COBIT framework.
Additionally, implementation efforts must be added to
the previous restriction because effort depends on the current
situation and the organizational context. Thus, assessment is
a very important tool in order to devise effort levels that can
be otherwise undistinguishable. In CEAACES assessing
different scenarios enabled a formal definition of both
service-oriented and business-continuity activities as
implementation strategy.
V. DISCUSSION OF RESULTS AND FUTURE WORK
As a result of having applied the model, and after an
evaluation against COBIT Capacity Model, processes with
low development, but with higher priority were deeply
analyzed. For example the IT organizational structure was
redefined, including the responsibilities for strategic
alignment. Also, both the Committee of Technology and the
Committee of Information Security were created in
CEAACES, based on COBIT and ISO 27001 (APO013.
DSS05), respectively. Basic practices to support a Business
Continuity Management System (BCMS) were developed
based on COBIT, ISO 22301 and ISO 27001 (DS04,
DSS03), and an ICT Strategic Plan was developed
(APO02.05) to reflect the strategic alignment.
The model applicability was demonstrated through the
experiences taken from the case study at CEAACES. In the
model specification, contributions from different theoretical
models that were previously analyzed (Section II) can be
explicitly differentiated, as well as the practical approach
that was taken from professional frameworks. The
theoretical models depict the state-of-the-art practices that
have been developed in recent years, whereas the
experiences. The model was developed using a systematic
approach, having the aforementioned contributions as
baseline. Then, after testing the model in three different
scenarios, the necessary steps for implementing it are
explained, using the experiences of its application in
CEAACES. The results of its application are obviously
reflecting changes at strategic and operational levels; in order
to validate the model‘s public and political contribution, a
more extended assessment period should be considered.
However, we believe that through the alignment of business
strategy and IT goals, activities carried out in CEAACES for
evaluation, accreditation and categorization of Higher
Education Institutions could show a very different result in
the medium and long term.
As it can be noticed in Table IV, for ITM
implementation, selecting ITM Guidelines and Norms, as
implementation mechanisms is a required and
complementary activity. In order to blend this mechanisms
with other aspects in the model, it is necessary to consider
the integration details in between mechanisms and the
granularity level that the implementation must have, as
processes need to be decomposed into simple and traceable
activities. This requirement is an opportunity for further
work in this topic, taking into account that some interesting
research has been already done in order to overcome
operative difficulties [42] [43] [44], which have not been yet
devised in the proposed model.
Finally, the preponderant usage of COBIT in the model,
shows a logical result which, unlike the academic
alternatives that were analyzed, demonstrate its applicability
and flexibility for strategic alignment using its Goal Cascade
within different scenarios in order to identify and prioritize
the results. This corroborates the statement made by De
Haes, Van Grembergen, and Debreceny [24] regarding the
research opportunities around COBIT, which are not
restricted by the framework practices, but can be expanded
and combined into an ecosystem of complementary
frameworks and guidelines (COBIT, ITIL, ISO 38500, and
others) which can guide the live cycle of IT processes for
strategic alignment.
 VI. REFERENCES
[1] Constitutional Assembly, "Ecuadorean Constitution," 2008.
[Online]. Available:
http://www.asambleanacional.gov.ec/documentos/constitucion_de_b
olsillo.pdf. [Accessed 30 July 2015 (in Spanish)].
[2] Labor Ministry. Ecuador, "Organic Structure of the Ecuadorean
Public Sector," 31 January 2015. [Online]. Available:
www.relacioneslaborales.gob.ec. [Accessed 30 July 2015 (in
Spanish)].
[3] SNAP. Ecuador, "National e-Government Plan 2014-2017," 2014.
[Online]. Available: http://www.gobiernoelectronico.gob.ec/5-
SNAP-DGGE-PLANDEGOBIERNOELECTRONICO-V01-DL.pdf.
[Accessed 30 July 2015 (in Spanish)].
[4] MINTEL. Ecuador, "Telecommunications and Information Society
Ministry," March 2015. [Online]. Available:
www.telecomunicaciones.gob.ec/estadisticas/. [Accessed 30 July
2015 (in Spanish)].
[5] CGE. Ecuador, Internal Control Norms for Public Sector Entities,
Institutions and Legal Entities of Private Rights with Access to
Public Funds, Quito, 2011.
[6] SNAP. Ecuador, "Compulsory Usage of the Ecuadorean Technical
Norms NTE INEN ISO/IEC 27000 for Information Security
Management," Quito, 2013.
[7] President's Office. Ecuador, Official Registry. Supplement Year IV,
N° 895. Agreement 118. Technical Norm for Process-Based
Administration, Quito, 2013.
[8] ISO, "ISO/IEC 27001, Information technology — Security
techniques — Information security management systems," ISO/IEC -
BSI, 2013.
[9] A. Hevner, S. Ram, S. March and J. Park, "Design Science in
Information Systems Research," MIS Quarterly, vol. 28, no. 1, pp.
75-105, 2004.
[10] P. Webb, C. Pollard and G. Ridley, "Attempting to Define IT
Governance: Wisdom or Folly?," in Proceedings of the 39th Hawaii
International Conference on System Sciences, 2006.
[11] M. Gerrard, "Defining IT Governance: Roles and relationships,"
Gartner. ID Number: G00139986, 2006.
[12] L. Wilbanks, "IT Management and Governance in Equal Parts," IT
Professional, vol. 10, no. 1, pp. 60-61, 2008.
[13] S. De Haes and W. Van Grembergen, Information Technology
Governance: Models, Practices, and Cases, IGI Publishing, 2008.
[14] R. Almeida, R. Pereira and M. Mira da Silva, "IT Governance
Mechanisms: A Literature Review," in Proceedings of the 4th
International Conference IESS, Porto, Portugal. February 7-8, 2013.
[15] ISO, ISO/IEC 38500: 2008. Corporate governance of information,
2008.
[16] A. Moura, J. Sauvé and C. Bartolini, "Business-Driven IT
Management —Upping the Ante of IT: Exploring the Linkage
between IT and Business to Improve Both IT and Business Results,"
IEEE Communications Magazine, vol. 46, no. 10, pp. 148-153, 2008.
[17] C. Bartolini and C. Stefanelli, "Business-driven IT Management," in
IFIP/IEEE International Symposium on Integrated Network
Management (IM), Dublin, Ireland. 23-27 May, 2011.
[18] J. Sauvé, A. Moura, M. Sampaio, J. Jornada and E. Radziuk, "An
Introductory Overview and Survey of Business-Driven IT
Management," in The First IEEE/IFIP International Workshop on
Business-Driven IT Management, BDIM '06, 2006.
[19] ITIL, "ITIL. Best Management practices product," TSO, 2011.
[20] ISACA, COBIT 5.0. A business framework for governance and
management of IT, 2012.
[21] C. Bartolini, "Management, Business-driven IT. DOTTORATO DI
RICERCA IN SCIENZE DELL‘INGEGNERIA INGEGNERIA
DELL‘INFORMAZIONE," Università degli Studi di Ferrara, 2009.
[22] R. Pereira and M. Mira da Silva, "Towards an Integrated IT
Governance and IT Management Framework," in IEEE 16th
International Enterprise Distributed Object Computing Conference,
2012.
[23] R. Pereira and M. Mira da Silva, "A Literature Review: IT
Governance Guidelines and Areas," in ICEGOV '12, Albany, 2012.
[24] S. De Haes, W. Van Grembergen and R. Debreceny, "COBIT 5 and
Enterprise Governance of Information Technology: Building Blocks
and Research Opportunities," Journal of Information Systems, vol. 7,
no. 1, p. 307–324, 2013.
[25] M. Knahl, "A Conceptual Framework for the Integration of IT
Infrastructure Management, IT Service Management and IT
Governance," World Academy of Science, Engineering and
Technology, vol. 52, pp. 438-443, 2009.
[26] ITG - OCG, Alineando Cobit 4.1, ITIL v3 e ISO/IEC 27002 en
beneficio del negocio, Rolling Meadows, 2008 (In Spanish).
[27] ITG, Cobit Mapping: Mapping of Itil v3. with Cobit 4.1, Rolling
Meadows, 2008.
[28] D. Avison, J. Jones, P. Powell and D. Wilson, "Using and validating
the strategic alignment model," Journal of Strategic Information
Systems, vol. 13, pp. 223-246, 2004.
[29] F. Amarilli, "A Framework for Business IT Alignment in Turbulent
Environments," Athens Journal of Technology Engineering, vol. 1,
no. 2, pp. 103-118, 2014.
[30] A. Cataldo, R. McQueen and J. Hardings, "Comparing Strategic IT
Alignment versus Process IT Alignment in SMEs," in Journal of
Research and Practice in Information Technology,, 2012.
[31] Y. Chan and B. Reich, "State of Art. IT alignment: what have we
learned?," Journal of Information Technology, vol. 22, pp. 297-315,
2007.
[32] J. Henderson and N. Venkatraman, "Strategic alignment: leveraging
information technology for transforming organizations," IBM
Systems Journal, vol. 32, no. 1, pp. 4-16, 1993.
[33] I. Kurti, E. Barolli and K. Sevrani, "Effective IT Governance in the
Albanian Public Sector – A critical success factors approach," The
Electronic Journal of Information Systems in Developing Countries,
vol. 63, pp. 1-22, 2014.
[34] C. Juiz, C. Guerrero and I. Lera, "Implementing Good Governance
Principles for the Public Sector in Information Technology
Governance Frameworks," Open Journal of Accounting, vol. 3, pp.
9-27, 2014.
[35] J. Denford, G. Dawson and K. Desouza, "An Argument for
Centralization of IT Governance in the Public Sector," in 48th
International Conference on System Sciences, Hawaii, 2015.
[36] T. Winkler, "IT Governance Mechanisms and Administration/IT
Alignment in the Public Sector: A Conceptual Model and Case
Validation," in 11th International Conference on
Wirtschaftsinformatik, 2013.
[37] P. Weill, "Don't just Lead, Govern: How Top-Performing Firms
Govern IT," MIS Quartely Executive, vol. 3, no. 1, pp. 1-17, 2004.
[38] S. De Haes and W. Van Grembergen, "An Exploratory Study into
the Design of an IT Governance Minimum Baseline through Delphi
Research," Communications of the Association for Information
Systems, vol. 22, no. 24, 2008.
[39] R. Saraiva de Almeida, "Implementing IT Governance. Thesis to
obtain the Master of Science Degree," Técnico Lisboa, Lisboa, 2013.
[40] President's Office, Ecuador, "Official Registry. Year III. N° 773.
Resolution. Organic Statutory Instrument for Process-Driven
Organizational Management," Quito, 2012 (In Spanish).
[41] SENPLADES - National Secretariat for Planning and Development,
"National Plan 2013-2017," 2013. [Online]. Available:
http://www.buenvivir.gob.ec/ (In Spanish). [Accessed 25 August
2015 ].
[42] SEI, "The PrIME Project," 2011. [Online]. Available:
www.sei.cmu.edu/process/research/prime-details.cfm. [Accessed
 2015]. 2004, pp. 1-36.
[43] C. Pardo, "A Framework to Support the Harmonization between
Multiple Models and Standards," Universidad de Castilla - La
Mancha, Ciudad Real (España), 2012a.
[44] Z. Kelemen, Process Based Unification for multi-model Software
Process Improvement, Budapest (Hungary): Ritter nyomda, 2013.
[45] W. Van Grembergen, S. De Haes and W. Guldentops, "Structures,
Processes and Relational Mechanisms for IT Governance," in
Strategies for Information Technology Governance, Idea Group Inc.,
 Experiencing Game-Based Learning through Mobile
Devices
R.K. Sungkur1, U.G. Singh2 and A. Adaya3

1,3
Computer Science and Engineering Dept, University of Mauritius, Mauritius
r.sungkur@uom.ac.mu
2
School of Management, Information Technology and Governance,
University of Kwa-Zulu Natal (Westville campus), South Africa
Singhup@ukzn.ac.za

Abstract—The introduction of game-based learning (GBL)
into the pedagogical processes and curriculum design can
increase student engagement in the learning process. There are a
range of game based learning approaches available, but, so far,
limited adoption of serious games has been recorded. The digital
habits of learners should be studied carefully, to better
understand the way current technology-savvy students learn, and
integrate socially. With the wide-spread use of Mobile Devices
today, GBL has a vital role to play in the educational landscape.
This research analyses the potential usage of Mobile Devices to
enhance the learning process, through Game-Based Learning.
Keywords—Game Based Learning; Mobile Devices; Serious
games; assessment; student feedback
focus time of between 10 and 20 minutes in a normal face to
face session [17]. Beyond this, the potential for students to
absorb more content, decreases. Thus, if an important concept
is presented after this ‘prime focus time’ students do not
absorb this material. Furthermore, certain modules require
greater practical exposure; but due to the nature of our current
curriculum, these modules are restricted to being taught
theoretically only. Quoting a Chinese proverb – “I hear and I
forget; I see and I remember; I do and I understand,” unless
students are involved in practical sessions, thereby engaging
with the material taught, comprehension of the theory taught,
is often difficult.

I. INTRODUCTION
Adopting Game Based Learning (GBL) as the basis of
pedagogical processes, and curriculum design, is a way of
promoting learner engagement in the education cycle. A
variety of GBL approaches exist but, actual, current usage is
restricted to serious games [1], [2], [3]. To fully exploit the
potential that GBL offers, the digital habits of learners should
be understood fully. This could provide a better understanding
of the way this ubiquitous generation of young students learn,
and integrate socially. Furthermore, learning games presently
available may be limiting, but not as restrictive as the syllabus
into which they are currently being integrated.
In simple terms, Game Based Learning refers to the concept of
learning a specified module through a game [4], [5], [6].
Too many ‘sage on the stage’ type lectures, on a daily basis,
can be monotonous, non-engaging, and unstimulating, hence
may not be the most effective way to teach students. In this
scenario, teaching is present, but the question often arises
whether learning has actually taken place, following the
teaching ‘session’. As early as 1976, research of student
attention patterns during lecture sessions suggested that
student attention declines steadily during a lecture [15].
Students' attention span lasts as long as their interest [16].
Pedagogically speaking, the average human mind has a prime
The question is, what does GBL offer, that traditional lectures
cannot? Ideally, GBL encourages the learner to explore the
solution space and ask, “What if I do this???”. The theory
behind gaming is that decisions are taken at regular intervals.
This forced decision making process ensures that critical
thinking is propelled, and reflection harvested. Being forced to
take decisions the learning experience increases by leaps and
bounds.
II. LITERATURE REVIEW
A. Pedagogical approaches to Learning
Every educator develops his/her own way of teaching. To be
effective this method should be adapted to suit his/her students
learning, while still supporting the personal pedagogy of the
teacher. There are different kinds of pedagogies to learning,
which include m-learning pedagogies [7], inclusive
pedagogies [8], deep and surface pedagogies [9], and
participatory approaches [10], to name a few. Three specific
approaches are explained briefly below, that relate directly to
the current research:
1. Learning by doing [11], [12]
This is a very practical form of learning as it requires ‘hands-
on experience’. This is usually adopted in technical and
science studies where students acquire knowledge, and test
this knowledge in laboratories and workshops, through
practical engagement. With the advent of web technology,
students can perform simulations in an online environment
enabling them to write-up their findings immediately. Since
this learning experience is group-based, the student can
collaborste with other learners to get answers to their queries,
as well as get feedback from their seniors, during the learning
process. The ‘learning by doing’ approach is highly effective
in environments where students can create their own artefacts.
2. Learning done through discussion and debate [11]
Traditionally discussion and debate learning was arranged in
‘real settings with real people’, and success was achieved
through a small class, and effective instructor. In recent times
discussions and debate are hosted in online environments,
using web technology to facilitate the communication process.
The introduction of web technology into this learning
approach, allows students to chat with each other using the
discussion forums, and thus share ideas in a very collaborative
and flexible manner, without the constraints of time and place.
The effectiveness of this learning approach is dependent on
two factors; how big the student group is, and the role of
instructor
3. Blended Learning [11], [13], [14]
Advocates of blended learning state that social and human
interactivity such as body language, welcome, socializing, and
face to face contact are necessary in education. In this context,
the blended learning approach has its roots in technology-
supported learning. de Boer (2004) defines blended learning as
an approach that blends different kinds of delivery and
learning methods, which can be enabled by technology and
supported by traditional teaching methods. Thus it offers the
better of both worlds.
B. Game Based Learning
Tang, Hanneghan and El-Rhalibi (2009) define GBL as:
“the innovative learning approach derived from the use of
computer games or software applications, that use games for
learning and education purposes, such as learning support,
teaching enhancement, assessment and evaluation of
learners.”
As presented in Figure 1, GBL not only offers a unique
concept to complement traditional teaching strategies, but also
infuses teaching with energy, encourages innovative thinking
and provides diversification in teaching methods [18]. GBL
encourages creative behavior, divergent thought (Fusard,
2001), and can serve as an excellent ice breaker. GBL can act
as a learning trigger, to induce lively discussion on learning
concepts amongst students following game play.
Zyda (2005) defines a ‘serious game’ as ’a mental contest,
played with a computer, in accordance with specific rules, that
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
uses entertainment, to further education and training’ in
various sectors. Sawyer and Smith’s (2008) taxonomy of
serious games expands this definition to include games for
health, training, science and research, production and
marketing. Serious Games is a sub-category of GBL, however
it should be noted that the terms are sometimes used
synonymously, (Corti, 2006).
C. What is GBL and how does it work?
Advocates of effective GBL state that it is a highly effective
and interactive platform, which motivates and engages
learners in the continuous learning process [25]. This trend has
been enhanced, with video game designers producing and
refining highly encouraging environments, for their players to
enjoy, as depicted in Figure 1.
Figure 1: Learning with GBL [25]
In an environment of effective game based learning, the
principal focus is on a pre-defined goal, choosing actions that
will help achieve the goal, as well as experiencing the
consequences on those actions, be they correct or incorrect
actions, all along the way. Mistakes are made in a minimal
risk setting, hence the participant actively learns from these
mistakes, and through repeated practice and experimentation,
the right actions are sought. In this way, the participant is
constantly alert, and engaged in exercising behavior and
thought processes, which can easily be transferred from a
simulated environment to a real life scenario.
D. GBL as Empirical Evidence
The benefits of games and simulations for educational has
been researched extensively. 68 comparisons were examined
by Randel, Wetzel and Whitehill comparing simulation/game
approaches to learning and traditional instruction, in relation
to student performance. Their main discoveries were:
 56% found no difference; 32% favored
simulations/games; 7% favored simulations/games
with control questionable; 5% favored conventional
instruction.
 In so far as retention is concerned, it was found that
simulations/games induced better retention over time
than conventional approaches.
 With regard to interest, 86% of the respondents
showed a much greater liking for games and
simulations over conventional approaches.
E. Traditional Learning [19]
Traditionally, learning adopted narrow procedures where
students were expected to memorize content, and were then
evaluated on what they could regurgitate. Even if students
passed a module successfully, all this indicates is that they
have only just managed to memorize and reproduce the
lesson’s facts and procedures. In reality, they are not being
assessed in the application of their learning to true life
situations. Even if real-world situations are built into the
curriculum in the traditional approach to learning, no matter
how comprehensive the module may be, it will not be possible
to cover procedures of every complex eventuality that students
may encounter in life. Thus, in contrast to traditional learning,
GBL allows the students to not only learn the facts, but also
teaches them the important principals of how and why, which
is gathered through practice and experimentation. Game Based
Learning prepares students to perform effectively and
consistently, even in new and unforeseen situations.
Cost-Effective
GBL can be cost-effective, since multiple students will have
access to the single GBL application on their respective
mobile phones, instead of having one teacher per class.
Furthermore, updating GBL software can be cheaper and
faster than training staff to adapt to changes in educational
materials. This is particularly applicable in the field of
Computer Science as learning materials have to be updated
regularly, due to the changing nature of technology.
Learning as leisure activity
Learning can either be an interesting or boring process
depending on the subject matter-, learning methods adopted,
and learning materials provided. Learning itself can be
stressful, if an individual has to focus for a prolonged period
of time to grasp the material. Eventually, this stressful learning
method can have a direct and negative impact on health. Thus
GBL offers an antidote to anxiety and depression as it
incorporates the ‘fun’ element in learning.

F. GBL v/s Traditional Learning Table 1: Comparison of Traditional Learning and GBL [27]

The effectiveness of hands on learning is not new [20]. Thus, Traditional Learning Game-based
properly designed GBL has many benefits over traditional (lectures, online Learning
learning methods, as presented in Figure 2 - including being tutorials)
less costly and having minimal risk. The main advantage, Cost-effective
however, lies in its significantly engaging learning process X X
[21]. A precise set of circumstances can be attempted several Low physical X X
times, thus enabling the student to explore the consequences risk/liability
of different actions. In the context of pedagogy GBL are
Standardized
useful to enlighten practically oriented teaching topics, as well X X
as dealing effectively with problem solving. Research shows assessments allowing
that ’GBL has a singular role in building students’ self- student-to-student
confidence’ and it can reduce the gap between quicker and comparisons
slower learners’ (Fuscard, 2001). The comparison between Highly engaging
GBL and Traditional Learning is summarized in Table 1. X

Learning pace X
customized to individual
student
Immediate feedback in X
response to student
mistakes
Student can without X
difficulty transfer
learning to real-world
scenarios
Learner is dynamically X
Figure 2: GBL as a Pedagogical Device
engaged
G. Advantages of GBL [22],[23]

Like with other forms of technology adopted in the teaching H. The Four Principles of Learning Games [13],[24]
process, GBL also offers advantages. Some of these
advantages are briefly presented below. Since GBL environments created are not suitable to everyone,
often the games that we find uninteresting are the ones which
cast a bad experience on our learning phase. For GBL to be
effective, it must be structured in an environment that is
conducive to student learning. In this context, the Carnegie
Mellon’s Eberly Center for teaching excellence, has devised
four key principles that describe successful GBL learning
approaches, as presented in Figure 3.
processes and skills. In a GBL environment, as learners
advance in the game, they also learn to experience immediate
in-game consequences, especially with regard to their
mistakes.
III. METHODOLOGY AND PROPOSED SOLUTION
A. Overview of Proposed System
The proposed system consists of developing mobile games
for the module titled ‘Network System Administration’, and is
depicted in Figure 4.
Games are designed in the form of quizzes and puzzles,
while at the same time, making interesting use of graphics and
videos.

Figure 3: GBL Design Model

Principle 1
The first principle relates to a student’s prior knowledge of the
game. Students who have interacted with GBL, or similar
environment previously, have an advantage of familiarity over
those students who are adopting it for the first time. In
traditional learning, buried student’s misconception might
arise only at test time, whereas in a gamed based approached,
the weaknesses of the learner is immediately made apparent

Principle 2
The second principle is a student’s motivation. It is one of the
key factors that sustain students on what they do to learn.
Some may even play a particular game repeatedly, until they
achieve a perfect score. Without knowingly realizing it,
learners learn how to operate within a game environment,
actively think, experiment and learn how to achieve their
goals. Through repeated practice the lessons learned, help
them to develop and master consistent and fruitful thought
processes.

Principle 3
Students need to acquire component skills from GBL and
know when to apply these skills. Learning is a step by step
process, and each learner works at different pace. This means
slower students often struggle, and faster students become
bored in passive training programs. On the contrary, game
based learning can be tailored to each learner. A learner
normally begins with the basic concepts and then gradually, as
he begins to master his interim goals, eventually moving on to
more advanced challenges.
Figure 4: System Structure
Principle 4
Students quality of learning can be enhanced by a goal
The State Transition of the proposed system is shown in
directed practice, coupled with targeted feedback. In this
Figure 5.
context, traditional learning, when compared to GBL, does not
provide learners with the opportunity to practice thought
 software e.g. Tablets and Smartphones (see Figure 6),
and
 An Android Emulator is required using the latest
APIs, to run the GBL application on a computer (see
Figure 7).

Figure 6: Emulator

C. Some Features of the Application implemented

Figure 5: State Transition
B. Evaluation of Proposed Tools and Technologies
Android Version 4.2
Android is a Linux-based OS designed primarily for mobile
devices, and has a large community of developers, writing
applications, that extend the functionality of devices, written
primarily in a customized version of the programming
language.
The mobile application had a number of interesting features
that were used to demonstrate Game-Based Learning.
Activities such as quiz and jigsaw puzzles are shown in Figure
6. These games included counters, scores, levels and time, to
make the learning experience more interesting and
challenging. The correct answers were displayed to students
and they were also provided with feedback. All these
interactive activities added a new dimension to the teaching of
the module ‘Network System Administration’. Students were
now able to practice the skills associated with the
administration of a network, which were taught theoretically
in a class, and now experienced through the GBL application
developed. A screenshot the GBL application is shown in
Figure 7.

The IDE (Integrated Development Environment) supports

Eclipse using the ADT (Android Development Tools) plug-in.
Android applications run in a sandbox, which means that
access to system resources are not permitted, without specific
permission granted to user. To explain this concept the
following example is described. When a game is installed on a
device, the user can only use the application and not view its
related resources (codes, images, database etc).

OpenGL and UI can be used for graphics and animations to

enhance the visual aspect of the game.

Specifications Required
The different tools that will be used are:
 Any device using an android platform [26] not
beyond version 4.2 (Jelly Bean) to run the GBL

Figure 7: Features of the GBL Mobile Application developed
IV. DISCUSSION
The Game-Based Learning Application for Mobile
Devices was successfully created and effectively used to
demonstrate how the teaching of a module on ‘Network
Systems Administration’ can be made more effective and
interesting. A number of interactive features such as quiz,
puzzles, jigsaw puzzles, counter, score, levels, graphics and
videos were adopted to make the mobile enrich the learning
experience. The GBL environment provided learners with
videos for enhancing their learning. The feedback facility
provided them with not only the correct answers to their tasks,
but also learning feedback on concepts that they
misunderstood. The pedagogical aspects of Game-Based
Learning on Mobile Devices were considered while
developing the application. Thus, the mobile application
developed was self-explanatory, intuitive and user friendly.
It is understood, that the GBL application developed is not
meant to replace the lecture class, but rather to complement it,
by providing students with a blended learning environment.
The GBL appeared to be much more popular and effective, for
revision purposes, with the students. The application has been
tested with 40 students following the above-mentioned module
and it was unanimously agreed that it definitely enhances the
process of learning. Further enhancements to the application
would include simulation of a network, speech recognition,
and the possibility of incorporating multiple players, to make
the games more challenging.
V. CONCLUSION
Game based learning helps develop in the individual, the
sense of perseverance, which is latent in him. When people are
provided with interesting materials, it becomes easier for them
to achieve their potential, in a particular field of subject. The
sense of competition within GBL software is an enabler to
foster inventiveness. The GBL application that has been
developed, not only allows for mobile learning but also allows
the students to fully use the concept of GBL to enhance their
learning experience. This has proved to be very motivating to
the students where they become actively engaged in the
learning process. Moreover, it can connect students sharing a
common interest, thus encouraging creativity, as well as
developing teamwork skills, which is crucial in the job market.
Increasing time in the classroom to attempt to teach students
how to think and perform in the face of real-world challenges,
will be ineffective. GBL can help students to better remember
what they are learning, through the experiences and
interactions with the application developed. This provides
motivation to the students as they are able to see the
connection between their learning experience and real-life
scenarios.
REFERENCES
[1] Dagnino, F., Ott, M., Pozzi, F., & Yilmaz, E. (2015). Serious games
design: reflections from an experience in the field of intangible cultural
heritage education. E-learning & Software for Education.
[2] Arnab, S., Berta, R., Earp, J., De Freitas, S., Popescu, M., Romero, M.,
& Usart, M. (2012). Framing the Adoption of Serious Games in Formal
Education. Electronic Journal of e-Learning, 10(2), 159-171.
[3] Stapleton, A. J. (2004). Serious games: Serious opportunities. In
Australian Game Developers‟ Conference, Academic Summit,
Melbourne.
[4] Blunt, R. (2007). Does game-based learning work? Results from three
recent studies. In Proceedings of the Interservice/Industry Training,
Simulation, & Education Conference (pp. 945-955).
[5] Van Eck, R. (2006). Digital game-based learning: It's not just the digital
natives who are restless. EDUCAUSE review, 41(2), 16.
[6] Pivec, M., Dziabenko, O., & Schinnerl, I. (2003). Aspects of game-
based learning. In 3rd International Conference on Knowledge
Management, Graz, Austria (pp. 216-225).
[7] Lindsay, L. (2015). Transformation of teacher practice using mobile
technology with one‐ to‐ one classes: M-learning pedagogical
approaches. British Journal of Educational Technology
[8] Spratt, J., & Florian, L. (2015). Inclusive pedagogy: From learning to
action. Supporting each individual in the context of ‘everybody’.
Teaching and Teacher Education, 49, 89-96.
[9] Boughey, C. (2015). Approaches to large-class teachinglarge-class
pedagogy: Interdisciplinary perspectives for quality higher education,
David J. Hornsby, Ruksana Osman, Jaqueline de Matos-Ala (Eds.): book
review. South African Journal of Science, 111 (1 & 2), 4-5.
[10] Anderson, V., mckenzie, M., Allan, S., Hill, T., McLean, S., Kayira, J.,
& Butcher, K. (2015). Participatory action research as pedagogy:
investigating social and ecological justice learning within a teacher
education program. Teaching Education, 26(2), 179-195.
[11] Tadesse, T., & Gillies, R. M. (2015). Nurturing Cooperative Learning
Pedagogies in Higher Education Classrooms: Evidence of Instructional
Reform and Potential Challenges. Current Issues in Education, 18(2).
[12] Yates, T. (2015). Learninglearning and Teaching Community-Based
Research: Linking Pedagogy to Practice by Etmanski, C., Budd, LH, and
Dawson, T.(ed.) 2014. University of Toronto Press. Toronto, ON.
388pp. Engaged Scholar Journal: Community-Engaged Research,
Teaching, and Learning, 1(1).
[13] Jabbar, A. I. A., & Felicia, P. (2015). Gameplay Engagement and
Learning in Game-Based Learning A Systematic Review. Review of
Educational Research, 0034654315577210.
[14] Roy, A., & Sharples, M. (2015). Mobile Game Based Learning: Can It
Enhance Learning Of Marginalized Peer Educators?. International
Journal of Mobile and Blended Learning (IJMBL), 7(1), 1-12.
[15] Johnstone, A. H., & Percival, F. (1976). Attention breaks in lectures.
Education in chemistry, 13(2), 49-50.
[16] Matheson, C. (2008). The educational value and effectiveness of
lectures. The Clinical Teacher, 5(4), 218-221.
[17] Read, B. (2005). Lectures on the Go. The Chronicle of Higher
Education, 52(10), A39.
[18] Pitt, M. B., Borman-Shoap, E. C., & Eppich, W. J. (2015). Twelve tips
for maximizing the effectiveness of game-based learning. Medical
teacher, (0), 1-5.
[19] Learning in Immersive Worlds
http://www.jisc.ac.uk/media/documents/programmes/elearninginnovatio
n/gamingreport_v3.pdf [Accessed 15 August 2015]
[20] Chen, C. H., Wang, K. C., & Lin, Y. H. (2015). The Comparison of
Solitary and Collaborative Modes of Game-based Learning on Students'
Science Learning and Motivation. Journal of Educational Technology &
Society, 18(2), 237-248.
[21] Furió, D., Juan, M. C., Seguí, I., & Vivó, R. (2015). Mobile learning vs.
Traditional classroom lessons: a comparative study. Journal of Computer
Assisted Learning, 31(3), 189-201.
[22] Hung, C. Y., Sun, J. C. Y., & Yu, P. T. (2015). The benefits of a
challenge: student motivation and flow experience in tablet-PC-game-
based learning. Interactive Learning Environments, 23(2), 172-190.
[23] Tsai, F. H., Tsai, C. C., & Lin, K. Y. (2015). The evaluation of different
gaming modes and feedback types on game-based formative assessment
in an online learning environment. Computers & Education, 81, 259-
269.
[24] Shi, Y. R., & Shih, J. L. (2015). Game Factors and Game-Based
Learning Design Model. International Journal of Computer Games
Technology, 2015.
[25] Jian, M. S., Shen, J. H., Huang, T. C., Chen, Y. C., & Chen, J. L. (2015).
Language Learning in Cloud: Modular Role Player Game-Distance-
Learning System Based on Voice Recognition. In Future Information
Technology-II (pp. 129-135). Springer Netherlands.
[26] Android 4.2 apis| Android Developers, 2013. Android 4.2 apis| Android
Developers. [Online],
http://developer.android.com/about/versions/android-4.2.html [Accessed
on 05 August 2015]
[27] Game Based Learning vs Traditional Learning, Game-Based Learning:
What it is, Why it Works, and Where it’s Going. [Online],
http://www.newmedia.org/game-based-learning--what-it-is-why-it-
works-and-where-its-going.html [Accessed on 10 August 2015]
[28] Sawyer, B. & Smith, P. (2008). Serious Game taxonomy. Paper presented
at the Serious Game Summit 2008, San Francisco, USA
[29] Corti, K. (2006) Games-based Learning; a serious business application.
PIXELearning Limited.
[30] Boer, W.F. de (2004). Flexibility support for a changing university.. PhD
dissertation, Faculty of Behavioural Sciences, University of Twente,
Enschede, NL.
[31] Tang, S., Hanneghan, M., & El Rhalibi, A. (2009). Introduction to games-
based learning. In T. Connolly, M. Stansfield, & L. Boyle
(Eds.), Games-based learning advancements for multi-sensory human
computer interfaces (pp. 1–17). Hershey, PA: IGI Global
[32] Zyda, M. (2005). "From Visual Simulation to Virtual Reality to
Games", IEEE Computer, vol. 38, no. 9, pp.25 -32
[33] O’NEIL, H. F., WAINESS, R. and BAKER, E. L. (2005). Classification
of learning outcomes: evidence from the computer games literature.
Curriculum Journal, 16(4): 455 – 474.
 Privacy Challenges in Proximity Based Social
Networking: Techniques & Solutions
Asslinah Mocktoolah
Department of Computer Science and Engineering
University of Mauritius
Mauritius
ashlinahmee@gmail.com
Abstract—The development of the Proximity based Mobile
Social Networking (PMSN) has been growing exponentially
with the adoption of smartphones and introduction of Wi-Fi
hotspots in public and remote areas. Users present in the
vicinity can interact with each other using the embedded
technologies in their mobile devices such as GPS, Wi-Fi and
Bluetooth. Due to its growing momentum, this new social
networking has also aroused the interest of business people and
advertisers. However, due to a lack of security in these
networks, several privacy concerns were reported. Users are
more reluctant to share their locations and to address this
issue, some initial solutions to preserve location privacy were
implemented. The aim of this paper is to present a clear
categorization of the different privacy threats in PMSN.
Moreover, the location privacy enforcement policies and
techniques used to ensure privacy are outlined and some
solutions employed in existent systems are presented and
discussed. To the best of our knowledge, this is the first study
done outlining several categories of PMSN privacy challenges
and their solutions in this new type of social networking
services. Finally, some privacy research challenges and future
perspectives are proposed.
Keywords—proximity based social networking; location
privacy challenges; security; privacy solutions; categorization of
privacy challenges
I. INTRODUCTION
The 21st century has witnessed an array of technological
advances in the world of Internet changing the nature of
socialization and communication, especially with the advent
of social media networks. According to statistics carried out
in March 2015, Online Social Networking (OSN) prevails
as the most popular networks visited worldwide [1]. Earlier,
even with the notion of freedom of speech, it was not
obvious to share views or pass on a message easily. With
this new phenomenon, voicing out opinions is as easy as a
piece of cake, for example, the hashtag #JeSuisCharlie
which translates to „I am Charlie‟ allowed millions of
people from all over the world to share their solidarity
messages just a few hours after the terrorist attack in Paris
[2].
During the recent years, with the rapid evolution of
technology and upsurge of smartphones, traditional OSN
websites such as Facebook [3] and Twitter [4] have
extended their services to mobile applications. Users can
Kavi Kumar KHEDO
Department of Computer Science and Engineering
University of Mauritius
Mauritius
k.khedo@uom.ac.mu
now enjoy the benefits of social networks by
communicating with their friends and sharing information
while they are on the move. Taking advantage of the
embedded location services such as GPS in mobile devices,
this concept was further exploited to create a new category
of mobile social networks called Proximity based Mobile
Social Networking (PMSN) targeting mainly users in
vicinity.
PMSN refers to the social interaction of physically
proximate mobile users through their mobile devices via
Bluetooth or Near Field Communication (NFC) techniques.
While OSN has been accused of promoting anti-social
behavior, PMSN on the other hand has bridged the gap
between the virtual and physical worlds. Different services
for users within close proximities are provided such as
connecting to known friends or strangers having similar
interests, selecting nearby restaurants having good reviews
or setting reminders at explicit places. However, sometimes,
these benefits are overwhelmed by the cost in revealing
locations and many users share a mutual apprehension of
sharing their location information. According to a recent
report by Pew Research Center, a slight drop is observed in
the number of smartphones users who use location based
services and around 46% of teen users and 35% of adult
users have turned off their location tracking features due to
privacy concerns [5]. The reluctance of adopting these
services are mainly due to the dangers associated with the
disclosed location coordinates such as physical stalking,
tracking usual routes of users or identifying sensitive
information of users‟ visits to some embarrassing places
such as clinics or clubs [6].
To minimize the risks related to the privacy issues and to
further encourage people to use PMSN services, many
attempts to preserve privacy have been made such as K-
anonymity algorithms, obfuscation techniques or
cryptographic schemes. However, no in-depth analysis of
the different privacy challenges and solutions of PMSN has
been made prior to the implementation of an effective
privacy preservation framework. Several recent studies on
mobile social networks have presented a categorization on
different privacy issues and outlined some possible
solutions but a survey on the PMSN privacy challenges and
the existing works on the solutions implemented is missing.
This study is important for further research in privacy and
security in PMSN networks by addressing the different
privacy concerns so as to assure the continuity of these
systems.
The rest of this paper is organized as follows. In Section
II, an overview of PMSN services is outlined and a
categorization of the privacy challenges related to these
platforms is presented in Section III. Section IV gives an
idea of the privacy enforcement policies and techniques
while Section V studies the existing privacy aware PMSN
solutions. Some privacy research challenges and future
perspectives are described in Section VI and the study is
summarized on a concluding note in Section VII.
II. PROXIMITY BASED MOBILE SOCIAL NETWORKING
(PMSN)
Social network giants such as Facebook focus on re-
creating a user‟s offline social graph on the net but usually
overlook changing social networks that users participate in
their daily lives, for example, people working out at the
gym, doing shopping for groceries or parents accompanying
their children to playgrounds. PMSN revolves around
connecting these users, who share the same interests or
activities at that same moment in time helping them to
discover new connections based on their physical
proximities.
PMSN applications such as Lokast [7] and Proxxi [8]
automatically discover users in proximity and help them to
interact with each other based on their geographical
locations. One of the popular PMSN application Foursquare
with more than 55 million of users worldwide adopted its
own location detection technology: Pilgrim by using GPS
and users‟ past check-in histories [9]. The users‟ locations
are also tracked in the background and push notifications are
sent when they are located at specific places. Although,
these tracking services can be helpful to users, privacy
issues may also arise since according to Foursquare privacy
policy, user information is leaked to third parties, e.g., to
display the user's current location on a map, the latitude and
longitude details are directed to the Google map server.
Actually, no privacy protection schemes are present for
large PMSN platforms such as Foursquare but the security
of users is only based on the privacy settings as managed by
the users.
These networks are also usually known as ephemeral
social networks due to the short duration for which they are
created at specific events [10]. For example, when attending
conferences, these applications can be very advantageous to
allow new attendees to connect to people sharing the same
research areas or having similar interests. PMSN has been
applied in other areas as well for businesses, advertising or
marketing and also for entertainment purposes such as
gaming applications where mobile game users adopt new
gaming experiences and different ways to interact with other
users around them. By diversifying into different application
areas, PMSN is paving the way to attract more users to
adopt this new trend.
Such applications are mostly based on a Peer-to-Peer
(P2P) architecture allowing mobile users to communicate
with each other in the absence of a central server through
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
wireless ad-hoc networks. The P2P architecture gives rise to
a greater number of privacy issues as compared to a client-
server architecture. The next section gives a broad
categorization of the different privacy challenges in PMSN
systems.
III. PRIVACY CHALLENGES IN PMSN
Together with the popularity of PMSN, an increasing
danger of user privacy is observed due to the repeated
release of location information of the users. Some efforts
have been made to give an overview of the privacy threats
of PMSN but a detailed and comprehensive categorization
of PMSN privacy challenges is missing. In this paper, the
main privacy challenges are categorized in four main
classes: location privacy, identity privacy, trust and
malicious attacks.
A. Location Privacy
In PMSN networks, it is imperative to share location
coordinates but this act is accompanied by several problems
such that it can cause user troubles ranging from being a
victim to be robbed to being sexually assaulted. Users‟ usual
routes can be predicted with advance attacks and even the
mode of transport can be inferred [11]. The location privacy
issues in these systems are further categorized into different
sub-classes: privacy disclosure, absence location privacy,
inferred location privacy, location data exploitation and
location cheating.
1) Privacy Disclosure
With the advent of location-enhanced technology,
disclosing location information to another person or service
can be valuable but yet dangerous. It can be risky if anyone
can access the information and in addition, location
information can be directly associated with other
information such as identities, preferences and social
relationships.
2) Absence Location Privacy
The absence location privacy issue refers to the
hypothesis that a person is not at a particular place at a
given moment. An example of such as privacy violation
refers to a theft occurred in Manchester in August 2010,
where social networking was used to verify the absence of
the victims before committing the theft at their residence
[12].
3) Inferred location Privacy
In this case, locations of other users shared in the
network are used to decipher the fact that a specific person
is currently at a place based on the relationships between
them. For example, a group of friends usually hang out
together, and if one of them share their locations, it can be
deduced that the others are located at same place.
4) Location Data Exploitation
Location information exploitation is related to the extent
of how PMSN applications or third parties use the data
shared and for which purposes it is being used. The privacy
policy of commercial PMSN applications such as
Foursquare clearly state that once information has been
shared, it is no longer safe and third parties can access same
[9].
5) Location Cheating
A location cheating attack may refer to deceiving the
service provider by claiming that he is at a particular place
when the user is in reality several miles away from the
shared location. Foursquare allows users to check-in the
application using a location information that is different but
nearby to the users‟ current location. The users attempt to
cheat on their locations usually for rewards such as
mayorships offered by the service. To prevent users from
cheating on their locations, Cheater Code was released by
Foursquare in 2010 [13].
B. Identity Privacy
Users‟ anonymity can be easily violated in PMSN
systems when their profile or location information is used to
identify their real identities. Sometimes, to protect their
identities, users opt for fake names on these networks.
However, the real identities of users can be inferred by
using the location co-ordinates that are shared even if they
use pseudonyms [14]. The different sub-categorizations
related to identity privacy discussed in this study are
anonymity issues, private information leakage and
neighborhood attacks.
1) Anonymity Issues
Beach et al. [15] introduced two classes of anonymity
issues: the direct anonymity problem where the identity of a
user can be inferred directly from his shared profile
information and the indirect or K-anonymity issue which
refers to the usage of the provided information to deduce or
crack the identity of the user. The user‟s anonymity is
further compromised as stated by Mascetti et al. [16] when
location information acts as a quasi-identifier when it is
associated with external data.
2) Private Information Leakage
Using the user location information leaked to tracking
sites or third party aggregation servers, data about the users‟
habits, usual activities, interests and even relationships can
be inferred.
3) Neighborhood attacks
The privacy of a user can be compromised if an
adversary has some knowledge about the neighbors of a user
in the network and the relationships between them.
C. Trust Relationships
Trust between users in online social networking sites
following a centralized architecture, is based mostly on the
real life social relationships such as friends, family members
or even colleagues. For PMSN services, the level of trust
between the users will be different since people
communicate with each other without any prior interactions
before. The interactions are only based on similar attributes
or activities which users have in common. According to a
recent study on the impact of trust in relationships, it is
observed that users interacting and sharing locations with
each other in an impromptu way as they move do not build
trust but damage trust [17]. Since communication is based
on a peer-to-peer architecture, trust management will be
more challenging due to the absence of a central server and
in addition to the mobility of the users.
D. Malicious Attacks
There are several types of malicious attacks that can take
place in PMSN platforms, whose goal is to violate the users‟
personal and location information and also to degrade the
network performance. The main attacks outlined in this
study correspond to eavesdropping attacks, spoofing and
replay attacks, wormhole attacks, Sybil attacks, the
Distributed Denial of Service (DDOS) attacks and the man-
in-the-middle (MITM) attacks.
IV. PRIVACY ENFORCEMENT POLICIES AND TECHNIQUES
This section outlines some privacy enforcement policies
and common used techniques that are usually employed on
these platforms to cater for the privacy concerns associated
with this social networking class.
A. User-Controlled Privacy
PMSN application such as Foursquare encourages
people to self-control their privacy settings by selecting the
audience with whom their posts or locations are shared to
protect their private information [9]. However, users may
not be accustomed of the level of privacy that they need
unless at some point in life, they are themselves victim of an
improper exposure or misuse of their location information
which can result in severe consequences [18].
PeopleFinder is an application based on Policy
Enforcing Agents (PEA) in which each user can specify a
set of rules that determine if their location information can
be accessed by others based on several conditions such as
defining groups of people who can view the locations, the
time during which locations can be shared or which
locations can be published [19]. Locaccino is a similar
location sharing system, which allows Facebook users to set
up privacy rules based on three criteria: who, time and
locations. The authors also observed that users are more
likely to share high entropy locations such as university
campus than low entropy locations such as their respective
classrooms.
B. Location K-anonymity methods
Location K-anonymity methods protect a user‟s
sensitive information by obscuring his personal and location
information so that it is difficult for at least K-1 other users
to determine the identity of the individual [20]. A trust
server known as the anonymizer is used to obtain K-
anonymous location privacy by removing the user‟s ID and
choosing an anonymizing spatial region (ASR) which takes
into consideration the user and at least K-1 users in the
vicinity. The users‟ locations are blurred so that they are
indistinguishable among other K-1 users [21].
K-anonymity techniques guarantee privacy for snapshot
locations only and since locations of PMSN users are
ephemeral and continuously changing, this technique might
not provide a complete security of location data. In addition,
a serious privacy issue may occur if the third party is
attacked by malicious users. P2P spatial cloaking algorithms
are usually used to ensure a safer platform by creating a
group with peers in the proximity and the user cloaks his
location into a spatial region that covers the network of the
other mobile users via single-hop communication or
multihop routing. The query is sent together with the cloak
area to obtain the desired services. In this technique, the
server does not even know the exact location of the user.
C. Obfuscation-based techniques
An alternative solution to location K-anonymity is to
employ obfuscation-based techniques where the location
privacy of users can be protected by decreasing the accuracy
of users‟ location so the real location of the user is not
disclosed. The main obfuscation methods refer to Deleting,
Randomizing, Discretizing, Subsampling and Mixing [22].
In the Deleting approach, high entropy locations are deleted
protecting sensitive locations of users. Randomizing applies
Gaussian noise to samples to generate new locations so that
others cannot determine the exact position. In its simplest
form, in the method Discretizing, a lower-accuracy grid or
square is defined which contains the user location and the
exact position cannot be determined but only that the user is
somewhere located in the square. Subsampling is to provide
locations at a coarser timescale where some data is deleted
so that there is a gap in time between the points. Mixing is
to reduce the resolution to achieve K-anonymity. In other
words, a user‟s data is mixed with others so that the user‟s
location can be confused with the locations of other users.
The notion of relevance can be used as a metric of
location information accuracy and privacy to minimize the
disclosure of personal location information based on spatial
obfuscation [23] [24]. Three obfuscation operators were
employed namely Enlarge, Reduce and Shift. Enlarge and
Reduce reduces the accuracy of the original location by
enlarging the radius and decreasing the radius respectively.
Shift, on the other hand, relates to shifting the center point
so that an obfuscated area, different from the original one, is
derived by using the distance between the two centers.
D. Cryptography Techniques
To prevent disclosure and leakage of information
between the interactions and discovery of proximity users,
data exchange can be encrypted with attribute policies so
that other users do not have access to the information. Most
common cryptographic algorithms used in PMSN systems
are public key cryptography and hash functions. When
recipients receive the cipher data, they decrypt the content
using private and public keys or the key-hash of the
message. However, the decryption process of the received
data adds an overhead and furthermore, additional
calculations are performed on the decrypted data. To
prevent the network‟s ID to be compromised, the use of
anonymous and secure identifiers which are generated using
cryptographic hash functions such as SHA-1 or elliptic-
curve cryptography can also be used. Some protocols to
ensure private matching of proximate users were based on
the Pailler cryptosystem so as not to disclose the private
information of the user profiles [25] [26]. Light-weight
cryptographic schemes can also be used to move the
functionality to the users‟ mobile devices where the keys are
store and users share encrypted data to interact with each
other [11].
V. PRIVACY AWARE PMSN SOLUTIONS
Protecting the privacy of PMSN users is considered to
be more important and urgent than in OSN since in addition
to the personal information, the physical locations of the
users are also leaked. Some security approaches employed
in real PMSN systems to tackle the aforementioned privacy
challenges are discussed in this section.
To ensure privacy in PMSN systems, most research have
applied cryptography techniques where pairs of devices in
proximity are established by using a public key
infrastructure which connects public keys with the user
identities using a certificate authority. Each communication
is encrypted and protected using a message authentication
code. Smile is based on the exchange of credentials to
establish trust between users in communication [27]. When
a message is broadcasted, the encounter key together with a
hash of the encounter key is posted. Users within the same
communication range and having same encounter key can
claim the message by verifying the key hash and then
decrypt same with the correct key. MeetUp is an advanced
prototype of Smile which uses visual authentication for
information exchange and verification [28]. The Tor
network is applied in MeetUp to obscure the identity and
location of user initiating the connection. The Tor network
is also incorporated in an encounter-based system in a recent
work to enable online anonymity and checks the presence of
eavesdropper throughout the network using a Ciphertext
Policy Attribute-Based Encryption (CPABE) scheme [29].
In addition to location privacy, VICINITYLOCATOR
supports dynamic-shape vicinities as well by introducing the
notion of vicinity region where it is assumed that two users
are in the proximity of each other if the vicinity region of
one user falls into the same location of another user [30].
The proximity detection service relies on both spatial
cloaking and encryption by referring the user‟s location to a
set of granules in his vicinity. To prevent published
resources from causing privacy concerns, the WYSE
technique allows users to specify their privacy preferences
and applies spatial and temporal generalization algorithm
and a cloaking technique to prevent users from disclosing
their locations as well as preventing others to infer their
locations via other shared locations and protecting absence
privacy location [31]. Sharp introduces a novel form of
location representation by building spatial-temporal location
tags using wireless signals in the surroundings [32]. Bloom
filters are used to better present the location tags. A
lightweight cryptography primitive is used to automatically
extract keys from users ensuring that their location
information is not revealed to the server and also this
information is hidden from users who are not in the vicinity,
thus preventing location cheating attacks.
Bettini et al. [33] presented some preliminary algorithms to
preserve K-anonymity such as the spatio-temporal
generalization algorithm and if the latter fails, the unlinking
technique can be used by changing the pseudonym of the
user. A mechanism to preserve both location data and K-
anonymity based on locality-sensitive hashing (LSH) is
proposed in [34]. The user locations are divided into groups
where each comprises of at least K users ensuring that
proximity information is preserved during mapping. Pan et
al. [35], on the other hand, showed that existing K-
anonymity location algorithms cannot prevent location-
dependent attacks on an effective way when users‟ locations
are continuously updated. They proposed a new algorithm
called the ICliqueCloak to protect against location-
dependent attacks.
The K-anonymity approach lacks a trusted third party
which relays the communication between the mobile users
and the LBS provider. To protect mobile users‟ location
information without the help of a third party, a new privacy-
preserving location authentication protocol called PriLA is
proposed in [36] to help in the authentication of user
location without compromising his location privacy in Wi-
Fi networks. PriLA exploits inherent physical layer
information named carrier frequency offset (CFO) to secure
the transmission and to exploit multipath profile.
Another method used to preserve the privacy of PMSN
users is to store the users‟ location information and profile
information separately. These techniques are applied in the
application Mobishare which uses two different servers: the
social network server and the location server to protect the
users‟ privacy ensuring that the adversary cannot link a
user‟s information to his location even if his identity is
known [37]. The Place-Its application protects people‟s
location privacy by ensuring that all computation is done on
the client‟s device only and not on the server [38]. The Koi
system also adapts the same principle by using two entities
ensuring that neither entity determines the association
between a user‟s ID and his location [39]. In addition, Koi
comprises of a privacy-preserving cloud-based location
matching service by allowing the application to specify a
location event of interest and when there is a location match,
a notification is sent. A novel privacy preserving and
fairness-aware friend matching protocol was introduced in
[40] ensuring that the matching of profiles takes place only
if the interests of both participants match others‟ profiles.
The fairness of this protocol ensures that private information
is not leaked such that no other extra information can be
accessed from the protocol except the information that the
user reveals. Zhang et al. [41] proposed a fine-grained
private matching protocol and argued that fine-grained
profiles enable a better distinction among users who have
different levels of interest for the same attribute.
MobiTrust is a trust model used to establish trust
relationships among mobile users and is based on three
factors: the user profile similarity, users‟ reputation and the
history of common friends between any two users [42].
However, to achieve a satisfactory performance, a high
amount of data transaction is required. A lightweight
trustworthy service discovery is proposed in [43] which
reduces the overall transaction cost.
Identity Server (IS) provides some solutions on the
privacy issues on PMSN networks making use of an
anonymous identifier (AID) as shown in Fig. 1 [15]. A new
AID is collected from the IS for each mobile device
discovered on the network using a cryptographic hash
function such as SHA-1 with a random salt value.
Fig. 1. Anonymous IDs and the Identity Server [15]
The encrypted information ensures that the
eavesdropping attack cannot take place and the use of AIDs
prevent spoofing, replay and wormhole attacks. To protect
against attacks such as Sybil, DDOS and MITM, several
techniques were proposed [44] [45] [46]. Table I gives an
overview of the different techniques employed to tackle
each challenge.
VI. PRIVACY RESEARCH CHALLENGES AND FUTURE
PERSPECTIVES
Though many attempts have been made in the recent
years to address the aforementioned privacy challenges in
PMSN systems, further researches are required to tackle
some overlooked technical issues and challenges. The
proposed privacy research directions in this section will help
to improve the efficiency, functionality and effectiveness of
future PMSN applications.
Identifying malicious content providers refers to a critical
trustworthiness challenge due to the absence of a centralized
server. Taking into consideration the mobility of the users in
addition to the privacy concerns, a proper strategy to manage
trust is imperative for a continuous availability of these peer
to peer services. Some proposed matching algorithms have
been implemented to ensure full anonymity measures,
however new techniques and algorithms should be
introduced to maintain a higher level of trust in the
communication and interaction of users since in the existing
ones, partial information about users is revealed.
A secure discovery service is indispensable in such
systems allowing users in the vicinity to discover each other
protecting the confidential information of the users and
ensuring the identification of the devices in communication.
The security of data exchange between the users requires a
full proof security and currently, cryptographic techniques Using technologies such as Bluetooth or Wi-Fi, intruders
are mostly used to ensure this privacy. However, the existing can easily penetrate and hack the system to steal information.

Challenges
Location Privacy Identity Privacy Malicious Attacks

Trust Relationships

Wormhole Attacks

Sybil Attacks
Direct Anonymity

K-Anonymity

Neighborhood Attacks

Spoofing & Replay

Man in the Middle

Information Leakage

Eavesdropping

Distributed DOS
Inferred Location

Location Exploitation

Location Cheating
Location Disclosure

Absence Location
Solutions
Techniques

Smile [27] Y Y Y Cryptographic hash functions

MeetUp [28] Y Y Y Y Authentication and Cryptography Scheme
VICINITY
Y Spatial Cloaking & Encryption
LOCATOR[30]
WYSE [31] Y Y Y Cloaking & Spatial-Generalization
Sharp [32] Y Y Y Lightweight Cryptography & location tags
Bettini et al. Spatio-Temporal Generalization Algorithm &
Y Y
[33] Unlinking Technique
Vu et al. [34] Y Y Y Locality-Sensitive Hashing (LSH)
ICliqueCloak
Y Y Clique-based Cloaking Algorithm
[35]
Authentication and Carrier Frequency Offset
PriLA [36] Y Y Y Y
(CFO) Encryption Algorithm
Obfuscation techniques/Identity data & location
MobiShare [37] Y Y
updates are stored at two separate entities
Privacy-preserving cloud-based matching
Koi [39] Y Y Y Y
service
Based on profile similarity, reputation and
MobiTrust [42] Y
friendship history
Identity Server Use of an anonymous identifier (AID) using a
Y Y Y Y Y Y
[15] cryptographic hash function SHA-1
relies on limited availability of real-world
SybilLimit [44] Y
friendship edges between node
Novel traffic profiling techniques : adaptive
D-Ward [45] Y
response & source-end deployment
Gollakota et al.
Y Tamper-evident pairing protocol
[46]
methods are not sufficient since it is known that encryption Future PMSN systems should cater these shortcomings and
algorithms can be deciphered easily and encryption keys can more robust algorithms ensuring security of user data are
be also stolen. Further works are required to go beyond necessary.
cryptographic methods to ensure a thorough data protection.

TABLE I. PRIVACY AWARE SOLUTIONS IN PROXIMITY BASED MOBILE SOCIAL NETWORKING

“Y” denotes which privacy challenge was addressed in each PMSN System

In addition, since in a peer to peer architecture there is no

central server to manage the P2P networks, appropriate VII. CONCLUSION
control schemes must be designed so as to handle peer An overview of PMSN is presented in this paper and the
discovery services for such systems. Future researches must related privacy challenges in this new social networking field
be carried out so as to address the above challenges as well are identified. The privacy challenges are categorized into
as focusing on broadcast or multicast within proximity, location privacy, identity privacy, trust relationships and
multihop and most importantly these protocols must be malicious attacks. A series of privacy aware PMSN solutions
designed in such a way so that they support multiple context are discussed highlighting how each challenge outlined were
based location applications. addressed by using different techniques such as K-
anonymity, obfuscation and cryptography. Lastly, some
promising research directions are proposed. This extensive
study on PMSN privacy challenges and solutions can be
further used by future researches on privacy preservation
frameworks on proximity based networking systems.
The next step in this work will be to set up the existing
solutions such as MeetUp [28], MobiShare [37] and
experimentally test the existing solutions against location
privacy, identity privacy, trust relationships and malicious
attacks of those solutions. The objective of the experiments
will be to confirm the extents to which the existing platforms
are shielded to the privacy concerns described in this paper.
The experiments will further allow us to identify loopholes
and limitations of the existing systems.
References
[1] Statista, 2015. The Statistics Portal. Leading social networks
worldwide as of March 2015, Available at
http://www.statista.com/statistics/272014/global-social-networks-
ranked-by-number-of-users/
[2] Mashable, 2015. #JeSuisCharlie: Messages of solidarity after Paris
attack. Available at http://mashable.com/2015/01/07/jesuischarlie-
charlie-hebdo/
[3] Facebook, 2015. Available at http://www.facebook.com
[4] Twitter, 2015. Available at http://www.twitter.com
[5] K. Zickuhr, “Location-based services”, Pew Research Center, Pew
Internet and American Life Project, September 2013
[6] M. Wernke, P. Skvortsov, F. Dürr, and K. Rothermel, “A
classification of location privacy attacks and approaches”, Personal
and Ubiquitous Computing 18 (1) (2014) 163–175, 2014.
[7] LoKast, 2015. Available at http://www.lokast.com/lokast-app/
[8] Proxxi, A Proximity-Based Social App, 2014. Available at
http://coenraets.org/blog/2014/01/
[9] The Foursquare Blog, 2014. Available at
http://blog.foursquare.com/post/87012827988/more-on-swarm-and-
the-future-of-foursquare.
[10] A. Chin and D. Zhang, “Mobile Social Networking - An Innovation
Approach”, Springer, 2013.
[11] K. Puttaswamy and B. Zhao, “Preserving privacy in location-based
mobile social applications”, In Proceedings of the Eleventh
Workshop on Mobile Computing Systems & Applications,
Annapolis, MD, USA, 22 – 23 February 2010, (pp. 1-6). ACM, 2010.
[12] WMUR Manchester, “Police: Thieves robbed home based on
Facebook”, September 2010, Available at
http://www.wmur.com/Police-Thieves-Robbed-Homes-Based-On-
Facebook-Social-Media-Sites/11861116#!bFENCI
[13] A. Wilhem, “Foursquare Releases „Cheater Code‟ To Stop People
From Checking In Where They Are Not”, April 2010, Available at
http://thenextweb.com/apps/2010/04/07/foursquare-releases-cheater-
code-stop-people-checking/
[14] A. Rault, “Privacy implications of geosocial proximity”, Technical
Report, Submitted on 24 August 2012, Université de Rennes.
[15] A. Beach, M. Gartrell, and R. Han, “Solutions to security and privacy
issues in mobile social networking”, in Proc. IEEE Int. Conf.
Comput. Sci. Eng., Aug. 2009, pp. 1036–1042.
[16] S. Mascetti, D. Freni, C. Bettini, X. S. Wang and S. Jajodia "Privacy
in geo-social networks: proximity notification with untrusted service
providers and curious buddies", CoRR, vol. 20, no. 4, 2011
[17] S.J. Fusco, R. Abbas, K. Michael and A. Aloudat, “Location-Based
Social Networking: Impact on Trust in Relationships”, Technology
and Society Magazine, IEEE, vol.31, no.2, pp.39, 50, Summer 2012.
[18] C. Mancini et al. “From spaces to places: Emerging contexts in
mobile privacy”, in Proceedings of the 11th international conference
on Ubiquitous computing. ACM, 2009, pp. 1–10.
[19] N. Sadeh, et al. “Understanding and Capturing People‟s Privacy
Policies in a Mobile Social Networking Application”, Personal and
Ubiquitous Computing, vol. 13, no. 6, Aug. 2009.
[20] K. El Emam and F.K. Dankar “Protecting privacy using k-
anonymity”, Journal of the American Medical Informatics
Association, Vol. 15, No. 5, pp. 627–637, 2008
[21] K. Vu, R. Zheng, and J. Gao, “Efficient algorithms for k-anonymous
location privacy in participatory sensing”, in IEEE INFOCOM, 2012
[22] A. B. Brush, J. Krumm, and J. Scott, "Exploring end user preferences
for location obfuscation, location-based services, and the value of
location," in Proceedings of the 12th ACM international conference
on Ubiquitous computing, ser. Ubicomp '10. New York, NY, USA:
ACM, 2010, pp. 95-104.
[23] C. Ardagna, M. Cremonini, S. De Capitani di Vimercati, and P.
Samarati, “An obfuscation-based approach for protecting location
privacy,” IEEE TDSC, vol. 8, no. 1, pp. 13–27, January-February
2011.
[24] P. Wightman, W. Coronell, D. Jabba, M. Jimeno, and M. Labrador,
“Evaluation of Location Obfuscation techniques for privacy in
location based information systems,” in Communications, IEEE
Latin-American Conference on, 2011, pp. 1–6.
[25] B. Krishnamurthy and C.E. Wills. Privacy leakage in mobile online
social networks. In Proceedings of the 3rd conference on Online
social networks (pp. 4-4). USENIX Association, June 2010.
[26] G. Konakanchi, R. Venkata. A New Privacy Preserving Private
Profile Matching Technique for Proximity-based Mobile Social
Networks. International Journal of Advanced and Innovative
Research, 2(11), (pp. 2278-7844), 2011.
[27] A. Mohaisen, E. Y. Vasserman, M. Schuchard, D. Foo Kune, and Y.
Kim, "Secure encounter-based social networks: requirements,
challenges, and designs," in CCS, 2010, pp. 717-719.
[28] A. Mohaisen , D.F. Kune , E. Vasserman , M. Kim and Y. Kim
"Secure Encounter-Based Mobile Social Networks: Requirements,
Designs, and Tradeoffs", IEEE Trans. Dependable Secure Comput.,
vol. 10, no. 6, pp.380 -393 2013
[29] B. Suganya, B and J.V. Priya. “A Secure Encounter-Based Key
Transmission over Tor Network In Mobile Social Networks for
Improving Security and Privacy”, International Journal of Innovative
Research and Development, vol. 3, no 4, April 2014.
[30] L. Siksnys, J. R. Thomsen, S. Saltenis, and M. L. Yiu, “Private and
flexible proximity detection in mobile social networks”, in Proc. 11th
Int. Conf. Mobile Data Manage., May 23–26, 2010, pp. 75–84.
[31] D. Freni, C. R. Vicente, S. Mascetti, C. Bettini, and C. S. Jensen,
“Preserving location and absence privacy in geo-social networks”, in
Proc. of CIKM‟10. ACM, 2010, pp. 309–318.
[32] Y. Zheng, M. Li, W. Lou, and T. Hou, “Sharp: Private proximity test
and secure handshake with cheat-proof location tags”, in Proc.
ESORICS, 2012, pp. 361–378
[33] C. Bettini, X. S. Wang, and S. Jajodia, “Protecting Privacy Against
Location-Based Personal Identification,” in Proc. of the Second
VLDB Workshop on Secure Data Management (SDM 2005).
[34] K. Vu, R. Zheng, and J. Gao, “Efficient algorithms for k-anonymous
location privacy in participatory sensing.” in IEEE INFOCOM, 2012
[35] X. Pan, J. Xu, and X. Meng. Protecting location privacy against
location dependent attack in mobile services. ACM CIKM, 2008
[36] Y. Chen, W. Wang, and Q. Zhang. “Privacy-preserving location
authentication in WiFi with fine-grained physical layer information”.
In Global Communications Conference (GLOBECOM), 2014 IEEE
(pp. 4827-4832). IEEE.
[37] W. Wei, F. Xu, and Q. Li. Mobishare: Flexible privacy-preserving
location sharing in mobile online social networks. In INFOCOM,
Proceedings IEEE, Orlando, FL, USA, 25-30 March 2012
[38] T. Sohn, K. A. Li, G. Lee, I. E. Smith, J. Scott, and W. G. Griswold,
“Place-its: A study of location-based reminders on mobile phones,” in
Ubicomp, 2005, pp. 232–250.
[39] S. Guha, M. Jain, and V. N. Padmanabhan. “Koi: A location-privacy
platform for smartphone apps”. In NSDI, 2012.
[40] R. Zhang, Y. Zhang, J. Sun and G. Yan. “Fine-grained private
matching for proximity-based mobile social networking”. In
INFOCOM, 2012 Proceedings IEEE, USA, 2012
[41] H. Zhu, S. Du, M. Li and Z. Gao. “Fairness-aware and privacy-
preserving friend matching protocol in mobile social networks”.
Emerging Topics in Computing, IEEE, 192-200, 2013.
[42] J. Li, Z. Zhang, and W. Zhang, “MobiTrust: Trust management
system in mobile social computing,” in Proc. IEEE 10th Int. Conf.
CIT, 2010, pp. 954–959.
[43] C. Chang, S. Ling, S. Srirama, Trustworthy service discovery for
mobile social network in proximity, in: Pervasive Computing and
Communications Workshops, 2014, pp. 478–483.
[44] H. Yu, P. B. Gibbons, M. Kaminsky, and F. Xiao, “Sybillimit: A
near-optimal social network defense against sybil attacks”, In IEEE
Symposium on Security and Privacy, pages 3–17, 2008.
[45] J. Mirkovic and P. Reiher, "D-WARD: A Source-End Defense
Against Flooding DoS Attacks," IEEE Trans. on Dependable and
Secure Computing, July 2005, pp. 216-232.
[46] S. Gollakota, N. Ahmed, N. Zeldovich, and D. Katabi, “Secure in-
band wireless pairing,” in USENIX Security, San Francisco, CA,
Aug. 2011.
 Cloud-based Technologies for Addressing Long
Vehicle Turnaround Times at Recycling Mills
Joubert van Tonder
Graduate School of Business Leadership (SBL)
University of South Africa (Unisa),
Midrand South Africa
Joubert.tonder@supergrp.com
Abstract—Transportation costs for road transport
companies may be intensified by rising fuel prices, levies,
traffic congestion, etc. Of particular concern to the Mpact
group of companies is the long waiting times in the queues at
loading and offloading points at three processing mills in the
KZN (KwaZulu-Natal) province in South Africa. Following a
survey among the drivers who regularly deliver at these sites,
recommendations for alleviating the lengthy waiting times are
put forward. On the strength of one of these recommendations,
namely the innovative use of ICTs, suggestions on how cloud-
based technologies may be embraced by the company are
explored. In the process, the value added by a cloud-based
supply chain, enterprise systems, CRM (Customer
Relationship Management) and knowledge management is
examined.
Keywords—Transportation company; Supply chain;
Cloud computing; Enterprise systems; CRM; knowledge
management
I. INTRODUCTION
South Africa with its large geographical distances between
establishments lends itself ideally to the use of road transport
services. Transportation costs, however, can be a significant
portion of a company„s overall logistics costs. With increases in
the price of fuel, the proportion allocated to transportation can rise
with up to fifty per cent. Such costs are usually passed on to the
customer, resulting in further increases in the price of goods [1].
Logistics costing on a macroeconomic level has increasingly
become more of a necessity in South Africa. Pressures on reducing
logistics costs internationally are even more acute in the South
African context, given the country‟s geographic location on the
globe. Hence, focusing on logistics value and cost drivers is vital
[2].
The proportion of commercial freight tonnage transported by road
in South Africa has steadily increased over the past ten years to
approximately 89%. Such increase is due mainly to fluctuating and
increasing needs of South Africa's supply chains and a number of
challenges faced by the National rail operator. These are low
service levels, lack of flexibility and capital investment delays.
Furthermore, owing to a decline in investment in road
infrastructure, huge strain is placed on the road infrastructure,
much of which is already in a bad state of repair, and results in
increased traffic congestion, and higher vehicle maintenance costs
[2] [3]. The impact of this is an increase in travel distances due to
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
John Andrew van der Poll
Graduate School of Business Leadership (SBL)
University of South Africa (Unisa),
Midrand, South Africa
Vdpolja@unisa.ac.za
vehicles taking detours, resulting in longer turnaround times,
which reduce overall supply chain efficiency. According to the
National Planning Commission, 20% of the paved road network is
currently classified as being in a poor or very poor condition [4].
Other transportation challenges include volatile fuel prices, toll
costs and stricter carbon requirements. Leading transportation
companies are addressing some of these issues by optimising
vehicle fleets, using more fuel efficient vehicles, implementing
strategic route planning, and advancing driver training [4].
An additional challenge experienced by transport companies is
potentially long waiting times during the offloading of goods and
the subsequent reloading of goods onto a transportation vehicle. Of
particular interest to the researchers is the impact such waiting
times have on the Mpact group of companies [5] which specialise
in the manufacture of paper. Trucks often have to wait for
extended periods in long queues during the offloading of timber
and the reloading of finished products.
A survey was conducted by the researchers to determine the effect
of the long queues on aspects such as turnaround times,
profitability, driver morale and loss of income. Some
recommendations to alleviate these challenges were formulated,
and one of the prominent enablers identified was the use of ICTs,
particularly the use of Cloud Computing to facilitate the
communication of road conditions, alternate routes, driver
schedules, etc.
Subsequently, the research questions (RQs) addressed in this paper
are (in the context of the Mpact group):
1.1 Research Questions
RQ1: What are the main causes of the long turnaround times
during the offloading and reloading?
RQ2: How may the use of ICTs reduce the turnaround times at
the recoverable fibre offloading sites?
In this paper, cognisance is given to the role cloud computing may
play in answering RQ2 above.
1.2 Layout of the paper
The operation of, and challenges experienced by the Mpact
transport companies are presented in Section 2, followed by a
definition of the research methodology in Section 3 and the data
collection methods in Section 4 respectively. Section 5 presents the
findings of the survey and subsequent recommendations. Supply
chain aspects and cloud service models are covered in Section 6,
followed by cloud solutions for the transport company in Section 7.
Section 8 presents a conclusion and considers future work in this
area.
II. IMPACT OPERATIONS AND CHALLENGES
The first part of the research was conducted to identify and
evaluate the reasons for the long turnaround times of trucks at the
three recycling processing mills which are based in the South
African towns of Felixton, Piet Retief and Merebank. The mills
make use of recycled paper to mainly produce carton board and
container board [5][6].
Figure 1: An Outlay of the Recoverable Fibre Supply Chain Process [6]
Efficient and quick truck turnaround times are essential not only to
ensure that customers receive their product on time, but also to
ensure that factories don„t build up unwanted inventory and costly
stock holding. An optimal supply chain involves suppliers who can
efficiently supply products to their customers, free of damage and
on time [15].
Naturally, transporters want to optimise the efficiency of their
trucks, with the truck drivers benefiting from up-time in
productivity due to them increasing their income potential as more
kilometres are covered. The aim is, therefore, to optimise
turnaround times [1].
Trucks arrive at any one of these mills and join the back of a
queue. Turnaround time is calculated by measuring the average
time elapsed between a truck's arrival at the particular facility and
its departure and is an indicator of the efficiency of the receiving
processes - the shorter the time elapsed, the more efficient the
operation.
In transport economics the value of time is the opportunity cost of
the time that a vehicle (truck) spends on its journey. In essence,
this determines the amount a customer would be willing to pay in
order to save time, or the amount they would accept as
compensation for lost time [1]. As depicted in Fig 1, recoverable
fibre are baled at a baling plant, then loaded onto a 34 ton interlink
truck and transported to the mills storage holding area for
production into paper or pulp.
The recoverable fibre transport matrix, as depicted in Fig 2 shows
that all three processing mills are based in Kwazulu Natal (KZN) –
Merebank; Piet Retief and Felixton. The recoverable fibre is
collected from various collection points in each of the provinces
and then transported to the recoverable fibre baling plants which
are situated in Cape Town, Port Elizabeth, East London, Tulisa
Park, Nelspruit and Maydon Wharf. The recoverable fibre is baled
into 1 ton blocks and subsequently transported to the various
processing mills in Kwazulu Natal for production into either pulp
or paper [6].
 Mpact Mills
Figure 2: Recoverable Fibre Collection Points and Recycling Mills (synthesised by researchers).
Recoverable fibre is moved to the individual processing plants
from various areas across the country. Bottle necks of vehicles
queuing at the offloading sites often occur. Currently the planning
of the vehicles is done off-line by an individual who receives an
order from the recycling plant to inform them of the various
recycled fibre available around the country as well as to which mill
the recyclable fibre needs to be transported. The scheduler does not
have the information with regard to the stock and consumption
levels of the mills needed to determine whether recyclable fibre
can indeed be sent to the mills. Over and above these factors the
scheduler also needs to determine the various travel times
associated with each of the loads as the distances to each mill
varies from the loading sites distributed across the country.
The above uncertainties lead to shortcomings in the planning of
routes, ultimately leading to trucks arriving at a mill with a long
queue, while there may be no queue at some of the other mills.
A part of a solution, the power of more information on the
processes, integrated communication, and real-time analytics may
well deliver savings. Logistics solutions that allow for real-time
monitoring make for accurate models of merge in transit possible
for the first time, viz. a process that was previously considered too
complex to accurately estimate, can now be forecasted and
monitored through a real-time, online dashboard in the cloud [16].
III. RESEARCH METHODOLOGY
Statistical sampling refers to the selection of individuals, units,
and/or settings from a population to be studied. Whereas
Recoverable Fibre
Collection Points
quantitative studies strive for random sampling, qualitative studies
often use purposeful or criterion-based sampling, that is, a sample
that has the characteristics relevant to the research questions [18].
Ultimately, all fieldwork culminates in the analysis and
interpretation of some dataset. Analyses involve categorising the
data into manageable themes, patterns, and relationships.
Subsequently one aims to understand the various constitutive
elements of the data through an inspection of the relationship
between concepts, constructs or variables, and to determine
whether there are any patterns or trends that can be identified or
isolated, or to establish themes (factors) in the data [19].
A survey instrument to assess the turnaround times of trucks at the
three recycling mills was developed and validated. The survey
relied not only on driver feedback, but also on data captured
through the use of on-board tracking devises, installed on each of
the trucks, and which was analysed to determine the actual
turnaround times of the vehicles at the various mills. The financial
implications for the recycling mills, the transporter as well as the
truck drivers were determined and analysed. Aspects of qualitative
and action research methods were integrated into the formal
process. Qualitative and quantitative data were collected from each
of the interviewees. Root causes were identified and meaningful
inferences, focusing on the objective (to reduce turnaround times)
were made.
The selection of a quantitative research approach was based on the
nature of the research questions, the sample of the population
within which the research was conducted as well as the expertise of
the researchers. The outcome of the questionnaire (Appendix A)
was analysed statistically [17].
 IV. DATA COLLECTION
Daily vehicle planning schedules from the transport companies
were analysed to determine whether the planning and scheduling of
the vehicles are contributing to the long turnaround times of
vehicles arriving at the mills at the same time, thus causing long
queues. Travel times from various pick up destinations to the mills
were calculated. Mill consumption rates, mill storage space, mill
maintenance shut downs, mill offloading operating hours are
examples of factors that were analysed to determine a solution.
The necessary ethical clearances were obtained and drivers who
were involved in the process from truck arrival to truck departure
were asked to participate in the survey. The data collected from the
completed questionnaires were analysed using Microsoft Excel.
The responses were assigned to a Likert scale option from which
the researchers calculated percentages.
THE RAW DATA
Figure 3: Data Flow Analysis (Adapted from the SA Office of Research, Development, and Technology).
V. FINDINGS & RECOMMENDATIONS
Analyses of the financial losses and the potential financial savings
indicated that by shortening the turnaround times to within the four
hour norm a savings of two million ZAR could be gained. The
results of the driver questionnaire highlighted the mills with the
worst offloading times as well as the areas within the offloading
process that required attention.
The study identified a number of measures that could be
implemented to reduce the time spent by a driver at the recoverable
fibre offloading sites:
The implementation of ICT planning software to take into account
variable factors that would determine the truck arrival dates and
times at specific mills, thereby facilitating planning to reduce
bottlenecks at the mills was identified as a strategic necessity.
Cloud computing could be used as the cloud allows for track and
trace functionality, with drivers submitting proof of delivery from
mobile devices. Customers could then access such data. Not only is
One of the aims of the questionnaire was to determine the general
impression of the truck drivers, related to their experience at the
various mills and to assess what the truck drivers believe are
causing the long turnaround times at the recycling mills. The data
collected was used to give feedback to the mills about what the
researches believed needed to be changed to induce cost savings to
mill customers as well as to improve on truck driver morale and
productivity [6].
The number of trucks in the Super Group fleet is 160, giving a
population of 160 drivers. All the drivers deliver recyclables to
each of the mills at least three times per annum. Drivers meet with
their debriefing managers fortnightly, hence the drivers were asked
to complete the questionnaires during these debriefing sessions.
Figure 3 illustrates the data collection process that was followed.
SURVEY REPORT
Researches data collection, analyses and
classification.
this a facility customers are coming to expect but it also alleviates
the administrative and communications burden on the scheduler
[21].
Through the survey it was also established that the use of flat deck
trailers contributed to the longer than normal turnaround times.
Subsequently the study identified that the use of tautliner trailers
instead of flat deck trailers would reduce the turnaround time due
to the time saving of not having to tarp and un-tarp the flat deck
trailer. It was recommended that tautliners be introduced into the
transporter fleet as part of their vehicle replacement programme.
Further discussion of this technical recommendation is beyond the
scope of this paper, but more information appears in [6].
The results of the survey furthermore revealed a correlation
between how helpful the drivers found the staff (EQ measurement)
at each mill and the turnaround time at such mill. Mills with more
helpful staff had a better turnaround time than those where the staff
were not perceived to be helpful. It is recommended that
companies educate their employees and create a service oriented
culture at all mills.

It is argued that cloud computing can assist a transport company

with the above aspects, as there are substantial cost, customer-
satisfaction and commercial benefits in subscribing to cloud
technology. Subsequently, it may be a proactive way of providing
the proverbial competitive edge. Section 0 elaborates on these
possibilities.

The discussion in this section answers our RQ1 above.

Next we discuss some cloud computing service models in the

context of a business‟ supply chain structures.

VI. CLOUD COMPUTING IN SUPPLY CHAIN

Cloud computing is generally described by three (3) service
models, namely IaaS (Infrastructure as a Service), PaaS (Platform
as a Service) and SaaS (Software as a Service). Fig 4 [7] illustrates
the differences and similarities among these approaches. Less
controlled is exercised by the subscriber moving from an IaaS
Figure 4: Cloud computing models [7].
option to PaaS to SaaS. Following a move from IaaS through to
SaaS, more information security concerns arise for both the user Digital supply chain (SC) networks, enabled by cloud technology
and the service provider. have four distinct advantages which, together, drive visibility,
insights and flexibility, while operating rapidly and at scale [8].

Fig 5 illustrates the relationships among various digital supply

chains.

Figure 5: Four Distinct Advantages of Digital Supply Networks [9]

Various advantages to the different cloud-based supply chains are
[10] [11]:
A connected supply chain has the following advantages:
 Real-time visibility: supply chains become more dynamic,
secure and interactive.
 Seamless collaboration: supply chain capabilities are
harmonised beyond physical boundaries.
 Highly evolved operating models: product/service delivery is
exponentially improved to meet customers‟ evolving
demands.
Advantages of an intelligent supply chain are:
 Actionable insights: innovative data analysis supports
advanced decision-making.
 Enhanced and accelerated innovation: digitally inspires and
supports creative advances in design, personnel, operations
and customer relationships (CRs)
Advantages to a scalable supply chain are:
 Improved efficiency: provides for the integration of people,
processes and technology
 Organisational flexibility: digital plug-and-play enablers
provide for natural “configure and re-configure” capabilities.
 Personalised experiences: channel-centric supply networks
help foster individualised products and services.
Rapid supply chain advantages are:
VII. CLOUD SOLUTIONS
In conjunction with existing supply chain structures, Cloud
solutions for transport companies may be provided for in a variety
of ways.
7.1 IT Infrastructures
The use of innovative ICTs [13] was identified as a solution to the
challenges experienced by the transport companies. To this end
cloud computing offers lucrative opportunities for transport
companies. Instead of acquiring expensive computing resources
(hardware, networking and software) they may opt for a utility-
based cloud solution. Depending on the specific need and their
level of ICT literacy, they may subscribe to any of the PaaS, SaaS
or SaaS service models through an appropriate CSP (Cloud Service
Provider) and utilise only the specific service(s) required.
The use of utility-based cloud in the form of the SaaS service
model (hardware, networking and software) has a number of
advantages for transport companies like Mpact discussed in this
paper. In particular it may:
 Lower capital investment on IT.
 Improve company cash-flow.
 Allow for the redeployment of existing IT staff.
Using scalable broadband connections, collaborative working from
a variety of locations is facilitated through subscribing to a cloud
service [20].
7.2 Using an Enterprise system and CRM
Vitally important in addressing IT infrastructures is the use of an
Enterprise system incorporating a modern CRM (Customer
Relationship Management) component. An Enterprise System is
defined as an integrated set of software packages with integrated
modules that pass common business transactions across groups,
divisions, and geographic locations in “real time” [22]. A CRM
module may addresses the aspect of staff morale in the context of
the EQ (Emotional Intelligence) of the staff at the three mills
identified in Section 0 above.
Subscribing to enterprise- and CRM software in the cloud offers
lucrative advantages to a transport company. Coupled with an
 Enhanced responsiveness: using better information and
sophisticated analytics to interpret and react speedily to
disruptions, including demand and supply signals.
 Proactive prevention: decision support, driven by predictive
analytics, helps to confirm reliability and rapid adaptability
 Last mile postponement: swift repurposing of organisational
assets at short notice helps to ensure that supplies always meet
changing demands.
Supply chain enabled cloud computing is lately recognised as a
decidable enabler, providing a route through which supply chain
executives can rapidly and efficiently access innovative supply
chain solutions [9] [10].
Next we discuss how cloud computing enablers may assist with the
transport company challenges identified earlier.
enterprise system may be the use of a tacit-driven [14] Knowledge
Management System (KMS) [23]. The tactical use of a cloud-
based KMS as part of an enterprise system could enhance the user
experience of the drivers arriving at any of the three Impact mills.
7.3 Cloud logistics
Cloud-integrated logistics could play a critical role in reducing a
transport company‟s logistics cost. It not only provides for more
information in real-time, it makes such information accessible to
the entire team, irrespective of location or time (provided the
necessary connections are available). Universal accessibility allows
for access to processes that require round-the-clock insights.
Logistics managers can observe processes in real-time from remote
locations and it allows for immediate resource deployment in case
response is needed [16].
While logistics optimisation and management remains a rather
complex task (refer0), cloud technology has the potential to make
these processes more efficient and improve on the user experience
(UX). Higher quality information delivered in real-time improves a
manager‟s control over critical processes while increasing the
flexibility of solutions [21].
As with IT infrastructures, transport company logistics may be
supported by a cloud SaaS service model.
The value added through the use of cloud computing by a transport
company, is summarised in T.
Table 1: Transport company cloud services.
CLOUD COMPUTING
UTILITY CLOUD SERVICE CLOUD SERVICE MODEL
IT Infrastructures SaaS
- Enterprise systems - Subscription for hardware,
networking and software
- CRM services.
- KMS
Transport Logistic
SaaS
 - Supply chain - Information in real time.
management (SCM)
- Improving user experience
(UX).
The information in this section provides an answer to RQ2.
VIII. CONCLUSIONS AND FUTURE WORK
Owing to the deregulating of freight movement on South African
roads and the decrease in rail volumes out of the mills, the net
result has been an increase in freight movement through paper
processing mills which have resulted in an increase in turnaround
times at such mills. The increased turnaround times incur
additional costs for the owners of the fleet, resulting in higher
transport tariffs for the customer of such services. One of the
challenges faced is the outlay of the recoverable fibre transport
matrix – with all the individual processing plants being located in
KZN and the recoverable fibre being transported from areas across
South Africa. Subsequently, personnel have difficulty in optimally
planning the movement/routes of the trucks.
In an attempt to alleviate the challenges, the researchers embarked
on a survey among the drivers delivering at these mills. A
REFERENCES
[1] S F. Lam. 2007. An accurate monitoring of truck waiting and flow
times at a terminal in the Los-Angeles ports. Final Report – Metrans
Project AR 05-01/.
[2] D De Jager. 2009. The assessment of improvements made in the
freight logistics costing methodology in South Africa from a
macroeconomic perspective. University of Stellenbosch Department
of Logistics.
[3] H. Steyn, C. Monismith, W. Nokes, T. Harvey, J Holland and N
Burmas. 2012. Challenges confronting road freight transport and
the use of vehicle-pavement interaction analysis in addressing these
challenges. Journal of the South African Institution of Civil
Engineering Vol 54 No 1, April 2012, Pages 14-21, Paper 814.
[4] N. Henderson. 2013. Meeting the challenges of transportation in
South Africa. CEO Barloworld Transport Solutions.
[5] Mpact Mills – Paper manufacturing. Available online at:
http://www.mpact.co.za/our-products/paper-business/paper-
manufacturing Accessed on 18 June 2015.
[6] J. van Tonder 2014. Improving on long vehicle turnaround times at
Mpact Mills, MBL dissertation, Unisa SBL, 2014.
[7] SmartFile. The differences between IaaS, SaaS, and PaaS. Available
online at: http://www.smartfile.com/blog/the-differences-between-
iaas-saas-and-paas/ Accessed 13 June 2015.
[8] D. Kumar. 2013. Challenges facing supply chains in South Africa.
KPMG Report
[9] R. Saideep and S. Aditya. 2014. Supply chain management in the
cloud How can cloud-based computing make supply chains more
competitive?Accenture.
[10] H. Schrodl. 2012. Adoption of Cloud Computing in Supply Chain
Management Solutions. Otto-von-Guericke University Magdeburg,
Magdeburg, Germany.
[11] B. Bhaskar and B. Narkhede. 2013. Exploring the Green Supply
Chain management: A Technical Review. International Journal of
questionnaire was distributed and it was confirmed that drivers
were paid on an incentive basis which is related to productivity and
hence the kilometres they drive per month. The research indicated
that driver morale is negatively affected by the long turnaround
times, leading to time lost on the road and subsequent loss in
income.
A further finding was that the use of innovative ICTs would be of
value. Some correlation was also revealed between the friendliness
of the personnel at the mills and the queue waiting times. The use
of a different kind of truck was also suggested.
On the strength of the recommendation for ICTs, the researchers
investigated how cloud-based technologies in the supply chain may
provide an answer to these. It was found that the SaaS service
model may be used for a cloud-based enterprise system embodying
a CRM (Customer Relationship Management) system and
knowledge management. As a result the company‟s IT
infrastructures and logistics management would be strengthened.
Further research would be required to study time and motion
properties of the waiting trucks. A cloud-based framework could
also be developed for road transport companies.
Application or Innovation in Engineering & Management (IJAIEM),
pp. 441 – 450, Volume 2, Issue 5, May.
[12] Z Muhamad. 2012. Green Supply Chaining. International Journal
of Managing Value and Supply Chains, pp. 1 – 18, Vol. 3, No. 1,
March
[13] J. Abrahamian. 2010. Risk Reducing Product and Process Design
During New Product Development. Delphi Corporation Advanced
Powertrain – USA Published 04/12/2004.
[14] W. Friedrich and van der Poll, J A. 2007. Towards a Methodology to
Elicit Tacit Domain Knowledge from Users. Interdisciplinary
Journal of Information, Knowledge and Management (IJIKM),
Volume 2, 2007, pp. 179 – 193. ISSN: Print 1555-1229. Available
online at: www.ijikm.org
[15] K.C. Laudon and J.P. Laudon. 2013. Management Information
Systems, 14th edition, Pearson.
[16] J Wyatt. 2013. 7 Benefits of Cloud-Based Logistics Management.
Supply Chain, 247. May 19.
[17] P. D. Leedy and J. E. Ormrod. 2005. Practical research: planning
and design. 9th edition. London: Pearson Education.
[18] J. Creswell. 2003. Research design: Qualitative, Quantitative and
Mixed Methods Approaches, 4th edition, Sage Publications.
[19] J Mouton. 1996. Understanding Social Research, van Schaik,
Pretoria.
[20] Wei Zhao, Yong Peng, Feng Xie and Zhonghua Dai. 2012. Modeling
and simulation of cloud computing: A review," Cloud Computing
Congress (APCloudCC), IEEE Asia Pacific, pp.20 – 24, 14-17 Nov
[21] P. Mousumi. 2012. Dynamic job Scheduling in Cloud Computing
based on horizontal load balancing. Department of CSE National
Institute of Technology, Durgapur.
[22] C V Brown, D W DeHayes, and J A Hoffer, E. 2014. Managing
Information Technology, 7th edition, Pearson.
[23] A Tiwana. 2002. The knowledge management toolkit. Upper Saddle
River, NJ: Prentice Hall.
 APPENDIX A – Driver questionnaire.
Mpact Recycling
Recoverable Fibre Turn Around Process
We are analysing the recoverable fibre turn around process and would appreciate your input.

Please enter your details below :

Name

1 Age

Telephone Number

E-Mail address

2.a. What time do you normally arrive to off-load at Felixton mill? 6:01 - 12:00 12:01 - 18:00 18:01 - 24:00 24:01 - 6:00

2.b. What time do you normally arrive to off-load at Piet Retief mill? 6:01 - 12:00 12:01 - 18:00 18:01 - 24:00 24:01 - 6:00

2.c. What time do you normally arrive to off-load at Merebank mill? 6:01 - 12:00 12:01 - 18:00 18:01 - 24:00 24:01 - 6:00

What time of the day would you consider the best to arrive at Felixton Mill - allows for quick
3.a. 6:01 - 12:00 12:01 - 18:00 18:01 - 24:00 24:01 - 6:00
turn-around

What time of the day would you consider the best to arrive at Piet Retief Mill - allows for
3.b. 6:01 - 12:00 12:01 - 18:00 18:01 - 24:00 24:01 - 6:00
quick turn-around

What time of the day would you consider the best to arrive at Felixton Mill - allows for quick
3.c. 6:01 - 12:00 12:01 - 18:00 18:01 - 24:00 24:01 - 6:00
turn-around

Enter the steel safety

Truck proceed to harness
Enter and proceed to gauge to take off the Forklift offload the Proceeds to the gate
4.a. What phase of the offloading process at Felixton mill takes the longest? area, for drivers to take of
the weighbridge straps securing the bales from the truck. to exit factory
the tarps.
bales.

Enter the steel safety

Truck proceed to harness
Enter and proceed to gauge to take off the Forklift offload the Proceeds to the gate
4.b. What phase of the offloading process at Piet Retief mill takes the longest? area, for drivers to take of
the weighbridge straps securing the bales from the truck. to exit factory
the tarps.
bales.

Enter the steel safety

Truck proceed to harness
Enter and proceed to gauge to take off the Forklift offload the Proceeds to the gate
4.c. What phase of the offloading process at Piet Retief mill takes the longest? area, for drivers to take of
the weighbridge straps securing the bales from the truck. to exit factory
the tarps.
bales.

5.a. How helpful do you find the Mpact personal at Felixton Mill? Very Helpful Helpful Unhelpful Very Unhelpful

5.b. How helpful do you find the Mpact personal at Piet Retief Mill? Very Helpful Helpful Unhelpful Very Unhelpful

5.c. How helpful do you find the Mpact personal at Merebank Mill? Very Helpful Helpful Unhelpful Very Unhelpful

6.a. How many trucks are on Avg before you in the queue at Felixton Mill? 0 to 3 4 to 7 7 to 10 10 or more

6.b. How many trucks are on Avg before you in the queue at Piet Retief Mill? 0 to 3 4 to 7 7 to 10 10 or more

6.c. How many trucks are on Avg before you in the queue at Merebank Mill? 0 to 3 4 to 7 7 to 10 10 or more

7 Do you feel that Mpact is trying there best to turn the trucks around quickly? Yes No

8 Are you incentivised and paid per kilometre you drive? Yes No

9 If you could choose not load for Mpact - would you? Yes No

10 Which of the three mills take the longest to turn your truck around? Felixton Piet Retief Merebank

11 Would the use of taut-liner trailers improve the turn-around time? Yes No

How much longer does it take on avg. to turn-around at Felixton Mill compared to other
12.a 0-1Hrs. 1-2Hrs. 2-3Hrs. 3 or more Hrs.
offloading points.

How much longer does it take on avg. to turn-around at Piet Retief Mill compared to other
12.b 0-1Hrs. 1-2Hrs. 2-3Hrs. 3 or more Hrs.
offloading points.

How much longer does it take on avg. to turn-around at Merebank Mill compared to other
12.c 0-1Hrs. 1-2Hrs. 2-3Hrs. 3 or more Hrs.
offloading points.
How to Detect Unknown Malicious Code Efficiently?
Jaehee Lee
Department of Information Security
Korea University
Seoul, Republic of Korea
foodlook88@gmail.com
Hyunsik Yoon
Department of Information Security
Korea University
Seoul, Republic of Korea
Hsyoon900827@gmail.com
Abstract—Recently, rapid developments of IT technology lead
to development of various platforms. With the development of
the new platforms, diverse malicious codes are created to target
the new platforms. These new malicious code means critical and
new threat to national infrastructure, especially the important
ones that can lead to social chaos. In Korea, Korea hydro and
nuclear power was hacked and blueprint was stolen, which was
later posted on-line. This created great problem as the place was
hacked was critical infrastructure. Thus, the vaccine related are
searched out as effective method to analyze the malicious codes
that are created every day uncontrollably. However, Personals
that manage malicious codes are limited compared to newly
create malicious codes. How to detect unknown malicious code
efficiently that remain to be unanswered? However, to answer
this question, malicious code analysis method has to be
concerned, especially the critical ones first. In order to analyze
the Unknown malicious codes effectively, Unknown malicious
code detection model was introduced in the previous paper.
However, this model sometimes treated normal file as malicious
code. This eventually decreased its effectiveness in finding and
analyzing the malicious codes. . Thus it became necessary to
decrease the misdetection rate in order to increase the
effectiveness of the model. As a result in this research, we created
specific conditions that lead to decrease the miss detection rate
significantly. Hence in this paper, we presented a method that
detects the Unknown malicious codes more efficiently.
Keywords—IT security; Unknown malicious code detection
model; Critical Infrastructure
I. INTRODUCTION
Information technology (IT) has become our life. It is
basically involved in our daily activities. Without technology,
we cannot live comfortably. With the advancement of
technology, issues related to data and file technology. With the
wide spread use of IT in society, various new IT based
platforms are being created. With development of new IT
platforms, malicious codes that targets the platform is also
being created and this makes preexisting Signature based
malicious code detection difficult. Unknown malicious code
has become important and advanced Persistent Threat to
Critical Infrastructure, one among them was hacking of
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Hyoungjun Kim
Department of Consulting
Ahnlab
Seoul, Republic of Korea
hyounjun.kim@ahnlab.com
Kyungho Lee
Department of Information Security
Korea University
Seoul, Republic of Korea
kevinlee@korea.ac.kr
Korean hydro and nuclear power incident which became great
national security threat. Unknown malicious codes are in
action consistently without being detected by security system.
To achieve its malicious goals, codes use variety of
mechanisms such as stealing the password information or
creating traffic in order to attack other host by using
connection of command control server [1]. However due to
limitation of the professional personals in the area, it is nearly
impossible to analyze every malicious code in short period of
time. Thus, analyzing the Unknown malicious code prior to
Computer System that causes severe problems has become
priority. . In order to solve the problem, Advanced Unknown
Malicious Code Detection Model deciding the priority of
malicious code was presented in previous paper [2]. However,
after analyzing one hundred normal files and one hundred
malicious files, cases were found which treated normal files as
malicious decreasing the effectiveness of the malicious code
detection. Thus, finding a method to decrease the identifying
of normal file as a malicious became a critical issue which led
to research in the field.
II. RELATED WORK
In this section, previous work regarding research method
for detection of malicious code is effectively introduced and
discussed. Further description on recent research is discussed
along with the problems that came with the previous method.
A. Static Analysis of Executables to Detect Malicious
Patterns
Detecting malicious code is an important part of
information security [3]. In this paper, they present a static
analysis of executables to detect malicious code patterns. In
malicious code detection, malicious code writer try to
obfuscate malicious code detection for evading anti-virus
software. Authors tested the efficiency of three commercial
anti-virus software against code-obfuscation. The anti-virus
software couldn’t detect that codes are obfuscated. So they
present a new architecture for detecting obfuscated malicious
code.
 In these days, the number of mobile malware has been
Group
increasing. Researchers tried to detect malicious java applets IDa Rule Description DRb IDGc
Description
by using static analysis [4].
But that alone is a static analysis to detect malicious code, Abnormal section form Malicious/Suspici
2 56%
there is a limit. So the research on dynamic analysis was detection ous Scan
conducted to complement the signature-based analysis. 2 Malicious/Suspici
Hidden/system/read only ous String
3 17% Packer
files detection
B. Dynamic analysis of malicious code
In this paper, they present a dynamic analysis tool for Unknown SECTION name
4 37%
analyzing behavior of Windows executables [5]. A dynamic detection Signature
analysis of executables is important part of malicious code detection
detection. Because a static analysis has limits. For example, if 5
Process search or control
2% 3
Process/Thread
an anti-virus software has not some malicious code patterns, it API usage detection test
cannot detect the malicious code. So it need to analyze and Patch/Hook
detection
monitor the malicious code’s action. They developed a 6
Process INJECTION API
10%
monitoring system for analyzing Windows executables. And usage detection
its accuracy was good, so it is used for dynamic analysis of
unknown malicious codes. But there is a limit to analyze the Execution compaction
7 46% Stack/Heap test
(known packer) detection
unknown malicious code by using only dynamic analysis. So File
researcher has used a static analysis and dynamic analysis at 4
generation/deletio
the same time [6]. And some researchers use Cuckoo sandbox Execution compaction n/modification
8 37%
(unknown packer) detection
for dynamic analysis of malicious code [7].
Process vulnerability attack
C. Recent researches on analysis of android maicious code attempt detection
9 (Attempt of finding its 5% Registry
Recently, Android phones suffered from a large malware execution place: Exploit generation/deletio
attack. Most of researches focus on detecting android Zero Call) 5 n/modification
Network
malware. The number of malicious code has rapidly increased Process vulnerability attack
connection
and numerous types of malicious code have been advanced 10 attempt detection ( HEAP 2%
SPRAY )
and progressed, so it is essential to require analysis for
malicious codes in order to defense system. It is hard to detect File concealment attempt
malicious code by its behavior. Malicious code’s registry must 11 5%
detection
be analyzed to detect. Hong et al proposed to a new approach Execution attempt
for Malware analysis method based on registry analysis [8]. Entropy
Attempt of generating thread
12 7% 6 calculation
in other process detection
Malicious/Suspici
III. ADVANCED UNKNOWN MALICIOUS CODE DETECTION Possession of service or ous Scan
MODEL 13 automatic start related 15%
registry handle detection
A. Applying Advanced Unknown malicious code detection Attempt of modifying
model 14 memory of other process 7%
detection Malicious/Suspici
Based on the malicious code detection model that was 7 ous String
presented in the past paper [2], actual malicious code detection IAT hooking attempt Packer
15 2%
program has been designed. The created detection program detection
helped in testing the effectiveness of the model. In the last
study, we tested the detection efficiency targets only unknown 16
Shell code API calling
2% 8
Signature
malicious code. Used at the time of the rule includes a static attempt detection detection
analysis and a Dynamic Analysis. We considered the
characteristics of each rule was the grouping [9]. Rule based PE generation in system path Process/Thread
17 44%
on the total of 24, we were measured malware detection detection test
efficiency of the rule results are as follows. 9 Patch/Hook
Compact executable file detection
18 54% Stack/Heap test
generation detection
TABLE I. GROUPING RESULT
Registration of registry in File
19 service or automation start 68% 10 generation/deletio
a b c Group
ID Rule Description DR IDG detection n/modification
Description
Registry
Suspicious IP address
Compaction of execution 20 5% 11 generation/deletio
Entropy connection attempt detection
1 file or encryption (above 29% 1 n/modification
calculation
entropy 7)
 TABLE III. CORRELATION ANALYSIS RESULT
Group
IDa Rule Description DRb IDGc
Description Severity Detection Rate
Key input information Pearson
21 interception attempt 10% 1 -.721**
correlation
detection Significant
Severity probability .000
System utility execution (2-tailed)
22 7%
blocking
Network N 24 24
12
connection
Execution in suspicious path Pearson
23 39% -.721** 1
detection correlation
Significant
Detection
Batch file execution attempt probability (2- .000
24 10% Rate
detection tailed)
a. N 24 24
Rule ID
b.
Detection Rate
c.
Group ID The above represents the results as a basis to Grouping
Reflect the rule of the characteristic used at the time was to Detection Rate and Severity correlation relationship equation
create a detection group with a total of 12 Group results of the as follows.
Grouping. We conducted a Focus Group Interview targeting
malicious code analysis of Ahnlab. Severity of infection was
calculated for when the action corresponding to the rule. We 
were able to achieve the same results to the following.
Y : Group Detection Rate
X : FGI Score
TABLE II. GROUP WEIGHT RESULT
C : Constant : 0.727
Detection
IDGd Group Description Severity
Rate
1 Entropy calculation 7.0 32% We created a formula that determines the Ordering score on
2 Malicious/Suspicious Scan 4.0 84% the basis of the expression.
3 Malicious/Suspicious String 3.3 51%
4 Packer 3.5 62% 
5 Signature detection 7.0 8% Y : Group Detection Rate
6 Process/Thread test 6.0 30% X : FGI Score

7 Patch/Hook detection 6.0 8% In other through the placement of malicious code detection
rule this expression was also as expected predict actual results
8 Stack/Heap test 6.0 3%
that were higher than the malware detection efficiency. We
File looked out the detection rate of infection consisting of 41
9 1.0 92%
generation/deletion/modification
Registry
samples when the arrangement order of the highest as
10 4.0 76% Ordering Score Group, sequentially combined. The results
generation/deletion/modification
11 Network connection 10.0 5%
were as follows.

12 Execution attempt 1.0 51%

TABLE IV. GROUP DETECTION RESULT
d.
Group ID

C(n)e 2 3 4~5 6~9 10~14 15

We used a statistical analysis in order to place in
consideration of the detection efficiency, the arrangement Detection
34 37 38 39 40 41
order of the group. In fact, as well as the infection detection number(n)
rate of the group to the arrangement at the same time taking
into account the severity of the group, because it is possible to Detection
82.9% 90.2% 92.7% 95.1% 97.6% 100%
rate(n/41)
more quickly determine the infection. Because, even if
malicious code, such as different degrees adversely affects the e.
Number of combination group
computer according to the severity being likely to exploit We saw configuring a series of steps to place the group of
different. Therefore, we analyzed the correlation was analyzing malware in Figure 1 in the process.
performed. The results were as follows.
Figure 1. Malicious code detection Group Ordering Process IV. EXPERIMENTAL RESULTS

A. Dataset
We planned experiments to improve the efficiency of the
analysis program developed by the infection. Advanced
Unknown malicious code detection model presented in the
previous work was applied to one hundred normal files and
one hundred malicious files for examining the effectiveness of
the model and finding and fixing the limitation. These files
were having variety of characteristics and were chosen
randomly and were provided by Ahnlab.

B. Implementation
We were each 100 each analysis to improve the program
and apply the ruleset Unknown malicious code normal files. A
total of 67 applied ruleset rule was applied by adding a rule for
each group at 24 ruleset who mentioned.

We have proposed an algorithm as shown in Figure 2 in
order to reflect the actual malware analysis program
implementing this process. We have developed a program that
detects unknown malicious code on the basis of this algorithm.
Figure 2. Advanced Unknown Malicious code Detection Model
C. Results
After experimenting on these sample files there were some
problems as expected. In some instances normal file was
treated with similar characteristics as malicious code and was
categorized as malicious code.
This resulted increased the miss detection rate. The result
of the experiment is as follows.

TABLE V. TEST RESULT

Total Test Number of Misdetection
Index
Trial detection rate
Malicious
100 100 0%
code
Normal file 100 45 45%

In the above experimental results showed a performance

better detect malware that year with a probability of 100% or
plain files can be found that the misdetection probability of
45%. The limitation of the previous model was that it had too
high miss detection rate. The main l aim of this study was to
increase the effectiveness of malicious code detection. To
reduce the Misdetection rate we saw check out what this rule
for determining normal files with malicious code. As a result,
the following rule has been determined to reduce the malicious
code detection efficiency.

TABLE VI. TROUBLED RULE

Number of
Rank ID Rule Description
Misdetection
compression or encryption (over
1 2 30
entropy 7) file detection

2 6 TLS call back function detection 11

detection of execution file

The main purpose of the program was to observe whether 3 24
compression by known packer
10
it can detect the Unknown malicious code effectively or not
for that reason, proving the effectiveness through testing was 4 51 detecting copy of own file 6
important.
Rank ID Rule Description
Number of malicious code that occur constantly with limited manpower
Misdetection advanced malicious code detection model can be this solution.
abnormal section characteristics And will continue to be an improvement over the algorithm is
5 5 5
detection also presented misdetection occur in the future dynamic
detection of execution file analysis. If malicious code that has not been analyzed come in
6 26 5
compression by unknown packer
to appearance, New Rule will be generated to increase the
detection of service or automatic
7 53 execution related registry 4 efficiency of the model. Also, when a new rule is added to
registration act increase the miss detection rate would be removed through the
same process it is complemented by the other rule. We are
planning a new study through the model we have developed.
Taken together, the above results also show that behavior
During the test, it was found malicious code that has dynamic
such as packing or encryption, copying, automatic execution
characteristics, act differently in different Windows. In next
file it showed that the general.
research problem, the reason of the action will be analyzed
and model that analyze the malicious code in different
D. Improved process
Windows.
We remove the rule and looked again conducted a test. The
result is shown below. Acknowledgment
This research was supported by the MSIP (Ministry of
TABLE VII. TEST RESULT 2
Science, ICT and Future Planning), Korea, under the ITRC
Index
Total Test Number of Misdetection (Information Technology Research Center) support program
Trial detection rate (IITP-2015-R0992-15-1006) supervised by the IITP (Institute
Malicious for Information & communications Technology Promotion)
100 100 0%
code
Normal file 100 5 5%
References
The results of the above misdetection rate of the infection [1] Sikorski, Michael, and Andrew Honig. Practical Malware Analysis: The
can be confirmed that a normal file that improved to 5-45% of Hands-On Guide to Dissecting Malicious Software. No Starch Press,
2012.
the original.
[2] Hyoungjun et al. “Advanced Unknown Malicious Code Detection
Model.” Research Briefs on Information & Communication Technology
V. CONCLUSION Evolution 1.1 (2015)
[3] Christodorescu, Mihai, and Somesh Jha. Static analysis of executables to
Unknown malicious code is modified version of the detect malicious patterns. WISCONSIN UNIV-MADISON DEPT OF
preexisting malicious codes or newly created. Thus, detection COMPUTER SCIENCES, 2006.
has to consider the signature and dynamic property as well. [4] Armando, Alessandro, et al. "SAM: The Static Analysis Module of the
This makes unknown malicious code detection difficult and MAVERIC Mobile App Security Verification Platform." Tools and
time consuming. But also the behavior of a normal file in the Algorithms for the Construction and Analysis of Systems. Springer
Berlin Heidelberg, 2015. 225-230.
dynamic analysis confirmed that there would be progress with
[5] Bayer, Ulrich, et al. "Dynamic analysis of malicious code." Journal in
malicious code similar pattern. By disregarding the Rule that Computer Virology 2.1 (2006): 67-77
lowered the efficiency of Unknown Malicious code Detection [6] Wang, Xiaolei, et al. "A Novel Hybrid Mobile Malware Detection
Model, the lowering of misdetection rate was successful. System Integrating Anomaly Detection With Misuse Detection."
Developing programs that analyze malware and will very Proceedings of the 6th International Workshop on Mobile Cloud
Computing and Services. ACM, 2015.
often occur in the process of doing the same operation as
[7] Vasilescu, Mihai, Lucian Gheorghe, and Nicolae Tapus. "Practical
above. We have to present a decision-making process to solve malware analysis based on sandboxing." RoEduNet Conference 13th
this reflected in the program for future improvement. Malware Edition: Networking in Education and Research Joint Event RENAM
detection solution development is very important for national 8th Conference, 2014. IEEE, 2014.
security. As well as the United States has invested heavily in [8] Lee, Sungjin. "New Malware Analysis Method on Digital Forensics."
order to train specialists analyze malicious code in several Indian Journal of Science and Technology 8.17 (2015).
countries. Nevertheless, the professional malware analysis [9] Ligh, Michael, et al. Malware analyst's cookbook and DVD: tools and
techniques for fighting malicious code. Wiley Publishing, 2010.
expertise is very low. In the process of analyzing the unknown
 Green ICT Maturity Models
Towards a general approach

Ravi Foogooa
rfoogooa@umail.utm.ac.mu

Chandradeo Bokhoree
sbokhoree@umail.utm.ac.mu

Kumar Dookhitram
University of Technology, Mauritius
La Tour Koenig
Pointe-aux-Sables, Mauritius
kdookhitram@umail.utm.ac.mu

Abstract—A lot of effort is still required in greening ICT and
the use of ICT to green organisations. However, Green ICT
initiatives are hard to sustain. In this context, Green ICT
maturity models help by providing a benchmarking tool and a
roadmap. However, several Green ICT maturity models have
been proposed by different researchers over the years with no
clear justification. This makes it difficult for companies to choose
which Green ICT maturity model to adopt. This research aims at
comparing the different Green ICT maturity models. This could
help companies manage their Green ICT initiatives in a more
sustainable way. It will analyse the evolution of Green ICT
maturity models and provide cues for further research in this
area.
Keywords—Green ICT, maturity models, sustainable
development
I. INTRODUCTION
Much effort has been put so far in sustainable development
and at least the rate of unsustainable growth has slowed down
in recent years. Nevertheless, a lot still remains to be done in
sustainable development as the world is getting less and less
sustainable. It is the duty of everyone to bring his contribution
to make the world a better place for ourselves and our future
generations. Research on the link between Information and
Communication Technologies (ICT) and sustainable
development has shown that ICT does account for 2% of
global carbon emissions [1] with its power consumption and e-
waste on the upper trend. However, ICT could also be used to
reduce the impact of human activities on the environment by
15% by 2020 [2] especially with applications geared towards
dematerialisation use such as electronic transactions, video
conferencing and ecommerce amongst others. However, ICT
can have a positive or negative impact on the environment
depending on how it is used [3]. There has been considerable
research on prescriptive directives on how to reduce the
impact of ICT on the environment such as in [4]–[7] amongst
others which are focused on reduced resource use. However,
getting reliable data for such assessments is a daunting task
[8]. Initial enthusiasm easily dies out in such cases. Thus,
there is need for a carefully devised Green ICT strategy to
craft an organisation‟s way to a harmonious development [4].
In this context, maturity models are very helpful. They can
help in assessing the current status of companies with respect
to best practices and they can also support in planning a
roadmap for promoting environmental sustainability
initiatives.
However, a number of Green ICT maturity models have
been developed over the years, for example in [9]–[13]. They
are based on different expectations, ambitions and assessment
criteria and it can therefore be difficult to choose which model
to adopt. An inappropriate choice could either not be
motivating enough for a company or further discourage it and
kill new Green ICT initiatives in the bud itself. In order to
overcome this difficulty, we set off to compare and contrast
several existing Green ICT maturity models in order to
understand the functioning and limitations of these models
through a cross comparative analysis. Subsequently, a generic
approach to Green ICT maturity is developed and presented
herein.
II. GREEN ICT MATURITY MODELS
Maturity models are not new. Way back in 1993, the
Software Engineering Institute of Carnegie-Mellon University
presented the Capability Maturity Model (CMM) for process
maturity [14] which over the years evolved in the Capability
Maturity Model Integrated (CMMI). It defined maturity in a
number of levels each characterised by a specific behaviour.
Organisations were encouraged to assess the level at which
they were and then follow the characteristics of the next level
to define initiatives for process improvement in their
organisation. The advantage of the maturity model was that it
allowed benchmarking across different organisations and also

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

provided the framework for a roadmap for process
improvement.
This approach seems to have inspired Graeme who
proposed a Green ICT maturity model [9]. This model
(referred to as Model 1) is based on the estimation and
optimisation of attitude, policies, practice, technology and
metrics in the effort of the organisation towards Green ICT. It
also looks at different categories such as equipment life-cycle,
end user computing, enterprise & data centre and ICT as a low
carbon enabler. The proposed maturity levels and their
characteristics are:
 0 – No Intention : No awareness
 1 – Initial : Some awareness but no implementation
 2 – Replicable: Ad-hoc implementation
 3 – Defined: Formal but immature programs
 4 – Managed: Methodical implementation with
proper measurement
 5 – Optimisation: All activities managed for optimal
performance
Desai & Bhatia also proposed another Green IT maturity
model (referred to as Model 2) [10]. It measures several
aspects of an organisation's "operations, behaviour and
achievements". It is centred on data centre and facilities, end
user computing, asset lifecycle, IT service management and
people activities. The different maturity levels are:
 Level 1 - Deficient : The company is evaluating green
IT benefits but has no proven implementations yet
 Level 2 - Monitored: The company measures the
carbon emissions of the IT activities in the
organisation and effectively reduces them by 10%
 Level 3 - Optimised: Carbon emissions from IT in the
organisation are reduced by at least half.
 Level 4 - Dynamic: The IT department is almost
carbon neutral and reports about it
 Level 5 - Sustainable: IT is expected to help the
business reduce it carbon footprint
The model prescribes a series of requirements for each
pillar for each maturity level. Moreover, it uses carbon
footprinting to assess maturity.
Donnellan, Sheridan & Curry proposed a Sustainable ICT
Capability (SICT) Framework (referred to as Model 3) [11].
The framework consists of the following actions:
 define the scope and objectives of ICT
 appraise the existing sustainable ICT maturity
 grow and monitor SICT capability building blocks
 evaluate and monitor evolution of SICT capability
The different maturity levels are:
 Initial - Little awareness of sustainable ICT
 Basic - Limited sustainable ICT strategy but
implementation is immature
 Intermediate - Presence of sustainable ICT strategy
but although it is included in life cycle of systems, it is
restricted to project level
 Advanced - Sustainability is a central theme of
business and IT life cycles.
 Optimising - Sustainable ICT practices are applied
across the whole supply chain for the organisation.
Hankel, Oud, Saan & Lago defended the need for a Green
ICT Maturity model because most Green ICT initiatives are
focused on reducing negative impacts of ICT [12]. They
pointed out that, on one hand, a large part of ICT carbon
emissions are with the end users and that, on the other hand,
the production of ICT equipment accounts for the majority of
their carbon footprint. They blamed current initiatives which
focus too much on low hanging fruits which can be good for
the short term but which prevents an optimal use of Green
ICT. They therefore built a Green ICT maturity model by
interacting with ICT experts over a workshop and then
validated the model through a survey. The feedback allowed
them to improve the model qualitatively although it did not
allow quantitative conclusions.
The model (referred to as Model 4) is based on CMM and
includes the following levels:
 Level 1: Initial
 Level 2 : Repeatable
 Level 3: Defined
 Level 4: Managed
 Level 5: Optimising
The model is centred on Green ICT in the organisation,
Greening of ICT and Greening of operations with ICT.
Although the model was built after consultation with Dutch
higher education, the model can be used for all types of
organisations.
DeMonsabert, Odeh & Meszarros defended the need for a
framework to avoid inconsistencies in Green ICT initiatives
and to enable an evaluation of same across organisations [15].
The authors built their model with objectives of environmental
protection, economic feasibility, social responsibility and
innovation. Their effort was a joint collaboration between a
university and a Green ICT Non-Governmental organizations.
. They then embedded an evolutionary approach in their model
- the first release was implemented in a company and feedback
sought to build the second version of the model. The second
version was oriented along the different dimensions of
sustainability, namely social, environmental, economic and
innovation.
This model is new in terms of its coverage of non-
environmental aspects of sustainability. Its social
sustainability criteria assess sustainability governance,
workforce, consumers and suppliers, local community and
society. However, it does not provide a maturity model per se
and thus was excluded from this study.
The UK Government proposed a model (referred to as
Model 5) which is based on CMMI model and is in line with
the Green ICT strategy of the UK government [13]. The
maturity levels are:
 Foundation - evidence and intelligence gathering to
inform actions, agreed plans
  Embedded - show commitment and basic initial
development, basic processes in place
 Practised - moving forward taking actions to improve,
repeatable actions
 Enhanced - pushing for new opportunities, adoption of
best practice, improving capability
 Leadership - taking control, having own vision,
optimising performance
They also defined 2 additional levels of „ad-hoc‟ (where
there is no clear strategy) and „not applicable‟ (where it is not
worth or not possible to have an assessment). It focuses on the
following areas (with a subdivision into further subcategories):
 Manage ICT Services
 Manage ICT Technology
 Changing ICT Services
 Exploiting ICT
It can be seen that the different Green ICT maturity models
have some features in common but are at first glance not
identical. However, none of the proposed models (except in
[12] which gave a superficial overview of the different
maturity models) compare their work with previous work nor
justify the need for a new maturity model. There is thus need
for a comparison between the different Green ICT maturity
frameworks to assess their quality and their appropriateness to
an organisation.
III. RESEARCH METHODOLOGY
We adopted a multiple case study approach for our
research. After a review of the materials on the available
Green ICT maturity models, it was obvious that it was not
easy to assess a company as per a maturity model. This was
because many of the models required assessments by
consultants although in some cases such as in [12], [13], there
were detailed guidelines to perform the assessment while in
others, third-party self-assessments such as in [16] were
available. We thus set out to devise a way to carry out a quick
preliminary assessment. The approach followed was to first
select a Green ICT maturity model and devise questions which
would allow us to verify if the company met the lowest
maturity level in that model. If this was the case, then same
would be applied for the next higher maturity level and so on
and so forth until the company did not match up to a certain
maturity level. The final assessment was the maturity level
just below the one which the company failed to meet.
We then prepared a list of fictitious companies with
different experiences with Green ICT generated from each
level of all the selected Green ICT maturity models. These
companies were then assessed with all the other Green ICT
maturity models to compare the different assessments. The
results were then analysed.
We are aware that the ideal situation would have been to
apply the Green ICT maturity model assessment with the
proper consultants on selected real companies with different
Green ICT experiences. However, for various reasons such as
lack of resources and time, we had to resort to an alternative
approach which would still be appropriate for our research
objectives.
IV. RESULTS & ANALYSIS
The assessment criteria devised for the different Green
ICT maturity models are as follows:
 Model 1 [9]
Level 0: Basic question on general Green ICT awareness.
Qualifies for the level if displays no awareness of Green
ICT.
Level 1: Some questions on Green ICT awareness and
implementation. Qualifies for the level if company has some
awareness but has not implemented any Green IT initiatives.
Level 2: Question on any formal Green ICT strategy.
Qualifies for the level if company has implemented some
Green ICT initiatives but has no formal Green ICT strategy.
Level 3: Questions on implementation of Green ICT
strategy. Qualifies for the level if company has a Green ICT
strategy but is still immature in its approach.
Level 4: Question on measurement and management of
Green ICT strategy. Qualifies for the level if makes use of
metrics in the implementation and control of the Green IT
strategy.
Level 5: Question on overall activity of the organisation.
Qualifies for the level if the impact on the environment is
optimised for all activities.
 Model 2 [10]
Level 1: Questions on Green ICT awareness and any
initiatives taken in this direction. Qualifies for the level if
company has no Green ICT implementations.
Level 2: Questions on measurement of achievement of
Green ICT initiatives. Qualifies for the level if company is
able to reduce the carbon emissions of the ICT activities by
10%.
Level 3: Questions on governance of Green ICT initiatives.
Qualifies for the level if there is a dedicated organisation for
driving Green ICT initiatives and if these result in a reduction
of 50% of the carbon emissions of the ICT activities of the
company.
Level 4: Questions on carbon emissions reporting due to
Green ICT initiatives. Qualifies if the ICT department is
almost carbon neutral and reports on it.
Level 5: Questions on carbon emissions reductions for the
whole organisation. Qualifies if Green ICT initiatives help
reduce carbon emissions for the whole organisation.
 Model 3 [11]
Initial: Questions on Green IT awareness and any
implementations. Qualifies for the level if company has only
little understanding and a few policies.
Basic: Questions on Green IT strategy and its
implementations. Qualifies for the level if company has a
Green IT strategy but implementation is immature with no
clear accountability.
Intermediate: Questions on sustainable IT strategy and on
the control of its implementation. Qualifies for the level if the
company has a full sustainable IT strategy with targets and
metrics at the individual project level.
Advanced: Questions on the importance of sustainability in
IT and business life cycles. Qualifies for the level if the
company puts sustainable ICT at the heart of IT and business
planning and both IT and the business drive the efforts
together.
Optimising: Questions on sustainable ICT across the
extended enterprise. Qualifies for the level if the company
adopts sustainable ICT practices across its full supply chain.
 Model 4 [12]
Level 1 – Initial: Questions on Green ICT awareness and
practice in the organisation. Qualifies for the level if there is
little awareness /practice, for example, in terms of e-waste and
energy efficiency and there is no Green ICT strategy for the
organisation.
Level 2 – Repeatable: Questions on Green ICT strategy
and scope of its implementation. Qualifies for the level if there
is only a basic Green ICT strategy and the scope of
implementation is restricted to the ICT department.
Level 3 – Defined: Questions on Green ICT strategy and
scope of its implementation. Qualifies for the level if there is a
clear Green ICT strategy geared towards resource
consumption and the ICT department is a regular partner in
greening of some processes across the organisation.
Level 4 – Manageable: Questions on the inclusion and
review of Green ICT strategy in the organisation and the role
of ICT department in greening of the organisation. Qualifies
for the level if the Green ICT policy is taken into
consideration in all business processes and it is itself reviewed
regularly.
Level 5 – Optimising: Questions on the role and scope of
the Green ICT strategy. Qualifies for the level if the Green
ICT strategy is broad, allows participants to moderate their
actions to optimise their impact on the environment and is
applied beyond the frontiers of the organisation to reach the
whole supply chain.
 Model 5 [13]
Level 0a – Not applicable: This is a decision by the
organisation that it is not worth or possible to carry out an
assessment.
Level 0 – Ad-hoc: Questions on Green ICT strategy.
Qualifies for the level if there is no agreed Green ICT plan in
place in the organisation.
Level 1- Foundation: Questions on Green ICT strategy.
Qualifies for the level if there is an agreed Green ICT plan in
the organisation and environmental issues are considered in
business decisions.
Level 2 – Embedded: Questions on Green ICT strategy and
committed resources. Qualifies for the level if there are
resources committed on the Green ICT strategy and initial
actions have started.
Level 3 – Practised: Questions on implementation of
Green ICT strategy across the organisation. Qualifies for the
level if the Green ICT strategy is agreed upon and its
implementation is growing across the organisation.
Level 4 – Enhanced: Questions on control of the
implementation of Green ICT strategy in the organisation.
Qualifies for the level if there is a consistent application of
Green ICT strategy with an emphasis to improve and learn
across the organisation.
Level 5 – Leadership: Questions on inclusion of Green
ICT strategy within the organisation strategy. Qualifies for the
level of effectiveness of Green ICT strategy is measured and if
sustainability concerns form part of business strategy.
We used the selected Green ICT maturity models to
generate the list of fictitious companies as follows:
 Model 1 – companies A-E
 Model 2 – companies F-J
 Model 3 – companies K-O
 Model 4 – companies P-T
 Model 5 – companies U-Y
This approach gave the following list of fictitious
companies with different experiences in Green ICT:
A. Company has no Green ICT awareness.
B. Company has some Green ICT awareness but has not
initiated any Green ICT actions
C. Company has implemented some Green ICT actions but
has no formal Green ICT strategy
D. Company has a formal Green ICT strategy and has started
only started implementing it.
E. Company has implemented its Green ICT strategy and
makes use of metrics to control its implementation
F. Company has no implemented any Green ICT initiatives
G. Company has implemented a few Green ICT initiatives
resulting in a reduction of carbon footprint of its ICT
activities by 10%
H. Company has a dedicated organisation for Governance of
Green ICT and achieves a reduction of carbon footprint of
its ICT activities by 50%
I. Company's ICT department is carbon neutral and reports
its carbon footprint
J. Company uses Green ICT to reduce carbon footprint
across the organisation systematically
K. Company has little awareness and no policies on Green
ICT
L. Company has a Green ICT strategy but implementation
has just started
M. Company has a Green ICT strategy and has proven
implementations with metrics to control it - Green ICT is
a concern in ICT projects. TABLE I. COMPARISON OF GREEN ICT MATURITY
ASSESSMENTS
N. Company has implemented its Green ICT strategy and ASSESSMENT
Green ICT concerns are at the heart of decisions across COMPANY Mod el 1 Mod el 2 Mod el 3 Mod el 4 Mod el 5

the whole business. Green ICT is driven by the business A

B
0
1
0
0
0
0
0
0
0
0
and ICT departments together. C 2 0 0 0 0
D 3 1 1 1 1

O. Company applies its Green ICT strategy across its entire E

F
4
1
2
0
2
0
2
0
2
0
supply chain G 2 1 1 1 1

H 3 2 2 2 2
P. Company has little awareness / practice in Green ICT and I 4 3 2 2 2
J 4 4 3 3 4
does not have a Green ICT strategy K 1 0 0 0 0
L 3 1 1 1 1
Q. Company has a basic Green ICT strategy but the scope of M 4 2 2 2 2

its implementation is restricted to the ICT department N 4 4 3 3 4

O 4 4 4 4 4

R. Company has a clear Green ICT strategy geared towards P 1 0 0 0 0

Q 3 1 1 1 1
resource consumption and the scope of this R 3 2 2 2 2
implementation has gone outside the ICT department (but S 4 4 3 3 4

not yet the whole organisation) T

U
4
0
4
0
4
0
4
0
4
0
V 3 1 1 1 1
S. Company has a practice of including Green ICT policy in W 3 1 1 2 2

all business decisions and reviews its Green ICT policy X 3 3 3 3 3

regularly as well.
T. Company has a practice to include Green ICT policy in all
its actions across its entire supply chain
U. Company has no green ICT plan
V. Company has an agreed Green ICT plan and
environmental issues are generally considered in the
company
W. Company has an agreed Green ICT plan and has
committed resources to it with some initial results
X. Company has an agreed Green ICT plan and its
implementation is growing across the company
Y. Company measures the effectiveness of the
implementation of its Green ICT strategy and
sustainability is considered as part of all business strategy
For the analysis, the results of the assessment of each
Green ICT maturity model was slightly modified for
comparison sake. As all the models used a five level scale, all
assessment levels were coded in the number range 0-4 to
facilitate comparison. A colour coding was also used to show
the difference in assessment of the same company by the
different Green ICT maturity models. The colour coding
followed is red, amber, yellow, blue and green for levels 0 to 4
respectively.
The table below shows the results of this analysis:
Y 4 4 3 3 4
The results show that the different models may agree on
low Green ICT maturity companies near the 0 level. However,
they differ on the medium to high Green ICT maturity
companies. The increasing expectations of all stakeholders in
recent times probably explains the difference in the models.
Indeed, the oldest model in [9], although encouraging for low
achieving (from the Green ICT maturity perspective), is less
ambitious and discerning as the other models. There is
however agreement on the need for a Green ICT strategy or
plan and for a rigorous control of its implementation across
the different models. There is also increasing use of metrics
and quantitative ways of controlling Green ICT initiatives for
higher maturity levels. Models such as the one in [10] include
carbon footprinting throughout as a means of assessment.
Although most of the models are ICT department centric with
many initiatives starting from the ICT department itself, they
also recognize that Green ICT cannot be restricted to the ICT
department and reward those that take it across the whole
organization. Some like in [11], [12] even expect high
maturity companies to extend Green ICT initiatives to the full
supply chain. All the different models focus mostly on the
environmental aspect of Green ICT with rare exceptions such
as the one in [15] which was unfortunately excluded from the
study. It is clear though that Green ICT maturity cannot focus
solely on the environmental aspect of sustainability any
longer.
V. GENERAL GREEN ICT MATURITY MODEL
In light of the above, we propose a general Green ICT
model encompassing the different models reviewed. The
model should be discerning at the lower levels to differentiate
between those who are not doing anything and those trying to
devise and implement initial Green ICT initiatives. We feel
that this is important to encourage those starting on the
journey of Green ICT. However, it should also set more
ambitious goals such as the inclusion of metrics at higher
maturity levels. Finally, Green ICT mature companies are
expected to promote Green ICT values not only within their
companies but rather across their full supply chain. Our
proposed general Green ICT model thus includes the
following maturity levels:
0. Reckless: At this level, companies have little or no
awareness of Green ICT and especially believe that
nothing can be done at the company level.
1. Ad-hoc: At this level, there is recognition of Green
ICT but initiatives are ad-hoc and not sustainable.
2. Conscious: At this level, the company has a Green
ICT plan and is committed to its implementation
across the ICT department
3. Responsible: At this level, the company has a Green
ICT plan and is committed to its implementation
across the whole organization. It is important for
metrics to be defined and monitored at this level.
4. Role model: At this level, the company includes
Green ICT issues at all levels of business decision
making and strives to implement it across its full
supply chain. Companies here are in a virtuous
improvement mode for sustainable development of the
business.
This model regroups the main directions and the levels
from several models and will help to provide a standard
guidance for companies wishing to properly manage their
Green ICT initiatives.
VI. CONCLUSION AND FUTURE WORK
The comparison attempted was not a statistical one and as
such, it is difficult to make far reaching generalising claims.
However, a qualitative assessment shows that there is some
agreement as well as clear differences among the different
Green ICT models.
Future work would involve confirming our results through
more quantitative approaches. It would also be interesting to
further develop and test the generic Green ICT maturity
model. There is also need for a taxonomy of Green ICT
maturity models which would give a direction for future
development of Green ICT maturity models.
References
[1] Gartner, “Gartner Estimates ICT Industry Accounts for 2 Percent of
Global CO2 Emissions,” 2007. [Online]. Available:
http://www.gartner.com/newsroom/id/503867.
[2] M. Webb, “SMART 2020 : Enabling the low carbon economy in the
information age,” 2008.
[3] L. Erdmann, L. Hilty, J. Goodman, and P. Arnfalk, “The Future
Impact of ICTs on Environmental Sustainability,” Speville /
Brussels, 2004.
[4] S. Murugesan, “Harnessing Green IT : Principles and Practices,”
IEEE IT Pro, no. February, pp. 24–33, 2008.
[5] S. Ruth, “Reducing ICT-related Carbon Emissions: An Exemplar
for Global Energy Policy?,” IETE Technical Review, vol. 28, no. 3,
pp. 207–211, 2011.
[6] K. Raza, V. K. Patle, and S. Arya, “A Review on Green Computing
for Eco-Friendly and Sustainable IT,” Journal of Computational
Intelligence and Electronic Systems, vol. 1, no. 1–14, pp. 3–16, Jun.
2012.
[7] G. Bekaroo, C. Bokhoree, and C. Pattinson, “Towards Green IT
Organisations: A Framework for Energy Consumption and
Reduction,” The International Journal of Technology, Knowledge,
and Society, vol. 8, no. 3, 2013.
[8] A. Shah, T. Christian, C. Patel, C. Bash, and R. Sharma, “Assessing
ICT ‟ s Environmental Impact,” IEEE Computer, no. July, pp. 91–
93, Jul-2009.
[9] Graeme Philipson, “A Green ICT Framework,” St Leonards, 2010.
[10] B. M. Desai and V. Bhatia, “Green IT Maturity Model : How does
your Organization Stack up?,” SETLabs Briefings, vol. 9, no. 1, pp.
49–57, 2011.
[11] B. Donnellan, C. Sheridan, and E. Curry, “A capability maturity
framework for sustainable information and communication
technology,” IT Professional, vol. 13, no. 1, pp. 33–40, 2011.
[12] A. Hankel, L. Oud, M. Saan, and P. Lago, “A Maturity Model for
Green ICT : The case of the SURF Green ICT Maturity Model,” in
28th EnviroInfo Conference, 2014.
[13] “UK HM Greening Government ICT Strategy Combined
Assessment Model 2015,” 2015. [Online]. Available:
https://www.gov.uk/government/publications/green-ict-maturity-
model.
[14] M. C. Paulk, B. Curtis, M. B. Chrissis, and C. V Weber, “Capability
Maturity Model for Software, Version 1.1,” Pittsburg, Pennsylvania
15213, 1993.
[15] S. DeMonsabert, K. Odeh, and J. Meszaros, “SustainaBits: A
framework and rating system for sustainable IT,” in International
Green Computing Conference (IGCC), 2012.
[16] R. Foogooa and K. Dookhitram, “A Self Green Maturity
Assessment Tool for SMEs,” in IST Africa, 2014.
 QoS-Aware Single Service Selection Mechanism for Ad-Hoc
Mobile Cloud Computing
Ayotuyi T. Akinola, Matthew O. Adigun, Alaba O. Akingbesote
Department of Computer Science, University of Zululand,
Private Bag X1001, KwaDlangezwa, 3886 South Africa.
Tel: +27781813402, Fax: +27359026569
ruthertosin@gmail.com
Abstract—Mobile technology has made notable progress
over the past few years, this is due to the development of new
handheld devices, improved wide area cellular coverage and
seamless integration of wireless data access into mobile devices.
These recent advances in wireless internet technologies have
given birth to the Mobile Cloud Computing paradigm. However,
intermittent internet disconnection and others lead to the
evolution of ad-hoc mobile cloud where mobile devices exposes
their computing resources and residence services to other
devices. One challenge in the ad-hoc mobile cloud is that of
service selection especially in a virtualized ad hoc mobile
environment. This is because the best service to be selected
might have left after discovery as a result of the dynamic nature
of the ad hoc mobile system. To resolve this challenge, we
propose a single service selection mechanism where the instances
or images pertaining to the information of community members
are stored temporarily on centralized virtual node. When a
selection is to be done, these images are searched instead of
searching individual node. A multi-criteria decision is proposed
as our solution approach. Experiment is conducted using
Execution time as the QoS parameter. Evaluation and analysis
are carried out. We first recorded an increase in execution time
as the number of services increases. But at a certain period, the
execution time increment is not well noticeable or proportional
to service increment. This is attributed to the service images that
are already in the virtual leading server thereby reducing the
execution time.
Keywords—Ad hoc cloud mobile; GUISET; service selection;
server mobile machine; SMMEs
I. INTRODUCTION
Cloud computing is an infrastructure service provisioning
paradigm which is based on pay per use mechanism. The
basic goals of this computing paradigm are to remove the
burden of providing infrastructure, platform and software
services from service consumers and also reduce costs. This
will allow service consumers to concentrate on the core
business.
One major improvement and advantage of cloud
computing to subscribers is the issue of mobile services. This
paradigm has allowed the mobile information to be moved to
the cloud and enhance data and information shareability. This
information or services could be obtained through a web
browser interaction with the wireless connection [1].
Mobile subscribers accumulate wealth of experience from
various uses of mobile services such as currency conversion
service and weather forecast service. These services operate
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
on the remote servers and/or on devices through wireless
networks. With the rapid growth in the new compelling era of
mobile computing, technological innovations are observed to
be occurring at an accelerated rate both in low and middle-
income nations [2]. Undoubtedly, ubiquitous demand for
mobile services would be increasing to the point where 60
percent of IP traffic would be originating from mobile
devices by 2016 [3].
The rich experience evolving from the integration of these
mobile devices with the cloud computing architecture through
the aid of wireless internet connections results in the recent
paradigm called Mobile Cloud Computing (MCC) [3]. MCC
simply refers to an infrastructure system that allows mobile
information or services to be moved to the centralized cloud.
This can them be exposed to any registered devices with the
central cloud. The benefits of cloud computing like reliability,
scalability, device-independent and on-demand services
brings a limelight to the functionality of resource constrained
mobile devices through this development.
This improvement motivated the idea of a use case
scenario in our project called Grid-based utility Infrastructure
for SMME Enabling Technology (GUISET) [5].
GUISET is a proposed middleware platform whose basic
goal is to provide utility infrastructure services to small
medium and macro enterprise (SMMEs), most especially in
the context of mobile services like m-Learning, m-Commerce,
m-Health [5]. Because of the richness of MCC, mobile
devices have become part and parcel of human being and
hardly can people walk without it.
Many studies are currently ongoing in different phases of
our GUISET project along the mobile cloud computing
paradigm for proper efficiency and effectiveness. Some of the
research works includes its implementation [6]–[8],
performance [5] and pricing strategy [9]. Going by the
numerous advantages brought out by MCC which includes
multi-tenancy, ease of integration, scalability and dynamic
provisioning of services in an on-demand manner, MCC also
incurs some challenges. Some of the challenges includes
intermittent dis-connectivity and long WAN latency
introduced [3]. Other challenges include low processing
capabilities, low memory, small displays and battery capacity
etc. [10]. It is with the view to mitigate for these challenges
that gave birth to an infrastructure-less mobile platform
named ad hoc mobile cloud [2].
 The ad-hoc mobile market is the environment that allows
the interconnectivity of mobile devices within community of
users that are ready to communicate and share services
between providers and consumers. The system allows devices
at will to join, leave or being available for providing services
in general to other mobile peers [11], [12]. The MCC
assumed mobile devices to be online always which is
practically untrue thus the offline solution for mobile devices
is the Ad hoc mobile cloud computing [3].
However, for mobile devices within a community to
harmonize together for resource sharing, there is need for
reliable process of service selection among the discovered
mobile peers within the community [2]. One area in the ad
hoc mobile environment where service selection has not been
fully explored is the ad hoc mobile virtualized environment.
Let us say we have a client who is in need of weather service
with a response time of say 10 seconds. This service may be
available from provider C when the discovery mechanism
searches. But when this service is to be selected it is no longer
available because of the dynamic nature of the ad hoc system
where members leave and join anytime. This then becomes a
great challenge. Our work explores the feasibility of
enhancing service selection based on QoS parameters in the
context of service-virtualized ad hoc mobile cloud
environment.
We propose a virtual storage infrastructure-less mechanism
that allows a temporary storage of instances or images
information pertaining to members of the ad hoc community
for a short period of time before been destroyed. When
selection is to take place then this images are then searched.
To get the single service, we first normalize the parameters
and calculate the aggregate scores based on user’s preferences.
The remainder of this paper is organized as follows:
Section II provides related works and Section III presents the
operating modes of ad hoc mobile cloud. In IV we discuss the
motivating scenario. QoS-aware service selection in ad hoc
mobile cloud was discussed in section V. Section VI provides
the result and discussions while section VII contains
conclusion and future work.
II. RELATED WORKS
The advancement in the mobile technology in the recent
years has influenced a notable progress in the mobile world
bringing about its limitless usefulness to human life. For
example a mobile salesman can use his mobile browser
application to obtain customer, latest product availability and
pricing information by issuing a simple service request while
on the move. It is however assumed that wireless connectivity
is always available to ensure such rich experience with the
use of mobile devices but many works in the literature are
addressing this assumption to proffer solution to mobile
offline mode.
The evolution of the ad hoc mobile cloud is one of the
research outputs addressing this issue where mobile devices
within a group/community utilizes and share resources among
themselves [2], [3]. Therefore, if mobile devices are to
function effectively within groups as depicted by ad hoc
mobile cloud, there is need for each mobile device to be able
to connect to one another [11], discover [13], and select peers
who could render such issued request [14]. The work of [11]
emphasises that mobile devices would only work effectively
and efficiently in ad hoc mobile cloud if and only if its
discovery and selection system were made spontaneous with
the interacting network. In [15] the authors present a
middleware that allows access from mobile devices to handle
a multimedia services exposed from other mobile devices.
But this work was silent about which device to select from
among all available ones to provide such multimedia service.
The Hyrax project [16] proposed an approach of using the
Hadoop framework on mobile devices to share data and
computation. It implements core functionality needed for an-
hoc mobile clouds, including global data access, distributed
data processing, scalability. However, the vast work was
designed for deployment on many servers leaving us behind
with the challenges of deploying it to run on mobile devices.
The work was also silent about what portion of computation
would be carried out by what device and how such a device
was selected.
The work of Huerta-canepa and Lee [11] present
guidelines for a framework to create a virtual mobile cloud
computing providers. It imitates local cloud providers using
neighbouring devices. The framework creates an on the fly
connection to stable discovered nodes in its vicinity which
avoids infrastructure-based facilities, thereby creating local
cloud of mobile devices in an ad hoc manner. In the authors
report, a better performance would have occurred if a good
service selection mechanism is in place.
On the issue of mobile web service selection, Wagh and
Thool [17] prove that Android mobile devices can host web
services, given us an insight into selection of web services
from mobile devices. He argued that since the role of mobile
devices have been extended to carrying out data transmission
through the recent mobile technological advancement; the
devices have been able to perform the dual roles of service
clients and providers. However, the service provisioning
capability of the mobile devices can only be meaningful in
the context of some selected scenarios such as m-Health, m-
Learning and m-Commerce etc. The research also shows that
REST web services have a shorter response time compares to
SOAP based web services, thereby showing the feasibility of
mobile service host.
Other works [18], [19] show that mobile devices can host
web services. This however gives us a good foundation for
the implementation of ad hoc mobile cloud service
provisioning. The authors in [20] identify three major
approaches of service selection based on the use of QoS
parameters. These are:
a. Functional based service selection approach
b. Non-Functional based service selection approach
c. User based service selection approach
Functional based approach utilizes functional descriptions
of a web service to match the virtual (requested) service to the
real services available in the mobile devices through web
service interfaces. The semantic search result generates more
than one service provider, offering similar services. The non-
functional based approach uses quality of services and the
semantic context for service selection [21], [22]. User based
approach implement selection through ranking among
numerous discovered services with the aid of feedbacks, trust
or reputation [23].
Various scholars have implemented all the approaches for
service selection in different paradigms. For example [21],
[22], [24] use only non-functional criteria for selection in
cloud computing field while [23] uses the customers’
reputations for selection of service in service oriented
architectures and [25] combines semantics and preferences
for service selection in cloud computing. However, most of
this approaches needed to be tested for it suitability in an ad
hoc mobile environment. Our work therefore contributes to ad
hoc mobile service selection through:
 The use of virtual service image within the server
mobile PC to reduce web service latency
especially when the service provider node is
farther away from the service consumer.
 Provisioning of extensible quality of service in
service selection within the context of ad hoc
mobile cloud;
III. OPERATING MODES OF AD HOC MOBILE CLOUD
Ad hoc mobile cloud evolves to mitigate the challenge of
intermittent disconnection to the cloud server. The two basic
modes [3] of ad hoc mobile clouds are:
a. Distributed processing mode and
b. Single Service request mode.
The distributed processing mode is a paradigm that allows
tasks from a mobile device to be processed in a shared and
cooperative fashion on all other mobile devices. The
feasibility of this architecture has been demonstrated in [11].
The architecture arranges close-by mobile devices that are
pursuing similar task into a virtual cloud to distribute
computational task. While the single service request mode
requires handling a specific task through one of the available
devices which acts as a server. Web services have non-
functional attributes like response time, throughput,
availability, reliability [26] through which clients’ interests
could be specified. This work is aligning with this second
mode of service operation in which a PC within the ad hoc
cloud acts as a server [3]. A reliable mobile laptop PC would
be within the community to act as a server. This mode is
modified further in our work to suit our idea.
IV. A MOTIVATING SCENARIO
This research is motivated based on the GUISET project.
The aim of the project is to provide cheaper services to
SMMEs in the form of m-learning, m-health etc. Our scenario
is based on m-Learning existing within a school or among
group of schools closer to one another within a community.
Our idea is that mobile devices in these schools could easily
form ad-hoc mobile cloud because the schools often consume
similar educational resources. We consider Martin as a
student from a particular high school. He uses some service in
the cloud to download some study materials and invoke other
educationally related services. He experiences intermittent
connections due to an unreliable internet connection. Every
time he initiates the service he loses connectivity during the
process and this repeated action has proven to be expensive.
He therefore connects to the local wireless access that is
shared by a laptop within the school premises. His device is
first authenticated and on confirmation, he was able to join
the virtualized ad hoc network. Martin then specifies the
service name and service QoS parameters he needed via a
GUI. The system first normalizes the QoS attributes and then
calculates the aggregate score which is being compared with
the services available in the cloud. The optimal service is
returned to Martin through the server PC based on Martin
preferences.
V. QOS-AWARE SERVICE SELECTION IN AD HOC MOBILE
CLOUD
A. QoS-Aware Selection Process
The general concept is to explain what constitutes an
optimal service selection based on QoS parameters [27].
Service selection is built by mapping relations to virtual
operations, which are invariably used in locating the actual
web service. The aggregation of the QoS for service selection
is based on the QoS of the individual services parameter. To
form QoS parameters in just a single formula, we use a
simple scaling or normalization technique. Our scaling helps
us to normalize distribution. This is done by considering the
parameter values and observe what happen to that service
when those criteria values increases or decreases. For
example, when the response time of a service decreases, then
it is at the benefit of the consumer. Therefore, we represent
this as 0 and an increase in reliability is at the advantage or
the consumer. Thus, this is represented as 1. Therefore, we
represent
𝑆𝑠 = set of ad-hoc mobile services of the same
functionality.
Let k1, k2, k3, and k4 represent the QoS metrics. That is
service cost, availability, response time and reliability.
This is represented in row by column form with 𝑆𝑠 as the
row and k1…..kn (Criteria) as the column.
𝑃 = 𝑃𝑖𝑗 ; 1 ≤ 𝐼 ≤ 𝑛 ; 1 ≤ 𝑗 ≤ 𝑚;
our quality vector is denoted as
𝑄𝑉 𝑣 = 0,1,0,1
in the order of cost, availability, response time, and reliability.
To normalize every QoS attribute we consider the formulae in
[28]–[30]. We use that of [28] because of its simplicity. We
applied equation 1 when 𝑄𝑣 = 0 and equation 2 when 𝑄𝑣 = 1
thus:
𝑄𝑗𝑚𝑎𝑥 −𝑄𝑖,𝑗
𝑖𝑓 𝑄𝑗𝑚𝑎𝑥 −𝑄𝑗𝑚𝑖𝑛 ≠0
𝑉𝑖𝑗 = 𝑄𝑗𝑚𝑎𝑥 −𝑄𝑗𝑚𝑖𝑛 (1)
𝑖𝑓 𝑄𝑗𝑚𝑎𝑥 −𝑄𝑗𝑚𝑖𝑛 =0
1
𝑄𝑖,𝑗 −𝑄𝑗𝑚𝑖𝑛
𝑖𝑓 𝑄𝑗𝑚𝑎𝑥 −𝑄𝑗𝑚𝑖𝑛 ≠0
𝑉𝑖𝑗 = 𝑄𝑗𝑚𝑎𝑥 −𝑄𝑗𝑚𝑖𝑛 (2)
𝑖𝑓 𝑄𝑗𝑚𝑎𝑥 −𝑄𝑗𝑚𝑖𝑛 =0
1
Where 𝑄𝑗𝑚𝑎𝑥 = 𝑚𝑎𝑥 𝑄𝑖,𝑗 , 1≤ 𝑖 ≤ 𝑛
𝑄𝑗𝑚𝑖𝑛 = 𝑚𝑖𝑛 𝑄𝑖,𝑗 , 1 ≤ 𝑖 ≤ 𝑛
and n is the number of web services
From this we have our algorithm 1 that generates a new
matrix thus
𝑉 = 𝑉𝑖,𝑗 ; 1 ≤ 𝑖 ≤ 𝑛, 1 ≤ 𝑗 ≤ 𝑚;
where 𝑄𝑣 = 0,1,0,1 .
The normalization algorithm for the quality of service is
shown in Table I. Based on the aggregate score generated;
selection is carried out using dynamic programing algorithm
which matches the request with the available solutions. The
selection is carried out using algorithm shown in Table II.
B. Mechanism description
Our proposed model is shown in Figure 1. Various mobile
devices are included in the system ranging from personal data
assistance (PDAs), smartphones and laptops. Each of these
devices within the mobile cloud performs functions higher
than one another due to their varying functionalities within
the system (e.g the function of a PC is higher than a tablet).
The various physical phone images (mobile devices) have a
replica version residing in a mobile PC alongside with the
service link within the ad hoc community system as shown in
Fig. 1.
The mechanism in Figure 1 performs the following sets of
steps towards service selection in an ad hoc mobile cloud:
a. Implements a protocol through which mobile
devices would be able to communicate and
indicates its status after connection.
b. Establishes authentication for the intending
mobile device willing to join the ad hoc cloud.
c. On authentication, allows devices to publish
their resources (e.g audio, video and web
service) with their QoS attributes to the network
which remain in the virtual phone farm.
d. QoS metrics DB stores the quality of service
parameters submitted by service providers.
e. However, in this work, we assumed that each
mobile node that uploads a service gives correct
values of service QoS parameters being
uploaded.
f. The service monitors help to categorize and
monitor various web services and data.
g. QoS aggregate of we services are used to select
services based on clients prescription.
h. History manager keeps track of the activities of
each mobile phone which could be referenced.
i. In a case where the device having the requested
service is not online, the history manager could
provide a means of retrieval.
C. Mechanism Components
The crux of the work occurs in the selection middleware
engine component. It also consists of virtual phone farm
which is accessible on the network but it is resident on server
PC. This virtualization system is introduced to reduce the
challenge that could accumulate due to latency in checking
various mobile devices for web services within the network.
Figure 1: Ad Hoc Mobile Cloud Selection Mechanism
i.)Physical Phone Image: This represents the various
mobile devices participating within the mobile cloud
community. They are the various smartphones which
could request for services and at the same time provide
services within the local cloud.
ii.)Virtual Phone Farm: This resides within the mobile
PC in the ad hoc cloud community. It houses a virtual
copy of uploaded web services within it, keeping track of
the content of each virtual copy through the service
monitors.
iii.) Selection Middleware Engine (sme): This is the
main component of the selection mechanism. The
handler mechanism resides within the engine to offer
QoS parameters collection offered by service providers.
The middleware keeps track of the various QoS attributes
of the web services and as well as the user history of the
various mobile clients within the network. The QoS
attributes are stored in the QoS metric DB for fast
retrieval during service selection.
TABLE I
WEIGHT NORMALIZATION ALGORITHM
𝑃 = 𝑃𝑖𝑗 ; 1 ≤ 𝐼 ≤ 𝑛 ; 1 ≤ 𝑗 ≤ 𝑚;
𝑓𝑜𝑟 𝑖 = 0; 𝑖 < 𝑚; 𝑖 + + 𝑑𝑜
𝑓𝑜𝑟 𝑗 = 0; 𝑗 < 𝑛; 𝑗 + + 𝑑𝑜
𝑖𝑓 𝑞𝑓 j 𝑒𝑞 0
𝑖𝑓 𝑑𝑖𝑓𝑓𝑞𝑜𝑠 j ! = 0
v i j ← 𝑞𝑚𝑎𝑥 j − 𝑝 i j / iffqos j ;
else
v i j ← 1;
end
else if 𝑞𝑓 j 𝑒𝑞 1
if 𝑑𝑖𝑓𝑓𝑞𝑜𝑠 j ! = 0
v i j ← 𝑝 i j − 𝑞𝑚𝑖𝑛 j / iffqos j ;
else
vi j ←1
endif
endif
 𝑉 = 𝑉𝑖,𝑗 ; 1 ≤ 𝑖 ≤ 𝑛, 1 ≤ 𝑗 ≤ 𝑚; Serial Parameters
Number Number of Web Services Execution duration (ms)
TABLE II
QOS SELECTION ALGORITHM 9 27 51
10 30 52
Set parameters (c,rt,r,a)
Input consumers required QoS
Call normalization
Set weights
Calculate score
if Ss(Score) available online, then
Ss(Score) ← service node
Print Ss(Score)
else
if Ss(Score) available in virtual farm, then
Ss(Score) ← service link
print Ss(score)
else
Print(nil)
end if end if Figure 2: User Login Page.
end.

VI. RESULT AND DISCUSSION

We use Java Netbeans IDE to simulate the service
selection mechanism we explained. Using Java 2 SDK v 1.4.1,
we install Apache Tomcat v 4.1.24. Also, using Jersey
environment framework setup, we developed web services
that were used to test our selection mechanism. We used four
mobile devices along with our server mobile machine. Figure
2 shows the log in interface of the ad hoc mobile cloud
system. The interface provides avenue for as many devices to
participate in resource sharing through the wireless
connection provided by the server mobile machine. All the
five mobile devices that participate in the mobile cloud are
shown in Figure 3. Each of the devices in the cloud possesses Figure 3: Interface showing connected devices.
services with similar functionality but different non-
functional parameters. Devices uploaded a copy of the service
to the server mobile machine with the uploaded service
carrying the IP address of the device that uploaded it as
shown in Figure 4. Due to the cost of increasing the mobile
devices in this experiment, we rather took the option of
increasing the number of web services within the mobile
cloud to test the performance of our mechanism. We started
with three web services in each of the devices and gradually
increase it to six, nine, and so on. The execution duration for
service invocation with each of the experiment conducted is
shown in Table III. Each of the mobile devices connected is
taken in succession for service invocation. In each case, four
devices were used and the average value of execution
duration is calculated and recorded as shown in the Table III
and Figure 5. Figure 4: Service specification and Query Interface.

TABLE III The execution duration increases with increasing number

EXECUTION DURATION OF SERVICE INVOCATION of web services but when it got to around twenty one, the
Serial Parameters
increment was not well noticeable compared to the initial one.
Number The incorporation of the virtual system accounted for this
Number of Web Services Execution duration (ms)
1 3 19
result. This has enhanced fast access to services with
2 6 23 increasing use of uploaded services based on the user history
3 9 27 within the server mobile machine.
4 12 30
5 15 35 VII. CONCLUSION AND FUTURE WORK
6 18 39
7 21 50 This paper presented a QoS-based simple service selection
8 24 51 mechanism in an ad hoc mobile cloud. The mechanism aims
at enabling service selection based on the influence of quality
of service.
60
Execution duration (ms)
50
40
30
20 execution
duration
10
0 Serv. Eng., pp. 252–256, Apr. 2014.
3 6 9 12 15 18 21 24 27 30
No of Web Services
Figure 5: Graph of execution duration against web services.
We performed an experiment to validate our proof of
concept using the service execution time as our evaluation
tool. A SQLite database was used within the server mobile
machine. The connecting mobile devices use wireless
connection setup by the server mobile machine to create a
local ad hoc mobile cloud scenario that allows mobile service
selection. An m-Learning scenario was used to illustrate the
applicability and usability of the proposed mechanism. The
users ultimately contribute to the system through acting as a
consumer (client) as well as a service provider. Consumers
specify the QoS of service request for the mechanism to
select the appropriate service. We recorded appreciable better
result due to the incorporation of the virtual server system.
The future work would like to incorporate an auto–switching
mechanism during mobile node interruptions (e.g. low battery)
that could automatically assign the server PC task to other
capable devices within the mobile cloud.
REFERENCES
[1] H. T. Dinh, C. Lee, D. Niyato, and P. Wang, ―A survey of mobile
cloud computing : architecture , applications , and approaches,‖ Wirel.
Commun. Mob. Comput., vol. 13, no. 18, pp. 1587–1611, 2013.
[2] D. Kovachev, Y. Cao, and R. Klamma, ―Mobile Cloud Computing: A
Comparison of Application Models,‖ in Networking and Internet
Architecture: Distributed , paralleland Cluster Computing, 2011.
[3] L. Fangming, S. Peng, J. Hai, D. Linjie, and Y. Jie, ―Gearing Resource
-Poor Mobile Devices with Powerful Clouds: Architectures,
Challenges, and Applications,‖ in IEEE Wireless Communications,
2013, no. June, pp. 2–10.
[4] A. U. R. Khan, M. Othman, S. A. Madani, and S. U. Khan, ―A survey
of mobile cloud computing application models,‖ IEEE Commun. Surv.
Tutorials, vol. 16, no. 1, pp. 393–413, 2014.
[5] A. O. Akingbesote, M. O. Adigun, S. Xulu, and E. Jembere,
―Performance Modeling of Proposed GUISET Middleware for Mobile
Healthcare Services in E-Marketplaces,‖ J. Appl. Math., vol. 2014, no.
4, p. 5, 2014.
[6] A. O. Akingbesote, M. O. Adigun, J. B. Oladosu, and E. Jembere, ―A
Quality of Service Aware Multi-Level Strategy for Selection of
Optimal Web Service,‖ IEEE Int. Conf. Adapt. Sci. Technol. ICAST,
2013.
[7] A. O. Akingbesote, M. O. Adigun, M. A. Othman, and I. R. Ajayi,
―Determination of Optimal Service level in Cloud E-Marketplaces
Based on Service Offering Delay,‖ 2014 IEEE Int. Conf. Comput.
Commun. Control Technol. (I4CT 2014) Langkwawi, Kedah, Malaysia,
no. 14ct, pp. 283–288, 2014.
[8] O. O. Ekabua and M. O. Adigun, ―GUISET LogOn : Design and
Implementation of GUISET-driven Authorization,‖ Cloud Comput.
2010 First Int. Conf. Cloud Comput. GRIDs Virtualization, no. c, pp.
1–6, 2010.
[9] M. E. Buthelezi, M. O. Adigun, O. O. Ekabua, and J. S. Iyilade,
―Accounting, pricing and charging service models for a GUISET grid-
based service provisioning environment,‖ Proc. 2008 Int. Conf. e-
Learning, e-Business, Enterp. Inf. Syst. e-Government, EEE 2008, pp.
350–355, 2008.
[10] E. Cuervo, A. Balasubramanian, and D. Cho, ―MAUI : Making
Smartphones Last Longer with Code Offload,‖ in 8th international
conference on Mobile systems, applications, and services ACM, 2010,
pp. 49–62.
[11] G. Huerta-canepa and D. Lee, ―A Virtual Cloud Computing Provider
for Mobile Devices,‖ in ACM Workshop on Mobile Cloud Computing
& Services: Social Networks and Beyond, 2010, pp. 35–39.
[12] D. Dev and K. L. Baishnab, ―A Review and Research Towards Mobile
Cloud Computing,‖ 2014 2nd IEEE Int. Conf. Mob. Cloud Comput.
[13] D. Kovachev, Y. Cao, and R. Klamma, ―Mobile Cloud Computing : A
Comparison of Application Models,‖ in Distributed, Parallel, and
Cluster Computing (cs.DC);, 2012.
[14] N. Kaushik and J. Kumar, ―A Literature Survey on Mobile Cloud
Computing : Open Issues and Future Directions,‖ Int. J. Eng. Comput.
Sci., vol. 3, no. 5, pp. 6165–6172, 2014.
[15] Y. Cao, M. Jarke, R. Klamma, O. Mendoza, and S. Srirama, ―Mobile
access to MPEG-7 based multimedia services,‖ Proc. - IEEE Int. Conf.
Mob. Data Manag., pp. 102–111, 2009.
[16] E. E. Marinelli, ―Hyrax : Cloud Computing on Mobile Devices using
MapReduce,‖ Master Thesis, Carnegie Mellon Univ., vol. 0389, no.
September, 2009.
[17] K. S. Wagh and R. . Thool, ―Web Service Provisioning on Android
Mobile Host,‖ Int. J. Comput. Appl., vol. 81, no. 14, pp. 1–7, 2013.
[18] M. Jansen, ―About an Architecture That Allows to Become a Mobile
Web Service Provider,‖ ICIW 2012 Seventh Int. Conf. Internet Web
Appl. Serv., no. c, pp. 90–96, 2012.
[19] F. AlShahwan, F. Carrez, and K. Moessner, ―Providing and evaluating
the mobile Web service distribution mechanisms using fuzzy logic,‖ J.
Softw., vol. 7, no. 7, pp. 1473–1487, 2012.
[20] M. Sathya, M. Swarnamugi, P. Dhavachelvan, and G. Sureshkumar,
―Evaluation of QoS Based Web- Service Selection Techniques for
Service Composition,‖ Int. J. Softw. Eng., vol. 1, no. 5, pp. 73–90,
2012.
[21] S. Liu, S. Guo, X. Chen, and M. Lan, ―A QoS Based Web Service
Selection Model,‖ 2009 Int. Forum Inf. Technol. Appl., vol. 3, p. 145,
2009.
[22] M. Swarnamugi and M. Sathya, ―Specification Criteria for Web
Service,‖ Int. J. Comput. Eng. Inf., no. May, p. 2034, 2010.
[23] S. Galizia, A. Gugliotta, and J. Domingue, ―A Trust Based
Methodology for Web Service Selection,‖ in IEEE, International
Conference on Semantic Computing (ICSC 2007), 2007, pp. 193–200.
[24] N. Keskes, ―Context of qos In Web Service Selection,‖ Am. J. Eng.
Res., vol. 2, no. 4, pp. 120–126, 2013.
[25] R. Iordache and F. Moldoveanu, ―QoS-Aware Web Service Semantic
Selection Based on Preferences,‖ Procedia Eng. 24th DAAAM Int.
Symp. Intell. Manuf. Autom., vol. 69, pp. 1152–1161, 2014.
[26] L. Arockiam and N. Sasikaladevi, ―Simulated Annealing Versus
Genetic Based Service Selection Algorithms,‖ Int. J. u- e- Serv. Sci.
Technol., vol. 5, no. 1, pp. 35–50, 2012.
[27] X. Yang and A. Bouguettaya, ―Efficient Access to Wireless Web
Services,‖ in Proceedings of the 7th International Conference on
Mobile Data Management (MDM’06), 2006.
[28] L. Zeng, B. Benatallah, A. H. H. Ngu, M. Dumas, J. Kalagnanam, and
H. Chang, ―QoS-Aware Middleware for Web Services Composition,‖
IEEE Trans. Softw. Eng., vol. 30, no. 5, pp. 311–327, 2004.
[29] M. Makhlughian, S. M. Hashemi, Y. Rastegari, and E. Pejman, ―Web
Service Selection Based on Ranking of QoS using Associative
Classification,‖ Int. J. Web Serv. Comput., vol. 3, no. 1, pp. 1–14,
2012.
[30] Y. Liu, A. H. H. Ngu, and L. Zeng, ―QoS Computation and Policing in
Dynamic Web Service Selection,‖ in proceedings of the 13th
International WWW Conference, ACM, 2004, pp. 66–73.
First Principle Leakage Current Reduction Technique
for CMOS Devices
Hippolyte Djonon Tsague
Council for Scientific and Industrial Research (CSIR)
Modelling and Digital Science (MDS)
Pretoria, South Africa
hdjonontsague@csir.co.za
Abstract—This paper presents a comprehensive study of
leakage reduction techniques applicable to CMOS based devices.
In the process, mathematical equations that model the power-
performance trade-offs in CMOS logic circuits are presented.
From those equations, suitable techniques for leakage reduction
as pertaining to CMOS devices are deduced. Throughout this
research it became evident that designing CMOS devices with
high-κ dielectrics is a viable method for reducing leakages in
cryptographic devices. To support our claim, a 22nm NMOS
device was built and simulated in Athena software from Silvaco.
The electrical characteristics of the fabricated device were
extracted using the Atlas component of the simulator. From this
research, it became evident that high-κ dielectric metal gate are
capable of providing a reliable resistance to DPA and other form
of attacks on cryptographic platforms such as smart card.The
fabricated device showed a marked improvement on the 𝑰𝒐𝒏 /𝑰𝒐𝒇𝒇
ratio, where the higher ratio means that the device is suitable for
low power applications. Physical models used for simulation
included 𝑺𝒊𝟑 𝑵𝟒 and 𝑯𝒇𝑶𝟐 as gate dielectric with TiSix as metal
gate. From the simulation result, it was shown that 𝑯𝒇𝑶𝟐 was the
best dielectric material when TiSix is used as the metal gate.
Keywords—Differential power analysis; High-K dielectric gate;
Smart card
I. INTRODUCTION
S mart Cards are a safe place to store valuable information
such as cryptographic private keys making them an
adequate medium capable of providing secure
authentication and storage of secret data. Recent development
in this field of study have redirected attention to a whole new
class of attacks developed to retrieve the secret key stored in
the card through the study and interpretation of the
information leaked by processor power dissipation. Such
attacks are known as Side Channel Attacks (SCA’s) and are
performed by taking advantage of the interdependence in
Complementary Metal Oxide circuits between dynamic power
dissipation on the processed data. Chief among those
techniques is Differential Power Analysis (DPA) which has
been widely shown to be a threat to the security of
cryptographic devices [1]. For that reason a large number of
side channel related countermeasures at various levels of
abstraction have been proposed. In particular, at the transistor
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Bhekisipho Twala
Faculty of Engineering University of Johannesburg (UJ)
Department of Electrical and Electronic Engineering
Science
Johannesburg, South Africa
level, countless alternatives to first order DPA attacks have
been developed and demonstrated. Researchers in [2] have
successfully demonstrated that in sub-100 nm and related
technologies, power leakages are as high as the dynamic
power of the device and hence the leakage (static) supply
current can be used as a new Side-Channel.
Off State leakages have been recognized by semiconductor
manufacturers as a major bottleneck for future microcontroller
integration. Off-state leakage is a static power current that
leaks through transistors even when such devices are turned
off. It is one of two principal sources of power dissipation in
today’s microcontrollers. The other source of leakage is of
course dynamic power, which is caused by the repeated
capacitance charge and discharge on the output of the hundreds
of millions of gates in today’s microcontrollers. Until recently,
only dynamic power was identified as a significant source of
power consumption, and Moore’s law has helped to control it
through shrinking processor technology. Dynamic power is
proportional to the square of the supply voltage; therefore
reducing the voltage significantly reduces the device’s power
consumption. Unfortunately smaller geometries aggravate
current leakages problems; static power begins to dominate the
power consumption equation in microcontroller design” [12].
CMOS technology was invented at Fairchild
Semiconductor in 1963 by Wanlanss Frank. The original idea
behind the technology was to design a low power alternative to
Transistor-transistor Logic (TTL). The early adopters of the
technology were watch designers who realized the importance
of low battery power consumption over the device processing
capabilities for electronic circuits. Nowadays, the CMOS
technology has grown substantially to become the dominant
technology integrated circuit design. “This is essentially
because area occupation, operating speed, energy efficiency
and manufacturing costs have benefited and continue to benefit
from the geometric downsizing that comes with every new
generation of semiconductor manufacturing processes. In
addition, the simplicity and comparatively low power
dissipation of CMOS circuits have allowed for integration
densities not possible in similar techniques such as bipolar
junction transistors (BJT)” [17]. Despite its many benefits,
authors of [12] argue that the power that such devices consume
has increased dramatically with increases in device speed and
chip density; and so has the number of attacks on cryptographic
devices.
II. POWER FUNDAMENTALS
Five mathematical equations govern the power performance
in the CMOS logic circuits according to [12]. In this paper, we
present them in a way that addresses the basics of physics and
logic circuitry design. The first mathematical equations related
to CMOS power fundamentals are the basics of low power
consumption [3] while the last two equations are more
concerned with sub-threshold and gate-oxide leakage
modeling in CMOS technologies.
A. Investigation of Frequency and Voltage Relationships
Equation (1) below depicts the supply voltage dependency
of the operating frequency of the device as computed in [12] :
𝑓 ∝ 𝑉 − 𝑉𝑡ℎ 𝛼 𝑉
(1)
In this equation, 𝑉 represents the transistor’s supply voltage
while 𝑉𝑡ℎ is the device’s voltage. The exponent α is an
experimentally derived constant with a value of 1.3
approximately [12]. Dynamic voltage scaling in CMOS
devices is used to control switching power dissipation in
battery operated systems. Also, power consumption
minimization techniques rely on low voltage modes and
lowered clock frequencies. In [12] authors have used the
relation derived in (1) to compute an equation that depicts the
relationship between frequency and supply voltage. The
derivation begins with the selection of the device’s working
voltage and frequency defined as 𝑉𝑛𝑜𝑟𝑚 and 𝑓𝑛𝑜𝑟𝑚
respectively. The quantities selected are normalized entities
depicting the relationship between the largest possible
device’s operating voltage 𝑉𝑚𝑎𝑥 and frequency 𝑓𝑚𝑎𝑥 . This
relationship is shown in (2) below:
𝑉 𝑡ℎ 𝑉 𝑡ℎ
𝑉𝑛𝑜𝑟𝑚 = 𝛽1 + 𝛽2 ∙ 𝑓𝑛𝑜𝑟𝑚 = + 1− ∙ 𝑓𝑛𝑜𝑟𝑚 (2)
𝑉𝑚𝑎𝑥 𝑉𝑚𝑎𝑥
From (1) it is evident that if f= 0, then (2) becomes:
𝑉𝑡ℎ
𝑉𝑛𝑜𝑟𝑚 = 𝛽1 =
𝑉𝑚𝑎𝑥
The value of 𝑉𝑛𝑜𝑟𝑚 can safely be approximated to 0.37. That
approximation closely matches present day’s industrial data
[12]. It is also worth mentioning at this stage that 𝑓𝑚𝑎𝑥 is
proportional to 𝑉𝑚𝑎𝑥 and that the frequency will drop to zero if
V is equal to 𝑉𝑡ℎ , as clearly shown (1).
B. CMOS Power Dissipation
Many origins of power consumption in CMOS devices exist
in CMOS devices and that what the third equation attempts to
clarify. The CMOS power consumption is the sum of dynamic
and static power as shown in (3):
𝑃 = 𝑃𝑑𝑦𝑛𝑎𝑚𝑖𝑐 + 𝑃𝑠𝑡𝑎𝑡𝑖𝑐 (3)
The first term of (3) can be broken off into two distinct
entities namely 𝑃𝑠ℎ𝑜𝑟𝑡 and 𝑃𝑠𝑤𝑖𝑡𝑐 ℎ . The first component 𝑃𝑠ℎ𝑜𝑟𝑡
is the power dissipated during gate voltage transient time
while the second component 𝑃𝑠𝑤𝑖𝑡𝑐 ℎ comes as a result of the
many charging and discharging of capacitances in the device.
The last term 𝑃𝑠𝑡𝑎𝑡𝑖𝑐 represents the power generated when the
transistor is not in the process of switching. Equation (3) can
be rewritten as:
𝑃 = (𝑃𝑠ℎ𝑜𝑟𝑡 + 𝑃𝑠𝑤𝑖𝑡𝑐 ℎ ) + 𝑃𝑠𝑡𝑎𝑡𝑖𝑐 = 𝐴𝐶𝑉 2 𝑓 + 𝑉𝐼𝑙𝑒𝑎𝑘 (4)
In (4) A denotes the number of bits that are actively
switching and C is the combination of the device’s load and
internal capacitance. It is worth mentioning at this stage that
for ease of simplicity, the power lost to spasmodic short circuit
at the gate’s output has been neglected.
From (4) it is evident that dropping the supply voltage leads
to an important decrease in the device power consumption.
Mathematically speaking, dividing the supply voltage by 2 or
halving it will reduce the power consumption by four. The
main drawback of that proposition is that it will reduce the
processor’s top operating frequency by more than half. A
better approach suggested in [12] relies on the use parallel or
pipelined techniques to compensate for the performance losses
due to supply voltage reduction.
C. Computing Leakage Current
Parallelism and pipelining techniques for power reduction
were first proposed by [19]. Since then researchers have
conducted studies aimed at optimizing the pipelining depth for
dissipated power reduction in CMOS devices. Furthermore,
researches have been conducted at a functional block level to
compare the performances of pipelining and parallelism to
find out which technique performs best when it comes to
minimizing total switching power. In (3) it was shown that
leakage current (source of static power consumption) is a
combination of subthreshold and gate-oxide leakage i.e.:
𝐼𝑙𝑒𝑎𝑘 = 𝐼𝑠𝑢𝑏 + 𝐼𝑜𝑥
Deriving Subthreshold Power Leakage: Authors of [19] present
an equation representing the direct relationship between a
CMOS device threshold voltage, its subthreshold leakage
current and the device supply voltage as follows:
 𝑛𝑉0 1− 𝑒 −𝑉 𝑉 𝜃
𝐼𝑠𝑢𝑏 = 𝐾1 𝑊𝑒 −𝑉𝑡ℎ (6)
In (6) K1 and 𝑛 are normally derived experimentally, W
represents the device’s gate width, and Vθ is its thermal
voltage. The quantity Vθ can safely be approximated to 25 mV
at room temperature (20 degrees Celsius). If Isub rises enough
to generate, Vθ will rise as well and in that process cause an
increase in Isub and this may result in thermal runaway. From
(6) it becomes clear that two ways exist for reducing Isub
which are (1) turning off the supply voltage and (2) stepping-
up the threshold voltage. In [12] it is argued that “since this
quantity shows up as a negative exponent, increasing that
value could have a dramatic effect in even small increments.
On the other hand, it is evident from (1) that increasing
Vth automatically creates a reduction in speed. The obvious
problem with the first approach is loss of state; as for the
second option, its major inconvenience relates to the loss of
performance”. The device’s gate width W, its gate length Lg,
the device’s oxide thickness Tox, and doping concentration
Npocket are other major contributors to subthreshold leakage in
CMOS based technologies. Processor designers often optimize
one or a few of those leakage components as a convenient
technique to reduce subthreshold leakage as will be seen in
subsequent paragraphs.
Subthreshold Power Leakage: Gate leakage mechanisms, such
as tunneling across thin gate oxide leading to gate oxide
leakage current become significant at the 90nm node and
smaller. Gate oxide leakage is not as well understood as
subthreshold leakage. For the purpose of this research, a
simplification of equations from the authors of [5] is sufficient
to illustrate the point:
2 3.75𝑒 13 atoms 𝑐𝑚−3 . In the third step, the deposited oxide
𝑉
𝐼𝑜𝑥 = 𝐾2 𝑊 𝑒 −𝛼𝑇𝑜𝑥 𝑉
(7)
𝑇𝑜𝑥
Where 𝐾2 and 𝛼 are derived experimentally. In (5) we draw
our attention to the oxide thickness, 𝑇𝑜𝑥 component of the
equation. “Increasing 𝑇𝑜𝑥 will reduce the gate leakage. However,
it also negatively affect the transistor’s efficiency since 𝑇𝑜𝑥 must
decrease proportionately with process scaling to avoid short
channel effects. Therefore, increasing 𝑇𝑜𝑥 is not a viable option”
[12]. A better approach to this problem may lie in the
development of high-κ dielectric gate insulators. This approach is
currently under heavy investigation by the research community.
III. REDUCING STATIC POWER CONSUMPTION
Many researchers as well research groups have developed
power models for reducing static power consumption for
embedded devices. Power gating [4, 16] is slowly becoming a
very popular design technique for decreasing leakage currents.
Although effective in reducing static power consumption in
many instances, its major drawback lies in its tendency to
introduce delays by adding extra circuitry and wires and also
uses extra area and power.
Another approach to static power reduction is based on the
utilization of multiple threshold voltage techniques. “Present
day’s processes typically offer two threshold voltages.
Microprocessor designers assign a low threshold voltage to
some of the few identified performance-critical transistors and
a high threshold voltage to the majority of less time critical
transistors. This approach has the tendency to incur a high
subthreshold leakage current for the performance-critical
transistors, but can significantly reduce the overall leakage”
[12].
Other techniques for reducing subthreshold leakage are
closely related to gate tunneling current, however, their effects
are still under investigations. Gate-oxide leakage has a
negligible dependence on temperature [12]. Therefore, as it
subsides with drops in temperature, gate-oxide related current
leakage become important.
IV. FABRICATION METHOD
For this paper, a 20nm NMOS MOFSET was fabricated
using the SILVACO Athena module and the device’s
electrical characteristic and performance were simulated and
evaluated using the Atlas module from SILVACO. The
specifications of the sample used in this experiment was p-
type (boron doped) silicon substrate with doping concentration
of 1.5𝑒 15 atoms 𝑐𝑚−3 and <100> orientation. The next step
consisted in developing the P-well by growing a 800 Å oxide
screen on top of bulk silicon. This technique makes use of dry
oxygen at a very high temperature (i.e. approximately 800°C)
followed by Boron as dopant with a concentration of
layer is etched and there after annealed to ensure that all boron
atoms are spread uniformly. This done at a temperature of
900°C using nitrogen followed by a futher rise in temperature
to 950°C using dry oxygen. The next step was to isolate the
neighbouring transistor by creating a shallow trench isolator
with a thicknesses of 130 Å. After that step, the wafer was
oxidized with dry oxygen for approximately 25 minutes at
temperature of 1000°C. Two important processes were
involved in the development of the STI namely, Low Pressure
Chemical Deposition (LPCVD) and reactive ion etching
(RIE). The LPCVD process starts with the deposition of a
1000 Å nitride layer on top of the STI oxide layer, followed by
a photo-resistor deposition on the wafer. The RIE process
consisted in etching the unnecessary part on the top of the STI
area. Both chemical and mechanical polishing was
implemented to strip away any extra oxide on the wafer. STI
was further annealed for approximately 15 minutes at a
temperature of 850°C. As a final STI step in the process, an
oxide layer was carefully deposited and etched to eliminate
possible defects that may have occurred on the surface.
It is important to mention at this stage that the deposition of
high-κ dielectric process with gate oxide thickness is selected
so that they have the same equivalent oxide thickness as 𝑆𝑖𝑂2 .
Furthermore, the length of the high-κ material was scaled so as
to get the equivalent 22nm gate length of the transistor. The
next step consisted of the deposition of Titanium Silicide
(TiSix) on top of the high-κ dielectrics (𝑆𝑖3 𝑁4 , 𝐻𝑓𝑂2 )
followed by halo implantation of indium dose to obtain the
optimum value of the NMOS device [6, 7]. The next step
involved the formation of the sidewall spacer that would serve
as source and drain electrodes for the device. In this case,
implantation with arsenic is followed by a dose of phosphor to
ensure an uninterrupted flow of current in the fabricated
NMOS device [8].
The next step consisted in the deposition of a layer (0.5 μm)
of Boron Phosphor Silicate Glass (BPSG) to act as the pre-
metal dielectric [9]. Once more, annealing was done at 950°C
on the wafer to strengthen the structure. The next step consisted
of compensation implantation with a layer of phosphorous. The
last step involved the deposition of aluminium layer which
served as metal electrode (source and drain). Once the model
structure design was completed we then proceeded with device
simulation using Atlas.
V. RESULTS AND DISCUSSIONS
The complete NMOS structure is shown in Fig. 1 below.
The fabrication process is the same for all high-κ devices
fabricated for this research except that the dielectric materials
were varied.
Fig. 1. Cross Section of the fabricated 22 nm NMOS device
Fig.2 on the other hand shows the doping profile of one of
the designed structure with gate length of 22nm NMOS.
Fig. 2. Doping profile of the fabircated device
Results of electrical characteristic simulation are obtained in
Fig. 3 below.The plots are also known as “𝑉𝑡 Curves”, because
devices designers use them extensively to extract the threshold
voltage (𝑉𝑡 ), which defines an approximation of when a
transistor is “on” and allows current to flow across the
channel.
For this research, the Figure presents the drain (𝐼𝑑 ) vs gate
voltage (𝑉𝐺𝑆 ) curves for 𝑆𝑖3 𝑁4 (k~29), 𝐻𝑓𝑂2 (k~21) and
aconventional device made of 𝑆𝑖𝑂2 . Typically, the fabricated
device’s drain voltage 𝑉𝐷𝑆 was fixed when (𝐼𝑑 ) vs (𝑉𝐺𝑆 ) was
plotted. The threshold voltage (𝑉𝑡ℎ ), state on current (𝐼𝑜𝑛 ) and
state off current (𝐼𝑜𝑓𝑓 ) can be extracted from the (𝐼𝑑 ) vs (𝑉𝐺𝑆 )
curve.
5.00E-05
4.00E-05
3.00E-05
Drain Current
(A/um)
2.00E-05
𝜀=3.9
1.00E-05 𝜀=21
𝜀=29
0.00E+00
0 1 2 3 4
-1.00E-05
Gate Voltage (V)
Fig. 3. Drain cureent Vs gate voltage curves (threshold curves)
A good doping concentration is one of the critical
factorsthat ensure that the transistor works well and emits
fewer leakage currentsso as to enhance gate control [10].
There are four factors that influence the threshold voltage;
these factors include (1) threshold voltage adjustment implant,
(2) halo implant, (3) channel implant and (4) compensation
implant. However for the purpose of this research, where the
main requirement was to investigate the leakage current
emissions of the device while varying the gate material used,
the threshold voltage adjustment implant technique was
utilised. To get a threshold voltage of 0.302651 as stipulated
by the ITRS the best doping concentration with boron was set
to 8.5763𝑥1013 𝑐𝑚−2 for 𝐻𝑓𝑂2 and
9.73654𝑥1013 𝑐𝑚−2 for𝑆𝑖3 𝑁4 . This doping concentration
variation is to account for the physically thicker and stronger
dielectric materials utilised. As expected, the drain current for
both 𝑆𝑖3 𝑁4 and 𝐻𝑓𝑂2 dielectric are decreased compared to the
drain current of the device made of 𝑆𝑖𝑂2 as can be seen from
Fig. 3 above.
Drain leakage current (𝐼𝑜𝑓𝑓 ) or sub-threshold leakage
current occurs when the gate voltage (𝑉𝐺𝑆 ) is lower than the
threshold voltage (𝑉𝑡ℎ ). In ideal case, when the transistor is
turned off, 𝑉𝐺𝑆 = 0 volt and 𝑉𝐷𝑆 = 𝑉𝐷𝐷 (voltage supply), there is
no current flow through the channel (𝐼𝑜𝑓𝑓 = 0). Again from Fig.
3 it is clear that the leakage current through 𝐻𝑓𝑂2 dielectric is
lowest compared to both 𝑆𝑖3 𝑁4 and 𝑆𝑖𝑂2 dielectrics. This
result suggests that 𝐻𝑓𝑂2 dielectric material is more
compatible with silicon and appears to be the most stable
oxide with the highest heat of formation. This observation
matches the findings of [12, 13].
Table 1 show the simulated results for 𝑆𝑖3 𝑁4 and for
𝐻𝑓𝑂2 dielectric materials with TiSix as metal gate for 22nm
NMOS. Clearly, the 𝐼𝑜𝑛 values from the simulation are larger
compared to prediction value.
TABLE I. SIMULATED RESULTS OF THE UTILISED DIELECTRICS
Parameter 𝑆𝑖3 𝑁4 𝐻𝑓𝑂2 ITRS 2011
prediction
𝑉𝑡ℎ (𝑉) 0.30226 0.30226 0.302
𝐼𝑜𝑛 3.03345−4 2.63345−4 1.03−7
𝐼𝑜𝑓𝑓 2.83345−15 1.03345−15 1.495−6
𝐼𝑜𝑛 1.07111 2.54111 6.6071−2
𝐼𝑜𝑓𝑓
From the Table it can be seen that the simulation results for
𝐼𝑜𝑓𝑓 are slightly lower than prediction value from the ITRS-
2011. This is an indication that the above selected high-κ
dielectric materials are suitable combination with metal gate
as well as compatible with silicon for building low leakage
transistors. Again as stated earlier 𝐻𝑓𝑂2 appears to be the
better option.
VI. CONCLUSION
NMOS structure with 22nm gate length were successfully
designed and simulated to study the electrical performance of
the device with a major emphasis on the device leakage current
emissions. Two high-κ dielectric materials namely 𝐻𝑓𝑂2 and
𝑆𝑖3 𝑁4 were investigated in this study. Throughout the study, it
became evident that materials with high-κ dielectric are
possible replacement to silicon oxide (𝑆𝑖𝑂2 ). Replacing 𝑆𝑖𝑂2
with materials such as 𝐻𝑓𝑂2 (Hafniumoxide) or
𝑆𝑖3 𝑁4 (Silicon nitride) could be some of the best mechanisms
to limit leakage current emissions in CMOS devices and help
in providing better resistance to attacks such as simple or
differential power analysis or even electromagnetic power
analysis that is gaining momentum in cryptographic devices
such as smart cards.
REFERENCES
[1] P.C Kocher, J. Jaffe, and B. Jun, “Differential power analysis,” in Proc.
Of CRYPTO ’99, pp. 388-397, 1999
[2] M. Alioto, L. Giancane, G. Scotti, A. Trifiletti, “Leakage PowerAnalysis
Attacks: a Novel Class of Attacks to NanometerCryptographic Circuits,”
IEEE Trans. on Circuits and Systems – partI, vol. 57, no. 2, pp. 355-367,
Feb. 2010
[3] H. Djonon Tsague and Bheki Twala, “Simulation and Parameter
Optimization of Polysilicon Gate Biaxial Strained Silicon MOSFETs,”
The Fifth International Conference on Digital Information Processing
and Communications (ICDIPC2015), Switzerland, Geneva 2015 in
press.
[4] Y. Shin, J. Seomun, K.M. Choi and T. Sakurai, “Power gating: circuits,
design methodologies, and best practice for standard-cell VLSI designs,”
ACM Trans. Des. Autom. Electron Syst. 15(4), 28:1–28:37 (2010).
[5] Integrated Circuit Characterization and Analysis Program (IC-CAP)
modeling software, available from http://www.home.agilent.com, [last
accessed: Sept. 2015]
[6] H.A Elgomati, B.Y Majlis, I. Ahmad, F. Salahuddin, F.A. Hamid, A.
Zaharim and P.R Apte, “Investigationof the Effect for 32nm PMOS
Transistor and Optimizing Using TaguchiMethod,” Asian Jounal of
Applied Science, 2011
[7] S.Yuan Chen et al., “Using simulation to Characterize Highperformance
65nm Node Planar,” International Symposium on NanoScience and
Technology, Taiwan, 20-21 Nov 2004.
[8] G. Darbandy, A. Jasmin, J. Seldmeir, U. Monga, I. Garduño, A.
Cerdeira, and B. Iñiguez, “Temperature Dependent Compact Modeling
for Gate Tunneling Leakage Current in Double Gate MOSFETs,” Solid-
State Electronics, 2013
[9] H. Wong and H. Iwai, “On the Scaling Issues and High-k Replacement
of Ultrathin Gate Dielectric for Nanoscale MOS Transistor,”
Microelectronic Engineering, pp.1867-1904, 2013.
[10] G. He and Z. Sun, “High-k Dielectrics for CMOS technologies,” John
Willey and son’s Inc. Aug. 2012
[11] M.G Priya, K.Baskaran and D. Krishnaveni, “Leakage Power Reduction
Techniques in Deep Submicron Technologies for VLSI Applications,”
International Conference on Communication Technology and System
Design. Vol. 30, pp 1163-1170, 2012
[12] N.S. Kim et al., “Leakage Current: Moore’s Law Meets Static Power,”
[13] S.Y. Chen et al., “Using simulation to Characterize High performance
65nm Node Planar,” International Symposium on Nano Science and
Technology, Taiwan, 20-21 Nov 2014.
[14] F.salahuddin, I.Ahmad, and A.Zaharim, “Impact of Different Dose and
Angle in Halo structure for NMOS Device,” ICMST 2010, pp. 81-85,
26-28 Nov 2010.
[15] K. Ashok “Optimization of Device Performance Using Semiconductor
TCAD Tools”, Silvaco International Product Description, Silvaco
International http://www.silvaco.com/products/descriptions.
[16] J Greer, A. Korkin and J.Lebanowsky, “Nano and Giga Challenges in
Microelectronics,” Molecular and Nano Electronics: Analysis, Design
and Simulation, 1st Edition, 24 Oct 2003
[17] S. Chattererjee, Y. Kuo, J. Lu, J. Tewg and P. Majhi, “Electrical
reliability aspects of HfO2 high-K gate dielectric with TaN metal gate
electrodes under constant voltage stress”, Microelectronics
Reliability,46, 69-76.
[18] N. Kim, T. Austin, D. Blaauw, T. Mudge, K. Flautner, J. Hu, M. J.
Irwin, M. Kandemir and V. Narayanan, “Leakage Current: Moore's
 Law Meets Static Power,” IEEE Computer, Vol. 36, no. 12, pp68-75,
2003
[19] Ganymede, Complementary Metal Oxide Semiconductor (CMOS),
[online]
http://ganymede.meccahosting.com/~a0006505/Dictionary/CMOS_Dict
ionary.htm, [accessed 20 October 2015]
[20] E. N. Shauly, “CMOS Leakage and Power Reduction in Transistors and
Circuits: Process and Layout Considerations,” Journal of Low Power
Electronics and Applications, open access January 2012
[21] A. Chandrakasan, et al., “Low-power CMOS digital design,” IEEE
JSSC, 27(4), pp. 473~484, Apr. 1992.
 Customer Engagement Manifestations on Facebook
Pages of Tesco and Walmart
Swaleha Peeroo
Department of Management
Université des Mascareignes
Mauritius
speeroo@udm.ac.mu
Brian Jones
Leeds Business School
Leeds Beckett University
United Kingdom
b.t.jones@leedsbeckett.ac.uk
Abstract—Social media have transformed the world in
which we live. Although several studies have uncovered forms
of customer engagement on social media, there is a scarcity of
academic research on customer engagement within the grocery
sector. This study therefore aims to address this gap in the
literature and shed light on the various ways customers engage
with grocery stores on Facebook. Netnography is used to gain
an understanding of the behaviour of customers on the
Facebook page of Tesco and Walmart. The findings of this
study reveal that cognitive, emotional and behavioural
customer engagement are manifested and that customers can
both create and destroy value for the firm. This study
contributes to knowledge by uncovering the various forms of
customer engagement on Facebook pages of Tesco and
Walmart.
Keywords—Social Media; Customer Engagement;
Relationship Marketing; Grocery Stores; Facebook
I. INTRODUCTION
Technological advances are dramatically and substantially
transforming the retail industry [1] [2]. The Internet enables
the transfer of information and knowledge worldwide in real
time to customers [3] who use these technologies to enhance
their shopping experience [4] To improve the satisfaction
levels of customers, retailers are increasingly using social
media, mobile and Internet technologies to enrich their
shopping experience [4] The advent of social media is
revolutionising marketing practices [5] [6].
With social media, it is now feasible for businesses to
have ongoing dialogues and exchange of experiences by
actively listening and responding to queries of customers.
However, in practice businesses seem to be using social
media just as any other communication medium and they
are not directly interacting and are not seeking to obtain
highly engaged customers through social media.
Consequently, it is critical that businesses understand the
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Martin Samy
Leeds Business School
Leeds Beckett University
United Kingdom
M.A.Samy@leedsbeckett.ac.uk
expectations of customers who are interacting with them
through social media.
Following a call for research on understanding what
customers seek when interacting with businesses on social
media [7], this paper aims to provide knowledge as to why
and how customers engage with companies on social media,
more specifically on Facebook pages of grocery stores.
Furthermore, some researchers examined the motivations of
customers for interacting with apparel retailers on social
media and suggested that further research be carried out to
understand motivations of customers who connect with
retailers on social media in different retail sectors [8]. This
paper will therefore focus on the grocery sector to
understand motivations of customers for interacting with
grocery stores on Facebook by analysing responses of
customers to customer and company initiated messages.
The purpose of this study is to examine the various ways
customers engage with grocery stores. The paper is
organised as follows. First, the main concepts of social
media, customer engagement and relationship marketing are
presented. Second, we outline the netnography method,
followed by an analysis and discussion of the findings. We
then present the conclusion, limitations, and suggestions for
future research directions.
II. LITERATURE REVIEW
A. Social Media and Facebook
Social media usage is exploding and online platforms
have become vital tools for marketing [9]. Facebook is the
social media platform the most widely adopted by brands
and companies [10]. These social media platforms have
converted the Internet from a platform for information to a
platform for influence [6]. Individuals leverage social
networks and blogs to create, recommend and share
information, hence extending the reach of marketing
influence [6]. Garretson [11] notes that consumers in the
digital age increasingly use the Internet and social media
platforms to not only find information about products and
services, but also to engage the corporations they purchase
from, and connect with other customers who may provide
valuable insights. Businesses during the past two years have
flocked to social media websites to interact with their
customers [12]. These businesses of all sizes within most
sectors are attracted by the potential benefits of adopting
these social media platforms [13]. Yet, many managers are
still unsure of the opportunities and threats pertaining to
social networks and this may be due to the lack of
knowledge of their structure and operation [14].
Nevertheless managers have rushed in the social media
world out of fear of losing the social media battle [15].
Social media have radically transformed the world
communication systems [16]. Among all the social media
platforms, Facebook is by far the most popular site boasting
1.23 billion active monthly users [17]. Facebook has now
turned into the most popular social media platform
worldwide [18]. It is an online community where people can
communicate with family, friends, and colleagues around
the globe [19].
Unlike other social media platforms, Facebook actively
seeks to attract businesses to use Facebook as a marketing
tool by enabling users to access the network through
applications and also gaining access to corporate websites
through social plug-ins [18]. Dekay [20] has identified
several reasons to explain why businesses have rushed to set
up their Facebook page. Being a global and popular
marketing and recruiting channel that allows direct
interaction with customers and employees, Facebook has
attracted businesses [20]. These data allow companies to
personalise the customer experience of visitors to their pages
[20]. Additionally Facebook allows companies to create and
develop applications (commonly known as ‗apps‘) which are
used to connect their Facebook page to their corporate
website. This eases the ordering process and other
transactions [20].
B. Social Media Marketing
Social media have brought about significant and pervasive
changes to communication between companies,
communities, and individuals [21]. Yet, several authors
have noted that organisations are still unsure of how to tap
the full potential of social media [22-24]. Traditional
methods of management are not suited to deal with
consumers on social media who expect firms to listen to
them, engage and respond accordingly [21]. Similarly,
marketers should shift from traditional campaign thinking
and to focus more on relationship building [25].
Social media marketing has been defined as ―a social and
managerial process by which individuals and groups obtain
what they need and want through a set of Internet-based
applications that enable interaction, communication,
collaboration of user generated content (UGC) and hence,
sharing of information such as ideas, thoughts, content and
relationships‖ [26]. Social networking sites such as
Facebook are increasingly being used by organisations to
establish their marketing network, since social media
platforms facilitate the establishment of communication and
ongoing real time conversations [27].
Marketers consider that product messages sent or
endorsed by friends are more believable and hence more
easily accepted by the network of friends [27]. This form of
electronic word of mouth (eWOM) has encouraged
companies to set up product pages on social networks and
urge customers to join these pages with the view of building
enduring long-term relationships [27]. When ―liking‖ or ―re-
tweeting‖ product related messages, these posts are
automatically sent to the user‘s newsfeed and consequently
the other members i.e. friends or followers in the network
would become aware and this may influence them to engage
in relationships with that product. The active transactional
or non transactional interactions of a consumer with an
organisation, with potential customers and with other
customers is referred to as ‗Customer Engagement‘ [28].
C. Customer Engagement
Social media enable interaction with customers and hence
companies have the opportunity to engage with customers
[29]. Interestingly, Web 2.0 technologies and tools seem to
facilitate the co-creation of value by the company and the
customer [30]. Interactivity between customers and a
business is at the heart of the engagement construct [31].
The term customer brand engagement denotes the ―level of
an individual customer‘s motivational, brand-related and
context-dependent state of mind characterised by specific
levels of cognitive, emotional and behavioural activity in
direct brand interactions‖ [32]. In this definition of customer
engagement, there are the three dimensions cited in the
literature, namely the cognitive, emotional and behavioural
perspectives [31], [33].
From a cognitive standpoint, customer engagement is a
positive state of mind that is embodied by energy, high
commitment, and loyalty towards an organisation [34], for
e.g. the user's level of concentration or engrossment in the
brand [32]. Behavioural customer engagement refers to
actions toward a firm that are not restricted to only
transactions [34], but also include participation, vigour and
interaction with the focal brand [33]. From an emotional
perspective, customer engagement may be characterised by
feelings of an individual towards a brand [35], for e.g. a
user‘s level of brand-related inspiration and/or pride [32].
Unlike authors who have used multidimensional
perspectives, van Doorn et al. [36] focus only on the
behavioural dimension of customer engagement. According
to these scholars, customer engagement consists of
behaviours, which go beyond transactions. This definition
highlights that behavioural manifestations do not only mean
purchases, but also include other activities of the customer
such as word of mouth, customer co-creation and
complaining behaviour [37] recommendations, helping
other customers, blogging, writing reviews and even
engaging in legal actions [36].
Customer engagement also incorporates customer co-
creation [36]. Value co-creation is defined as the mutual
collaborative activities by stakeholders participating in
direct interactions aspiring to add to the value that
materializes for either one or both parties [38]. Customers
can participate in the invention of the offering, or the design
and production of related products. Hence, co-creation
happens when the customer contributes through
spontaneous, discretionary behaviours that personalise the
customer-to-brand experience [36].
In virtual communities, conversations occur on prices,
performance, quality and personal experiences with specific
brands [39]. These conversations in virtual communities
illustrate customer empowerment and value co-creation
[39]. Interventions of marketers are accepted in online
communities only if they contribute to the community [39].
Customers tend to react negatively when marketers have
commercially driven communications. This research reveals
that customer engagement increases satisfaction, loyalty,
empowerment, emotional bonding, connection, trust and
commitment.
Social media enable customers to connect and interact
with other customers and non-customers in their social
networks and influence them [40]. Customers with strong
emotional bonds can become advocates for businesses in
peer-to-peer interactions with other customers and non-
customers and play a crucial role as co-creators of value in
the value adding process [40].
However, when organisations fail to engage customers
they have to face the potential threat of customer
enragement [41], a situation where customers can easily
become value destroyers instead of value creators for
companies [42] [43]. The open-comment platform of
Facebook and the anonymity offered by the Internet produce
the ideal conditions for public outrage to be vented on
corporate walls [44]. Furthermore, social media have
empowered customers and the public by giving them a voice
and weakened the position of companies by rendering them
vulnerable to customer attacks, negative publicity and
corporate reputation damage [45]. Additionally, social
media users can generate huge waves of outrage within a
short period of time when reacting to questionable activity
or statement of an organisation [46]. They qualify such a
phenomenon as an online firestorm, which they define ―as
the sudden discharge of large quantities of messages
containing negative WOM and complaint behavior against a
person, company, or group in social media networks‖ [46].
Similarly, these repetitive and systematic customer attacks
are referred to as shitstorm [45]. ―A shitstorm denotes
emotional and often irrational criticisms carried out by
many consumers. Rational negative opinions usually form
the basis for shitstorms, which eventually grow through
irrational and assertive content added by other dissatisfied
users.‖ [45]. Likewise, customers tend to express strong
emotions on the Internet more readily and easily owing to
the anonymity offered via this medium of communication
[44].
III. METHODOLOGY
To gain an understanding of the types of customer
engagement occurring on the Facebook pages of grocery
stores, a netnographic study was undertaken. Netnography,
developed by Dr. Robert V. Kozinets in the late nineties, is
a participant-observation research used for data collection to
research online communities. It is an interpretive method
formulated specifically to investigate the consumer
behaviour of communities and cultures present on the
Internet [47]. Netnography provides insights into virtual
communities similar to the ways that anthropologists try to
understand the norms, cultures and practices of traditional
offline communities. Virtual communities consist of online
gatherings of customers expressing interest in similar
lifestyles, brands, products and services [47]. Similar to
ethnographic research, netnography tries to provide
understanding of a community and the interactions and
communications within the community [48].
The purpose of this netnographic research is to observe
the interactions between hypermarkets and supermarkets
with their customers within a Web 2.0 platform. In order to
undertake netnographic research, the researchers have
observed interactions on the official Facebook pages of
Tesco and Walmart. Facebook has been selected as it is the
social platform, which is the most widely used by
companies to interact with their customers. The researcher
has opted for a non-participation observation during the
netnographic research similar to studies carried out by other
researchers [48], [49].
To ensure a rigorous and reliable research approach, the
researchers have followed the five stages and procedures
recommended by Kozinets [47]. The five sequential steps
are (1) making entrée, (2) data collection and analysis, (3)
providing trustworthy interpretation, (4) research ethics, and
finally (5) member checks.
For the entrée, the researchers have selected Tesco and
Walmart, the leading supermarkets and hypermarkets in the
world based on the March 2013 Global Food Retail report.
Moreover, both Tesco and Walmart have set up their official
Facebook pages where the language used is mainly English.
Another reason why these two grocery stores have been
selected is that they both have adopted an international
strategy and are operating in several countries. The official
Facebook pages of Tesco and Walmart meet the criteria that
have been set by Kozinets [47] in that they are relevant to
the topic of the research. On both Facebook pages there are
high traffic of postings and a large number of discrete
posters. More detailed and rich data are available on both
Facebook pages. And finally the two Facebook pages meet
the last criteria by enabling companies to communicate to
customers, customers to communicate back to the company,
and also customers communicating to other customers.
Data has been collected for a one-month period during
which saturation of data was reached as recommended by
Kozinets [47]. During this one-month data collection period,
the researcher downloaded conversations occurring on the
official Facebook pages of Tesco and Walmart. The data
were analysed using the qualitative software NVivo 7.0.
Therefore the researcher used qualitative content analysis to
elicit themes from the datasets. Similar to previous research
in social media in other sectors [50] [51], the unit of
analysis consisted of the content of the Facebook pages of
Tesco and Walmart and the coding units were the individual
posts and comments by customers of these grocery stores.
The third step of netnography is to provide trustworthy
interpretation [47]. Research is deemed to be reasonable and
trustworthy when conventional procedures of netnography
are followed while collecting and analysing data [47]. There
has been triangulation of data to enhance credibility of the
study. Triangulation involved the use of a wide range of
customers who have posted comments [52]. Viewpoints and
experiences of customers could be verified against opinions
and beliefs of other customers, thus building a rich picture
of needs, attitudes and behaviours of the users under
scrutiny [53]. Additionally the researcher has achieved site
triangulation by the participation of customers from two
distinct organisations (Tesco and Walmart) in order to
reduce the effect on the research of particular local factors
peculiar to one grocery store. Findings may be perceived as
more credible when similar findings emerge from two or
more different sites [53].
In this study, the ethical procedure recommended by
Langer and Beckman [54]) has been adopted because the
comments posted by customers of Tesco and Walmart on
the official Facebook page are not password restricted and
are available to the public.
The researchers did not carry out member checks in their
study as they argued that it was unnecessary to present the
findings back to members of the community who
participated in the research when it was conducted entirely
unobtrusively.
IV. FINDINGS AND DISCUSSION
Customers react to messages posted by companies in
various ways on Facebook pages of companies. Customer
behaviours such as liking, sharing and commenting on
social media pages are now used as measurements of
consumer engagement in social media [36] [55]. All three
dimensions of customer engagement have been observed on
the Facebook page of Tesco and Walmart.
A. Cognitive Engagement
From a cognitive standpoint, engagement is a positive
state of mind that is represented by high commitment,
energy, and loyalty towards a firm [34], for e.g. the person's
level of concentration or engrossment in the brand [32].
Cognitive engagement occurs at Tesco and Walmart,
whenever customers post comments in which they provide
information and help to other customers, or when they give
advice to other customers. Customers post comments on the
Facebook page of Tesco and Walmart when they want to
share some information to other members of the brand
community. For e.g. a happy customer posted a comment to
inform about the gift he got.
“Found a plane in my sons [sic] kinder surprise” .
This is consistent with findings of a previous study, which
observed that customers gain social benefits by sharing their
experiences with other customers on the social network,
which is essentially a social venue [56].
Another form of cognitive engagement is when customers
post comments about their loyalty towards the stores. A
customer of Walmart posted a comment in which she
expresses her loyalty to the store by mentioning that
Walmart is her favourite store:
“Walmart the best store in the whole world, everything
that I need is there, yeah [sic] because after looking in the
other stores website comparing prices at end Walmart is the
my favourite and I finished my day in Walmart.”
Prolonged customer engagement with a brand can result
in customer loyalty [57]. Additionally, customer loyalty is
triggered in several ways for e.g. through positive online
interactions with the brand and the community members, or
by actively defending the company when faced with negative
user generated content, or by getting good customer care
service after having experienced a bad service [28].
B. Emotional Engagement
From an emotional perspective, customer engagement
may be characterised by feelings of an individual towards a
brand [35]. In this study, four main emotions have been
identified: enthusiasm, humour, sarcasm and skepticism.
Enthusiasm mirrors an individual‘s intrinsic level of
interest and excitement about the online brand community
while enjoyment indicates the level of pleasure and
happiness derived by the customer when interacting with the
virtual community and its members [57]. Enthusiasm is
linked to positive emotions felt by customers who post
enthusiastic comments when they are happy or excited.
Thus they convey in words their hedonic shopping value
[33]. For instance customers of Tesco have expressed
happiness and excitement for the game it launched for
Easter. The following posts reveal the positive emotions felt
by customers who were very happy and excited to have won
after participating in the egg hunt organised by Tesco:
“Thank you tesco [sic] for running the competition! I
won a hudl from egg #17! Very excited for it to arrive”
Customers use humour in their comments when they
find an event or a situation amusing or funny. Walmart
posted an advertisement for yoghurt with a picture showing
a lady putting the yoghurt in her bag and the copy of the
advertisement reads as
“Meet your new office buddy - delicious Chobani Greek
Yogurt (5.3 oz.). Now only $1 on Rollback.” (Walmart,
posted on 9 April 2014, Walmart, Facebook page).
In response to this advertisement, a customer posted the
following comment:
“SHE'S STEALING YOGURT! LITERALLY PUTTING
IT IN HER PURSE AND NOT IN THE GROCERY KART
[sic]”
These humorous comments posted by customers are
ways by which they engage with companies. This form of
customer engagement brings positive value to the online
community [28].
However there are times when customers express
negative emotions, hence damaging the image of the
company. Comments are labeled as sarcastic when irony is
used to mock or to convey contempt. Customers post such
comments to express their anger and/or disappointment
following an action of the company or an event that has
occurred. This form of emotion brings negative value and so
can potentially harm the relationship between the two
parties. This study reveals that sarcasm is the form of
emotional expression mostly used by customers both on
Tesco and Walmart Facebook pages. Following a company
post in which Tesco asked about the most exotic food that
customers had not tasted yet, several customers posted
sarcastic comments referring to the horsemeat scandal as
illustrated below:
“Was going to say horsemeat but then I remembered, I
had it last year in place of beef !!! [sic]” (TDW, posted on
7 April 2014, Tesco Facebook page).
These findings are consistent with another study which
showed that information broadcast by a third party through
social media exacerbates publics‘ emotions such as disgust,
anger, and contempt when the reason for the crisis is
internal [58]. Furthermore, some customers seem to have
lost trust, which is one of the pillars of relationship
marketing Sarcastic comments may indicate that the
relationship between Tesco/Walmart and its customers have
been damaged.
Sceptic comments are those comments that incorporate
an element of doubt or lack of conviction about something.
Customers let companies know that they remain dubious or
have a feeling of incredulity about promises made by
companies. Sceptic comments are posted when customers
have lost trust in the company and this may damage or have
already damaged the relationship between the customer and
the company. At Walmart, customers post sceptical
comments whenever Walmart post comments, which they
consider to be dubious. For e.g. Walmart shared a link
praising the action of a Walmart employee. One customer
posted two comments to let the other members of the
community know that she has doubts about the truth of this
article.
“This is bogus!!!!” (WAB, posted on 23 April 2014,
Walmart Facebook page).
“This is the second article I've seen that a [sic]
employee of Walmart supposedly done something good. Like
I said before Bogus” (WAB, posted on 23 April 2014,
Walmart Facebook page).
When customers are emotionally engaged, the nature of
the relationship changes ([31]. Positive emotions associated
with enthusiasm and humour, tend to enhance the
relationship between the company and the customer, while
negative emotions associated with sarcasm and scepticism
may harm the relationship between the two parties.
C. Behavioural Engagement
From a behavioural viewpoint, engagement refers to
actions toward a firm that go beyond transactions [34], for
e.g. participation, vigour and interaction with the focal
brand [33]. This study reveals that customers engage with
the grocery stores by communicating back (C2B
communication) to seek for more information i.e. customer
query, for entertainment, to get additional incentives and to
participate by responding to posts of the stores.
Additionally, this study uncovers that customers
converse with other customers (C2C communication) on
Facebook, share advertisement, give advice to other
customers, get or provide feedback to other customers,
criticise other customers, help other customers, make
themselves or someone else known to other customers i.e.
reputation and also provide support and encouragement to
other customers. These forms of behavioural engagement
add value to the relationship. These types of customer
behaviour are consistent with findings of van Doorn et al.
[36] who categorise it as a form of behavioural customer
engagement, which incorporates customer co-creation.
Hence, co-creation happens when the customer contributes
through spontaneous, discretionary behaviours that
personalise the customer-to-brand experience [36].
Examples of customer engagement behaviours are when
customers suggest how to enhance the customer experience,
helping and training service providers, or simply helping
other customers to enjoy their customer experience [36]. For
e.g. a customer of Walmart posted a comment to advise
customers how to keep fresh strawberries for a longer period
of time:
“I found a better way for them to last longer, put them in
a bowl of water and a cup of vinegar and soak for 10 min.
and then rinse off and put in refrigerator. They stay fresh
LONGER AND THEY ARE CLEAN AND PESTICIDE
FREE!!! [sic]”
Furthermore, the researchers have identified five
positive actions resulting from comments posted by
customers which add value to the organisation: customer
referral, customer suggestions, customer defending the
company, customer defending employees of the company
and finally promotion by customer. Customers post positive
comments to recommend the brand or the company to their
friends. This type of customer behaviour is consistent with
findings of Sashi [40] who refer to these customers as
advocates. Advocacy is a form of consumer engagement,
which happens when consumers actively recommend
specific brands, and/or ways of using these brands [39].
Similarly, customer referral is a form of customer
engagement and is termed as endorsing [57]. Endorsing is a
form of behavioural customer engagement in which the
customer proactively recommends products and services to
the members of the online community ([57].
Customers at times make suggestions to the company
and these suggestions turn customers into co-creators of
value for the organisation [43]. These types of comments
are very valuable for any organisation as the customers are
readily informing the company about their needs and wants,
providing competitive intelligence for free, and tell
companies how to solve problems that they have
encountered.
Another form of behavioural customer engagement is
when customers defend the company or employees of the
company on the Facebook page. This is highly appealing for
the organisation as it shows high level of customer loyalty
[57]. In this study there are several occasions when the
company has posted on its official Facebook page and have
been criticised by its customers. Other customers who do
not agree with the latter have replied back to these
customers and defended the company. These findings are
consistent with researchers who argue that very often
customers respond to complaints before the companies do
by giving the company the benefit of the doubt [49]. For
example when Tesco posted a comment to gain insight into
the type of exotic food that its customers had not yet
consumed, several customers posted sarcastic comments
referring to the horsemeat scandal, which Tesco had
allegedly been selling in the past. A customer defended the
company by posting the following comment:
“Sick of hearing about horse meat comedy, I don't think
Tesco's would knowingly sell you anything that you did not
want... The supplier is at fault not Tesco
Analysis of data also shows that customers post
comments that are negative and unfavourable for the
company in five main instances: to complain, to criticise, to
provide information about competitors, to warn customers
against a product/service, and lastly to retaliate i.e. to inform
of actions they have taken or are about to take because of
their dissatisfaction. These forms of negative comments are
referred to as negative word of mouth [49]. Complaints are
the most common form of customer post containing an
unfavourable message in this study. Customers use the
Facebook page of the company to voice out their
dissatisfaction and discontent with the product or service of
the company. Customers post messages on the Facebook
pages of Tesco and Walmart to criticise actions taken by the
respective companies. When Walmart posted about its
initiative of empowering women in different parts of the
world, a customer posted a negative comment to inform
Walmart and the Facebook community that she does not
believe this company post, which she describes as hogwash
i.e. insincere speech or meaningless talk as Walmart does not
pay its employees decent wages
“What a bunch of hog wash [sic] !!!! Walmart you are
one of the hugest reasons the working poor stay poor!!!!
Shame on u [sic]!!! Pay living wages and give health
benefits and then you can make claims like this!!!!”
Such type of customer engagement activities may have
negative consequences for the organisation [36]. When
customers post negative reviews, the reputation of the
company may be damaged. Thus, co-destruction of value
arises during interactions between the company and its
customers where instead of creating value for both parties,
value is destroyed either for all parties or for one party [60].
V. CONCLUSION
The aim of this study was to analyse the various forms of
customer engagement by examining the reactions of
customers to company initiated messages on Facebook
pages of Tesco and Walmart. From the extant literature and
the findings of the study on the use of Facebook in the
hypermarket and supermarket sector, it is clear that
Facebook is influencing the way businesses are being
managed today. Within the hypermarket and supermarket
sector, co-creation in terms of new product development or
improving existing products are limited since these
businesses are merely selling products manufactured or
produced by suppliers. Nonetheless, this study has revealed
that customers do create value for the organisation by
interacting on the Facebook page of Tesco and Walmart,
thus adding to the literature on co-creation of value by
customers. Customers become co-creators of value when
they offer advice or help other customers within the
community, when they refer products or services to other
customers, when they make suggestions to the grocery
stores, when they defend the company and its employees.
The findings of this study provide considerable support to
customers destroying value for the grocery stores on their
Facebook pages. This study has revealed that customers
mainly use the Facebook pages of grocery stores to post
complaints and criticisms. Furthermore, they actively
recommend customers to boycott the grocery stores and to
patronise stores of competitors. Additionally, customers
provide information about products and services of
competitors, thus encouraging the other customers to shop
in other outlets. These actions of customers harm the
organisation as they destroy value [41] [42]. This study
therefore contributes to the body of literature in that it has
compiled various ways that customers can threaten and
harm grocery stores on Facebook.
REFERENCES
[1] S. Sands, E. Harper, and C. Ferraro, C., ―Customer-to-non-customer
interactions: Extending the ‗social‘ dimension of the store
environment‖, Journal of Retailing and Consumer Services, Vol. 18,
2011, pp. 438–447.
[2] R. Varadarajan, R. Srinivasan, G.G Vadakkepatti, M. Yadav, P.A
Pavlou,., S. Krishnamurthy, and T. Krause, T., ―Interactive
Technologies and Retailing Strategy: A Review, Conceptual
Framework and Future Research Directions‖, Journal of Interactive
Marketing, Vol. 24, No. 2, 2o11, pp. 96–110.
[3] M. O‘Hern, , and A. Rindfleisch, A., ―Customer co-creation: a
typology and research agenda‖, Review of Marketing Research, Vol.
6, 2009, pp. 84–106.
[4] S. Pookulangara, and. K. Koesler, ―Cultural influence on consumers‘
usage of social networks and its‘ impact on online purchase
intentions‖ Journal of Retailing and Consumer Services, Vol.18,
2011, pp. 348–354.
[5] C. Dubose, ‗The social media revolution‘, Radiologic Technology,
Vol. 12, No 1, 2011, pp. 30-43.
[6] R. Hanna, A. Rohm, V.L., Crittenden, ―We‘re all connected: The
power of the social media ecosystem‖, Business Horizons, Vol. 54,
2011, pp. 265—273.
[7] G. Tsimonis, and S. Dimitriadis, S., ‗Brand strategies in social
media‘, Marketing Intelligence & Planning, Vol. 32, No. 3, 2014, pp.
328-344
[8] R. M . Alexander, J.K., Gentry, ―Using social media to report
financial results‖, Business Horizons, Vol. 57, 2014, pp. 161—167
[9] L.I. Labrecque, ‗Fostering Consumer–Brand Relationships in Social
Media Environments: The Role of Parasocial Interaction‘, Journal of
Interactive Marketing, 2014, in press
[10] C. Ho, "Consumer behavior on Facebook", EuroMed Journal of
Business, Vol. 9, Iss 3, 2014, pp. 252 – 267.
[11] R. Garretson: Future tense: The global CMO. Retrieved September
29, 2010, from http://graphics.eiu. com/upload/Google%20Text.pdf
[12] S. Asur, ‗The Economics of Attention: Social Media and Businesses‘,
Vikalpa: The Journal For Decision Makers, Vol. 37, No. 4, 2012, pp.
77- 85.
[13] G.V. Valerio Ureña, D. Herrera Murillo, N. Herrera Murillo, FJ
Martínez Garza, ―Purposes of the communication between companies
and their Facebook followers‖, Revista Latina de Comunicación
Social, 70, 2015, pp. 110 to 121.
The limitation of the study is linked to the nature of the
netnography method, which restricted the analysis to those
customers who have posted comments online, thus
neglecting other sources such as offline customer feedback.
Despite choosing two main grocery stores, the researcher
has not done a comparative analysis of the use of Facebook
within the two grocery stores as the aim of the study was to
gain an understanding of the types of customer engagement
occurring within the grocery sector on Facebook. For future
research. a comparative analysis of the use of Facebook by
grocery stores may be undertaken to analyse the social
media strategies adopted by these firms. An additional
avenue for future research could be to carry out research by
focusing on one grocery store operating in different
countries. This would help the identification of differences
and similarities between social media communication in
various country contexts. Furthermore, this will allow the
researcher to look at how communication is tailored to
specific locations i.e. the importance of place and culture.
http://www.revistalatinacs.org/070/paper/1037mx/07en.html DOI:
10.4185/RLCS-2015-1037en
[14] A. Kaplan A., M. Haenlein, ―Users of the world, unite! The
challenges and opportunities of social media‖ Business Horizons
Vol.53, 2010 pp. 59-68
[15] Harvard Business School The New Conversation: Taking Social
Media From Talk to Action, Harvard Business Review Report, 2010
Harvard Business Publishing, Boston, MA, available at:
www.sas.com/resources/whitepaper/wp_23348.pdf (accessed
December 17,2012).
[16] D.E., Schultz, J., Peltier, ‗Social media‘s slippery slope: challenges,
opportunities and future research directions‘, Journal of Research in
Interactive Marketing, Vol. 7 No. 2, 2013, pp. 86-99.
[17] Facebook(2014).Newsroom–Keyfacts.http://newsroom.fb.com/Key-
Facts.
[18] H.G. Pereira, M.F Salgueiro,., I., Mateus, ‗Say yes to Faceook and get
your customers involved! Relationships in a world of social
networks‘, Business Horizons, 2014,
http://dx.doi.org/10.1016/j.bushor.2014.07.001
[19] E. Pöyry,., P. Parvinen, T. Malmivaara, T., ‗Can we get from liking to
buying? Behavioral differences in hedonic and utilitarian Facebook
usage‘, Electronic Commerce Research and Applications, Vol. 12,
2013, pp. 224–235
[20] S. H Dekay "How large companies react to negative Facebook
comments", Corporate Communications: An International Journal,
Vol. 17 Iss 3, 2012, pp. 289 – 299.
[21] J. Kietzmann, H. Hermkens, Kristopher & Mccarthy, Ian P. &
Silvestre, S. Bruno ―Social Media? Get Serious! Understanding the
functional building blocks of social media‖, Business Horizons, vol.
54, 2011, pp. 241-251.
[22] S. Fournier, S, J.Avery, ‗The uninvited brand‘, Business Horizons,
Vol. 54, 2011, pp. 193-207.
[23] P. LaPointe, , 'Measuring Facebook's Impact on Marketing', Journal
Of Advertising Research, 52, 3, 2012, pp. 286-287.
[24] W.G Mangold,. D.J., Faulds,), ‗Social media: The new hybrid
element of the promotion mix‘, Business Horizons, Vol. 52, 2009, pp.
357—365.
[25] J. Colliander, & M. Dahlén, 'Following the Fashionable Friend: The
Power of Social Media', Journal Of Advertising Research, Vol. 51,
Iss.1, 2011, pp. 313-320.
[26] N.L Chan,., B.D Guillet,., ‗Investigation of social media marketing:
How does the hotel industry in Hong Kong perform in marketing on
social media websites?‘, Journal of Travel & Tourism Marketing,
Vol. 28, No. 4, 2011, pp. 345-368, p. 347.
[27] K. S Coulter, A. Roggeveen, "Like it or not": Consumer responses to
word-of-mouth communication in on-line social networks",
Management Research Review, Vol. 35 Iss: 9, 2012 pp. 878 – 899.
[28] V. Kumar, L. Aksoy, B. Donkers, R. Venkatesan, T. Wiesel, and S.
Tillmanns, S., ‗Undervalued or Overvalued Customers: Capturing
Total Customer Engagement Value‘, Journal of Service Research,
Vol. 13, No. 3, 2010, pp. 297-310.
[29] N. Woodcock, A. Green, M. Starkey, ‗Social CRM as a business
strategy‘, Journal of Database Marketing & Customer Strategy
Management, Vol. 18, No. 1, 2011, pp. 50–64.
[30] T.M. Harrison, and M. Barthel, ‗Wielding new media in Web 2.0:
exploring the history of engagement with the collaborative
construction of media products‘, New Media Society, Vol. 11, No. 1-
2, 2009, pp. 155-178.
[31] R.J Brodie,. L.D., B., Juric, A, Ilic, , ‗Customer Engagement:
Conceptual Domain Hollebeek, , Fundamental Propositions, and
Implications for Research‘ Journal of Service Research, Vol. 14,
No.3, 2011a, pp. 252-271.
[32] L. Hollebeek, , ‗Demystifying customer brand engagement: Exploring
the loyalty nexus‘, Journal Of Marketing Management, Vol. 27. No.
7/8, 2011, pp. 785-807. p. 790
[33] L.D., Hollebeek, ‗The customer engagement/value interface: An
exploratory investigation‘, Australasian Marketing Journal, Vol. 21
Issue 1,2013, p17-24.
[34] C. Porter, N. Donthu, W. MacElroy, D. Wydra ‗How to foster and
sustain engagement in virtual communities‘, California Management
Review, Vol. 53, No. 4, 2011, pp. 80-110.
[35] S. Vivek, S. Beatty, & R. Morgan, 'Customer Engagement: Exploring
Customer Relationships Beyond Purchase',Journal Of Marketing
Theory & Practice, 20, 2, 2012, pp. 122-146
[36] J. van Doorn, K.N. Lemon, V.Mittal, S. Nass, D. PickP. Pirner, P.,
and P.C. Verhoef, ‗Customer Engagement Behavior: Theoretical
Foundations and Research Directions‘, Journal of Service Research,
Vol. 13, No.3, 2010, pp. 253-266.
[37] T.H Bijmolt, . A., Leeflang, P.S. H., Block, F., Eisenbeiss, M.,
Hardie, B. G. S., Lemmens, A., and Saffert, P., ‗Analytics for
Customer Engagement‘, Journal of Service Research, Vol. 13, No. 3,
2010, pp. 341-356.
[38] C. Grönroos,), ―Creating a relationship dialogue: communication,
interaction and value‖, The Marketing Review, Vol. 1, No. 1, 2012,
pp. 5-14.
[39] R.J., Brodie, A., Ilic, B. Juric, L. Hollebeek, ‗Consumer engagement
in a virtual brand community: An exploratory analysis‘, Journal Of
Business Research, Vol. 66, Iss. 1, 2011b, pp. 105–114.
[40] CM. Sashi, (Customer engagement, buyer-seller relationships, and
social media", Management Decision, Vol. 50 Iss: 2, 2012, pp. 253 –
272.
[41] P.S.H., Leeflang, P.C., Verhoef, P., Dahlström, T Freundt,.,
‗Challenges and solutions for marketing in a digital era‘, European
Management Journal, Vol. 32 (2014) 1–12.
[42] P. Verhoef, S., Beckers, & J. van Doorn, J.,), ―Understand the Perils
of Co-Creation‖, Harvard Business Review, Vol. 91, No. 9, 2013, p.
28-2.
[43] P.C Verhoef, W.J., Reinartz, and M. Krafft, ‗Customer Engagement
as a New Perspective in Customer Management‘, Journal of Service
Research, Vol.13, No.3, 2010, pp. 247-252.
[44] V.Champoux, J. Durgee, L. McGlynn, "Corporate Facebook pages:
when ―fans‖ attack", Journal of Business Strategy, Vol. 33, Iss 2,
2012, pp. 22 – 30.
[45] I Schulze Horn., T.Taros, S. Dirkes, L. Huer, M. Rose M., R.
Tietmeyer, E. Constantinides, ―Business Reputation and Social Media
- A Primer on Threats and Responses‖, IDM Journal of Direct, Data
and Digital Marketing Practice, Vol. 16, no. 3, 2015,
http://www.palgrave-journals.com/dddmp/index.html. p. 4
[46] J. Pfeffer, T. Zorbach, & K. M., Carley, ‗Understanding online
firestorms: Negative word-of-mouth dynamics in social media
networks‘, Journal of Marketing Communications, Vol. 20, Nos 1-2,
2014, pp. 117-128, DOI: 10.1080/13527266.2013.797778, p. 118.
[47] R.V Kozinets,., (2002), ‗The field behind the screen: using
netnography for marketing research in online communities‘, Journal
of Marketing Research, Vol. 39, pp. 61–72.
[48] A. Rageh, TC Melewar, A. Woodside, ―Using netnography research
method to revel the underlying dimensions of the customer/tourist ,
Qualitative Market Research: An International Journal Vol. 16 No. 2,
2013,pp. 126-149
[49] J. Colliander, and A.H., Wien, ,―Trash talk rebuffed: consumers‘
defense of companies criticized in online communities‖, European
Journal of Marketing,Vol. 47, No. 10, 2013, pp. 1733-1757.
[50] H. Hsieh, & S. Shannon, ‗Three approaches to qualitative content
analysis‘, Qualitative Health Research, Vol. 15, No. 9, 2005 pp.
1277–1288.
[51] C. Stavros, M.D Meng,., K. Westberg, F. Farrelly, ‗Understanding
fan motivation for interacting on social media‘, Sport Management
Review, Vol. 17, 2014 pp. 455–469
[52] Y.S Lincoln,., and E.G Guba,.‗Naturalistic Inquiry‘, Beverly Hills,
CA: Sage Publications, 1985.
[53] A.K Shenton,. ‗Strategies for ensuring trustworthiness in qualitative
research projects‘, Education for Information, Vol. 22, 2004, pp. 63–
75
[54] R. Langer, & S.C. Beckman, ‗Sensitive research topics: netnography
revisited‘, Qualitative Market Research: an International Journal,
Vol. 8, No. 2, 2005 pp. 189-203.
[55] J. Gummerus, V. Liljander, E. Weman, M. Pihlström, ,"Customer
engagement in a Facebook brand community", Management Research
Review, Vol. 35 Iss: 9, 2012, pp. 857 – 877.
[56] H. Park, Y.K Kim, ‗The role of social network websites in the
consumer–brandrelationship‘, Journal of Retailing and Consumer
Services, Vol. 21, 2014, pp. 460–467
[57] L. Dessart, C.Veloutsou, A. Morgan-Thomas, "Consumer
engagement in online brand communities: A social media
perspective", Journal of Product & Brand Management, Vol. 24 Iss
1, 2015, pp. 1 – 33.
[58] Y.Jin, B. Fisher Liu, & L. L Austin,.. ―Examining the role of social
media in effective crisis management: The effects of crisis origin,
information form, and source on publics‘ crisis responses‖,
Communication Research, Vol. 41, No.1, 2014, pp. 74–94.
[59] A. Palmer, & N. Koenig-Lewis, ―An experiential social network-
based approach to direct marketing‖, Direct Marketing: An
International Journal, Vol.3, 2009, pp. 162–176.
[60] L. Plé, R. Chumpitaz Cáceres, "Not always co-creation: introducing
interactional co-destruction of value in service dominant logic",
Journal of Services Marketing, Vol. 24, Iss 6, 2010, pp. 430 - 437
 The Impact and Opportunities of e-Tutoring in a
Challenged Socio-Economic Environment
Petra le Roux1; Marianne Loock2
School of Computing
University of South Africa (UNISA)
lrouxp@unisa.ac.za1; loockm@unisa.ac.za2
Abstract—As social network sites rise in popularity, the youth
spends large amounts of their time browsing the Internet and
interacting via social network sites. Furthermore, social network
sites creates opportunities that allow the youth to connect to
different learning environments and thus opens up options for
new dimensions to learn. This gives rise to research on a shift
from conversational to education content of social media.
However, due to social and economic challenged circumstances,
many children in third world countries cannot share in these
opportunities. Poor numeracy and literacy levels achieved in
basic education, predict huge stumble blocks for these learners
during their school career. Added the huge shortage of teachers
who can provide mother tongue education to non-English
speaking learners, a challenge is inevitable. But the rapid
technological changes in information and communication
technology enable people to help one another, even over distance.
Situated in this context, the broad aim of this ongoing research is
to investigate how the use of social networking tools as a platform
for cross-age e-tutoring addresses the social and educational
needs of social and economic challenged learners. The first part
of this ongoing research used an experimental study and the
conclusions drawn clearly indicates that the possibility exists for
adolescent tutors to develop a higher self-esteem when
administered in a position as tutors of younger tutees. The
younger tutee also indicated that it was a positive learning and
social experience. These results are enough reason to redo this
research by moving to a quasi-experimental research method
with a control group.
Keywords—social computing; mobile learning; e-tutoring;
identity development; information security awareness
I. INTRODUCTION
There are severe problems with basic numeracy and
literacy levels in third world countries [1], [2]. According to
the South African Department of Basic education [3] the
"quality of basic literacy levels is still well below what it
should be. Fewer than half of all learners in the country
perform at a level that indicates that they have at least partially
achieved the competencies specified in the curriculum. In
Grade 6, the results indicate that only around 30% of learners
fall into this category. The percentage of learners reaching at
least a „partially achieved‟ level of performance varies from
30% to 47%, depending on the grade and subject considered.
The percentage of learners reaching the „achieved‟ level of
performance varies from 12% to 31% ".
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
The severe problems with basic numeracy and literacy
levels in South Africa have a ripple effect that can be seen in
the low Grade 12 mathematics throughput. Mathematics offers
an account of a learner's ability to think logically and
systematically, reason, judge, calculate, compare, reflect and
summarize. Furthermore, mathematics requires consistency,
the knowledge of numbers learnt from Grade 1 and earlier. It
is generally accepted that addressing numeracy and literacy
problems during early years of education will reduce the
problem of high levels of failure and dropout in later grades.
This challenge is particularly important in communities that
are socially and economically challenged.
South Africa has a huge shortage of teachers - especially
teachers who can provide mother tongue education to non-
English speaking children. This problem has a direct effect on
class sizes and teachers need support that will make up for the
fact that they cannot attend to individual children sufficiently
during class time [3]. Many current teachers have had little
exposure to technology and do not feel comfortable using it.
They will require training (with patience) to get them to adjust
fully to teaching practices that rely on technology. Bringing
technology into the classroom may be met with resistance
from current, older teachers which can lead to alienation.
The problem stated above can be addressed by improving
the numeracy and literacy outcomes for the socially and
economically challenged learner without placing additional
load on the already overburdened teaching resources in the
formal learning environment [4]. A solution is to make use of
out-of-school time for learning where school and community
partners together to create academic and enrichment activities
to support learning. Expanded learning opportunities describe
these ranges of learner programs and activities that occur
beyond the traditional school hours [5]. The proposed solution
will endeavour to provide a secure expanded learning
opportunity for socio-economic challenged youths through e-
tutoring. This will be attempted by examining both the
opportunities offered by social networking tools and mobile
learning as well as the sociological and psychological impact
it has on the participants.
II. LITERATURE REVIEW
A. Social computing
Schuler [6] defines social computing as "any type of
computing application in which software serves as an
intermediary or a focus for a social relation" [6]. Thus, social
computing is a cross-disciplinary research and application
field that includes both computational and social sciences. One
application of social computing is the creation of an online
social network "to support effective online communications
for social communities" [7]. Social network sites open up new
options to learning and allow a learner to connect to formal,
non-formal and informal learning environments. Non-formal
learning is a continuous process taking place in multiple
places. In an education context, non-formal learning describes
learning that takes place away from the classroom
environment. There is plenty evidence of how people who
thing alike find one another and organise knowledge exchange
for educational purposes. Even young children, even those
with limited access to technology, are drawn to technology
and will be able to use social network sites to communicate
with others.
As the use of social network sites rise in popularity, the
youth spends large amounts of their time browsing the Internet
and interacting via social network sites. It is thus important to
be aware of the effect that this use has on their sociological
and psychological development. According to Wang et al. [7]
there is a move in social computing from what he calls social
information - features such as social relations, influence and
control - to social intelligence – motivations, intentions and
attitudes.
B. Educational potential of social networking sites
There is an increase to use social networks for non-formal
educational activities [8]. According to Lang [9] there is
evidence that social networks can sustain some kinds of
learning. Boyd & Ellison [10] identify four aspects of social
network sites that have implications for education: persistence,
searchability, replicability and invisible audiences. The
persistence of network sites refers to electronic text that can be
stored indefinitely. Social network sites allowed that all
content can be easily found (searchability) as well as duplicate
and change (replicability). The fact that the audience is
invisible allows that no one can tell who is online and how
thoughts are interpreted. These characteristics of social
networking sites hold significant potential in the context of
mobile technology.
C. Mobile learning
There is an increase to use social networks for non-formal
educational According to the mobilearn consortium [11],
"mobile learning is any sort of learning that happens when the
learner is not at a fixed, predetermined location, or learning
that happens when the learner takes advantage of the learning
opportunities offered by mobile technologies”. In mobile-
learning (m-learning) both the instructor and learner are
mobile and interact with portable technologies. The factors
that make m-learning convenient are that it is accessible from
virtually anywhere, sharing is almost instantaneous and
feedback and tips can be instant. The use of mobile tools to
create learning materials is becoming an important part of the
non-formal learning environment [12]. Technology and
mobile technology in particular, are identified as central in
addressing the problems in basic education in the developing
world and the shortage of skilled teachers.
D. Cross-age e-tutoring
Cross-age e-tutoring refers to the experience of an older
child working with a younger child [13]. The interest in using
children as instructional resources for each other is based on
the capability of children to be active in their education. In the
learning process a child is no longer a passive receiver of
instruction, but an active participant in the teaching-learning
process. One of the most powerful, proven and economical
liable way to help children to learn, both academically and
socially is cross-age tutoring.
E. Social networking sites as an environment for Identity
Development
According to Tazghini and Siedlecki [14] self-esteem
"refers to the extent to which a person prizes, approves, likes
and values him or herself". Also as stated by Smith and
Mackie [15] "for an individual to direct and regulate his or her
thoughts, feelings and behaviours, an individual needs to
know who he or she is". For a person to know him- or herself
a person must have self-knowledge which is made up by two
components: the self-concept and self-esteem. The self-
concept is what one thinks about oneself. Some aspects of
one‟s self-concept can be evaluated and these evaluations, be
it positive of negative, determine how one feels about oneself
(self-esteem). Self-esteem originate from three major sources:
family experiences, interpersonal feedback and social
comparisons (peer-acceptance). Both Valkenburg and
Schouten [16] and Pujazon-Zazik and Park [17] confirm that
research showed the importance of peer acceptance and
interpersonal feedback on the youth. However, both these
influences are predictors of self-esteem in the youth [18]-[20].
Therefore, self-esteem is used as the measure to examine the
effect that the use of social networking site has on the
sociological and psychological development of participants.
F. Social networking sites and Information Security
Awareness
The cross-age tutoring model for this research resulted in a
group of younger children (tutees) being introduced to mobile
computing. These tutees come from socio-challenged
environments where mobile computing might not be generally
accessible. This immediately puts these tutees in the
vulnerable position of being possible targets of Information
Security crimes via mobile computing. This possibility gave
the research the opportunity to test the tutees‟ Information
Security Awareness levels on a specific concept in the pre-
tests. This Information Security Awareness concept was then
explained by the tutors to the tutees and the tutees‟ awareness
regarding the specific concept were again tested in the post-
tests, to measure any possible changes in awareness.
 III. PURPOSE OF THE RESEARCH
This research wants to find a way to create an online
learning community and change the conversational content
from social to educational. For this reason, the choice was a
cross-age e-tutoring environment where both the tutor and tutee
will provide input. The cross-age e-tutoring framework will
address the literacy and numeracy skills, and the relationship
between the tutor and tutee to accommodate learning in the
context of a social networking site is central. Although the
Information security awareness concepts that were also
introduced in the pre-tests, and the awareness thereof tested
again in the post-tests, can be seen as educational, it is the
opinion of this research that any higher awareness level of the
tutees are of an educational as well as a personal safety value.
IV. METHODOLOGY
The aim of this research is to propose an environment
where cross-age e-tutoring through social network sites can
occur.
A three phase approach was used. Phase 1 was initiated by
an in-depth study of existing literature on issues pertaining to
social networking sites, mobile learning and self-esteem in
youth in order to compile a body of knowledge needed to
investigate the contribution of cross-age e-tutoring through
social networking tools on the self-esteem of both the tutor
and tutee. Phase 2 included an investigation and proposal of
available models for e-tutoring, social networking tools,
tutoring materials and self-esteem measuring instruments.
Lastly, Phase 3 was used to establish and verify the proposed
model and the subsequent analysis of the data.
The approach used was a mixed methods research
method approach. As stated by Johnson, Onwuegbuzie and
Turner [21] "mixed methods research is an intellectual and
practical synthesis based on qualitative and quantitative
research; it is the third methodological or research paradigm
along with qualitative and quantitative research".
As the purpose of this study is to understand the
phenomenon of self-esteem of youth in a specific
environment, a quasi–experimental design was deemed
suitable. However, quasi-experimental studies that engage
treatment and control groups matched on pre-tests and post-
tests of key outcome variables were not included in this
review. At this stage only an experimental study with the
treatment group are documented. The pre-tests and post-tests
of only the treatment group also form part of this review. The
pre-test and post-test studies where a treatment group is
compared to a control group were not included and forms part
of future research.
Sampling is fundamentally concerned with finding a
representative set of the unit of analysis. Purposive sampling
will be used in this study and is a non-probability sampling
method. In purposive sampling, personal judgement and
willingness to participate are used to decide on the selection of
participants. This is to ensure that participants will stay
involved.
For data collection and verification several methods were
used. The Rosenberg Self-Esteem Scale (RES), a ten item
Likert scale, was used to gain insight into the self-esteem of
each tutor whereas a survey was used to gain insight into the
self-esteem and learning of the tutee. Other data collection
instruments used includes interviews, observation, and the
keeping of experience journals. These gave rise to
descriptions, quotations, excerpts and social data, resulting in
narrative descriptions.
The Ethics Review Committee (ERC) of the institution
ensured that all research carried out has been subjected to the
necessary ethical reviews. The ERC has prescribed guidelines
for obtaining ethical clearance and these procedures were
followed diligently
V. IMPLEMENTATION OF THE METHODOLOGY
The proposed methodology was used in a week long study
where the focus was on five grade 11 adolescents (tutors) and
five learners from unfavorable socio-economic circumstances
(tutees). Essentially the model was created to accommodate
the tutoring of the learners by adolescents using social
networking tools. Several factors were considerate when the
environment was compiled, which include availability and
cost, Windows support and security. In order to create such an
environment, several possibilities were created, tested and
integrated before the environment was established.
The main aim of this project is to add a human touch to
tutoring when face-to-face tutoring is not a possibility. Thus,
the environment must enable a tutee and tutor, at their
different locations, to see and hear one another. Thus, the
minimum software environment consists of a peer-to-peer
VoIP application that specializes in providing real-time
multimedia services over the Internet and tutoring material.
The first service that the VoIP application must provide is
synchronous audio-video communication, collective named
videotelephony. Videotelephony allows for the simultaneous
transmission and reception of audio-video signals at different
locations. This video service will enable the tutee and tutor at
their different locations to see and hear one another. The
second service needed is a means for asynchronous
communication. Short Message Service (SMS) is an element
of any mobile communication device or system and involves
the exchanging of short text messages via standardised
communications protocols. This SMS service is needed to
allow both the tutor and tutee to setup time slots for tutoring
sessions. A third service needed is a desktop sharing facility
for real-time collaboration. Desktop sharing is a technology
that allows remote access or collaboration on a computer
through a graphical terminal emulator. Online collaboration is
an important component of multimedia communications and
allows for the tutor to share his or her screen with the tutee,
and vice versa. The final required service it a whiteboard. In
the context of computer software applications the term
whiteboard refers to features that simulate traditional
whiteboards. "Virtual whiteboards" has a simulated canvas on
which participants can write or draw. This will enable the tutor
to explain concepts to the tutee. Video chat, short messaging,
desktop sharing and a virtual whiteboard will create the notion
of virtual space where the tutor and tutee can socialise and
tutoring can take place. The minimum hardware requirements
are 2GB or more of RAM, microphone and speakers and a
webcam. A webcam is a relatively low cost device which can
provide live video and audio streams via a laptop or tablet, and
is used for video chats.
The SkypeTM implementation was selected as it is currently
one of the most popular VoIP applications available on the
web. The services Skype provides include the minimum
services explained above and it is Microsoft compatible. As
cost is a huge consideration bearing in mind the socio-
economic circumstances of the tutee, the software had to be
free. SkypeTM is based on a freemium model. A freemium is
one of the many pricing strategy available for online products
or services. SkypeTM is provided free of charge, but money is
charged for other functionalities and proprietary features.
As the system that this research will employ will augment
the literacy skills of socially and economically challenged
children in South Africa via social network sites the sample
will consist of tutors and tutees. Five tutors working with a
tutee participated in the study. The tutors were 16 to 17 year
old adolescents. The tutors have sound academic backgrounds
as well as technological skills. An important prerequisite is
that they must have the same mother tongue as the tutees. The
tutees will be underprivileged grade 2 to 3 children who live in
care centers where after school support is not possible or
I feel that I am a person of worth, at least on an equal plane
with others.
All in all, I am inclined to feel that I am a failure.
I am able to do things as well as most other people.
I feel I do not have much to be proud of.
I take a positive attitude toward myself.
On the whole, I am satisfied with myself.
I wish I could have more respect for myself.
Fig 1: Rosenberg Self-Esteem Scale (RES)
As this research is still ongoing, intermediate results are
presented here. The pre-test was conducted before any of the
participants started with the e-tutoring. After the week long
participation in the e-tutoring, the tutors complete another
RSES. Table 1 depicts the scores obtained of one tutor. The
overall pre-test score is 23 and post-test score is 25; indicating
a high self-esteem as well as an increase in the self-esteem
RES dataset of a
participant(3)
4
Response scale 1 to 4

nominal. Tutorial material included e-books, mathematic 3

worksheets and comprehension reading worksheets.
2
The project was started with an initial face to face meeting
1
where the tutees was familiarised with the technology and
software application by the tutors. An interactive session via 0
1 2 3 4 5 6 7 8 9 10
tablets was established. During this first session, although
physically not distributed, an online tutoring session took pre-test 1 3 3 3 2 2 3 2 2 2
place to familiarise all role players. Five tutoring sessions was post-test 3 1 3 2 3 2 3 2 3 3
agreed upon by each tutor and tutee. The tutees were assisted RES questions
during these tutoring sessions by caretakers. A final face-to-
face session was conducted after the completion of the e-
tutoring sessions. The tutors completed a RES questionnaire score.
before and after embarking on the tutoring process. The tutors TABLE 1: TABLE REPRESENTING DATA COLLECTED VIA RSES - TUTOR 3.
kept an experience journal throughout the process. Interviews
were also conducted.

VI. DATA
QUANTITATIVE

The pre- and post-tests involve the completion of the

Rosenberg Self-Esteem Scale (RES) by the tutors. The RES
was developed as a measure of self-esteem. As discussed by
Ciarrochi and Bilich [22] the RES items represent a continuum
of self-esteem statements ranging from statements that are
endorsed even by individuals with low self-esteem to
statements that are endorsed only by persons with high self-
esteem. Rosenberg (1979) scored his 10-question scale that
was presented with four response choices, ranging from
strongly-agree, agree, disagree to strongly-disagree as
depicted in Figure 1.
TABLE 2: TABLE REPRESENTING DATA COLLECTED FOR ALL TUTORS

Table 2 depicts the pre- and post-test scores of all the
tutors. There was an increase in the self-esteem score with all
other tutors except 1 (tutor 2). Examining the experience
journals each tutor kept, it was clear that technical issues
hindered the e-tutoring session of tutor 2 which could have
contributed to the score.
The final table, table 3 depicts the overall increase in self-
esteem during the experiment.
RES score impact
someone talked to me", "I will never forget you", "can we
please do it again". And all but one could recall the security
issues as explained at the beginning of the study.
All tutors kept an experience journal for the duration of the
study and were asked to write down their experiences with the
words situation, feelings and thoughts as a guide. Situation
entails a detailed description of who, what, when, where to
help the tutor to remember his or her feelings and thoughts.
Feelings prompted the tutee to write down their feelings in a
single word. If a sentence is needed to describe their feelings
it is most probably a thought. They were also encouraged to
give the emotion a value between 0% and 100%. Finally they
had to describe their thought with word or pictures. An
extract form an experience journal is depicted in Figure 2.

22.5 Situation Feeling Thought

Meeting with my Good Was a different and new
tutee today. We (80%) experience for me, but
22 chatted and read. enjoyed it.
At home - think Interested She enjoys biology and has
21.5 about my tutee's (80%) a good general knowledge
RES Score

knowledge. of it. She struggles a bit to

21 post-test read. Hope I can help her.
1st Skype session: Disappointed The 1st Skype session did
post-test Try to do some (50%) not go as planned.
20.5
maths problems. Struggled with the internet
connection. Only did 8
20 problems in 25 minutes.
My tutee battles with basic
19.5 addition.
In my room - and Have to do some more
Mean think about our maths problems. As she is
next session. so interested in biology, we
will read a poem of nature.
TABLE 3: TABLE REPRESENTING THE MEAN DIFFERENCE IN THE SCORES (PRE- Back at children's Satisfied I am glad she enjoyed the
AND POST-TEST). home. (80%) week and felt that she had
learned something despite
the Skype problems.
QUALITATIVE Fig 2: Extract from an experience journal.

An initial questionnaire tried to determine the tutors'

concept of social media and the time spend using any form of
social media. Their use of social media is mostly restricted to
the WhatsApp application and Instagram. WhatsApp is a
messaging app that allow for text, pictures and video to be
sent anywhere at no cost. Instagram is a photo-sharing app for
smartphones and is based on the sharing of photos between
users. Although the respondents all indicated that there are
other uses of social media they were not able to clarify.
Post study interviews were conducted with the tutees.
Questions relating to technology use, content of the tutoring
sessions, security and overall experiences were asked. As the
tutees did not have access to the necessary technology they
were supplied with tablets for the duration of the study. There
was no learning curve with technology acceptance and usage.
With reference to tutoring content they found the reading
lessons very helpful. Their overall experiences were very
positive and included responses like "the nicest was that
VII. CONCLUSION
The first set of conclusions drawn from this experiment
clearly indicates that the possibility exists for adolescent tutors
to develop a higher self-esteem when administered in a
position as tutors of younger tutees. This result is enough
reason to redo this research by moving over to a quasi-
experimental research method with a control group.
Benefits that also came from this experiment were the
experience that the socially and economically challenged
tutees gained from the experiment by using social networking
tools. The exposure to information security concepts important
to social networking tools could have a positive result on the
information security awareness of the tutees. This concept will
be measured in the next round of research.
 REFERENCES [22] J. Ciarrochi, L. Bilich, “Acceptance and Commitment Therapy”, School
of Psychology, University of Wollongong, revised 2006, pp. 61-62.
[1] W. Gould, People and Education in the Third World. Google Books,
2014.
[2] S. Nag, S. Chiat, C. Torgerson, and M. J. Snowling, “Literacy, foundation
learning and assessment in developing countries. Education rigorous
literature review.” 2014.
[3] DBE, “Report on the annual national assessment of 2011.” Pretoria, 2011.
[4] G. Masters, “A shared challenge : improving literacy, numeracy and
science learning in Queensland primary schools.” Australia, 2009.
[5] D. Blyth and L. LaCroix-Dalluhn, “Expanded learning time and
opportunities: key principles, driving perspectives, and major
challenges.” New Dir. Youth Dev., no. 131, pp. 15–27, Jan. 2011.
[6] D. Schuler, “Social Computing” Commun. ACM, vol. 37, no. 1, pp. p28–
29, 1994.
[7] F.-Y. Wang, K. M. ; Carley, D. ; Zeng, and W. Mao, “Social Computing:
From Social Informatics to Social Intelligence” Intell. Syst., vol. 22, no.
2, pp. p79 – 83, 2007.
[8] C. Madge, J. Meek, J. Wellens, and T. Hooley, “Facebook, social
integration and informal learning at university: „It is more for socialising
and talking to friends about work than for actually doing work” Learn.
Media Technol., vol. 34, no. 2, pp. 141–155, Jun. 2009.
[9] A. Lang, “Exploring the potential of social network sites in relation to
intercultural communication” Arts Humanit. High. Educ., vol. 11, no. 1–
2, pp. 120–139, Jun. 2011.
[10] D. M. Boyd and N. B. Ellison, “Social Network Sites: Definition,
History, and Scholarship” J. Comput. Commun., vol. 13, no. 1, pp. 210–
230, Oct. 2007.
[11] C. O‟Malley, G. Vavoula, J. P. Glew, J. Taylor, M. Sharples, and P.
Lefrere, “MobiLearn: WP- Guidelines for Learning/Teaching/Tutoring
in a Mobile Environment.
http://www.mobilearn.org/download/results/guidelines.pdf,” 2003.
[Online]. Available:
http://www.mobilearn.org/download/results/guidelines.pdf.
[12] T. Udanor and C. Nwodoh, “A Review of M-learning Models. Indian
Journal of Computer Science and Engineering.” Indian J. Comput. Sci.
Eng., vol. 1, no. 4, pp. p426–435, 2010.
[13] J. Fogarty and M. Wang, “An investigation of the cross-age peer tutoring
process: Some implications for instructional design and motivation”
Elem. Sch. J., vol. 82, no. 5, 1982.
[14] S. Tazghini and K. L. Siedlecki, “A mixed method approach to examining
Facebook use and its relationship to self-esteem” Comput. Human
Behav., vol. 29, no. 3, pp. 827–832, May 2013.
[15] E. Smith and D. M. Mackie, Social Psychology, Third. Hove AND New
York: Psychology Press, Taylor & Frances Group, 2007.
[16] P. M. Valkenburg, J. Peter, and A. P. Schouten, “Friend networking sites
and their relationship to adolescents‟ well-being and social self-esteem.”
Cyberpsychol. Behav., vol. 9, no. 5, pp. 584–90, Oct. 2006.
[17] M. Pujazon-Zazik and M. J. Park, “To tweet, or not to tweet: gender
differences and potential positive and negative health outcomes of
adolescents‟ social internet use.” Am. J. Mens. Health, vol. 4, no. 1, pp.
77–85, Mar. 2010.
[18] T. F. Heatherton and C. L. Wyland, “Assessing Self-Esteem” pp. 219–
233, 1998.
[19] D. R. Kille and J. V Wood, Self-Esteem, 2nd ed. Elsevier Inc., 2012.
[20] V. Zeigler-Hill, “Discrepancies between implicit and explicit self-esteem:
implications for narcissism and self-esteem instability.” J. Pers., vol. 74,
no. 1, pp. 119–44, Feb. 2006.
[21] R.B. Johnson, A.J. R Onwuegbuzie, L.A. Turner,
“Towards a definition of Mixed Methods Research” J. Mixed Methods
Research., vol. 1, no. 2, pp. 112-133, Apr. 2007.
Weighting Technique on Multi-timeline for Machine
Learning-based Anomaly Detection System
Kriangkrai Limthong∗ , Kensuke Fukuda†, Yusheng Ji† , and Shigeki Yamada†
∗ Computer Engineering Department, Bangkok University, Prathumthani 12120 Thailand.
† National Institute of Informatics, Tokyo 101-8430 Japan.
kriangkrai.l@bu.ac.th, {kensuke,kei,shigeki}@nii.ac.jp
Abstract—Anomaly detection is one of the crucial issues
of network security. Many techniques have been developed
for certain application domains, and recent studies show that
machine learning technique contains several advantages to detect
anomalies in network traffic. One of the issues applying this
technique to real network is to understand how the learning
algorithm contains more bias on new traffic than old traffic.
In this paper, we investigate the dependency of the time period
for learning on the performance of anomaly detection in Internet
traffic. For this, we introduce a weighting technique that controls
influence of recent and past traffic data in an anomaly detection
system. Experimental results show that the weighting technique
improves detection performance between 2.7-112% for several
learning algorithms, such as multivariate normal distribution, k-
nearest neighbor, and one-class support vector machine.
Keywords—weighting technique, machine learning, multiple
timeline, anomaly detection.
I. I NTRODUCTION
Because of mobile applications, access and backbone In-
ternet traffic has exponentially grown every year. In addition,
available software and tools for novice attackers is easily
available [1], so that they can simply use these software to
attack and intrude into target networks. These attack tools have
been developed very well to imitate normal packets, flows and
easily evade existing detection systems. Some attackers have
been applied advanced techniques, so conventional detection
methods hardly detect the traffic from such malicious software.
For these reasons, it makes daily operation difficult for network
administrators to monitor unusual incidents or anomalies pass-
ing through their own systems.
Many anomalies in current computer networks are more
complicated than prior ones. We generally categorize anoma-
lies on computer systems into two groups. The first group
is anomalies caused by threats or human intention, such as
attacks, viruses, worms, scanning, and spamming. The second
group is anomalies caused by accidents, including outages,
misconfigurations, and flash crowds. The desirable detection
system should discover both types of anomalies with high
accuracy and produce low error rate.
In past years, researchers have studied various detection
techniques to compete against new malicious software and new
types of anomalies in computer networks. A study by Denning
[2] classified detection systems into host-based and network-
based detection systems. Host-based detection systems are
installed locally on many different types of machines, namely
servers, workstations, notebook computer, or even mobile
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
devices. Host-based detection systems analyze and pass traffic
through a particular host if there are not potentially malicious
packets. An example of host-based detection systems is Port-
Sentry [3]. Network-based detection systems are strategically
positioned in a network to monitor all traffic and detect any
unusual traffic on the hosts of that network. Network-based
detection systems need to be a very fast to analyze traffic,
because it needs to monitor all packets or flows passing though
the network. Examples of network-based detection systems are
RealSecure, SecureNet [4], and Snort [5].
Recently, studies on machine learning [6] show promising
results in anomaly detection. They have been mainly classified
into two techniques: classification and clustering techniques
[7]. Classification is a learning technique that requires labeled
instances as training data to generate a function for classi-
fying a new instance. Detection systems applying supervised
techniques need to label packets or flows which are normal or
anomalous. Several learning algorithms based on classification
technique have been studied, such as k-nearest neighbors [8],
support vector machines [9]. Clustering is a learning technique
that tries to find hidden structure in data. We could detect
anomalies in network traffic on the basis of the assumption
that major groups are normal traffic and minor groups are
anomalies. Many studies have employed clustering techniques,
such as [10] and [11].
As an extension of network anomaly detection with ma-
chine learning, we proposed a multi-timeline detection system
[12] that use multiple time series of network traffic cor-
responding to previous days behavior, as input to machine
learning algorithm. This technique helps the system to detect
anomalies more accurate than those of sigle timeline, as shown
in the empirical results [12]. Nevertheless, this multi-timeline
technique treats all traffic or timelines equally, but some
network environments need a weight on particular timelines,
for example, more on recent timelines. Thus, a use of weight
among different time series would improve the detection
performance.
In this paper, we propose a weighting technique to influ-
ence the multi-timeline detection system over recent timelines.
We conduct experiments on real network traffic to compare
detection performance between weighting and non-weighting
technique for several types of attacks with three learning algo-
rithms: the multivariate normal distribution, k-nearest neighbor,
and one-class support vector machine.
The following sections provide explanation of our detection
system and show experimental results on a campus network
 Network Feature Feature Weighting or by using the range of feature as
Traffic Extraction Scaling Process
x−μ
Training Data x = , (2)
max(x) − min(x)
Detector
Module Classifier Alarm where x is the normalized value derived from x an original
Test Data
value, μ is the average value, max(x) and min(x) are maxi-
mum and minimum values of x respectively. These two scaling
functions have the same time complexity as O(n) where n is
Fig. 1. Block diagram of detector module with weighting process. the number of feature values, so these would take a very short
time of computation for real-time systems.
traffic. Experimental results also show that the proposed de- 3) Weighting Process: Weighting is our proposed process
tection system has the capability to detect several types of for the multi-timeline detection module during the learning
anomalies without prior knowledge of such attacks. process. This module guides the learning algorithm to gener-
ate a decision function which relies on particular timelines.
II. P ROPOSED S YSTEM D ESIGN Consequently, network operators could weight on one or more
timelines to bias the decision function of detection system.
This section provides explanation of our weighting tech- For example, recent timelines should have a strong influence
nique for applying to the multiple timeline-based anomaly on the classifier than other older timelines. There are various
detection system. The original detector module [12] consists weighting techniques for different purposes that could be plug
of three primary functions, namely feature extraction, feature into this module, such as by performing a sum, integral,
scaling, and classifier. The weighting process is a newly added average or even calculus [14]. In our experiments, we adapted
module between feature scaling and classifier during learning a gradual weighting function to recent timelines and we will
process as shown in Figure 1. describe more details of our weighting function later.
1) Feature Extraction: The main function of feature extrac-
4) Classifier: The major role of classifier is distinguishing
tion is selecting important features of network traffic. Extracted
between normal and anomaly in network traffic. The classifier
features should directly or indirectly associate with expected
in detector module is created by an algorithm using training
anomalies. In our study, we can extract feature of network
data during learning process. As a result, effectiveness of clas-
traffic by using information on packet header and aggregate
sifier depends highly upon three factors: learning algorithm,
packets or flows on interval basis. Table I shows traffic features
representation of input data, and training set. In detecting
obtained by aggregating information of packet header.
process, the classifier receives a test data as an input, then
TABLE I. F UNDAMENTAL FEATURES OF NETWORK TRAFFIC BY produces a label of test data as an output to alarm system.
AGGREGATING INFORMATION OF PACKET HEADER . One of our purposes to multi-timeline learning is enhancing a
capability of detector module. In addition, the detector module
f# Feature Description could be applied in various schemes to detect different types
f1 Packet Number of packets of anomalies in network systems.
f2 Byte Sum of packet size
f3 Flow Number of flows In classifier block, conventional detection systems represent
f4 SrcAddr Number of source addresses input data with sigle timeline as shown in Figure 2. For this
f5 DstAddr Number of destination addresses representation, we aggregate traffic into five intervals; xt−4 ,
f6 SrcPort Number of source ports xt−3 , xt−2 , xt−1 , and xt occur on a single timeline or on
f7 DstPort Number of destination ports the same day. We intend to classify a test interval whether it
f8 ΔAddr |SrcAddr − DstAddr| is an anomaly instance or not by using the decision function,
f9 ΔPort |SrcPort − DstPort| the test instance is xt for example. The decision function is
a mathametical function which takes a dataset as input and
2) Feature Scaling: Feature scaling is a process to stan- gives a decision (a test interval is normal or anomalous) as
dardize a wide range of feature values to the same range. output. The decision function f (xt ) can use information from
This process may be unnecessary for other techniques for the rest of instances, from xt−4 to xt−1 (it may include the
anomaly detection in network traffic; however, feature scaling test instance xt as well) as input. As shown in the figure, we
or feature normalization is indispensable for the multi-timeline depict the bold line to represent a test connection between the
detection system that relies mainly upon machine learning test instance xt and the decision function. We also draw the
algorithms. Most learning algorithms will not work properly dash lines to represent learning connection between the rest of
without feature scaling. instances and the decision function. In this figure, we assume
that the decision function does not have learning connection
Even though many scaling functions have been proposed
with the test instance xt .
[13], the common goal of scaling functions is to independently
normalize each feature component to the [0,1] or [-1,1] range. The multi-timeline representation as shown in Figure 3
We used following two scaling functions that would be appro- is different from the single timeline. Suppose we intend to
priate for fast computation. We normalize feature values by classify the test instance xpt on the present timeline p and have
using the max value as other two timelines, p − 2 and p − 1 with the same length. We
x also depict intervals on timeline p − 2 with xp−2 p−2
t−2 , xt−1 ,and
x = , (1) p−2
max(x) xt , and so forth on others timelines. The algorithm uses
 3) Neptune attack, a denial of service attack involving a
f (xt )
SYN flood at one or more destination ports.
4) PortSweep attack, a surveillance sweep through many
xt−4 xt−3 xt−2 xt−1 xt ports to determine which services are supported on a
single host.
Timeline 5) Smurf attack, an amplified attack using an ICMP echo
Fig. 2. Single timeline representation of network traffic. reply flood.
We also measured the volume of normal traffic, which
xp−2
t−2 xp−2
t−1 xp−2
t is the aggregated traffic volume of all packets from/to the
computer center, and anomaly traffic in both datasets. The
minimum, maximum, and average volume of normal traffic
xp−1
t−2 xp−1
t−1 xp−1
t f (xpt )
are approximately 496 bit/sec., 394 kbit/sec., and 13 kbit/sec.
respectively. The average volume of Back, Ipsweep, Neptune,
xpt−2 xpt−1 xpt PortSweep, and Smurf are approximately 560 kbit/sec., 11
kbit/sec., 32 kbit/sec., 576 kbit/sec., and 8 Mbit/sec.
Timeline
Fig. 3. Multi-timeline representation of network traffic. B. Feature Extraction and Feature Scaling
We focused on nine interval-based features of network
multiple timeline from the past as input at the same interval traffic listed in Table I. The interval-based features are char-
of test instance xpt . The learning algorithm also generates the acteristics of network traffic that occur in a particular time
decision function f (xpt ) by learning from the same interval interval, the number of packets, the number of IP addresses,
as the test instance. In this figure, for the test instance xpt , and the number of ports occur, for example. In Table I, the
we show a detecting connection has been represented by the first column (f #) denotes the abbreviation of each feature; the
bold line, and learning connection by dash lines. The algorithm second column lists the feature name; and the last column
generates the decision function for xpt by learning from prior shows the description of each feature.
intervals from timelines p − 2 and p − 1, which are represented The main function of the feature extraction process is
by xp−2
t and xp−1
t , respectively. extracting interval-based features from network traffic. We
The previous study [12] shows that the multi-timeline divided one-day network traffic data into each individual
system outperforms single timeline system in most cases. How- time interval. Every time interval had an index number to
ever, this technique treats all input timelines equally, we add distinguish it. The number of time intervals or index numbers
the weighting technique in the multi timelines. To validate the in a day depended on the interval value selected. For example,
effectiveness of the weight, we conduct experiments on multi- if we set the interval value at 1 second, the number of time
timeline system to compare detection performance between intervals in a day was 86,400, so the index numbers were
weighting and non-weighting system. from 0 to 86,399. If we changed the interval value to 60
seconds, the number of time intervals in a day was 1,440.
In our experiments, we set the interval value at 10 seconds.
III. M ATERIALS AND M ETHODS
During feature scaling, we normalized the wide range of
In this section, we describe data sets, all setting variables different features into a standard range of 0 to 1. We scaled
for detector module, learning algorithms employed in our features according to
experiments, and the evaluation metric. xi,j
x i,j = , ∀i=1...f ∧ ∀j=1...m , (3)
maxj (xi,j )
A. Data Sets
where x i,j is a scaled feature, maxj (xi,j ) is the maximum
The entire network data comprise 55 days of normal traffic value of the data in the i-th feature, f is the number of the
from a relatively controlled campus network.We divided the feature, and m is the number of samples in the training data.
data into two sets; training and testing data, 39 days of normal
traffic as the training set for the classifiers and the remaining C. Weighting Technique
16 days as the test set. We selected five types of attacks
from the Lincoln Laboratory at the Massachusetts Institute of This process will guide the learning algorithm to gener-
Technology [15], and then we combined each type of attack ate a decision function which relies on particular timelines.
into the test set to create separate test set for each type of Consequently, network operators could weight on one or more
attack. timelines to bias the decision function of detection system.
Figure 4 shows the weighting process in our detector module
The description of selected attacks are as follows: during the learning process.
1) Back attack, a denial of service attack through port 80 In our implementation, we conduct experiments on the
of the Apache web server in which a client requests detector module both with and without the weighting pro-
a URL containing many backslashes. cess. There are two parameters for weighting process, namely
2) IpSweep attack, a surveillance sweep involving either weight length and weight value; the weight length is the
a port sweep or ping on multiple IP addresses. number of weighted training days and weight value is the
maximum number of replication. We also made an assumption
that recent data contains more useful information than older
ones. Figure 4 shows an example of weighting process; m is
the number of training days, and setting weight length to 6
and weight value to 3. As a result, we add a weight from day
m to m − 5 where weight 3 added to m and m − 1, weight 2
added to m − 2 and m − 3, weight 1 added to m − 4 and m − 5
sequentially. For weight value, we replicate training samples as
the number of weight value. For example, we replicate samples
3 times for weight value 3 and so on. We call this technique
as linear gradual weighting, used in a study by [16]. We could
modify by reducing the weight value exponentially called
exponential gradual weighting. In our experiments, however,
we apply only linear gradual weighting technique because it
is faster than exponential weighting.
D. Learning Algorithms
We employed three well-known algorithms of machine
learning [17]: namely the multivariate normal distribution, k-
nearest neighbor, and one-class support vector machine, to
work with our proposed system.
1) Multivariate Normal Distribution (MND): The MND is
a generalization of the Gaussian or normal probability density
function (pdf) in high dimensions [18]. In the f -dimensional
space, the pdf is given by
 
1 1 T −1
p(x) = exp − (x − μ) Σ (x − μ) ,
(2π)f /2 |Σ|
1/2 2 In the input space, a binary support vector machine decision
(4)
where μ = E[x] is the vector of mean value. Σ is the f × f
m 
covariance matrix defined as
Σ = E[(x − μ)(x − μ)T ],
where |Σ| denotes the determinant of Σ.
To classify test data, we defined an adaptive threshold
 
1 1 2
ε= exp − ρ , (6)
(2π)f /2 |Σ|
1/2 2 basis function (RBF) kernel of the form
2) k-Nearest Neighbor (KNN): KNN is an instance-based
learning for classifying data point based on closest learning
examples in the f -dimensional space [19]. In our experiment,
the nearest neighbors of data are defined by the standard Eu-
clidean distance. More precisely, let test instance x comprising
f features be described by the feature vector (x1 , x2 , ..., xf ),
where xi denotes the value of the i-th feature of data x. The
Euclidean distance between two instances x and y is defined
by 
 f

d(x, y) =  (xi − yi )2 . (8)
i=1
To classify test data, we constantly specified the parameter
k = 3, and defined the classify function of test data x as
anomaly if amount of training data nearest
f (x) = to x is less than k in the distance D,
normal otherwise.
(9)
Thanks to the feature scaling step, we can vary a constant value
D on a logarithmic scale between 10−6 and 100 for selection
of the best detection performance.
3) One Class Support Vector Machine: The support vector
machin maps training data into a high-dimensional feature
space. Therefore, we construct a separated hyperplane by
maximizing the margin or distance from the hyperplane to the
nearest training data points.
function is presented by the following formula:

f (x) = sgn αi yi .k(x, xi ) + b . (10)
i=1
(5)
Here x is the feature vector; α and y are the weights of the
support vectors; having y as a positive or negative class mark
(+1 or -1) and b is the bias; function k() is the kernel function.
Training vectors for which αi = 0 are called support vectors.
In our study, we use the Libsvm [20] tools with a radial

where ρ is a parameter to get the proportion of maximum prob- k(x, x ) = exp(−γ x − x ). (11)
ability, where smaller values of ρ produce higher probabilities. Each support vector thus becomes the center of a RBF, and
We varied ρ between 2 and 4 on a linear scale for selection of γ determines the area of influence that the support vector has
the best detection performance. We define the classify function over the data space. We varied the γ value between 10−5 and
of test data x as 104 to observe a change for the best accuracy.

anomaly if p(x) < ε, To classify testing data, we used the Svm-Predict function
f (x) = (7)
normal otherwise. from the Libsvm to determine an unknown vector sample x,
which belongs to the positive or negative class. It returns +1 or
-1 as the result of classification and provides to y the result of
w3 w3
the sum from the SVM decision formula Eq. 10. If the result
w2 w2 w2 w2 Weight
value
at a particular time interval is -1, we classify that time interval
w1 w1 w1 w1 w1 w1 as an anomalous interval. While, if the result is +1, we classify
Training that time interval as a normal interval.
day
m-5 m-4 m-3 m-2 m-1 m
E. Performance Metric
Weight length
We used F-score [21] as a single measure for evaluating the
Fig. 4. An example of weighting process with weight length = 6 and weight detection performance of our proposed technique. The F-score
value = 3. is widely used to evaluate the quality of binary classifications,
 Detection performance of MND with weighting process Detection performance of OSVM with weighting process

0.7 0.68 0.6 0.55

0.6 0.5
0.66 0.5
0.5 0.4 0.45
F-score

F-score
0.4 0.64
0.3 0.4
0.3 0.62
0.2 0.2 0.35
0.6 0.1
0.1 0.3
0 0.58 0 0.25
0.56 0.2

25 25
20 20
15 9 15 9
Weighting length (days) 10 7 Weighting length (days) 10 7
5 5
5 3 5 3
Weighting value Weighting value
0 1 0 1

Fig. 5. Detection performance of multi-timeline module with weighting Fig. 7. Detection performance of multi-timeline module with weighting
process by using MND. process by using OSVM.

Detection performance of KNN with weighting process

from a confusion matrix [22]. The acceptable F-score in our
experiments is 0.5.
0.8 0.73
0.7
0.6 0.72 IV. R ESULTS
F-score

0.5 0.71
0.4
0.3 0.7
We conduct the experiment to explore detection perfor-
0.2 0.69 mance of both the multi-timeline detector module with and
0.1
0 0.68 without weighting process. Network operators intend to give
0.67 some recent traffic more weight on recent traffic behavior. We
25
use a weighting process on the multi-timeline detector module
20 and comparing detection performance to those without the
15
7
9 weighting process.
Weighting length (days) 10
5
5
0 1
3
Weighting value We perform experiments with a linear gradual weighting
technique. First, we set the weight length and weight value as 1
for weighting process. Second, we select the best feature from
Fig. 6. Detection performance of multi-timeline module with weighting the first experiment for each type of anomaly by using first
process by using KNN. learning algorithm, MND. After that we measure F-score for
every types of anomalies and compute the average performance
of multi-timeline detector module for MND. Third, we alter
especially when the sizes of two classes are substantially the weight length from 1 day to 5, 10, 15, 20, and 25 days,
skewed. The F-score, which considers both the precision and then average detection performance for all weight length.
recall [22] to compute the score, assigns a value ranging Next, we change the weight value from 1 to 3, 5, 7, and
between 0 and 1, where 1 represents the perfect detection and 9, then follow the procedure from the first to third step
0 represents the worst detection. We measured the precision, and compute the average detection performance for all value.
recall, and F-score based on entire intervals. The precision, Finally, we switch the learning algorithm from MND to KNN
recall, and F-score are derived by Eqs. 12-14 respectively: and OSVM respectively, and plot all computed average values
on three-dimensional graphs to compare trends in detection
TP performance between these learning algorithms.
precision = , (12)
TP + FP
Performance results from this experiment on weighted
TP timeline are shown in Figures 5-7 for MND, KNN, and OSVM,
recall = , (13)
TP + FN respectively. The x-axis represents the weight values from 1
precision × recall to 9, and the y-axis indicates the weight length during the
F-score = 2 × , (14) learning process. The z-axis shows the F-score that contains
precision + recall a value between 0 and 1, where 0 represents the worst and 1
represents the best detection performance.
where TP is the number of true positives (the number of
anomalous intervals that were correctly detected), FP is the Our experimental results indicates the advantage of weight
number of false positives (the number of normal intervals process in most cases. Results of MND (Figure 5) show that
incorrectly identified as anomalous intervals), and FN is the the multi-timeline detector module with the weighting process
number of false negatives (the number of anomalous intervals produces 3.2-16.6% improvement than the module without
that were not detected). TP, FP, and FN were directly derived the weighted process. The case of KNN (Figure 6) indicates
that adding the weighting process randomly produces a small
effect between 2.7-6.4% improvement. Employing OSVM in
Figure 7, however, the trend in detection performance is quite
different from those by using MND and KNN. The detection
performance with OSVM is improved linearly for increase of
the weight values and weight lengths. The improvement of
performances by using OSVM are between 3.5-112%. From
these three figures, OSVM is a suitable algorithm for the multi-
timeline detector module with the linear gradual weighting
technique.
V. D ISCUSSION AND C ONCLUSION
Empirical results from our experiment reveal that the
detection performance of multi-timeline system with weight-
ing relies on two main factors. One factor is dependency
of two parameters in the weighted process. The weighting
technique in this experiment is timeline replication, so both
the number of training data and the number of replication as
a weighting value mainly affect the F-score. Another factor
is the learning algorithm employed to the detection module.
Our results indicate that the weighting technique strengthens
the role of OSVM. The weighting process with MND has
slightly improved for increase of the number of training data
and weighting value. The results from KNN show random and
small change when we added the weighting technique to the
multi-timeline detection module.
In summary, we proposed a weighting technique over the
multi-timeline detection system so that we can apply any
machine learning algorithms or use any traffic feature to detect
network anomalies. We conducted two experiments to examine
capabilities of the multi-timeline, detection performance over
different volumes of background traffic and weighting process.
Experimental results strongly confirm that the multi-timeline
detector module with weighting process outperform those
without weighting process, especially for OSVM.
For our future work, we intend to apply the multi-timeline
detection system to a real network environment. Although,
our experimental results show that the multi-timeline detection
system with some learning algorithms detected several types
of anomalies with promising performance, there are many
factors in network traffic that might adversely affect detection
performance of the system. Moreover, to fulfill an essential
requirement of anomaly detection in real time, we intend to
develop an automatic inspector who provide full details of
anomalies after it have been detected by the multi-timeline
detection system.
ACKNOWLEDGMENT
We gratefully acknowledge the funding from the Faculty
Members Development Scholarship Program of Bangkok Uni-
versity, Thailand. The authors would like to thank all of the
anonymous reviewers for their excellent suggestions that have
greatly improved the quality of this paper.
R EFERENCES
[1] H. F. Lipson, “Tracking and tracing cyber-attacks: Technical challenges
and global policy issues,” Software Engineering Institute, Carnegie
Mellon University, Pittsburgh, Pennsylvania, Special Report CMU/SEI-
2002-SR-009, November 2002.
[2] D. E. Denning, “An intrusion-detection model,” IEEE Trans. Softw.
Eng., vol. 13, no. 2, pp. 222–232, Feb. 1987.
[3] B. Toxen, Real World Linux Security, 2nd ed. Prentice Hall Profes-
sional Technical Reference, 2002.
[4] P. Spirakis, S. Katsikas, D. Gritzalis, F. Allegre, J. Darzentas, C. Gi-
gante, D. Karagiannis, P. Kess, H. Putkonen, and T. Spyrou, “Securenet:
A network oriented intrusion prevention and detection intelligent sys-
tem,” in Proceedings of the 10th International Conference on Informa-
tion Security, IFIP SEC94, The Netherlands, May 1994.
[5] M. Roesch, “Snort - lightweight intrusion detection for networks,” in
Proceedings of the 13th USENIX conference on System administration,
ser. LISA ’99. Berkeley, CA, USA: USENIX Association, 1999, pp.
229–238.
[6] C.-F. Tsai, Y.-F. Hsu, C.-Y. Lin, and W.-Y. Lin, “Intrusion detection by
machine learning: A review,” Expert Systems with Applications, vol. 36,
no. 10, pp. 11 994–12 000, 2009.
[7] V. Chandola, A. Banerjee, and V. Kumar, “Anomaly detection: A
survey,” ACM Comput. Surv., vol. 41, no. 3, pp. 15:1–15:58, Jul. 2009.
[8] S. Manocha and M. A. Girolami, “An empirical analysis of the prob-
abilistic k-nearest neighbour classifier,” Pattern Recogn. Lett., vol. 28,
no. 13, pp. 1818–1824, Oct. 2007.
[9] W.-H. Chen, S.-H. Hsu, and H.-P. Shen, “Application of svm and ann
for intrusion detection,” Comput. Oper. Res., vol. 32, no. 10, pp. 2617–
2634, Oct. 2005.
[10] S. Jiang, X. Song, H. Wang, J.-J. Han, and Q.-H. Li, “A clustering-
based method for unsupervised intrusion detections,” Pattern Recogn.
Lett., vol. 27, no. 7, pp. 802–810, May 2006.
[11] A. Kind, M. Stoecklin, and X. Dimitropoulos, “Histogram-based traffic
anomaly detection,” Network and Service Management, IEEE Transac-
tions on, vol. 6, no. 2, pp. 110 –121, june 2009.
[12] K. Limthong, K. Fukuda, Y. Ji, and S. Yamada, “Unsupervised learning
model for real-time anomaly detection in computer networks,” IEICE
Transactions on Information and Systems, vol. E97.D, no. 8, pp. 2084–
2094, 2014.
[13] S. Aksoy and R. M. Haralick, “Feature normalization and likelihood-
based similarity measures for image retrieval,” Pattern Recognition
Letters, vol. 22, no. 5, pp. 563 – 582, 2001, image/Video Indexing
and Retrieval.
[14] J. Grossman, M. Grossman, and R. Katz, The first systems of weighted
differential and integral calculus. Archimedes Foundation, 1980.
[15] R. Lippmann, D. Fried, I. Graf, J. Haines, K. Kendall, D. McClung,
D. Weber, S. Webster, D. Wyschogrod, R. Cunningham, and M. Ziss-
man, “Evaluating intrusion detection systems: the 1998 darpa off-line
intrusion detection evaluation,” in DARPA Information Survivability
Conference and Exposition, 2000. DISCEX ’00. Proceedings, vol. 2,
2000, pp. 12 –26 vol.2.
[16] S. Amasaki and C. Lokan, “The effects of gradual weighting on
duration-based moving windows for software effort estimation,” in
Product-Focused Software Process Improvement, ser. Lecture Notes in
Computer Science, A. Jedlitschka, P. Kuvaja, M. Kuhrmann, T. Mnnist,
J. Mnch, and M. Raatikainen, Eds. Springer International Publishing,
2014, vol. 8892, pp. 63–77.
[17] X. Wu, V. Kumar, J. Ross Quinlan, J. Ghosh, Q. Yang, H. Motoda, G. J.
McLachlan, A. Ng, B. Liu, P. S. Yu, Z.-H. Zhou, M. Steinbach, D. J.
Hand, and D. Steinberg, “Top 10 algorithms in data mining,” Knowl.
Inf. Syst., vol. 14, no. 1, pp. 1–37, Dec. 2007.
[18] S. Theodoridis and K. Koutroumbas, Pattern Recognition, Fourth Edi-
tion, 4th ed. Academic Press, 2008.
[19] T. M. Mitchell, Machine Learning, 1st ed. New York, NY, USA:
McGraw-Hill, Inc., 1997.
[20] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector
machines,” ACM Transactions on Intelligent Systems and Technology,
vol. 2, pp. 27:1–27:27, 2011.
[21] C. J. V. Rijsbergen, Information Retrieval, 2nd ed. Newton, MA, USA:
Butterworth-Heinemann, 1979.
[22] J. Davis and M. Goadrich, “The relationship between precision-recall
and roc curves,” in Proceedings of the 23rd international conference on
Machine learning, ser. ICML ’06. New York, NY, USA: ACM, 2006,
pp. 233–240.
Algorithm Performance Indexing through Parallelism
Peter Manyere Andren.L. Nel
University of Zimbabwe University of Johannesburg
Electrical Department Mechanical Department
P. O Box MP 167, Mt Pleasant P O Box 524, Auckland Park 2006
Harare, Zimbabwe Johannesburg, South Africa
Email: pmanyere@eng.uz.ac.zw Email: andren@uj.ac.za

Abstract—Image processing of Spotlight Synthetic Aperture II. MOTIVATION

Radar (SSAR) involves large amount of data to be processed and
the time required to handle such data is relatively large.
Evaluation of an algorithm in terms of processing speed (run
time) and space (data structures) is therefore a critical step in
ascertaining suitability of an algorithm for a particular
application. Major limitation in physical processing speed is
principally associated with the complex computations required to
provide solutions to large size problems (Jaja, 1992). With an
increase in massive parallelism of processors, the overall
processor operational speed has drastically improved especially
in image processing. The best choice of an algorithm on a
particular processor and for a particular task is based on the
understanding of the performance of the algorithm and
hardware at hand. In this paper, we address this problem by
providing a means to know the characteristics of a given
algorithm by evaluating its performance. Two segmentation
based SSAR algorithms, namely the 1-D and 2-D data
segmentation are evaluated and algorithm performance indexes
drawn. In both cases, image data was split into segments and
parallel processed by individual asynchronous processors. Based
on the relationship between the non-overlapping data segment
size and execution time, the performance index of an algorithm
was determined.
Keywords—Image Processing; Parallelism, SSAR Algorithm
The difference between processor speed and memory speed
presents some tremendous performance challenges. In most
cases, the processor has to idle or wait for the memory and
other elements to accomplish their tasks for the next task.
Some tedious computations such as the Fast Fourier
Transformation (FFT) require some time to execute (Gonzalez
et al, 2008). In serial version, FFT reuses the results of certain
computations. In parallel computing, the results cannot be
reused since different processing elements generate the results.
Some computations have to be performed multiple times on
different processing elements at the expense of time. A
detailed understanding of the performance of processors and
algorithms of choice is particularly important where large data
has to be handled. Hence the need to index the performance of
an algorithm in series or parallel processors, which is why this
paper was written.
III. SSAR DATA
The SSAR data used in the research was simulated from the
general expression of motion compensated, sampled and
stabilized received SSAR data (Jakawatz et al, 1996; Mahafza,
2004; Carrara et al, 1995):
N p 1 K 1

I. INTRODUCTION
Sds (n, k )  at  e 
n 0 k 0
j 1 ( n ,k )
(1)

where
Parallel computing systems can be evaluated in terms of 4 fc k 2Ro 2Ra Ts Tp
1 (n, k )   (      )( Rt  Ra ) (2)
execution time. The execution time is expressed as a function
c  Fs c c 2 2
of the size of the input, the number of processors used, their
relative computation and inter-process communication speed Ra =slant range of (antenna phase center) APC to scene center
(Grama et al, 2003). A parallel computing system comprises
of the algorithm and architecture. An evaluation of the
Rt is the distance from the location of the APC to the point
algorithm alone may not be conclusive due to some loss in target at time 𝑡.
accuracy. The assessment of performance of parallel
𝑇𝑝 =pulse duration
processors is based on speed-up and efficiency. The reason for
having multiple processors is to execute a given problem 𝑛 =pulse number
faster. One would expect a system to run twice as much when Ro =real time range
twice as many hardware resources are used. Practically, this 𝑇𝑠 =sampling period
will not be the case due to some overheads related to
parallelism which include inter-process interaction, idling and 𝐹𝑠 =sampling frequency
excess computation.

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

𝑘 =sample number 2-D Segmentation Parallel Processed Image for 36 Piont Targets

𝛾 =chirp rate -20

60

𝑐 =velocity of electromagnetic wave -15

𝑎𝑡 =radar cross-section (RCS) -10
50

Down Range(m)
-5 40

The simulated six target raw SSAR data is depicted in 0

figure 1. Noiseles SSAR Raw Data 5
30
(m)

10 20

15
Range

10
Do 20

wn 25
-30 -20 -10 0 10 20 30
Ran Cross Range (m)
Down

ge Fig. 3 Images for simulated 36 point target grid

(m)

Cross Range (m)

IV. 1-D SEGMENTATION AND PARALLEL
FORMULATION
Fig 1. SimulatedCroSSAR
s Range (m) Raw Data
The 1-D Segmentation algorithm (fig. 4), segments data in
After processing and filtering, images in figures 2 were either 1st or 2nd dimension, each into multi-row or multi-
obtained during the indexing process. 𝑚 𝑛
column segments containing a block of or consecutive
𝑁𝑝𝑐 𝑁𝑝𝑐
rows or columns respectively, from 𝑚 × 𝑛 size SSAR data,
where 𝑁𝑝𝑐 was the number of processors in the grid.

Fig 2.a 1-D Algorithm Image

2-D Segmegmentation Proceesed and Average Filtered Image
-50
60
-40

-30
50
-20 Fig 4. 1-D Segmentation Algorithm
Down Range(m)

-10 40

0
30
The segments were parallel processed into sub-images that
10
were recombined after processing to form a composite image.
20 20

30

40
10 V. PARALLEL FORMULATION BY 2-D BLOCK
50
SEGMENTATION
-60 -40 -20 0 20 40 60
Cross Range (m)

Fig 2.b 2-D Algorithm Image The choice of the shortest path to solve the problem is critical.
Two methods of achieving the fastest path of solving a
Simulation of 36-point target data that was processed by 2-D problem uses the vertices of data matrix in what is referred to
Segmentation Algorithm, generated an image in fig 3. a ‘All-Pairs Shortest Paths’ (Dijkstra, 1983), Source-
Partitioned Formulation and Source-Parallel Formulation. The
Source-Partition Formulation partitions the vertices among n
processors. Each processor computes the shortest path for all
vertices assigned to it. On the contrary, Source-Parallel
Formulation allots each vertex to a set of processors and
applies the parallel formulation of a single source algorithm in
order to solve the problem on each set of processors (Grama et
al, 2003). The major limitation of Source-Partition Parallel
Formulation is that only n processors are kept busy doing
useful work while the rest will be idle. Parallel Formulation function To , and is given by:
was therefore adopted in this paper to improve the
performance of Source-Partition Parallel Formulation (Floyd, To  N pcTN p  TS (3)
1962). Two-dimensional (2-D) block segmentation was
The speedup S , captured the relative benefit of solving a
therefore implemented in this research (fig.5), a useful tool for
problem in parallel. For n  n data segmented by edge
segmenting SSAR data for parallel processing.
detection, the efficiency and the corresponding values of the
speedup for the parallel algorithm can be determined. The
2
whole operation on a serial computer would take 9tc n
seconds where tc represents the time for each multiply-add
operation. A simple parallel algorithm to solve this problem
would require the partitioning of the image data equally across
the processing elements with each element applying the
template to its own image segment (sub-image). In this case, a
processing element allotted a vertically sliced image data of
dimensions n  (n / N pc ) , has to access a single layer of n
Fig 5. 2-D Segmentation Algorithm pixels from the processing element to the left and right. The
total time for the parallel algorithm becomes:
The 2-D technique partitioned data A( k ) into N pc blocks or
9tc n 2
segments of size (n / N pc )  (n / N pc ) . Each segment TN p   2(ts  twn) (4)
N pc
was assigned to one of the N pc processors in parallel for
with ts as the startup time for data transfer and t w being the
processing and hence the number of segments was equivalent
per word transfer time.
to number of processors. In practice, not all data is square (
n  n ) and this calls for data resizing prior to segmentation. It The value of speedup becomes:

could be assumed that the N pc processors were arranged

9tc n 2 9tc n 2
logically in a grid of size N pc  N pc so that a process on S  . (5)
TN p 9tc n 2
 2(ts  tw n)
the ith row and jth column could be referred to as Pij . The N pc
processor Pij was then allotted a segment of A( k ) with an
Parallel time speedup can also be expressed as a function of
upper left vertex or corner ((i  1)n / N pc  1,(j 1)n / N pc  1) and the problem size and number of processing elements given by:
the lower-right corner (in / N pc , j n / N pc ) . Therefore, each
n
processor had to update its part of the matrix at each kth S . (6)
iteration. The 2-D Segmentation, was based on the n
segmentation of 𝑚 × 𝑛 data in both 1st and 2nd dimensions into
 2 log N pc
N pc
𝑚 ×𝑛 two dimensional data segments for Ideally, the increase in speedup with increasing processor
𝑁𝑝𝑐 𝑁𝑝𝑐
elements is supposed to be linear, but, this is not the case due
parallel processing into sub-images recombined to form a
to some overheads. An ideal parallel system containing N pc
composite SSAR image.
processing elements can potentially deliver a speedup equal to
VI. PERFORMANCE PARAMETERS N pc . Practically, due to some inefficiencies, speed-up would
T1
A number of parameters can be used to analyze the normally be less than N pc , so that  N pc . This compels
performance of parallel systems depending on the desired TN pc
outcomes. Such parameters include execution time, total
parallel overhead function, speedup, efficiency and cost us to re-define efficiency as a fraction of the time for which a
processing element is usefully employed. In mathematical
(Grama et al, 2003). The serial run time is denoted by TS and
form, the efficiency is given by:
the parallel run time by TN p where N pc is the total number of
processors. In parallel systems, it is possible to represent the
total parallel overheads by one expression called the overhead
 S 1
E  (7)
N pc 2 N pc (ts  tw n) For this algorithm the best execution time of about 0.600
1 seconds was achieved with 8 processors in both environments
9tc n 2 and data segment sizes of 256 x 256 samples. Polar Format
In terms of the problem size and for a given number of Algorithm (PFA) under same noise free environment took
processing elements, the efficiency is alternatively expressed 3.085 seconds. Therefore 1-D Segmentation algorithm
as: improved the processing speed by an effective 5.142 times
1 speedup. Under the noisy environment, the 1-D algorithm cut
E (8) the processing time from 3.683 seconds for PFA to 0.600
2 N pc log N pc seconds, thus yielding a 6.138 times speedup. The results are
1 summarized by comparison in Table 2. The results show
n
better performance by 1-D Algorithm in both noise free and
noisy environments as compared to PFA.

VII. RESULTS Table 2. Relative Speedup Performance for 1-D Segmentation Algorithm
Best Running Time (seconds)
Algorithm
7.1 Performance Index for 1-D Algorithm
Noise-Free Noisy
Environment Environment
The 1-D Segmentation algorithm for SSAR was evaluated in
terms of its execution time in both noisy and noise-free PFA 3.085 3.683
environment. The number of parallel processors as well as
data segment sizes were varied and the mean execution times 1-D 0.600 0.600
recorded as in Table 1. Segmentation
Speedup (No. of 5.142 6.138
times faster)
Table 1. Mean Run Times for 1-D Segmentation Algorithm
Segment No. of Parallel Noise-Free Noisy 7.2 Performance Index for 2-D Algorithm
Length Processors Environment Environment
(samples) Run time (s) Run time (s)
1024 2 0.7635 0.7903 The algorithm was evaluated in terms of its data processing
512 4 0.6351 0.6517 speed both in noisy and noise-free environments. The number
of parallel processors in the grid was varied and mean
256 8 0.5972 0.5973
execution time recorded (Table 2). The mean run times show a
128 16 0.6361 0.6456
general decrease in execution time as the grid size increases to
64 32 0.7300 0.7334
some point. The minimum mean execution time is obtained
with 16 processors. A further increase in number of processor
The actual performance index of the 1-D Algorithm is tend to reduce the efficiency of the parallel system. Figure 7
presented in figure 6. It is observed that as the number of shows the performance index for 2-D algorithm.
parallel processors was increased from 2 to 8, the algorithm 1.4
2-D Segmentation Algorithm Performance Index

performance improved as seen from the execution time that

1.35
decreased drastically. However, a further increase in Noise-free Environment
Noisy Environment
processors steadily increased the execution time. 1.3
General Algorithm Performance Trend

1.25
Run Time (s)

The same trend was also observed in noisy environment 1.2

(figure 6). The behavior was generally characteristic of most 1.15

parallel systems. 1.1

1-D Segmentation Algorithm Performance Index
0.8
1.05

0.75
1
0 5 10 15 20 25 30 35 40
Parallel Processors

0.7
Run Time (s)

Fig. 7 2-D Segmentation Algorithm

Performance Index
0.65
Noise-free Environment
Noisy Environment
General Algorithm Performance Trend
0.6
The performance index suggests the algorithm’s better
0.55
performance in noise-free environment than in noisy
0 5 10 15
Parallel Processors
20 25 30 35
environment, a general common trend for most algorithms.
Fig. 6 1-D Segmentation Algorithm When compared with the PFA under the same environment,
Performance Index
we obtain results shown in Table 3. The two algorithm
execution performances were determined at their best. At best, IX. FUTURE SCOPE
the PFA took 3.085 seconds to process the SSAR data in noise
free environment. The 2-D algorithm took 1.0497 seconds Improvement of execution time can be achieved by
with 16 processors, to process the same data under the same implementation of Field Programmable Gate Arrays (FPGAs)
environment, an effectively 2.939 times faster than the PFA which are faster than ordinary digital signal processors.
under the same noise-free environment. In noisy environment, However, FPGA implementation would require external
the 2-D algorithm was 3.344 times faster than the traditional storage memory (RAM) to store the bulky received SSAR
PFA (Table 3). phase history, the segmented data and output from the range
and azimuth FFT, and the output from the DCT in the 2-D
Table 3. Mean Execution Times for 2-D Segmentation Algorithm Segmentation. These implementations are likely to improve
No. of Computer Noise-Free Noisy both algorithms’ performances to unprecedented levels.
Processors Grid Size Environment Environment
Run Time (s) Run Time (s)

4 2 2 1.1492 1.3147 REFERENCES

9 3 3 1.0778 1.1430 [1] Jakawatz, C.V. Jr., Wahl, D. E., Elchet, P. H., Chiglia, D. C., and
16 4 4 1.0497 1.1013 Thompson, P. A.(1996). Spotlight-Mode Synthetic Aperture Radar: A
Signal Processing Approach, Sandia National Laboratories, Albuquerque,
25 5 5 1.0603 1.1063 New Mexico, Boston, London, Dordrecht, Kluwer Academic Publishers.
[2] Grama, A., Gupta, A., Karypis, G. and Kumar V. (2003). Introduction to
36 6 6 1.1381 1.1635 Parallel Computing, 2nd Ed, Addison-Wesley, Pearson Education, New
York.
[3] Carrara, W. G., Goodman, R. S. and Majewski, R. M. (1995). Spotlight
Synthetic Aperture Radar Signal Processing Algorithms. Boston, London,
Table 4. Relative Speedup Performance for 2-D Algorithm Artec House.
Best Running Time (seconds) [4] Mahafza, B. R. (2004). MATLAB Simulations for Radar Systems Design.
Decibel Research, Inc, Huntsville, Alabama, Chapman &Hall/CRC.
Algorithm
[5] Chapman, S. J. (2002). MATLAB Programming for Engineers,
Noise-Free Noisy Brooks/Cole, Thompson Learning, 2nd Ed, USA.
[6] Gonzalez, R. C. and Woods, R.E. (2008). Digital Image Processing, 3rd
Environment Environment
Ed, Pearson International Edition, Prentice Hall, New Jersey.
[7] Jaja, J. (1992). An Introduction to Parallel Algorithms, Addison-Wesley
PFA 3.085 3.683 Publishing Company, New York.

2-D 1.0497 1.1013

Segmentation
Speedup (No. 2.939 3.344
of times faster)

VIII. CONCLUSION

In this paper, 1-D and 2-D Segmentation based algorithm

performances were analyzed and indexed through parallelism.
In both cases, the run times decreased drastically as the
number of parallel processors were increased to a certain
level, beyond which the processing time began to rise
irrespective of the number of processors added. The
performance index suggests the algorithms’ better
performance in noise-free environment than in noisy
environment, a general common trend for most algorithms.
The results obtained validate the fact that the system does not
run at twice as much when twice as many hardware resources
are used. This was as a result of some overheads associated
with parallelism which include inter-process interaction, idling
and excess computation.
 Designing of Licensed Number Plate Recognition
System Using Hybrid Technique from Neural
Network & Template Matching
Mr. Binay Binod Kumar Mr Mohit Bansal
Research Scholar, ECE Deptt. Assistant Professor, ECE Deptt.
Gateway Institute of Engineering Gateway Institute of Engineering &
& Technology, Sonipat, India. Technology, Sonipat, India.
bbinay2010@gmail mohit.bansal0903@gmail.com
Abstract—License Plate Recognition (LPR) is an
automatic system which is able to recognize a license number
plate. It is an effective monitoring method that uses optical
character recognition (OCR) on images to read vehicle number
plates. Genetic neural network used such type of system for
developing many challenges. The system used in highway
computerized toll aggregation and surveilling the movements
of road traffic. Licensed Number Plate Recognition System is
very important component of Intelligent Transportation
System. Using morphological & connected component element
the image of car number plate has been extraction to show the
character segmentation and at last each individual character is
matched with neural data base and that vehicle has been
recognized. Feed forward Back-propagation neural network
(FFBPNN) is selected as most powerful tool to perform the
recognition process. The main focus in this paper on three
module: license plate location (LPL), characters segmentation
and characters recognition using neural network.
The research work begins by the Pre-processing, extraction of
an image, connected component element of an image and
finally recognition of image by the feed forward Back
Propagation Neural Network.
Keywords—Template matching; Neural Network; Feed
Forward Back Propagation; License Number Plate
I. INTRODUCTION
Intelligent transportation systems (ITSs) act as an
important role in people’s routine life as their main
objective is to improve safety and mobility of transportation,
also to enhance the productivity by the use of new
technologies.[1]
In this work, a character recognition algorithm for a
Vehicle license plate recognition system is introduced to be
used as brain of an intelligent infrastructure like
computerized payment systems like highway toll collection
system, collection fee for parking, and Arial management
systems for traffic monitoring.[1] Template matching is
used which is based on Minimum Euclidean distance for
visual similar characters such as ’5’ & ’S’ and ’0’ &
’D’.[12]
Vehicle license plate recognition (VLPR)
algorithm consists of the following steps:
(i) Captured the car's number plate
(ii) Used Morphological & connected component element,
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Mr. Puneet Verma
Assistant Professor ECE Deptt.
HMR Institute of Technology &
Management, New Delhi, India
puneetverma7@gmail.com
(iii) Extraction of character image from license plate,
(iv) To extract alpha numeric characters from license plate
image,
(v) Character recognition of license number plate with the
help of neural network and identify the vehicle.
A neural network is an information-processing capability
that is influenced by the same way as the biological nervous
system, like brain. ANN is composed of large amount of
highly interconnected processing element and neurons.
Every neuron has their local memory and the output of each
neuron depends only upon input signal arriving at the
neuron.
The basic architecture of neuron layers are: input
layer, hidden layer and output layer. The information flows
from input units to output units, strictly in a feed-forward
manner. There is no feedback connection and data can be
processed using multiple layers of units. The dynamical
properties in feed forward of the network are important. In
other applications, the significant changes due to output
neurons activation, such that the transient behaviour
constitutes the output of the neural network. Some other
artificial neural network architectures are adaptive
resonance theory maps, Elman network, competitive
networks, etc. depending on the properties and necessary
condition of the application. [2]
Individual weight is multiplied with every input
neuron at the entrance. Intermediate section of neuron is a
sum function which sums all weighted inputs and bias. At
output side, transfer function is calculated using the
summation of previously weighted inputs and bias is
passed via activation function (Fig. 1.).
Fig.1. Working of an artificial neuron.[8]
 For image to text processing Optical Character Recognition
(OCR) is the most effective method.
In complex and huge images, it is required to segment the
image and then characters are extracted with the use of
character segmentation method.
Then OCR engine further process the segmented
character as shown in fig. 2. Captured image of number
plate is further processed to find the ROI i.e. region of
interest. As shown in fig 2 (b), the number plate region is
detected using an image segmentation process. Every
alphanumeric character should be clipped from the
segmented Number plate to identify the vehicle number.
Character segmentation method can be used to accomplish
the tasks. In fig 2 (c), the segmented characters are shown
[3].
Fig.2 segmentation of Image and segmentation of character (a) Original
input Image (b) Segmented binary Image (c) Image with segmented alpha
numeric characters [3]
II. PROPOSED METHODOLOGY
a) Capturing of video by Camera so as to convert it into
frames or number of images.
b) Conversion of video into images or frames. These
frames are examined so as to extract those frames
having number plate.
c) Pre-processing of images having number plate for
better quality.
 RGB to gray-scale conversion.
 To define structuring element for dilation.
 Dilate using above structuring element.
 To define another structuring element for erosion.
 Application of erosion using above structuring
element.
 Display of image after morphological operation.
 Obtaining threshold level for binarization using
ostu's method.
 Binaries the input image using above threshold
level.
 Closing the binary elements using morphology.
 Display of the image after binarisation.
d) Object Segmentation
 Segmentation of different objects in binarized
image using region props.
 Loop against every segmented object to find proper
Aspect ratio to locate number plate.
e) Number plate localization
 Find bounding box of each object.
 To select threshold values for aspect ratio.
 Threshold the value for area of object.
 Plot rectangle on selected number plate.
 To Crop an image according to bounding box.
 Display of cropped image.
 Find the gray-scale conversion level using ostu's
method.
 Binaries image using above level.
 Define constructing Element.
 To erode image using constructing element.
 Define constructing element.
 Open an images or filling holes.
 Erode image.
 Again segmentation of different objects in
binarized image using region props.
 Finding of object having maximum size.
 Finding of location points of this object.
 Display of object and plotting of bounding box
around the object.
 Cropping of image according to selected area given
by bounding box.
 Display of cropped image.
f) Character Segmentation
 Acquire threshold level for binarization using ostu's
method.
 Binaries the image using above level.
 Define constructing Element.
 To erode image using constructing element.
 Define constructing Element.
 Open an images or filling holes.
 Erode image.
 Display of erode image.
 Object segmentation using region-props.
 Calculation of aspect ratio.
 Resize cropped image for exact matching.
 Filter the bounding box as per aspect ratio.
 Threshold the bounding box by area and aspect
ratio.
 Plotting of boundaries on each character so as to
segment it.
g) Taking out of matrix of each character so as to
compare it with that of data base characters.
h) Creation and training of feed forward back propagation
neural network.
 Declaration of loop according to number of
training images in data base.
 Calling of image files one by one.
 Reading of images.
 Conversion of images into gray scale image.
 Finding of darkest portion of image i.e. Black color
pixels.
  Conversion of darkest portion into a standard size
i.e. (16 x 16).
 Reversing of contrast.
 Accumulation of all template matrixes into one
variable.
 Accumulation of location address of each character
under one variable.
 Creation of feed forward and back propagation
neural network vector according to the size of
template matrix.
 Declaration of training parameters of neural
network i.e. Number of epochs, goal etc.
 Training of neural network using defined array of
matrices having templates and matrix having
location of templates.
 Writing and saving of trained neural network.
i) Recognition of character using trained feed forward
back propagation neural network.
 Calling and reading of segmented characters of
number plate.
 Conversion of images into gray scale image.
 Display of segmented character.
 Finding of darkest portion of image i.e. Black
colour pixels.
 Conversion of darkest portion into a standard size
i.e. 16 x 16.
 Reversing of contrast.
 Accumulation of all template matrixes into one
variable.
 Simulation of neural network and processed matrix
of segmented characters to select the location of
characters.
 Display of character according to its location.
j) Display of matched characters in command window.
k) Classification
 Calculate and print correct classification.
 Calculate and print incorrect classification.
INPUTTING OF ORIGINAL IMAGE
MORPHOLOGICAL OPERATION WITH INPUT
IMAGE
PLATE SEGMENTATION USING CONNECTED
COMPONENT ANALYSIS
CHARACTER SEGMENTATION USING
CONNECTED COMPONENT ANALYSIS
CHARACTER RECOGNITION USING NEURAL
NETWORK
TEMPLATE MATCHING
OUTPUT
Fig.3. Designing of licensed number plate recognition system.
III. MORPHOLOGICAL OPERATION WITH
INPUT IMAGE.
Morphological operators often take a binary image as
input and combine them using a set of operators. Image
restoration reconstructs the damaged regions of an image by
description of region shape, boundaries, skeletons, and the
convex hull are extracted by Morphological image
processing.[11] It can be used for filtering, thinning and
pruning on sets of pixels. Erosion and dilation are the main
morphological operations where the image growing and
shrinking such as small holes are filled and disjoint objects
are connected. These types of morphological operations are
very much effective to extract the exact objects. [4]
IV. PLATE SEGMENTATION USING
CONNECTED COMPONENT ANALYSIS
First the captured image are converted into digital form
then the image is converted into gray level image and lastly
it uses pre processing top-hat and bot-hat techniques which
improve overall contrast of an image.[5][10]
Connected components are generally a group of
neighbouring pixels in contact with each other. By labelling
the connected components each region may be identified
and select each region with a number that helps in their
recognition. The region areas of these components are
calculated and the ratio technique is utilized which defines
the aspect ratio. The required segmentation of characters is
done with the help of OCR. The major concept of connected
components analysis is utilized since every alpha-numeric
character on vehicle license number plate is just a connected
component with a combination of neighbour pixels in
contact with one another. Every neighbouring pixel is
labelled with a value depending on the module to which it
was assigned. The image is changed from colour or RGB to
binary image, connected components are labelled with a
particular number and their coordinates are taken out. Using
these coordinates, each connected component in the binary
image was extracted.[6]
Segmenting the Alphabets and the number from the
license plate we carry out the steps they are cropping of
number plate, plate normalization and quantization of plate.
The constructing fundamentals of dilations are pixel
horizontal or vertical lines [7]
V. CHARACTER SEGMENTATION USING
CONNECTED COMPONENT ANALYSIS
CCA is a technique based on pixels connectivity of
binary image components. It analyzed to filter out unwanted
region from an image. Histogram is a statistical
representation of the distribution of gray level pixel values
on a given range of gray image using histogram
representation. The change between the neighbouring pixels,
mainly at the corner of characters or number will be very
high. Region are also discarded on the basis of dynamic
threshold and find out the region with high probability of
licence plate image.[1] Lines Function is used to divide text
and crops black letter with white background . Cropped
images are resized and stored in an array. [5][9]
VI. TEMPLATE MATCHING
This technique is one of the effective image
processing techniques for comparing the small parts of an
image with original image which is based on strong
features. It is widely used in to detect edges in images.
There are two main approaches for template matching
A. Feature based approach
It is based on strong features matching of an image with
large resolution. This approach does not consider the entire
region of an image for matching purpose
B. Template based approach
It is based on without strong features matching of an
image with multi-resolution. This approach does consider
the entire region of an image for matching purpose. it is
more effective and reduces the number of sampling point
VII. CHARACTER RECOGNITION USING
FEED FORWARD BACK PROPAGATION
NEURAL NETWORK.
Character recognition is the main part of this
system for the identification of character. Here the
traditional optical character recognition is replaced by
feed forward, Back Propagation neural network for a
given system. To adjust the connection weights BPLAs
use the gradient-decent search method. In feed-forward
neural network information moves only in one forward
direction. The direction is from the input layer via the
hidden layer, to the output layer. Loops are not present
in this network. The limitation of single layer
perceptrons has been overcome by applying multilayer
feed-forward network with one or more hidden layers.
It can be trained using back propagation algorithms.
This increases the overall recognition rate due to multi
layer feed forward network. Randomly, weights are
chosen, the needful corrections can be computed using
back propagation algorithm. This algorithm doesn’t
works when the value of error function becomes
negligible small.
In this paper work a scheme is presented for the
automatic number plate recognition. All the simulation has
been performed on MATLAB R2013a using generalized
MATLAB toolbox and image processing toolbox. The
elements such as noise or environmental changes play an
intruding role in license plate verification. These factors can
make the extracted region of the image incomplete. The
recovery stage is applied to reconstruct the license plate
image before the recognition step is taken. A typical
algorithm used for the recovery process is vertical intensity
projection, by which the average width and height of
alphanumeric characters in the segment are measured, and
according to this information, the whole plate is recovered.
A data set of 7 number plate images has been taken for
evaluation of performance of proposed algorithm as shown
in table1. Although, there are many steps from inputting of
image to number extraction. These steps are shown by
figures below. Fig.4. shows the original input test image.
Fig.5. is the snapshot of number plate after morphological
operations, binarization and thresholding. Fig.6. is the
snapshot of number plate with boundary on selected region
for segmentation. Fig.7. shows segmented alphanumeric and
characters present on the number plate. After this process,
segmented numbers and characters will be tested for
recognition purpose. Table1 is showing number plates with
their respective segmented characters and numbers. Also,
final message box containing a string comprised of an
alphanumeric number.
Fig.4. Capturing of input image.
Fig.5. Application of morphological operatin and Conversion into
gray scale image.

VIII. RESULT AND DISCUSSION

Fig.6. Segmented of image using connected component analysis.
Simulation conditions: The simulation conditions
considered during the simulation of the presented scheme
are as follows.

i). Input image must have a recognized format i.e.

(.jpg, .tif, .png, .gif etc).
ii). Input image must not be captured totally in front of Fig.7. Extracting of characters from number plate using connected
camera i.e. not at a great angle to the camera. component analysis.
iii). The number plate must not be captured in rotated
direction.
 Table1. Data base of 6 original captured input image with segmented and recognized characters.
IX. CONCLUSION & FUTURE SCOPE
A licensed number plate recognition system based on
neural network and template matching has been proposed in
this paper. It can be concluded from the experimental result
that the proposed system is 100% efficient in segmentation
as well as recognition of characters. A combination of
neural network and template matching is performing well as
compared to both of them alone. The future scope of this
system can be used for high security recognition of number
plates.
It can be redesign for multinational car licensed plate with
two-rows existing on it. The research work is done on the
stationary vehicles with white background of licensed plate.
Further work can be done on image of moving vehicles.
Also, recognition of colored background of number plates
can be considered in future. Number plate with two-rows
will also be a typical problem to sort in future.
REFERENCE
[1] Harpreet Kaur and Naresh Kumar Garg, “Number Plate Recognition
Using Neural Network Classifier and KMEAN,” International Journal
of Advanced research in Computer Science and Software Engineering,
vol 4, Issue 8, pp.429-434, August 2014.
[2] Ajith Abraham, Artificial Neural Networks, Oklahoma State University,
Stillwater, OK, USA, pp. 900-908, (2005).
[3] Chirag Patel, Atul Patel and Dipti Shah “A Review of Character
Segmentation Methods,” International Journal of Current Engineering
and Technology, Vol.3, No.5, pp. 2075-2078, December 2013.

[4] A. M. Raid, W. M. Khedr, M. A. El-dosuky and Mona Aoud “IMAGE

RESTORATION BASED ON MORPHOLOGICAL OPERATIONS,”
International Journal of Computer Science, Engineering and
Information Technology (IJCSEIT), vol. 4, No.3, pp.9- 21, June 2014.

[5] Anuja P. Nagare “License Plate Character Recognition System using

Neural Network,” International Journal of Computer Applications
(0975 – 8887), vol 25– No.10, pp.36-39, July 2011.

[6] Sarmad Majeed Malik and Rehan Hafiz “Automatic Number Plate
Recognition based on connected component analysis technique,” 2nd
International Conference on Emerging Trends in Engineering and
Technology (ICETET'2014 London (UK), pp.33-36), May 30-31,
2014.

[7] V. Karthikeyan, V. J. Vijayalakshmi and P. Jeyakumar “License Plate

Segmentation Using Connected Component Analysis,” IOSR Journal
of Electronics and Communication Engineering (IOSR-JECE), , pp.18-
24 Volume 4, Issue 5 (Jan. - Feb. 2013).

[8] Andrej Krenker, Janez Bešter and Andrej Kos, “Introduction to the
Artificial Neural Networks”.

[9] Danilo Octavio, “optical-character-recognition-ocr”, 2009.

[10] Antonio Albiol, Jose Manuel Mossi, Alberto Albiol, Valery Naranjo,
“Automatic License Plate Reading Using Mathematical Morphology”,
In proc. of Spanish Ministry of Science and Technology, 2002.

[11] R. C. Gonzales, R. E. Woods, “Digital Image Processing”, 2-nd

Edition, Prentice Hall, 2002.

[12] Xifan Shi, Weizhong Zhao, Yonghang Shen, "Automatic License Plate
Recognition System Based on Color Image Processing” ICCSA, pp.
1159-1168, 2005.
 A Classification Method to Detect if a Tweet Will be
Popular in a Very Early Stage
ZHAO Xianghui1, PENG yong1, YAO Yuangang1
1 2
China Information Technology Security Evaluation Center
Beijing, China
e-mail: zhaoxh@itsec.gov.cn
Abstract—Timely prediction of the popular tweet is of great
value in monitoring public opinion, marketing, emergency
detection, personalized recommendations and other areas. This
paper makes two improvements of predicting popularity of a
tweet in Microblog. On one hand, we proposes some dynamic
features, such as the retweet depth, retweet width, the total fans
of the retweeters, to improve the prediction accuracy. On the
other hand, we sharply shorten the time to detect whether a tweet
will be popular by putting forward a method called LR-DT
which combining the linear regression and the decision tree. We
firstly use the linear regression method to predict the dynamic
features' amount an hour later after the tweet transferred and
then combine them with some static features into the decision
tree classifier to detect the popularity of tweets. Our experiments
are based on the real data set from Sina Weibo, which is the most
popular micro blog service platform in China . The results show
that we proposed method effectively identify the popularity of a
tweet in less than 5 minutes while with little lose of accuracy.
Keywords—Microblog; retweet; popularity; prediction.
I. INTRODUCTION
Microblog is a platform for information interaction, sharing
and forwarding based on users' relationships. Because of the
originality of its content, the convenience of its broadcasting
and other features, Microblog has provided people a brand
new way of social communication. The number of its
registered users showed the explosive growth and it has
reached over 0.4 billion in China. With so many users, we can
observe public opinions, detect hot topics and analyze surveys
and perform marketing from popular microblogs. Thereby, it
is important to detect if a tweet will become hot in the early
stage.
Researchers worldwide have done few jobs at tweet-level
forwarding and then analyze if a tweet will be popular[1]. Zhu
X, Tang X and Wang X studied the facts influenced the
popularity of tweets respectively [2,3,4]. Their research
results showed that not only the content based features but
also the context features, such as the number of verified
retweeters, impacts much on the popularity of a tweet. Zaman
T R suggested that authors and the retweeters' features are of
importance at predicting whether a tweet will be tweeted [5].
However experiments conducted by Kai W [6] and Yang Z[7]
and Luo Z[8] showed that the tweets are easier to be
transferred among people sharing the common interests. Wu
Z[9] and Kong S[10]found that the similar tweets deserve the
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
WANG Xiaoyi2, ZHENG Zhan2,
International School of Software, Wuhan University
Wuhan, China
similar retweet counts. Can E F[11] suggested that whether the
content contained pictures ,tags and other nice visual features
do have some impacts on the final retweet counts. There are
also some researchers directly studied the popularity of a
tweet. For example, Hong L[12] defined the tweet popularity
problem as a classification task. Peng B[13] showed that
besides some static features, the dynamic transferring features
also impact on the tweet popularity. Gupta J P[14] proposed
an approach to recognize the human activities through gait,
and contributed the use of Model based approach for activity
recognition with the help of movement of legs only. Kong
S[15] argued the tweet popularity could be analyzed from two
aspects: tweet lifespan and the final report count. In
summarization, some researchers have studied the tweet
popularity problem. However, they didn’t pay much attention
to the dynamic features of tweets especially the transferring
characteristics. In addition, previous research focus on the
features after a tweet has been created longer than an hour. So
they could not make timely prediction of tweet popularity in a
very early stage.
In this paper, we analyzed dynamic features of tweets and
proposed a novel approach by combining linear regression and
decision tree to detect if a tweet will be popular. Contrary to
previous study, our research work aims to find a way to detect
the tweets' popularity at a very early stage. Firstly we add four
dynamic tweet features (repostDepth, repostWidth,
vUserCount, totalFollowersCount) into traditional static tweet
features. Next, by analyzing the forwarding history of tweets,
we apply feature selection and choose ten features mostly
related to the tweet's popularity by information gain measure.
Among them, we find some dynamic features especially
respostCount, repostWidth and vUserCount almost present
linear growth with time going. Finally, we combine the linear
regression and the decision tree as LR-DT to detect tweet's
popularity at a very early stage. We tested our approach on a
tweet dataset crawled from Sina weibo. It covers 14376 tweets
and 708534 retweets related to 70073 users. Experimental
results showed that our approach with new features can
achieve higher prediction accuracy than previous studied
features. More important, the detection accuracy are achieved
in just 5 minutes after tweets are created.
The contributions of this paper are concluded as two sides.
One hand, we added some dynamic features which changes
linearly over time. The propagation characteristics like
repostWidth, vUserCount, and other features have a closer
relationship with the tweet forwarding effect than the TABLE I. THREE GROUPS OF FEATURES
traditional static features such as the context characteristic. User features Tweet features Propagation
Experiments showed that they are helpful to achieve a better features
prediction accuracy. On the other hand, we proposed the LR- homePage user depth
DT method which is able to predict if a tweet will be popular in Name createTime width
a very early stage. Img url vUserCount
userCardKey content repostCount
Verified source followersCount
II. METHODOLOGY userType repostsCount
A tweet in Microblog is propagated through retweeting by Gender commentsCount
user's fans and the fans of fans. However, different tweets will followersCount likeCount
friendesCount pic
come to different propagation hot levels because of the type of statusesCount parentMid
tweet’s user or the appealing content or some other factors. favouritesCount parentUid
Without these factors, a tweet generally has its lifespan which decription
is no more than 72 hours. In this paper, we aim to predict Tags
constellation
whether a tweet will be popular during its lifespan by location
considering all possible related factors. Although previously Edu
some relevant experiments have been conducted, the proposed Job
methods generally have two common disadvantages, one is level
that the selected factors mainly focus on user’s features or
tweet’s features while ignoring dynamic propagation features, From Table 1, we see that there are about 34 features in
total. However, these features have different scale of ability for
the other is that previous method predict the tweet popularity
tweet popularity prediction and some of them might be not
after it is generated for more than one hour and can't predict informative. Therefore, we apply information gain measure to
the popularity at a very early phase. select the most informative features before prediction.
In this paper, we aim to improve the tweet popularity
prediction from the above two perspectives. Firstly, we will B. Prediction by Integration of Decision Tree and Linear
analyze various features’ changes as time going and we can Regression
observe a good set of feature combination and a good timing We apply classification algorithm on the tweet popularity
to build a tweet propagation prediction model. Next, we prediction problem: classifying a tweet as popular or non-
categorized the tweets into two classes: popular or not popular category by a training set modeling. There are many
popular. Thereby, the prediction task is transformed as a classification algorithms in data mining fields, such as Bayes,
classification problem. support vector machine - SVM, decision tree and so on. In this
A. Feature Selection paper, we choose decision tree algorithm as our classification
method because of its following advantages: 1) decision tree
We first introduce the features selected in our approach. By can produce some easy understandable classification rules. 2)it
summarizing previous research results, we selected two types has relative faster speed. 3)it achieves relative high accuracy.
of features related to tweets’ popularity: user information and Among all the decision tree algorithms, C4.5 is a classical
tweet information. representation. Thereby, we choose C4.5 as our base
User information contains almost all the user's information prediction method.
we can observe from the Microblog platform. The user here The previous studies showed that by observing tweet
not only just represents the tweet's creator but all the users’ propagation more than one hour, it is possible to extract
forward tweets. Because the tweets are generally retweeted by features and predict tweet popularity accurately. However, the
user's followers, it is apparently a user with more fans will model achieves less accuracy if a tweet created within one
have more power and result in more retweeters. hour. It poses us a challenge that how to predict tweet
In addition, we know that different content of tweets will be popularity in a very early stage.
appealing to different scale of followers. Thereby, we also In this paper, after analyzing both static features and
extract features from tweet information. dynamic features of tweets, we found that some dynamic
Besides the two groups of features, we observe the dynamic features linearly grow with time going. It gives us the chance
features during the tweets’ propagation process have important to estimate the dynamic feature values after one hour by using
impacts on tweet popularity. Especially, a dynamic topological linear regression based on values within one hour. Thereby,
graph describing a tweet’s propagation show clearly how large we proposed a prediction algorithm by integration of linear
size of the propagation changes. It is an important indicator for regression and decision tree (denoted as LR-DT). It mainly
tweet popularity. Thereby, we include the propagation features has three steps,
in our study. 1)Based on the training set with feature values at 1 hour to
The three groups of features selected in our research are construct a decision tree classification model;
summarized in Table 1.
2)Estimate quantity of dynamic features at 1 hour by using
linear regression based on values of dynamic features within n
minutes;
 3)Combine the estimated feature values at 1 hour and some To further illustrate the functionality of dynamic features,
static features together as our features, and then use decision we designed a control experiment with these groups of feature
tree to predict a tweet popularity. settings (Shown in Table 3). Base group contains user features
and tweet features showed in previous study. Total group
C. Evaluation Measures
contains all three groups of features and the selected group
As for the prediction evaluation metrics, we choose the contains only the top 10 informative features selected by
most popular classification measures in data mining information gain.
applications: Accuracy, errorRate, recall, precision and F1.
We define the number of popular tweet as P, the number of TABLE III. THREE GROUPS OF FEATURE SETTINGS
non-popular tweet as N, TP represents the number of popular Base user tweet
tweet classified as popular, TN represents the number of features features
popular tweet classified as non-popular. Similarly, FP Total user tweet Propag
represents the number of non-popular tweet classified as features features -ation
features
popular and FN represents number of non-popular tweet Repost Followers vUser verified width depth
classified as non-popular. In this way, the definition of the five Selected Count Count Count
measures are reported in Table 2, Follower Statuses Friends Content
Count Count Count Length

TABLE II. EVALUATION MEASURES In the experiments, we apply 10-fold cross validation. The
dataset is randomly partitioned into 10 subsets. Each subset
Parameter name Formula
accuracy or recognition rate
has roughly equal size. We rationality select one subset as test
set and the remaining data sets as training. A decision tree
error rate or misclassification rate classification model will be learned from a training set and
recall or true positive rate apply it on the corresponding test set. Then we can obtain the
final prediction results. They are reported in Table 4.
Precision
F1 or harmonic mean of precision TABLE IV. PREDICTION RESULTS BY 10-FOLD CROSS VALIDATION
and recall Groups accuracy errorRate recall precision F1
Base 85.03% 14.07% 86.22% 75.98% 80.63%
III. EXPERIMENTAL RESULTS Total 86.68% 13.32% 91.80% 82.00% 86.47%
Selected 86.67% 13.33% 91.79% 81.99% 86.46%
We tested our approach on a tweet dataset crawled from
Sina weibo. It covers 14376 tweets and 708534 retweets
related to 70073 users. From Table 4, we see that the total group including dynamic
For each tweet, we extract 34 features illustrated in Table 1 features achieves the best performance. The selected group
during its propagation process. Next, we apply information with only top 10 informative features almost don’t loss any
gain as a measurement to select the best 10 features showed in classification accuracy. It shows that the dynamic features of
figure 1. tweets have major effects on the tweet propagation while basic
features of tweets affect far less.
0.4 Through the experiments above, we concluded that with
0.35
0.3 dynamic and static features, decision tree method can predict
gain rate

0.25
0.2 the tweet’s popularity well. However, these experiments are
0.15
0.1 based on the features extracted after 1 hour from the tweet
0.05 creation. We aim to make the prediction at a very early stage.
0
followersCou…

Thereby, we analyze the features in a very early stage during

width
repostCount

followerCount
vUserCount

depth

friendsCount

contentLength
verified
statusesCount

its propagation. Static features won’t have any influence as

Figure.1 The top 10 informative features
From figure 1, we can see clearly that two dynamic
features, repostCount and vUserCount are two most
informative features for tweet’s popularity. Actually, the top
five features all are tweet’s propagation dynamic features
while the users’ information and tweet’s features contribute
less to the tweet popularity.
time going. While for dynamic features, the situation might
change. We draw five pictures (Shown in figure 2) to compare
the five dynamic features’ change as time going between
popular tweets and non-popular tweets. From figure 2, we see
that in both popular tweets and non-popular tweets, the
quantity of repostCount, width and vUserCount almost have
linear relationship with time while followersCount and depth
not. It confirms that by linear regression of some dynamic
features in a very early stage, we can estimate their values
after a tweet propagation for one hour.
Thereby, we inducted the following experiments by using
LR-DT as prediction method. Here, the dynamic features
quantity are estimated by using linear regression based on the
values after a tweet propagation only few minutes. In details,
we run our experiments with feature values recorded every 5
minutes during tweet propagation. We selected the values of
dynamic features after one hour propagation as base group.
The prediction results are reported in figure 3. We can easily
draw conclusion that LR-DT is able to predict tweet’s
popularity at a very early stage (even only after 5 minutes
from tweet creation) without loss much accuracy.
Figure.3 Prediction results by using LR-DT approach
IV. CONCLUSION
In this paper, we make two improvements for predicting
popularity of a tweet in Microblog by comparing with
previous study. On one hand, we proposes some new dynamic
features, such as the retweet depth, retweet width, the total
fans of the retweeters, to improve the prediction accuracy. On
the other hand, we sharply shorten the time to detect whether a
tweet will be popular by putting forward a method called LR-
DT which combining the linear regression and the decision
tree. The experiments based on the real data set extracted
from Sina Weibo show that our proposed method
effectively identify the popularity of a tweet in less than 5
minutes while with little loss of accuracy.
Our future work will focus on two aspects. Firstly, we
observe that the tweet creation time also affect the final
popularity and we will explore to build customization models
based on different time phase. Secondly, in this paper we just
classified tweets as popular and non-popular classes, in the
future, we would like to classify it into more detailed
categories according to different level of popularity.
ACKNOWLEDGMENT
This work was supported by a grant from the National High
Technology Research and Development Program of China
(863 Program) (2012AA012903). The second contribution is
equivalent to the first author.
REFERENCES
[1] Liang H, Xu N S, Lu G. Analysis of users and users’ Weibo information
in Sina Weibo. Computer Engineering and Applications, 2015,
51(7):141-148.
[2] Zhu X, Jia Y, Nie Y P, et al. Event propagation analysis on microblog
[J]. Journal of computer research and development, 2015, 52(2): 437-
444.
[3] Tang X, Quan Y, Miao Q, et al. Information Propagation with Retweet
Probability on Online Social Network[M]//Intelligent Computation in
Big Data Era. Springer Berlin Heidelberg, 2015: 326-333.
[4] Wang X, Zhang C, Wu M. Sentiment Classification Analysis of Chinese
Microblog Network[M]//Complex Networks VI. Springer International
Publishing, 2015: 123-129.
[5] Zaman T R, Herbrich R, Van Gael J, et al. Predicting information
spreading in twitter[A].Workshop on computational social science and
the wisdom of crowds[C], nips. 2010, 104(45): 17599-601.
[6] Wu K, Ji X S, Liu C X. Modeling information diffusion based on
behavior predicting in microblog[J].Application research of
computers,2013,30(6):1809-1812.
[7] Zhang Y, Lu R, Yang Q. Predicting retweeting in microblogs[J]. Journal
of Chinese information processing, 2012,26(4):109-114,121.
[8] Luo Z, Osborne M, Tang J, et al. Who will retweet me? finding
retweeters in twitter[A].Proceedings of the 36th international ACM
SIGIR conference on Research and development in information
retrieval[C]. ACM, 2013: 869-872.
[9] Wu Z, Liao J, Zhang L.PREDICTING ON RETWEETING OF HOT
TOPIC TWEETS IN MICROBLOG[A].Proceedings of 2013 5th IEEE
International Conference on Broadband Network & Multimedia
Technology[C].2013:119-123.
[10] Kong S, Ye F, Feng L. Predicting Future Retweet Counts in a
Microblog[J]. Journal of Computational Information Systems, 2014,
10(4): 1393-1404.
[11] Can E F, Oktay H, Manmatha R. Predicting retweet count using visual
cues[A].Proceedings of the 22nd ACM international conference on
Conference on information & knowledge management[C]. ACM, 2013:
1481-1484.
[12] Hong L, Dan O, Davison B D. Predicting popular messages in
twitter[A].Proceedings of the 20th international conference companion
on World wide web[C]. ACM, 2011: 57-58.
[13] Bao P, Shen H W, Huang J, et al. Popularity prediction in microblogging
network: a case study on sina weibo[A].Proceedings of the 22nd
international conference on World Wide Web companion[C].
International World Wide Web Conferences Steering Committee, 2013:
177-178.
[14] Gupta J P, Singh N, Dixit P, et al. Human activity recognition using gait
pattern[J]. International Journal of Computer Vision and Image
Processing (IJCVIP), 2013, 3(3): 31-53.
[15] Kong S, Feng L, Sun G, et al. Predicting lifespans of popular tweets in
microblog[A].Proceedings of the 35th international ACM SIGIR
conference on Research and development in information retrieval[C].
ACM, 2012: 1129-1130.
Figure. 2 Comparison of five dynamic features change as time going between popular and non-popular tweets (The left is popular, the right is non-popular)
 Load Frequency Control of a Nonlinear Two-Area
Power System
R. Ramjug-Ballgobin, S. Z. Sayed Hassen and S. Veerapen
Faculty of Engineering
University of Mauritius
Reduit, Mauritius
Abstract—In an interconnected power system, all the
generators must run at an appropriate capacity to meet the
demand in power. Loss of synchronism between the generators
and/or too much frequency fluctuations may cause protective
equipment to trip. Load frequency control (LFC) is necessary to
balance the power generation and the load, by monitoring
frequency and power changes between interconnected power
systems in tie-lines. In this paper, power systems from previous
research works are analysed for stability, and different types of
controllers are designed and validated through simulation and
compared with a Proportional-Integral-Derivative (PID) -
controlled power system. Three types of controllers are
considered, namely Fuzzy, Fuzzy-PID and Adaptive Neuro-
Fuzzy controllers. The first power system considered is a linear
identical non-reheat two-area system. However, a linear system
does not model an actual power system completely because of
neglected nonlinearities. Hence, two main sources of nonlinearity
(generation rate constraint (GRC) and governor dead band
(GDB)), which arise due to practical constraints are considered
and included in the model of the system.
Keywords—Load Frequency Control, Fuzzy, Adaptive Neuro
Fuzzy, Fuzzy-PID, Governor Deadband, Generation Rate
Constraint.
I. INTRODUCTION
An electrical power service aims at providing reliable and
uninterrupted electricity. Reliable and uninterrupted mean that
the supply must be of constant RMS voltage and frequency.
In real life, of course this is not possible and hence a tolerance
range according to the service provider is used. This range
must be within a certain norms. For example, if there is a
voltage drop of 10-15% or a frequency drop of a few hertz,
there is a risk of stalling in motor loads but usually, generators
trip on voltage or frequency before this happens. An electrical
grid involves a lot of planning and simulation and its
operation is complex. Hence, automatic control is used in the
system instead of human control because fast reaction speed
is needed.
In an electrical grid, many synchronous generators with
different voltage ratings are connected to bus terminals which
have the same frequency and phase sequence as the
generators. All generators connected in parallel must be run at
the appropriate capacity to meet the demand. If a generator
loses synchronism, there will be fluctuations in the voltage
and frequency supply. It is essential to synchronise the bus
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
with the generators throughout transmission for stable
operation. This is called power system stability.
Power system stability, also known as synchronous stability
refers to the ability of a system to return to synchronism after
any disturbance such as a sudden change in loading
conditions. Supply frequency and voltage must always be
within a certain limit to ensure safe and reliable operation of
electrical equipment and apparatus both at the consumer
premises and during transmission and distribution. Thus it is
essential to be able to monitor and keep the voltage and
frequency within limits.
After a perturbation involving a net change in power, the
system will enter a transient state which is normally
oscillatory and reflected by fluctuations in the power flow
over transmission lines. This is called the dynamic system
performance. In a tie-line connecting one group of generators
to another, these oscillations may build up and be reflected by
excessive fluctuations in power flow in the tie line. This will
cause protective equipment to trip [1].
A stable system is one in which after any perturbation, the
synchronous machines remain in synchronism at the end of a
finite transient period. Moreover, the amplitude of the
oscillations in the transient period must be kept within a
control. Besides, if there is any prolonged change, as long as
these changes are within a predefined limit, the system must
remain stable and both frequency and voltage kept constant.
To accomplish LFC, a controller is used to make the system
return to synchronism after any load change. As input, the
controller needs the error signal which is the difference
between the desired and the actual output value. The
controller will then generate a control signal which will affect
the system’s output value by amplifying it or attenuating it in
an attempt to obtain the desired output. Controllers can be
adjusted to improve transient response by decreasing
maximum overshoot and settling time and also to improve
steady state response by removing steady state error.
Electrical energy is one of the most important resources and
hence optimization in power systems is essential. LFC is thus
important and being able to design optimal but cheap
controllers is of upmost priority. There is a wide variety of
controllers designed using different methods, each offering
distinct advantages and disadvantages. There are conventional
controllers like PID and other interesting concepts like fuzzy
controllers. It is useful to compare the actual impact of these
controllers on a system and subsequently move towards
optimal control.
The aim of this paper is to design a controller to perform
LFC. At first an analysis of different systems from previous
research works was performed. The Matlab/Simulink
platform was used for all simulation results obtained. The
stability of the systems was investigated and the transient and
steady state response of the systems were simulated before
any controller was designed.
The control objective is to reduce overshoot, settling time and
steady state error. In LFC, deviation in frequency and tie line
power should return to zero in the minimum time and with
minimum overshoot after a load change [2]. Controllers were
designed and tested to achieve this. Different types of
controllers were simulated in Simulink for the system namely
the PID, Fuzzy, Neuro-Fuzzy and Fuzzy-PID to improve its
performance. Finally a comparative analysis of the results
from the different controllers was done and the most
appropriate one was chosen.
A. Fuzzy Controller
The design of a fuzzy controller involves three main
processes:
1. Fuzzification which requires mapping input data to fuzzy
sets.
2. Inference process which describes the decision making
block. In this process, fuzzy inputs are mapped on
corresponding fuzzy outputs based on their membership.
3. Defuzzification, a process where the fuzzy output from the
inference process is converted to a non-fuzzy output which
will then be the controller output [3], [4].
The fuzzy logic approach is similar to the decision making
process of a human operator who uses his experience to
process information and take appropriate measures.
1) Fuzzy logic and Membership functions
A fuzzy controller uses fuzzy logic. In a classical set theory,
an individual is either a member or not a member of the set. In
fuzzy logic, an individual can be in a set partially and partially
in another set. This is illustrated in Figure 1.
In fuzzy sets, members have a partial membership values
which show the degree to which they belong to each set. The
fuzzy sets are described by membership functions which gives
the membership, in the form of a continuous value between 0
and 1, of any individual. The membership function can be
anything but most common ones are the triangular function
shown in Figure 2 and the Gaussian membership functions
[5].
Figure 1 Fuzzy Numbers [4]
The input to a fuzzy controller is usually a crisp value. In the
process of fuzzification, the membership functions are used to
map the crisp values onto the corresponding fuzzy values with
associated partial memberships [6]. Each fuzzy set can be
described by linguistic variables such as ‘very fast’ or ‘slow’.
These linguistic variables can then be used in the construction
of fuzzy rules.
Each triangular function can be described by a
linguistic variable
Membe
rship
(Contin
uous
values
betwee
n 0 and
1) Members range
Figure 2. Membership Functions
The advantage of fuzzy logic is that it tries to imitate human
reasoning by even using linguistic variable and is hence easier
to understand for the designer [7], [8]. Fuzzy logic also allows
for uncertainty, approximations and vague, incomplete and
inconsistent description of a member, it is often more suitable
for practical problems than conventional methods. Most data
cannot be classified to be completely in one set without any
connection to another set.
2) Fuzzy Rules Construction
The fuzzy rules in the controller describe how the controller
should react under each combination of inputs [9]. Operations
in fuzzy logic rules are AND, OR, NOT, implication and
equivalence.
Using the linguistic variables, fuzzy rules can then be
constructed. The general form of the fuzzy rules is the ‘IF-
THEN’ rule. The inference process makes use of the rules to
make decisions. For example, ‘IF Very fast AND Raining
THEN low output’ where inputs ‘Very fast’ and ‘Raining’
and output ‘low output’ are linguistic variables.
3) Defuzzification
The output of the fuzzy controller will be in the form of fuzzy
linguistic variables. The defuzzification process converts
these fuzzy outputs into crisp outputs. It combines all the
possible control outputs into an average formula and creates a
crisp overall control signal.
The defuzzification method used in this dissertation for
Mamdani FIS is the center-of-gravity (centroid formula)
given by:
μ i x .x dx
u t = (1)
μ i x dx
where,
µi (u) is the aggregated membership function,
x is the output variable
There are many other methods of defuzzification [5], [10].
A. PID Controller
A proportional integral derivative (PID) controller involves
three parameters, the proportional, integral and derivative
gain.
Often not all three control actions are needed to optimally
control a system with minimal cost and hence P, I, PI and PD
controllers can also be used. PID controllers are very popular
in the industrial field because of their simplicity and low cost
while effectively improving transient and steady state
responses even when the actual working principles and
parameters of the system are unknown.
B. Fuzzy PID Controller
The Fuzzy-PID controller uses a fuzzy and a PID controller
together to improve the control action. The membership
functions and fuzzy rules are difficult to fine tune to obtain
the best response. Besides, both fuzzy and PID controllers are
not adaptive, meaning that if the set-point is changed, the
system may not be optimally controlled anymore or even
become unstable. Using a fuzzy and a PID as cascaded
controllers such that the fuzzy controller adjusts the
parameters of the PID controllers depending on the actual
response of the system results in an online and more optimal
control.
C. Neuro-fuzzy Controller
In Neuro-fuzzy controllers, a set of training data representing
the desired response of the controller is used to train the fuzzy
controller’s rules and membership functions. This is done
using the Adaptive Neuro-Fuzzy Inference system, ANFIS
function in Matlab/Simulink.
For this type of controller, the ANFIS training routine trains a
Sugeno-type fuzzy inference system using hybrid training
algorithm (a combination of backpropagation gradient descent
method and least-squares method) to identify the parameters
of the FIS.
II. METHODOLOGY
A linear two-area system and a nonlinear two-area system
considered for LFC. All systems are analysed to observe their
initial response.
After this various types of controllers are designed for the
two-area systems such as PID, Fuzzy, Fuzzy-PID and
adaptive Neuro-Fuzzy so that a comparative analysis can be
later performed. The systems are implemented in the form of
block diagrams in the s-domain and simulated. Controllers are
designed using the control system toolbox which can be used
to tune predefined controllers such as PID and Fuzzy so as to
adapt them to a particular system.
A. Systems
The stability of the system is checked by analysing the
uncontrolled system’s response to confirm if the frequency
deviation and incremental tie line power responses tend to
finite values.
1) Linear Two Area System
The two-area interconnected power system proposed by [11]
is used with a load change of 0.01 pu MW for area 1 and 0.02
pu MW for area 2. The areas are both large systems of
1000MW, 60Hz each [14].
2) Non-Linear Two Area System
Considering an uncontrolled non-linear system [12],[13], a
Generation Rate Constraint (GRC) of 10% pu MW min-1
(0.0017 % pu MW s-1) and a Governor Deadband (GBD) of
0.05 % were added to the linear two area system for each area
[15], [16].
B. Fuzzy Controller
A fuzzy controller is now designed for each area of the two-
area system. Matlab/Simulink inbuilt fuzzy block is used. To
create the FIS which defines all the parameters of the fuzzy
block, ‘fuzzy’ is typed in the Matlab command window so as
to open the FIS editor. Before starting the simulation, the FIS
are exported to the workspace and the name of the
corresponding FIS is used as parameter for the fuzzy block.
ACEi = BiΔfi + ΔPtie,I (2)
where,
ACEi is the Area Control Error,
i is the Area number,
Δfi is the frequency deviation,
Bi is the frequency bias parameter [17-19]
The variables being controlled are ACE and ACE input
variable and there will be one output variable, the control
action. The area control error is often used as the controller
input.
By using ACE as our controller input, deviations in both Δf
and ΔPtie can be reduced.
Next the membership functions are defined using the Fuzzy
Inference System (FIS) membership function. Seven
Triangular membership functions are used for both the input
and control variables. The ranges of the membership
functions are initially set according to the ranges of the
chosen input variables in the uncontrolled two-area system.
Afterwards, the linguistic variables are assigned.
The defuzzification method is taken as centroid. The rules are
designed approximately by logic and then fine-tuned by trial
and error to obtain the best control action. The weight of each
rule is kept at 1. The ranges of the membership functions are
adapted to obtain a better control action and the rules are
tuned again to improve response.
C. Fuzzy-PID Controller
The same method as in fuzzy controllers is used to design the
fuzzy rules and set the fuzzy membership functions. The
output of the fuzzy controller is connected to a PID controller.
D. Adaptive Neuro-Fuzzy Controller
The same system as for the Fuzzy controlled system is used
except that the FIS is a Sugeno-type designed using the
Adaptive Neuro Fuzzy Inference System (ANFIS) editor.
The AND method is set as product and weighted average is
used for defuzzification.
A set of training data representing the inputs and desired
controller output is obtained from the two-area system with a
PID controller. The training data set is loaded in the ANFIS
editor and an initial FIS is generated using grid partition. The
number of membership functions for the inputs is set to 7 and As can be observed from Table I, even though the same step
the output is set to be linear. The FIS is then trained for 10 load changes are applied in all systems, there are
epochs using the hybrid optimization method and tolerance 0. discrepancies when nonlinearities are added to the system.
When GRC is added, the steady state error, the settling time
III. DATA ANALYSIS AND FINDINGS and the maximum overshoot increases for ∆f and ∆Ptie in both
The systems were first analysed for a step load change of 0.01 areas. This is because GRC limits the rate of change of power
pu MW for the first area and 0.02 pu MW for the second area. generation. When both GDB and GRC are present, settling
The same step changes were initially used for the different time increases by a much greater amount due to the very
controllers to be able to compare their performances. Then, oscillatory response of system. On the other hand, steady state
different load changes were applied to the controlled systems error and maximum overshoot decrease slightly compared to
to investigate their responses under different conditions. system with GRC only. This is because GDB adds a range of
inputs in the system for which there is no reaction.
A. Uncontrolled two area system
Table I Uncontrolled Two Area System
Two-area system with GRC
Linear two-area system Two-area system with GRC
and GDB
Area 1 Area 2 Area 1 Area 2 Area 1 Area 2
ΔPtie ΔPtie/ ΔPtie ΔPtie/ ΔPtie
ΔPtie/p
Δf/Hz /pu Δf/Hz pu Δf/Hz /pu Δf/Hz pu Δf/Hz /pu Δf/Hz
u MW
MW MW MW MW MW
Steady
- 0.00 - - - 0.00 - - - 0.00 -
state -0.004
0.035 5 0.035 0.005 0.327 5 0.327 0.005 0.292 4 0.292
value
Maximum 0.00 0.01 0.01
0.044 0.052 0.006 0.331 0.330 0.010 0.365 0.368 0.010
Overshoot 6 0 0
Settling 154.
5.55 6.13 5.44 6.13 9.87 17.3 8.95 17.3 68.3 67.1 154.7
time/s 7

B. Controlled Linear Two Area System
It can be observed from Table III, that in terms of settling
time, adaptive Neuro-fuzzy controller perform best with a
decrease in settling time of over 40% for all responses. It also
has good performance in terms of maximum overshoot
reduction which is above 70%. However it has the poorest
performance in terms of steady state error. The PID controller
gave zero steady state error and greatly decreased maximum
overshoot by above 80% in all areas but had poor
performance in terms of settling time (reduction of 0.082% to
11.41%) compared to other controllers.
On the other hand, the fuzzy controller provided good
performance in terms of settling time (above 30%) but little
reduction of maximum overshoot (around 40% for frequency
deviation in both areas) and steady state error was still
present.
The Fuzzy-PID controller designed completely removed
steady state error and gave more than 80% reduction in
maximum overshoot for all responses. Besides, though not the
best, it gave good enough reduction in settling time of more
than 25% for both Δf and ΔPtie.. Different load changes are
then applied to the Fuzzy-PID system to check if it can
perform well in different conditions.
As can be observed in Figure 3, the Fuzzy-PID system
performs well for other load changes including negative load
changes.
C. Controlled Non-Linear Two Area System
As can be observed in Table IV, steady state error could not
be completely removed though reduced to very small amount.
This is because of the oscillatory nature of the system when
GDB was added. In terms of settling time, both types of
controllers offered similar improvement. However, the Fuzzy-
PID controller gave better response in terms of maximum
overshoot.
The controlled system is now tested under different load
changes. As can be seen from Figure 4, the system responds
well even with other load changes.
Table II Controller Types
Controller
Controller type
number
1 PID
2 Fuzzy with Δf and ΔPtie as inputs
3 Fuzzy with ACE and ACE as inputs
4 Fuzzy-PID
5 Adaptive Neuro-Fuzzy
Table III Controlled Linear Two Area System Table IV Controlled Non-Linear Two Area
Δ
Δ
Change in Change Change in Δ stead
maximum in steady maximu
settlin y
overshoot/ settling state m
g state
% time% error/% overshoo
time% error
t/%
1 -90 -0.082 -100 /%
2 -72.27 -3.52 -99.7 -
PID -15.14 -62.96
Area Area 99.86
3 -39.55 -45.98 -99.4
1 1 Fuzzy -
4 -90.45 -26.62 -100 -25.99 -65.13
-PID 99.82
5 -72.05 -59.52 -66.86 Δf
Δf -
PID -16.73 -62.42
1 -84.2 -11.1 -100 Area 99.96
2 -82.86 15.69 -99.7 2 Fuzzy -
-26.8 -64.86
Area -PID 99.78
3 -39.43 -43.42 -99.43
2 -
4 -82.48 -33.86 -100 PID 3.33 -83.36
98.67
5 -76.57 -62.53 -66.86 ΔPtie
Fuzzy -
2.08 -82.24
1 -93.3 -11.41 -100 -PID 99.23
2 -11.86 -21.06 0
ΔPtie 3 -89.77 -30.08 -98.39
4 -90.55 -32.39 -100
5 -96.37 -41.54 -97.94

Figure 3. Linear two area system with Fuzzy-PID controller Figure 4. Non-Linear two area system with Fuzzy-PID
response under different load changes controller response under different load changes
 IV. CONCLUSION
From simulation results obtained, it can be concluded for the
linear two area system, the Fuzzy-PID controller offered the
best overall performance with elimination of steady state
error, more than 80% reduction of maximum overshoot and
over 25% reduction in settling time. Also, it was found that by
adding GRC and GDB as non-linearities to the system leads
to larger settling times, higher maximum overshoot and
steady state errors. Moreover, upon designing controllers for
the non-linear two area system, the Fuzzy-PID controller only
gave slightly better performance than the anti-windup
controller.
Hence, it can be concluded that Fuzzy-PID controllers should
be considered and following the trend of mostly using PID
controllers in industrial environment may not always lead to
the best control action. However, it was also observed that
there is not a single controller which will give the best control
action in all situations. Instead, each controller gives different
performances when applied to the different systems under
consideration.
International Journal of Engineering and Science, 2(11), pp. 06-
13.
[14] Beaufays, F., Abdel-Magid, Y. & Widrow, B., 1994. Application
of neural networks to load-frequency control in power systems.
Neural Networks, 7(1), pp. 183-194.
[15] Panda, G., Panda, S. & Ardil, C., 2009. Automatic generation
control of interconnected power system with generation rate
constraints by hybrid neuro fuzzy approach. International Journal
of Electrical and Electronics Engineering, 3(9).
[16] Kumari, N. & Jha, A. N., 2013. Effect of generation rate
constraint on load frequency control of multi area interconnected
thermal systems. Journal of Electrical and Electronic Engineering
Research, 5(3), pp. 44-49.
[17] Elgerd, O. I., 1973. Electric Energy Systems Theory: An
Introduction. New Delhi: Tata Mc-Graw Hill Publishing
Company ltc.
[18] Ndubisi, S. N., 2010. An intelligent fuzzy logic controller applied
to multi-area load frequency control. American Journal of
Scientific and Industrial Research, 1(2), pp. 220-226.
[19] Prakash, S. & Sinha, S. K., 2011. Load frequency control of three
area interconnected hydro-thermal reheat power system using
artificial intelligence and PI controller. International Journal of
Engineering, Science and Technology, 4(1), pp. 23-37.

REFERENCES
[1] Anderson, P. M. & Fouad, A. A., 1977. Power System Control
and Stability. Ames, Lowa: The LOWA State University Press.
[2] El-Saady, G. et al., 2013. A new robust technique LFC of multi-
area power system using SMES. International Journal of Control,
Automation and Systems, 1(2).
[3] Babulu, K. & Kumar, K., 2012. Fuzzy self-adaptive PID controller
design for electric heating surface. International Journal of
Engineering Inventions, 1(5), pp. 10-21.
[4] Sharma, D., 2011. Designing and modeling fuzzy control systems.
International Journal of Computer Applications, 16(1), pp. 46-53.
[5] Chen, G. & Pham, T. T., 2001. Introduction to Fuzzy Sets, Fuzzy
Logic, and Fuzzy Control systems. New York: CRC Press.
[6] Ramesh, S. & Krishnan, A., 2010. Fuzzy rule based load
frequency control in a parallel AC-DC interconnected power
systems through HVDC link. International Journal of Computer
Applications, 1(4).
[7] Altas, I. H. & Neyens , J., 2006. A fuzzy logic load-frequency
controller for power systems. Turket, s.n.
[8] Vavilala, S. K., Srinivas, R. S. & Machavarapu, S., 2014. Load
frequency control of two area interconnected power system using
conventional and intelligent controllers. Journal of Engineering
Research and Applications, 4(1), pp. 156-160.
[9] Datta, M., Senjyu, T., Yona, A. & Funabashi, T., 2011. A fuzzy
based method for leveling output power fluctuations of
photovoltaic-diesel hybrid power system. Renewable Energy: An
International Journal, Volume 36, pp. 1693-1703.
[10] Sivanandam, S. N., Sumathi, S. & Deepa, S. N., 2007.
Introduction to Fuzzy Logic using Matlab. New York: Springer.
[11] Venkata, P. B. & Kumar, J. S., 2005-2008. Load frequency
control for a two area interconnected power system using robust
genetic algorithm controller. Journal of Theoretical and Applied
Information Technology, 4(12), pp. 1204-1212.
[12] Jang, R. J., 1993. ANFIS: adaptive-network-based fuzzy
interference system. IEEE Transactions on Systems, Man and
Cybernetics, 23(3), pp. 665-685.
[13] Loganathan, C. & Girija, K. V., 2013. Hybrid learning for
adaptive neuro fuzzy interference system. Research Inventry:
 Observer-based Control for Biomass Regulation in
Wastewater Treatment Plants
R. Ramjug-Ballgobin and H. C. S. Rugooputh K. Busawon and R. Binns
Faculty of Engineering Faculty of Engineering and Environment
University of Mauritius Northumbria University
Reduit, Mauritius Newcastle upon Tyne, NE1 8ST
United Kingdom

Abstract—In this paper, an estimation and control design In this paper, we propose an observer-based control
methodology for biomass concentration in bioreactors is to regulate the biomass concentration for a bioreactor that
presented. For this, a feedback linearising control was designed emulates the dynamics of a wastewater treatment system.
to perform biomass regulation. Since the controller depends on First, we show that such control should be bounded and
the measurements of the biomass concentration, an observer should obey some practical constraint. Consequently, the
was designed to estimate the latter. After that, an observer and
control law has to be designed such that these constraints
an observer-based controller were applied to the system.
Simulation study showed the good convergence features of the are obeyed at all times. Next, a simple observer is design
proposed observer-based controller using Matlab/Simulink. based on the structure of the system. The controller and the
observer are then combined to produce an observer-based
Keywords—bioreactor; biomass concentration; observer; control. A simulation study is carried out to show the
controller. efficacy of the proposed observer-based controller using
Matlab/Simulink. Finally, some conclusions are drawn.
I. INTRODUCTION
During the last decades, biological and biotechnical
processes have gained significant importance in industry. II. SYSTEM MODEL AND ANALYSIS
Some common applications are the production of certain A. Wastewater Treatment System
chemical compounds by microorganisms, the cultivation of The study of wastewater treatment control can be
a specific biomass for its utilization, the extraction of its reduced to studying the following simple microbial growth
metabolites and the degradation of pollutants; such as in reaction given by:
biological wastewater treatment plants. As a result,
bioreactors must use sophisticated control procedures to  X (t )   (t ) X (t )  D(t ) X (t ) (1)

ensure a satisfactory and efficient performance. The S (t )  D(t )( S in  S )  Ys  (t ) X (t )
regulation of the bioreactor is, however, a complex problem
since the exact biological models are not available in most where X is the biomass concentration; S is the substrate
cases. One way to overcome this difficulty is to use the concentration; µ, the biomass specific growth rate; D is the
mass-balance-based modelling [1] where the unknown dilution rate; Sin, the influent substrate concentration and Ys
biological components are included in the kinetics of the is the yield coefficient for substrate concentration.
bioreaction.
In this work, we focus on biological wastewater
treatment plants. The absence of a reliable sensor for the on-
line measurement of water quality parameters is a major
problem that needs to be tackled. The difficulty resides in
the fact that many of these parameters cannot be measured
by on-line sensors and the correctness of the actual
hardware sensors is either inadequate or too costly. Also, the
latter gives rise to maintenance problems which can result in
the system's malfunction. This has led to an increased
interest among researchers to focus on the problem of
observer design [2-6] for the estimation of biomass
concentration in wastewater treatment plants (see e.g. [1],
[7-10]). For instance, since wastewater treatment control
amounts to studying a simple microbial growth reaction, Figure 1 A schematic of the bioreactor
nonlinear observer-based estimators have been proposed for
on-line estimation of kinetic rates inside bioreactors [7].

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

Generally, the specific growth rate, µ, is not specifically Ks D (8)
known but the most commonly used model is the Monod Sin  0
max  D
model:
Therefore, the maximum allowable value of D that we can
 max S (2) employ is

Ks  S

where µmax is the maximum specific growth rate and KS is the

 max Sin (9)
saturation constant. Dmax    max
Before designing a controller for the above system, it is K s  Sin
important to carry out some analysis and highlight the Also, from equation (7), if µmax =D then X(t) will tend to -∞
essential properties of the above system. and there will be no steady state solution. This occurs when
the dilution rate, D, is too high and the cells cannot grow fast
tS
First note that, in steady state, we have X t0 , enough to reach steady state. The cells will just "wash out".
so that This is, in fact, called the wash out condition and generally
0   (t )  D(t ) X (t )
occurs if D>Dmax. From this, one can conclude that if one
 (3)
 has to design a controller for the above system, it is
0  D (t )(Sin  S (t ))  Ys  (t ) X (t ) necessary to make sure that D< Dmax for the result to be
This means that we have two equilibriums: i) X(t)=0 and ii) meaningful.
µ(t)=D(t).
Now, from the second equation, we have Next, it is obvious that for different expressions of growth
rate, µ(t) the steady state values of biomass concentration, X,
D(t )(S in  S (t ))  Ys  (t ) X (t )  Ys D(t ) X (t ) (4)
and substrate concentration, S, will be different. The
This implies that difficulty therefore resides in the appropriate model for
growth rate, µ(t). Since in practice the model of µ(t) is not
known in general, we shall consider the issue of model
Sin  S (t )
X (t )  (5) uncertainty in the expression of µ(t) and attempt to provide a
Ys controller in order to regulate the concentration of the
biomass and provide an estimate of the latter.
To find the value of S in steady state, it is important to III. OBSERVER AND CONTROLLER DESIGN
choose the expression for the specific growth rate µ(t). If µ
is chosen as the Monod's law, then from the 2nd equilibrium A. Control design for biomass regulation
condition, we have Consider again system (1). The main objective here is to
derive some methodologies in order to regulate the biomass
concentration, X. For this consider the biomass equation of
 max S
 D the bioreactor:
Ks  S
X  X  DX (10)
so that
The easiest solution would be to design a feedback
linearising control, D, such that:
K sD
S S  #
max D (6) X  DX  kX  v (11)
where v is an additional control and k>0 is the controller
gain. As a result, the expression of the controller is given by
By replacing the last expression in (5), we get
v
1  Ks D  Dk 
X   Sin    X X (12)
Ys   max  D  (7)
Therefore, in closed-loop, we have
X  kX  v (13)
Here the notation X and S denotes the values of X and S
in steady state. Since X is always positive, It is clear that, for k>0, the system (13) is stable. We can
employ an open-loop control strategy to find v. For this,
Ks D suppose that we wish to regulate X(t) at a constant reference
Sin  0 value Xref when t +∞. Then, when t +∞, v = k Xref.
 max  D
In the limit case, Finally, the overall control law is given by:
 where
kX ref
D( X )  k   (14)  2 1  0 
X F   2  and    .
 0    D  2 
The next step would be to find the value of the gain, k, such
that the condition D< Dmax is satisfied; as mentioned in the
previous section. Therefore, by setting One can easily check that the eigenvalues of F are -
which shows that F is stable. Now let = where
=diag(1,1/). Then, one can check that
 X 
0  D  k 1  ref     k    Dmax
 X    
Choosing,  Ys F1   (17)
  2 1 1
0 k Dmax   Ys      
  1 0  
One can see that the above controller requires a division by X
and that is not recommended in practice; especially if the
value of X is too small. Additionally, this means that X needs It is therefore clear that if we choose  large enough the
to be measured in order for D to be implemented in practice. right hand term will tend to zero and the first stable term will
This requires an estimation of X and is the purpose of the be predominant. Consequently, the above observer
next section. converges asymptotically to zero.
B. Estimation of X
We assume that the substrate concentration, S, is C. Observer-based control design for biomass regulation
measured so that an observer for system (1) can be written
In this section, we combine the controller designed in
as:
Section III, A together with the observer designed in the

 
. previous section. More precisely, the estimate of the biomass
Sˆ  Ys Xˆ  D( S in  y )  k1 S  Sˆ (15) is fed into the controller (14) so that we obtain:

 
.
Xˆ    D  Xˆ  k 2 S  Sˆ kX ref
D( Xˆ )  k   (18)
where k1 and k2 are the gains of the observer, which are Xˆ
chosen such that the overall error dynamics are stable.
Setting However, it is not all guaranteed that the overall system
would stay stable under the feedback (17). Consequently, we
1  S  Sˆ need to proceed for a stability analysis of the overall system
 2  X  Xˆ controlled via the observer (separation principle). For this
consider the closed loop system:
the error dynamics of the observer is given by: X  X  D( Xˆ ) X
1  k1 1  Ys  2  X  D( X ) X  D( X ) X  D( Xˆ ) X (19)
(16)
2  k 2 1    D  2  kX  v  ( D( X )  D( Xˆ )) X
That is,

 1    k1  Ys     1   0  Also,
         
 2    k 2 0    2     D  2 
D( X )  D( Xˆ )X  kX ref
1 1
 ˆ  X
(20)
X X
The gains of the observer are chosen as follows:
 
 kX ref  1 
k 1 2Ys   Xˆ 
k 2  2 Ys 
Now since  and X̂ are bounded, one can conclude that the
overall closed-loop system is stable.
so that,

 1    2Ys   Ys     1   0 
           
 2     Ys 
2
0    2     D  2 
 2 1   1   0 
 Ys   2      
 0    2     D  2 
 Ys F  
 IV. RESULTS AND DISCUSSION D=0.8 h-1, Sin=10 g/l, μmax= 0.9 h-1, Ks= 9 g/l, θ=5, Ys=2,
A simulation of the bioreactor model given in Figure 2 and xref=1.5, k=1
the controller are carried out by using the following
simulation parameters:

Figure 2 Illustration of system architecture

Figure 3 gives the profile of the biomass concentration, X, for The figure shows good convergence of the observer.
the original bioreactor model before the introduction of the
observer and controller.

2.5
Biomass Concentration(g/l)

1.5

0.5

0
0 5 10 15 20 25 30 35 40
Time(Days) Figure 4 Profiles of X and 𝑋

Figure 3 Original profile of X

When the observer of equation (28) is inserted in the system,
the profiles of biomass concentration, X, and estimated biomass
concentration, 𝑋, are as given in Figure 3. The initial
conditions were chosen as X
04 and X04.
Figure 5 shows the profile of X when the controller given by
equation (14) was applied. It can be seen that X converges to its
reference value, Xref, in about 4 days. The convergence can be
made faster by modifying the value of k as shown in Figure 6
which shows a convergence to time of 1 day when k is set to 5.
 3
V. CONCLUSION
In this paper, we propose an observer-based control design
strategy for the regulation of biomass concentration in
wastewater treatment plants. The proposed controller takes
Biomass Concentration(g/l)

2.5 into account the practical physical constraints in order to avoid

wash-out conditions. Simulation study has shown the good
performance of the proposed observer-based controller in
regulating the biomass concentration. As a future work, it
2 would be recommendable to design controller that does not
depends on the estimation of the biomass concentration but on
that of the substrate.

1.5 VI. REFERENCES

0 1 2 3 4 5 6 7 8 9 10
Time(Days) [1] JE Bailey, DF Ollis; Biochemical Engineering Fundamentals.
McGraw-Hill, New York (1986)
Figure 5 Profile of X under feedback linearising control [2] M. Farza, M. M’Saad, M.L. Fall, E. Pigeon, O. Gehan and K.
with Xref = 1.5 and k =1 Busawon, Continuous-discrete time observers for a class of MIMO
nonlinear systems, Accepted in IEEE Transactions on Automatic
Control, 2013.
3
[3] Y. Khaled, J-P. Barbot, D. Benmerzouk and K. Busawon, A new
type of impulsive observer for hyperchaotic systems, IFAC
Conference on Analysis and Control of Chaotic Systems, CHAOS
2012, Cancun, Mexico, 2012.
Biomass Concentration(g/l)

2.5
[4] W. Haddad, V. Chellaboina and S. Neresov, Impulsive and hybrid
dynamical systems: stability, dissipativity and control, Princeton
University Press, 2006.
[5] H. Hammouri, M, Nadri and R. Mota, Constant gain observer for
continuous-discrete time uniformly observable systems,
2
Proceedings of the 45th IEEE Conference on Decision & Control,
Manchester Grand Hyatt Hotel, San Diego, CA, USA, December
13-15, 2006.
[6] Lasson Karafyllis and Costas Kravaris, From Continuous-Time
Design to Sampled-Data Design of Observers, IEEE Transactions
1.5 on Automatic Control, Vol. 54, no. 9, 2009.
0 0.2 0.4 0.6 0.8 1
Time(Days)
1.2 1.4 1.6 1.8 2
[7] M. Farza, K. Busawon and H. Hammouri, “Simple Nonlinear
Observers for On-line Estimation of Kinetic Rates in Bioreactors”,
Automatica, Vol. 34, No. 3, pp. 301-318. (1998).
Figure 6 Profile of X under feedback linearising control [8] J-L. Gouzé and V. Lemesle, “A bounded error observer with
with Xref = 1.5 and k =5 adjustable rate for a class of bioreactor models”, In Proceedings of
the European Control Conference, ECC 2001. Porto, Portugal.
Figure 7 shows the profile of X using the observer-based (2001).
control. One can see that X goes to its target reference value, [9] S.Nunez, F. Garelli, H. De Battista, “Second-order sliding mode
Xref, as expected. The transient oscillations are due to the observer for biomass concentration and growth rate estimation in
dynamics of the observer. batch photo-bioreactors”, International Journal of Hydrogen
Energy 2014, 39: 8772-8779.
[10] D. Tingey. An observer design for state affine systems with
3
application to a bioprocess, Proceedings of Control 2004, Bath,
UK. (2004).
2.5
Biomass Concentration(g/l)

1.5

0.5
0 5 10 15 20 25 30 35 40 45 50
Time(Days)

Figure 7 Profile of X using the observer-based control

 Rough Set Theory Based Method for Building
Knowledge for the Rate of Heat Transfer on Free
Convection over a Vertical Flat Plate Embedded in a
Porous Medium
Hossam A. Nabwey
a) Department of Engineering Basic sciences, Faculty of
Engineering, Menofia University, Menofia, Egypt.
b) Department of Mathematics, Prince Sattam bin
Abdulaziz University, College of Science and
Humanity Studies, Alkharj, Saudi Arabia.
E-mail: eng_hossam21@yahoo.com
Abstract—Thermal convection and fluid flow in porous media
has gained increasing research interest in recent years due to the
presence of porous media in many engineering applications.
Rough set theory has been regarded as a powerful feasible and
effective methodology in the performance of data mining and
knowledge discovery activities. This paper introduce a method
for building knowledge for the rate of heat transfer (Local
Nusselt Number) considering the free convection flow over a
vertical flat plate in a fluid-saturated porous medium in the
presence of heat sources or sinks and with nonlinear density
temperature variation. First, Solutions for a class of coupled
nonlinear equations are obtained by using the fourth-order
Runge-Kutta method with shooting technique; second, Numerical
calculations of rate of heat transfer for different parameters such
as; variable suction/blowing, variable wall temperature exponent,
heat source/sink, nonlinear density temperature variation (NTD)
for both uniform and variable permeability cases are made and
presented in tabular form (decision table). Finally, a set of
maximally generalized decision rules is generated by using rough
sets methodology. The effectiveness of the obtained results will be
illustrated by comparing it with previously published work and
are found to be in excellent agreement. The proposed method
effectively decreases the time and complexity of process to
obtaining the rate of heat transfer.
Keywords—Data mining - Knowledge discovery - Rule
Induction - Rough sets –free convection - porous medium- variable
permeability - heat source/sink.
I. INTRODUCTION
During the past several decades the subject of thermal
convection in porous media has gained increasing research
interest [1, 2]. This interest stems from the presence of porous
media in a wide range of very important geophysical and
engineering applications. These applications include, but are
not limited to, ground water pollution, geothermal energy
extraction drying processes (wood and food products), solid
matrix heat exchangers, thermal energy storage, nuclear waste
repositories, heat pipes, separation processes in chemical
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
M. Modather M. Abdou
a)Department of Mathematics, Prince Sattam bin
Abdulaziz University, College of Science and
Humanity Studies, Alkharj, Saudi Arabia.
b) Department of Mathematics, Faculty of Science,
Aswan University, Aswan, Egypt
industry, filtration processes heat transfer enhancement
especially in high heat flux applications such as cooling of
electronic equipment, building insulation [3,4,5,6,7].
In recent years, many approaches have been proposed for
extracting hidden relationships holding among pieces of
information stored in a given database [8]. Rough set theory is
a relatively new mathematical and AI technique proposed by
Pawlak and Skowron [9, 10] to handle imprecision,
uncertainty and vagueness. As an effective method to feature
selection, rough sets can preserve the meaning of the features
It has been widely applied in many fields such as machine
learning [11,12], data mining [13], stock market analysis [14].
The main advantage of rough set theory is that it does not
require any preliminary or additional information about data-
like probability in statistics, basic probability assignment in
DS theory or the value of possibility in fuzzy set theory.
The rest of the paper is organized as follows. Section 2
discusses the basic concepts of Rough set theory, Section 3
analysis of the proposed problem, section 4 results and
discussion, and section 5 concludes the paper.
II. ROUGH SETS
In this context some essential definitions from rough set
theory that are used for extracting decision rules will be
recalled.
A. Decision Tables
Data are often presented as a table, columns represent
variables or observations are labeled by attributes while rows
represent (cases or events) are labeled by objects of interest
and entries of the table are attribute values. Formally, a
decision table is a triple S = (U, C, {D}) where U is the
universe, D  C is the decision attribute and elements of C
are called conditional attributes.
B. Indiscernibility relations
Indiscernibility is central to the rough sets and the objects
in a decision table are classified into equivalent classes which
are called concepts, this methodology is quite close to the
thought process of a human brain that conceptualizes the real-
world objects through granulation on the basis of their
similarity.
In decision system S = (U, A, V, f), every attributes subset,
an indiscernible relation IND(B) defined in the following way:
C. Reduction and Core
Consider the decision system S= (U, A, V, f), Let
b  B and B  A .
 If POSB(D) = POSB −{b}(D) then the attribute b is
redundant to B, which relatives to D, otherwise the
attribute b is indispensable.
 If IND(B) = IND(A) and POSB (D) ≠ POSB −{b} (D) ,
then B is called a reduction for information system S.

IND( B)  { ( x , y )  U a  B , a(x)  a(y) } (1)  the intersection of these reduction sets is called core
denoted as: CORE  Reducts
Once the indiscernibility in the condition concepts is found,
the equivalence classes are used to classify the objects
considering available information. So we need the concept of III. ANALYSIS
set approximation.
A. Formulation of the Problem
Consider the universe is U and identify the partition induct
by relation R and items as R(x) Then: According to the assumptions and analysis mentioned in
[6] the ordinary differential equations which describe the
problem are:
 The Lower Approximation of X: is the set contain all
objects which with certainty belong to the set X. it can

be defined as (Fig.1 (b)): f \  1      -1 e  1   0 (5)
RX  R( x) : R( x)  X  (2)
xU

 The Upper Approximation of X: is the set contain all

 1   \
 \\  
 2 
\

 f    f  0   0  (6)
objects which possibly belong to the set X. it can be
defined as (Fig.1 (c)):

R ( x ) : R ( x )  X   
And the boundary conditions are:
RX  (3)
xU
f  0  fw ,   0  1 (7)
 The Boundary Region of X: is the difference between
upper and lower approximation. it can be defined as : f \   0 ,    0 (8)
BN R  X   RX  RX (4) Where
 : is the permeability parameter
When a boundary region exists; i.e. when RX  RX  0
 : is the nonlinear density temperature (NDT) variation
then X is a Rough Set [15]
parameter
 0 : is the heat source/sink parameter

f w : is the suction/blowing parameter

 : is the range of exponent

The fourth-order Rung- Kutta method was used to solve

Fig. 1. Illustration of the lower approximation, the upper approximation and
the boundary region
equations (5) and (6) with the corresponding boundary
conditions in (7) and (8). The numerical results of the
variation of rate heat transfer of the local Nusselt number
Nu x Rax1/2   '  0  corresponding to the several values
of  ,  ,  , f w ,  are summarized in the decision Table I.
0
TABLE I. THE VALUES OF RATE OF HEAT TRANSFER IN TERMS OF THE
LOCAL
1/ 2
NUSSELT NUMBER Nu x Ra x
'

  0 FOR SELECTED VALUES
X36 -0.2 0.2 0.5 0 2 0.67649
X37 0 -0.2 0 0 2 0.66379
OF ,, , f , .
0 w X38 0 -0.2 0.2 0 2 0.81433
X39 0 -0.2 0.5 0 2 1.00721

fw 0   Nusselt
X40 0 0 0 0 2 0.49465
U  X41 0 0 0.2 0 2 0.67484
number
X42 0 0 0.5 0 2 0.89321

X1 -0.2 -0.2 0 -0.5 2 0.56544 X43 0 0.2 0 0 2 0.26944

X2 -0.2 -0.2 0.2 -0.5 2 0.68309 X44 0 0.2 0.2 0 2 0.50269

X3 -0.2 -0.2 0.5 -0.5 2 0.83361 X45 0 0.2 0.5 0 2 0.76062

X4 -0.2 0 0 -0.5 2 0.37643 X46 0.2 -0.2 0 0 2 0.72422

X5 -0.2 0 0.2 -0.5 2 0.52685 X47 0.2 -0.2 0.2 0 2 0.88464

X6 -0.2 0 0.5 -0.5 2 0.70548 X48 0.2 -0.2 0.5 0 2 1.09294

X7 -0.2 0.2 0 -0.5 2 0.10266 X49 0.2 0 0 0 2 0.5594

X8 -0.2 0.2 0.2 -0.5 2 0.32322 X50 0.2 0 0.2 0 2 0.74818

X9 -0.2 0.2 0.5 -0.5 2 0.55095 X51 0.2 0 0.5 0 2 0.98118

X10 0 -0.2 0 -0.5 2 0.622 X52 0.2 0.2 0 0 2 0.34171

X11 0 -0.2 0.2 -0.5 2 0.74915 X53 0.2 0.2 0.2 0 2 0.58082

X12 0 -0.2 0.5 -0.5 2 0.91427 X54 0.2 0.2 0.5 0 2 0.85192

X13 0 0 0 -0.5 2 0.4373 X55 -0.2 -0.2 0 0.5 2 0.64632

X14 0 0 0.2 -0.5 2 0.59573 X56 -0.2 -0.2 0.2 0.5 2 0.80938

X15 0 0 0.5 -0.5 2 0.78808 X57 -0.2 -0.2 0.5 0.5 2 1.01352

X16 0 0.2 0 -0.5 2 0.17252 X58 -0.2 0 0 0.5 2 0.48647

X17 0 0.2 0.2 -0.5 2 0.3969 X59 -0.2 0 0.2 0.5 2 0.6788

X18 0 0.2 0.5 -0.5 2 0.63646 X60 -0.2 0 0.5 0.5 2 0.90756

X19 0.2 -0.2 0 -0.5 2 0.68211 X61 -0.2 0.2 0 0.5 2 0.28286

X20 0.2 -0.2 0.2 -0.5 2 0.81969 X62 -0.2 0.2 0.2 0.5 2 0.52239

X21 0.2 -0.2 0.5 -0.5 2 1.00092 X63 -0.2 0.2 0.5 0.5 2 0.78672

X22 0.2 0 0 -0.5 2 0.50252 X64 0 -0.2 0 0.5 2 0.70397

X23 0.2 0 0.2 -0.5 2 0.66995 X65 0 -0.2 0.2 0.5 2 0.87567

X24 0.2 0 0.5 -0.5 2 0.87754 X66 0 -0.2 0.5 0.5 2 1.09347

X25 0.2 0.2 0 -0.5 2 0.24855 X67 0 0 0 0.5 2 0.54733

X26 0.2 0.2 0.2 -0.5 2 0.47765 X68 0 0 0.2 0.5 2 0.74712

X27 0.2 0.2 0.5 -0.5 2 0.73033 X69 0 0 0.5 0.5 2 0.98886

X28 -0.2 -0.2 0 0 2 0.60659 X70 0 0.2 0 0.5 2 0.34853

X29 -0.2 -0.2 0.2 0 2 0.74812 X71 0 0.2 0.2 0.5 2 0.59363

X30 -0.2 -0.2 0.5 0 2 0.92694 X72 0 0.2 0.5 0.5 2 0.86991

X31 -0.2 0 0 0 2 0.43376 X73 0.2 -0.2 0 0.5 2 0.76462

X32 -0.2 0 0.2 0 2 0.60627 X74 0.2 -0.2 0.2 0.5 2 0.94576

X33 -0.2 0 0.5 0 2 0.81135 X75 0.2 -0.2 0.5 0.5 2 1.17846

X34 -0.2 0.2 0 0 2 0.2024 X76 0.2 0 0 0.5 2 0.61167

X35 -0.2 0.2 0.2 0 2 0.43054 X77 0.2 0 0.2 0.5 2 0.81976
X78 0.2 0 0.5 0.5 2 1.0757 X120 0 -0.2 0.5 0 5 1.37089
X79 0.2 0.2 0 0.5 2 0.4187 X121 0 0 0 0 5 0.66832
X80 0.2 0.2 0.2 0.5 2 0.67007 X122 0 0 0.2 0 5 0.93906
X81 0.2 0.2 0.5 0.5 2 0.95941 X123 0 0 0.5 0 5 1.27715
X82 -0.2 -0.2 0 -0.5 5 0.66625 X124 0 0.2 0 0 5 0.48514
X83 -0.2 -0.2 0.2 -0.5 5 0.85675 X125 0 0.2 0.2 0 5 0.79953
X84 -0.2 -0.2 0.5 -0.5 5 1.10629 X126 0 0.2 0.5 0 5 1.17141
X85 -0.2 0 0 -0.5 5 0.49265 X127 0.2 -0.2 0 0 5 0.87965
X86 -0.2 0 0.2 -0.5 5 1.07773 X128 0.2 -0.2 0.2 0 5 1.13416
X87 -0.2 0 0.5 -0.5 5 0.99213 X129 0.2 -0.2 0.5 0 5 1.46507
X88 -0.2 0.2 0 -0.5 5 0.24566 X130 0.2 0 0 0 5 0.74004
X89 -0.2 0.2 0.2 -0.5 5 0.5313 X131 0.2 0 0.2 0 5 1.02071
X90 -0.2 0.2 0.5 -0.5 5 0.85536 X132 0.2 0 0.5 0 5 1.37464
X91 0 -0.2 0 -0.5 5 0.73 X133 0.2 0.2 0 0 5 0.56635
X92 0 -0.2 0.2 -0.5 5 0.93219 X134 0.2 0.2 0.2 0 5 0.88792
X93 0 -0.2 0.5 -0.5 5 1.19903 X135 0.2 0.2 0.5 0 5 1.2738
X94 0 0 0 -0.5 5 0.56308 X136 -0.2 -0.2 0 0.5 5 0.82737
X95 0 0 0.2 -0.5 5 0.79561 X137 -0.2 -0.2 0.2 0.5 5 1.09703
X96 0 0 0.5 -0.5 5 1.08951 X138 -0.2 -0.2 0.5 0.5 5 1.4375
X97 0 0.2 0 -0.5 5 0.33122 X139 -0.2 0 0 0.5 5 0.69476
X98 0 0.2 0.2 -0.5 5 0.3969 X140 -0.2 0 0.2 0.5 5 0.98995
X99 0 0.2 0.5 -0.5 5 0.96057 X141 -0.2 0 0.5 0.5 5 1.35228
X100 0.2 -0.2 0 -0.5 5 0.79654 X142 -0.2 0.2 0 0.5 5 0.53451
X101 0.2 -0.2 0.2 -0.5 5 1.01088 X143 -0.2 0.2 0.2 0.5 5 0.86655
X102 0.2 -0.2 0.5 -0.5 5 1.29574 X144 -0.2 0.2 0.5 0.5 5 1.25787
X103 0.2 0 0 -0.5 5 0.63648 X145 0 -0.2 0 0.5 5 0.89156
X104 0.2 0 0.2 -0.5 5 0.87991 X146 0 -0.2 0.2 0.5 5 1.17102
X105 0.2 0 0.5 -0.5 5 1.19088 X147 0 -0.2 0.5 0.5 5 1.52659
X106 0.2 0.2 0 -0.5 5 0.41982 X148 0 0 0 0.5 5 0.76298
X107 0.2 0.2 0.2 -0.5 5 0.71758 X149 0 0 0.2 0.5 5 1.06708
X108 0.2 0.2 0.5 -0.5 5 1.06951 X150 0 0 0.5 0.5 5 1.44387
X109 -0.2 -0.2 0 0 5 0.74896 X151 0 0.2 0 0.5 5 0.60931
X110 -0.2 -0.2 0.2 0 5 0.98185 X152 0 0.2 0.2 0.5 5 0.94847
X111 -0.2 -0.2 0.5 0 5 1.28018 X153 0 0.2 0.5 0.5 5 1.35302
X112 -0.2 0 0 0 5 0.59898 X154 0.2 -0.2 0 0.5 5 0.95782
X113 -0.2 0 0.2 0 5 0.86022 X155 0.2 -0.2 0.2 0.5 5 1.2475
X114 -0.2 0 0.5 0 5 1.1831 X156 0.2 -0.2 0.5 0.5 5 1.6188
X115 -0.2 0.2 0 0 5 0.40618 X157 0.2 0 0 0.5 5 0.83328
X116 -0.2 0.2 0.2 0 5 0.71378 X158 0.2 0 0.2 0.5 5 1.14669
X117 -0.2 0.2 0.5 0 5 1.07225 X159 0.2 0 0.5 0.5 5 1.53857
X118 0 -0.2 0 0 5 0.81313 X160 0.2 0.2 0 0.5 5 0.68609
X119 0 -0.2 0.2 0 5 1.0566 X161 0.2 0.2 0.2 0.5 5 1.03276
X162 0.2 0.2 0.5 0.5 5 1.45114 X204 0 0 0.5 0 1 0.77044
X163 -0.2 -0.2 0 -0.5 1 0.53672 X205 0 0.2 0 0 1 0.20898
X164 -0.2 -0.2 0.2 -0.5 1 0.63175 X206 0 0.2 0.2 0 1 0.41523
X165 -0.2 -0.2 0.5 -0.5 1 0.75047 X207 0 0.2 0.5 0 1 0.63254
X166 -0.2 0 0 -0.5 1 0.3447 X208 0.2 -0.2 0 0 1 0.67708
X167 -0.2 0 0.2 -0.5 1 0.47346 X209 0.2 -0.2 0.2 0 1 0.80536
X168 -0.2 0 0.5 -0.5 1 0.62074 X210 0.2 -0.2 0.5 0 1 0.96966
X169 -0.2 0.2 0 -0.5 1 0.06615 X211 0.2 0 0 0 1 0.50518
X170 -0.2 0.2 0.2 -0.5 1 0.26787 X212 0.2 0 0.2 0 1 0.66258
X171 -0.2 0.2 0.5 -0.5 1 0.46544 X213 0.2 0 0.5 0 1 0.85191
X172 0 -0.2 0 -0.5 1 0.59057 X214 0.2 0.2 0 0 1 0.27579
X173 0 -0.2 0.2 -0.5 1 0.69387 X215 0.2 0.2 0.2 0 1 0.4865
X174 0 -0.2 0.5 -0.5 1 0.82539 X216 0.2 0.2 0.5 0 1 0.71535
X175 0 0 0 -0.5 1 0.40185 X217 -0.2 -0.2 0 0.5 1 0.59282
X176 0 0 0.2 -0.5 1 0.53715 X218 -0.2 -0.2 0.2 0.5 1 0.71939
X177 0 0 0.5 -0.5 1 0.69614 X219 -0.2 -0.2 0.5 0.5 1 0.87467
X178 0 0.2 0 -0.5 1 0.13034 X220 -0.2 0 0 0.5 1 0.42594
X179 0 0.2 0.2 -0.5 1 0.3342 X221 -0.2 0 0.2 0.5 1 0.58328
X180 0 0.2 0.5 -0.5 1 0.54143 X222 -0.2 0 0.5 0.5 1 0.76407
X181 0.2 -0.2 0 -0.5 1 0.6481 X223 -0.2 0.2 0 0.5 1 0.21209
X182 0.2 -0.2 0.2 -0.5 1 0.76075 X224 -0.2 0.2 0.2 0.5 1 0.42029
X183 0.2 -0.2 0.5 -0.5 1 0.9068 X225 -0.2 0.2 0.5 0.5 1 0.63849
X184 0.2 0 0 -0.5 1 0.46362 X226 0 -0.2 0 0.5 1 0.64718
X185 0.2 0 0.2 -0.5 1 0.60659 X227 0 -0.2 0.2 0.5 1 0.78127
X186 0.2 0 0.5 -0.5 1 0.77906 X228 0 -0.2 0.5 0.5 1 0.94859
X187 0.2 0.2 0 -0.5 1 0.201 X229 0 0 0 0.5 1 0.48259
X188 0.2 0.2 0.2 -0.5 1 0.40807 X230 0 0 0.2 0.5 1 0.64604
X189 0.2 0.2 0.5 -0.5 1 0.62657 X231 0 0 0.5 0.5 1 0.83802
X190 -0.2 -0.2 0 0 1 0.56535 X232 0 0.2 0 0.5 1 0.27174
X191 -0.2 -0.2 0.2 0 1 0.67692 X233 0 0.2 0.2 0.5 1 0.48398
X192 -0.2 -0.2 0.5 0 1 0.81487 X234 0 0.2 0.5 0.5 1 0.7123
X193 -0.2 0 0 0 1 0.38726 X235 0.2 -0.2 0 0.5 1 0.70482
X194 -0.2 0 0.2 0 1 0.53104 X236 0.2 -0.2 0.2 0.5 1 0.84739
X195 -0.2 0 0.5 0 1 0.69591 X237 0.2 -0.2 0.5 0.5 1 1.02829
X196 -0.2 0.2 0 0 1 0.14792 X238 0.2 0 0 0.5 1 0.54317
X197 -0.2 0.2 0.2 0 1 0.35067 X239 0.2 0 0.2 0.5 1 0.71373
X198 -0.2 0.2 0.5 0 1 0.55796 X240 0.2 0 0.5 0.5 1 0.91845
X199 0 -0.2 0 0 1 0.61949 X241 0.2 0.2 0 0.5 1 0.33656
X200 0 -0.2 0.2 0 1 0.7389 X242 0.2 0.2 0.2 0.5 1 0.55373
X201 0 -0.2 0.5 0 1 0.88921 X243 0.2 0.2 0.5 0.5 1 0.79361
X202 0 0 0 0 1 0.44377
X203 0 0 0.2 0 1 0.59416
  An increasing of the values of the
B. The Rough Set Framework for Rule Induction parameters  ,  , f w and  enhances the rate of
The complete process to generate a set of decision rules the heat transfer.
then classified to assess their performance using rough sets
methodology are shown in fig. 2.  As the (NDT) parameter  0 increased the range
of local Nusselt number reduced.
Physically it means that as ε increases the rate of the heat
transfer increases in magnitude because the increased near
wall permeability allows the fluid to advocate heat away more
quickly than it would for   1 , there by thinning the
boundary layer and increasing the temperature gradient.

Table II shows part of the rule set extracted by using

rough set methodology which explained in figure 1. The tool
which used for generating the minimum rules is ROSETTA
which is an RST analysis toolkit.
Tables III presents a comparison of rate of heat transfer in
terms of the local Nusselt number at selected values of f w ,
 0 ,  ,  ,  between the results those reported earlier by
(EL-Kabeir et al. [4]) and the proposed methodology. The
result of the comparison shows that there is an excellent
agreement between the two methodologies which lends to
confidence in the correctness of the proposed method

Tables IV presents the values of rate of heat transfer in

Fig. 2. Complete Steps to Extract Decision Rules
IV. RESULTS AND DISCUSSION
From the decision table I it can be noted that:
terms of the local Nusselt number at different values of f w ,
 0 ,  ,  ,  which are not calculated in the previous work
of EL-Kabeir [6] and calculated with the proposed method. It
can be seen that the results behave the same behavior
mentioned above i.e. as the values of the parameters increase
there is an enhancement in the rate of the heat transfer.

TABLE II. PART OF THE GENERATED RULE SET

Rules
.....

.....

 f w  "(-Inf,-0.1)")  (  0  "(-0.1,0.1)")  
 
Rule 85 IF (   "(-Inf,0.1)")  (   "(-Inf, 0.25)")   THEN  Nu={0.49265}

 
 (  "(3.5,Inf)" 
 
  f w  "(-0.1,0.1)")  (  0  "(-Inf,-0.1)")  
 
Rule 118 IF (   "(-Inf,0.1)")  (   "(-0.25,0.25)")   THEN  Nu={0.81313}

 
 (  "(3.5,Inf)" 
 

 f w  "(0.1,Inf)")  (  0  "(-0.1,0.1)")  
 
Rule 212 IF (   "(0.1,0.35)")  (   "(-0.25,0.25)")   THEN  Nu={0.66258}

 
 (  "(-Inf,1.5)" 
 

 f w  "(0.1,Inf)")  (  0  "(0.1,Inf)")  
 
Rule 243 IF  (   "(0.35,Inf)")  (   "(0.25,Inf)")   THEN  Nu={0.79361}
 
 (  "(-Inf,1.5)" 
 

TABLE III. THE VALUES OF RATE OF HEAT TRANSFER IN TERMS OF THE LOCAL NUSSELT NUMBER COMPARED WITH THE RESULT OF (EL-KABEIR ET AL. [4])
METHOD AND THE PROPOSED METHODOLOGY

fw 0 
U   Nusselt Number
(EL-Kabeir et al. [4] )
Nusselt Number
( proposed methodology )

X1 -0.2 -0.2 0 -0.5 2 0.56544 0.56544

X2 -0.2 -0.2 0.2 -0.5 1 0.63175 0.63175

X3 -0.2 -0.2 0.5 -0.5 5 0.85675 1.10629

X4 0 0.2 0.2 -0.5 2 0.39690 0.3969

X5 0 0.2 0.5 -0.5 1 0.54143 0.54143

X6 0 0.2 0.5 -0.5 1 0.54143 0.54143

 REFERENCES
TABLE IV. THE VALUES OF RATE OF HEAT TRANSFER IN TERMS OF THE
[1] K. Javaherdeh, Mehrzad Mirzaei Nejad, M. Moslemi, “Natural
LOCAL NUSSELT NUMBER FOR DIFFERENT VALUES OF f ,  , , ,  convection heat and mass transfer in MHD fluid flow past a moving
w 0
vertical plate with variable surface temperature and concentration in a
porous medium” , Engineering Science and Technology, an International
Journal, Volume 18, Issue 3,, Pages 423-431, September 2015
[2] K Asma Khalid, Ilyas Khan, Arshad Khan, Sharidan Shafie , “Unsteady
U
fw
0    Nusselt Number MHD free convection flow of Casson fluid past over an oscillating
( present work ) vertical plate embedded in a porous medium” , Engineering Science and
Technology, an International Journal, Volume 18, Issue 3,, Pages 309-
317, September 2015
[3] D. A. Nield and A. Bejan, Convection in Porous Media, 4th edition,
X1 0.1 -0.2 0 -0.5 2 0.68211 Springer, New York, 2013.
[4] K. Vafai (Ed.), Handbook of Porous Media, Marcel Dekker, Marcel,
X2 0.3 -0.2 0.2 -0.5 2 0.81969 2000.
[5] I. Pop, D.B. Ingham, “Convective heat transfer: mathematical and
computational modelling of viscous fluids and porous media” ,
X3 0.6 -0.2 0.2 -0.5 2 0.81969 Pergamon, Oxford, 2001.
[6] S. M. M. EL-Kabeir and A.M. Rashad, "Influence of variable
X4 -0.6 -0.2 0.2 -0.5 2 0.68309 permeability on free convection over vertical flat plate embedded in a
porous medium", 2006
X5 -0.1 -0.2 0.2 -0.5 2 0.74915 [7] Hossam A. nabwey, S. M. M. EL-Kabeir and A.M. Rashad, “Lie group
analysis of effects of radiation and chemical reaction on heat and mass
transfer by unsteady slip flow from a Non-Isothermal stretching sheet
X6 -0.2 -0.3 0 -0.5 2 0.56544 Immersed in a porous medium”. Journal of Computational and
Theoretical Nanoscience, vol. 12, no. 10, pp. 1–7, Jan. 2015.
X7 -0.2 -0.4 0.2 -0.5 2 0.68309 [8] Hossam Abd Elmaksoud Mohamed, “An algorithm for mining decision
rules based on decision network and rough set theory”. T.-h. Kim et al.
(Eds.): UCMA 2011, Part I, CCIS 150, pp. 44–54, 2011.
X8 -0.2 0.35 0.5 -0.5 3 0.55095 [9] U. Fayyad, K. Irani , “ Multiinterval discretization of continuous-valued
attributes for classification learning “. Proc. of the 12th International
X9 0 0.6 0.2 -0.5 2 0.3969 Joint Conference on Artificial Intelligence ,1022–27, 1993
[10] U. Fayyad, On the Induction of Decision Trees for Multiple Concept
Learning. Ph.D. thesis, University of Michigan (1991).
X10 -0.2 -0.2 0 -0.5 4 0.66625
[11] Hossam A. Nabwey, “A Probabilistic Rough Set Approach to Rule
Discovery”. International Journal of Advanced Science and Technology
X11 -0.2 -0.2 0.2 -0.5 6 0.85675 Vol. 30, May, 2011
[12] Hossam A. Nabwey, Mahdy S. El-Paoumy, “An integrated
X12 -0.2 -0.2 0.3 0.6 3 0.80938 methodology of rough set theory and grey system for extracting decision
rules”, International Journal of Advances in Applied Sciences (IJAAS),
Vol. 1, , 2013
X13 -0.2 -0.2 0 -0.8 2 0.56544 [13] Y. Wang, “Mining stock price using fuzzy rough set system,” Expert
Syst. Appl., vol. 24, no. 1, pp. 13–23, Jan. 2003.
X14 -0.2 -0.2 0.2 -0.8 1 0.63175 [14] Ching-Lai Hor, , Peter A. Crossley and Simon J. Watson, “Building
knowledge for substation-based decision support using rough sets”,
IEEE Transactions On Power Delivery, Vol. 22, No. 3, July 2007
[15] Fabrizio Gizzi, Nicola Masini, Maria Rosaria Potenza, Cinzia Zotta,
Lucia Tilio, Maria Danese, Beniamino Murgante, “Assessing
macroseismic data reliability through rough set theory: The Case of
V. CONCLUSIONS Rapolla (Basilicata, Southern Italy), Lecture Notes in Computer
This paper suggests the use of rough set theory to process ScienceVolume 6016, 2010, pp 320-330, 2010.
and extract rules for rate of heat transfer (Local Nusselt
Number) on free convection flow over a vertical flat plate
embedded in a fluid saturated porous medium in the presence
of heat sources or sinks and with nonlinear density
temperature variation. The results of this study were in good
agreement with previous studies of uniform permeability case.
In the variable permeability case, the results show that the rate
of heat transfer is much higher than in uniform permeability
case. The technique has been simplified logic-based rules,
reduces the time and resources required to building
knowledge. The comparison with calculated results show
promise for practical applications.
 Full-Half Duplex Mixed Strategy for Capacity
Maximisation in MIMO Interference Channels
Emmanuel Migabo, Thomas Olwal and Thato Phate
Department of Electrical Engineering and F’SATI
Tshwane University of Technology
Pretoria, South Africa, 0001
Email: {migabo.emmanuel, thomas.olwal and thatophate}@gmail.com
Abstract—Existing wireless channel Multiple-Input Multiple-
Output (MIMO) interference models demonstrate that full duplex
communication in a k-link MIMO system suffers from both self-
interference and inter-user interference [1]. However, the fact
that nodes exchange information simultaneously (Transmission
and reception activities) in a bi-directional manner on each
of the k considered links; comes with considerable potential
gain in terms of channel capacity. On the other hand, half-
duplex communication between the k MIMO links comes with
the advantage that it does not suffer from self interference.
Half duplex MIMO communication however comes with the
drawback that it only achieves single link channel capacity. It also
can still result in inter-user interference. This paper proposes a
mixed strategy which takes advantage of the Full-duplex MIMO
capacity gain while reducing self-interference by tapping into the
Half-duplex potential and also reducing inter-user interference by
interleaving Full-duplex and Half-duplex communications on the
k MIMO links. This is done by designing optimum transmitter
and receive filters for Weighted Sum Rate (WSR) maximisation
using the Rosen’s gradient projection method. After proper
modeling and MATLAB simulation, the obtained results show
that the Full-Half duplex Mixed strategy enhances the overall
channel capacity as compared to the k-link MIMO Full-duplex
WSR maximisation approach.
Index Terms—Full-duplex, Half-duplex, Multiple-Input
Multiple-Output (MIMO), self-interference, inter-user
interference, mixed strategy, filter design, Weighted Sum
Rate (WSR).
I. I NTRODUCTION
We live in an age of globalization. The world is becoming
a small village. This has triggered the need to interconnect
everything through what has started to be called Internet
of Things (IoT) [2]. The proliferation of such applications
has created a serious demand in terms of high data rates
which has therefore ignited the quest for achieving more and
more spectral efficiency. Research works have been trying to
investigate ways to push the boundaries of the wireless channel
utilization by proposing so many possible solutions in order
to use as much capacity of the wireless channel as possible.
Multiple Input Multiple Output (MIMO) technologies as well
as opportunistic communication such as cognitive radio have
demonstrated to be very powerful technologies that help
improve the efficiency of the wireless spectrum utilization.
The Radio spectrum is becoming increasingly scarce as
more and more devices go wireless. Opportunistic commu-
nication technologies such as cognitive radios such as MIMO
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
technologies, are promising technologies in wireless commu-
nications for enhancing the spectrum efficiency of the channel.
It is also important to note that the ideas of MIMO and
opportunistic communication are not exclusive. There exists
wireless communication systems which consist of both a
cognitive radio and MIMO topology with the aim to achieve
higher spectral efficiency [5].
The MIMO technology provides spatial multiplexing by
using multiple transmit and receive antennas instead of a single
pair. It has therefore proven to be capable of tremendously
improve communication date rates capabilities in more than
one current applications ranging from sensor networks to
cellular systems. However, the use of multiple antennas comes
with its own challenges.
The Interference, between the multiple transmissions re-
mains one of the main challenges faced by MIMO technolo-
gies. The advent of MIMO technologies has shifted the focus
of the wireless channel modeling from a fading channel to
an interference channel. The latter approach pays much more
attention to the multiplicative channel impairments such as in-
terference while the fading channel modeling mainly considers
the additive effect of noise on the transmitted signal.Therefore,
handling interference at both the transmitter and the receiver
design level has become the purpose of many research works
[3]. The purpose of this work is to tap into the potential of
both the Full-duplex and the Half-duplex models as a possible
way to enhance spectral-efficiency by interleaving them on a
single k links MIMO channel.
II. R ELATED WORK
Most existing MIMO technologies have been so far fully
Half-duplex [4] or fully Fully-duplex [1]. Most of them
have been either Time-Division Duplex (TDD) or Frequency-
Division Duplex (FDD). This means that despite the increase
in channel capacity provided by MIMO technologies, most of
them have either simply provided possibility for only using
half capacity of a link as they have not been able to allow
transmission and reception to occur concurrently or they have
evaluated the impact of using a Full-duplex communication
for all the k links. The channel model is depicted in figure 1
it clearly shows the different considered channel impairments
at each link.
 data rates in k-links MIMO as a result of the limited
ሺ૛૚ሻ ni (2) Dynamic Range (DR) of the input circuitry.
ඥ࣋࢏࢏ ࡴ࢏࢏ (b)
• PT and Pi are the total(sum) available power of the
Node i (1) + Node i (2)
ඥࣁ࢏࢏ ࡴ࢏࢏
ሺ૚૚ሻ system and individual available powers for all nodes b
ሺ૛૛ሻ
ඥࣁ࢏࢏ ࡴ࢏࢏ (Ni, Mi) respectively.
(Ni, Mi) ඥ࣋࢏࢐ ࡴ࢏࢐
ሺ૛૚ሻ
The performance of Full-duplex MIMO systems is usually
+ ሺ૚૛ሻ limited due to the strong self-interference at the front end
ඥ࣋࢏࢏ ࡴ࢏࢏
ሺ૚૚ሻ of the receive antenna. This self-interference results in the
ඥ࣋࢏࢐ ࡴ࢏࢐
ሺ૛૛ሻ combination of the signal received from a transmitting antenna
ni (1) ඥ࣋࢏࢐ ࡴ࢏࢐
on a different node and the leaked signal from the transmitting
Node j (1) Node j (2) antenna of the node itself as both radios happen to be switched
ሺ૚૛ሻ on at the same time [1]. However, designing a transmit filter
ඥ࣋࢏࢐ ࡴ࢏࢐
(Nj, Mj) (Nj, Mj) is not the only existing method or approach for minimizing
(reducing) self-interference at the transmitter. Some work, both
theoretical [7] as well as experimental [8] have been performed
for reducing self-interference by smart antenna designs for
Fig. 1. Channel behaviour analysis of a Full-duplex 2-links MIMO [1] instance. Most of these smart antenna designs are based on
• Ni and Mi are the numbers of transmit and receive antenna beamforming and target achieving as minimum self-
antennas on link i respectively. interference as possible. This approach is usually referred to
√ as the spatial domain self-interference suppression technique
• ρii is the channel’s average attenuation on the link i as
derived from the channel’s average power gain ρii . [9].
(ab) The operation of a k-links MIMO in Full-duplex mode
• In general,Hkm is the channel matrix component from
the links k to m, between the transmitting node b and the is conditional to achieving an acceptable level of self-
(12) interference cancellation at the transmitter. This is normally
receiving node a. An example is Hii .
(1) (2) done by means of a well-designed transmit filter designed in
• ni and ni are the distortion Additive White Gaussian
such a way to minimize the effect of self-interference as much
Noise (AWGN) signals at node 1 and node 2 on link i
as possible [6].
respectively.
Self-interference does not affect half-duplex systems be-
After proper modeling and mathematical derivations in [1],
cause of its temporal multiplexing. However, there is a price
based on the interference scenario as depicted in figure 1
to pay which is reduction in capacity. However due to the
above, a non-convex optimisation problem was formulated in
multiple links (k-links in this case) nature of a MIMO system,
terms of the Weighted Sum Rate maximisation under either
Half-duplex suffers from inter-user interference which affects
individual or sum power constraints as follows:
K  2 the spectral efficiency performances. Even in relay MIMO
 (a) (a)
MaxV (b) tr{μi Ii } technologies, the Half-duplex MIMO channel capacity perfor-
i
i=1 a=1 mance is lower than the Full-duplex case [13]. The increase
(b) (b) (b) in transmit power in the Half-duplex MIMO mode results in
Subject to : tr{Vi (Vi )H } ≤ Pi , ∀ (i, b)
improving the quality of the link while for the Full-duplex
K 
 2
(b) (b) (b) mode, it results in improving the quality of the link and at the
or subject to : tr{Vi (Vi )H } ≤ PT
same time causes an increase in self-interference as illustrated
i=1 a=1
(1) in figure 2 below.
where :
(b)
• Vi is the precoding matrix of the transmitter’s data
stream’s signal vector xi which is used as the optimisation
variable of the problem as derived in equation (10) of [1].
• K is the number of MIMO links considered. This varies
from as little as 2 links to more links.
• a represents the considered block of nodes which could
be the transmitting or receiving block as a Full-duplex
(bi-directional) scenario is considered. Improved Induced self-
Link quality interference
• tr represents the computation of the trace of a matrix.
(a)
• μi denotes the weight values and are always positive
(a) Fig. 2. Effect of transmit power increase on MIMO channel interference
values (μi ≥ 0).
a (a)
• Ii is the lower bound of the achievable rate of node i Additional limitations to full-duplex operation arise due to
under Gaussian signaling as determined by a study in imperfections of radio devices. Thus it can be seen that in re-
[11] which analyses the maximum achievable (limited) ality, self-interference cannot be cancelled completely. Hence
the optimisation approach instead of cancellation approach has
gained much more research attention. The idea being proper
formulation of an optimisation problem in order to find the
best possible results under considered major constraints such
as power limitation for example.
The exploitation of both spatial and temporal freedoms
of the source covariance matrices of MIMO links has been
used in studying the sum-rate maximisation problem for
full-duplex bidirectional MIMO channels for slow and fast
fading channels. A weighted sum-rate maximisation, signal-to-
leakage-plus-noise maximisation (SLNR), total transmission
power minimisation and distributed sum-rate maximisation
but with different variances. Considering a transmission from
node 1 to node 2 on link i, the variance of the transmitter and
receiver distortions are respectively modeled as follows:
(1) (2)
V ar(ni ) = δ ∗ ET x and V ar(ni ) = γ ∗ ERx (2)
where δ and γ are constant values multiplying the ran-
domly varying energy of the transmitted and the received
signals ET x and ERx respectively.
The idea behind the Full-Half duplex mixed strategy as
proposed by this paper is depicted in figure 3 below. As it
can be seen, the communication is set in such a way that for
a K links MIMO, there will be K 2 Full-duplex links and 2
K

problems for bi-directional full-duplex systems were studied Half-duplex links in case K is even or (K+1)
2 and (K−1)
2 Full-
in [10]. Furthermore weighted sum-rate maximisation for full- duplex and Half-duplex links respectively. Assuming that data
duplex systems under multiple pairs of nodes or full-duplex streams are transmitted from node i(a) , it first goes through a
MIMO interference channels has been considered in [1]. It (a)
transmit filter Vi also referred to as the precoding matrix and
comes with the expectation of doubling the link capacity as a is modelled as a Gaussian distributed and zero mean random
major advantage. However it suffers from a very strong self- vector and given by:
interference at each the links’ nodes. Therefore the present (a) (a) (a)
work proposes a strategy which will allow to still gain in terms xi = Vi di (3)
of channel capacity while minimizing the sel-interference where:
negative effect at the front-end of the used receiver antennas. (a)
• di is the complex, zero-mean and i.i.d transmitted data
III. M IXED STRATEGY: M ODELING AND ALGORITHM streams at node i(a) .
(a)
This section discusses the proposed model which leads to • xi is the Ni X1 signal vector transmitted by node i(a) .
the formulation of the optimisation problem. It also briefly ሺ૛૚ሻ n1 (2)
ඥ࣋૚૚ ࡴ૚૚
explains how the formulated problem is solved by using Node 1 (1) Node 1 (2)
ሺ૚૚ሻ +
the gradient projection optimisation algorithm for inequality (N1, M1) ඥࣁ૚૚ ࡴ૚૚ (N1, M1)
FD
ሺ૛૛ሻ
ሺ૛૚ሻ ඥࣁ૚૚ ࡴ૚૚
constraints. ඥ࣋૚૛ ࡴ૚૛

Notation: + ሺ૚૛ሻ
ඥ࣋૚૚ ࡴ૚૚
The complex conjugate of a matrix A is noted as AH through- ሺ૚૚ሻ
ඥ࣋૚૛ ࡴ૚૛ ሺ૛૛ሻ
ඥ࣋૚૛ ࡴ૚૛
out the modeling process. n1 (1)
Node 2 (1) ሺ૚૛ሻ + Node 2 (2)
A. Modeling of the system
HD
ඥ࣋૚૛ ࡴ૚૛
(N2, M2) n2 (2) (N2, M2)
This study proposes the utilisation of full-duplex MIMO +
interference channels as considered in [1], interleaved with
...

n2 (1) ...
half-duplex MIMO systems. There are two nodes on each link Node K (1) K odd: FD Node K (2)
and each full-duplex link has two neighbouring half-duplex (NK, MK) K even: HD (NK, MK)

links. Some of the expressions used in this modelling are

closed to Monte Carlo simulations. The full-duplex nodes on Fig. 3. Mixed-strategy Full-duplex and Half-duplex k-links MIMO
each pair are affected by self-interference, due to concurrent (a)
information exchange but in the half duplex case there is no After transmission, the signal xi will be modified by
self-interference. However both systems are affected by inter- the different channel impairments of which self-interference,
user interference due to simultaneous transmission on all links inter-user interference as well as the additive noise due to
(a)
[1]. Two types of power constraints are considered and these the transmitter’s distortion (ni ). The additive noise signal at
are; sum power constraints and individual power constraints. the transmitter and receiver can be modelled in terms of their
Channel impairments which will normally affect a signal respective covariance matrices cai and eai as follows:
(a) (a) H
(a)
xi transmitted from a node a on a link i are the “transmitter cai ∼ CN (0, δ diag (Vi (Vi ) )) (4)
distortion” as well as the “receiver distortion” signals. These
(a) (a) (a)
distortion signals are random by nature and are caused by eai ∼ CN (0, γ diag (ψi ) with: ψi = Cov(μi ) (5)
the fact that the transmitter and receiver designs used in
(a)
practice are non-ideal amplifiers, oscillators, Analog-to-Digital Considering that μi is the undistorted received signal mean-
Converters (ADCs) as well as Digital-to-Analog Converters ing that it is the equivalent to the received signal with no
(DACs). A study in [12] has proven experimentally by phys- receiver’s distortion eai added to it.
ical measurements that the transmitter and receiver distortion A receiving antenna of a node b on the same Full-duplex
(1,2) (a)
signals can be modelled as AWGN (ni ) with zero mean ink i as the signal xi is transmitted throughout the wireless
channel in the presence of both self-interference and inter-user the Full-duplex one in this paper) links respectively can be
interference can be modelled as follows: modeled as follows:
(a) √ (ab) √ (aa) (a) (a) (a)
yF Di = ρi Hii (xbi + cbi ) + ηii Hii (xi + ci ) (a) (ab) (b) (b) H (ab) H
−1
IF Di = log2 |IM i +ρi Hii Vi (Vi ) (Hii ) ( ) |
K  2  F Di
+
(ak) (k) (k) (a)
(nij )Hij (xj + cj ) + ei + nai (a) (ab) (b) (b) H (ab) H
(a) −1
IHDi = log2 |IM i +ρi Hii Vi (Vi ) (Hii ) ( ) |
j=i k=1 HDi
F or i ∈ {1, 2, ..., K}, (a, b) ∈ {1, 2}, a = b (6) (11)

The signal which is received by a node on a Half-duplex link where:

j = 2, as in figure 1 for example, will not consist of any - IM i is a M iXM i identity matrix,
(a) (a) (a)
self-interference effects and can be modelled as follows: - ( F Di ) and ( HDi ) are the covariance matrices of zF Di
(a)
K  2  and zHDi respectively and can be modelled by taking into
(a) √ (ab)
 (ak) (k) (k)
yHDi = ρi Hii (xbi +cbi )+ (nij )Hij (xj +cj ) account all the channel’s impairments as follows:
j=i k=1 (a)     
(ab) (b) (b)H (ab)
+
(a)
ei + nai ≈ ρi δHii diag Vi Vi Hii
F Di
  H 
F or i ∈ {1, 2, ..., K}, (a, b) ∈ {1, 2}, a = b (7) (aa)
+ ηii δHii diag Vi
(a)
Vi
(a)
(Hiiaa )
H

(a) 
The undistorted received signal μi
in both cases (Full-  H  H 
(ab) (b) (b) (ab)
duplex and Half-duplex) can be modeled as follows: + γρi diag Hii Vi Vi Hii
(a) (a) (a) (a)   H 
μF Di = yF Di − eai and μHDi = yHDi − eai (aa) (a) (a)
+ γηii diag Hii Vi Vi
(8)
In order to design optimal transmit and receive filters for K 
 2      

(ac) (c) (c)H (b) (b)H

both the Full-duplex and Half-duplex links, it is important to + ηij Hij Vi Vi + δdiag Vi Vi
know the desired output (received) signals for both beforehand. j=i c=1
K 
 2    H 
Since the self-interference component occurs at the transmit (ac) (c) (c) (ac)
filter which means that it occurs before transmission, this could + γηij diag Hij Vj Vj Hij + IM i
j=i c=1
be measured and therefore can be assumed to be known by
(a) (12)
the transmitter. The self-interference received cancelled ỹF Di
(a)
and ỹHDi signals can therefore be modelled as: and:
(a) (a) √ (aa) (a) (a)     
ỹF Di = yF Di − ηii Hii xi (ab) (b) (b)H (ab)
≈ ρi δHii diag Vi Vi Hii
√ (ab) (b) (a)
HDi
  H  H 
= ρi Hii xi + zF Di and (ab) (b) (b) (ab)
(a) (a) √ (aa) (a) + γρi diag Hii Vi Vi Hii
ỹHDi = yF Di − ηii Hii xi
√ (ab) (b) (a) K 
 2      

= ρi Hii xi + zHDi (9) (ac) (c) (c)H (b) (b)H

+ ηij Hij Vi Vi + δdiag Vi Vi
Unlike self-interference which is known by the transmitter j=i c=1
filter prior to the transmit filter design, the inter-user inter-  2
K     H 
(ac) (c) (c) (ac)
ference plus the noise components which are also part of the + γηij diag Hij Vj Vj Hij + IM i
received signal both in Full-duplex and Half-duplex modes are j=i c=1
(a) (a) (13)
unknown and random components zF Di and zHDi which can
be computed as: By using the derived lower bound of achievable rate of node
(a) √ (ab) √ (aa) (a) (a) (a) i(a) in equation (11) above, an overall Full-duplex and Half-
zF Di = ρi Hii cbi + ηii Hii ci + ei + ni
duplex WSR (Weighted Sum Rate) optimisation problem for
K 
 2 
(ak) (k) (k) maximum spectral efficiency can be formulated as follows:
+ (nij )Hij (xj + cj ) and
j=i k=1 K 
2
√  (a) (a) (a)
(a) (ab) (a) (a) tr{μi (IF Di + IHDi )}
zHDi = ρi Hii cbi + ei + ni Max(V (b) (b)
F Di +VHDi )
2 
K  i=1 a=1
 (ak) (k) (k) (b) (b) (b)
+ (nij )Hij (xj + cj ) (10) Subject to : tr{Vi (Vi )H } ≤ Pi , ∀ (i, b)
j=i k=1 K  2
(b) (b) (b)
The lower bounds of achievable rates of a node i(a) under or subject to : tr{Vi (Vi )H } ≤ PT
i=1 a=1
Gaussian signaling for the Full-duplex (as systematically de-
rived in [11]) and the Half-duplex (as derived with respect to where: Vib = VFb Di + VHDi
b
(14)
Due to the solution difficulty caused by the non-convexity • gi ((Vib )n ) is the individual or even the sum power
of the WSR maximisation problem, a change of objective constraint as these two have been proven to yield more
function from WSR maximisation to the Weighted Mini- or less the same results in [1].
mum Mean-Squared-Error (WMMSE) minimisation results in Let us consider the objective function in problem (15) as:
a non-convex problem which is quite easier to solve. This (b) K  2
(a) (a) −1 (a) −1
change of objective function results in the following WMMSE f (Vi ) = tr{Wi (log2 |(EF Di ) (EHDi ) |},
i=1 a=1
minimisation problem: The gradient projection algorithm is a modified version of
K  2
(a) (a) −1 (a) −1 the steepest descent algorithm in which only solutions that
MinV (b) tr{Wi (log2 |(EF Di ) (EHDi ) |} lie in a closed bounded domain are valid and is thoroughly
i
i=1 a=1
(b) (b) (b)
detailed in [14]. The gradient projection algorithm for solving
Subject to : tr{Vi (Vi )H } ≤ Pi , ∀ (i, b) the WSR maximisation problem (14) through WMMSE
 2
K  minimisation problem (15) is given in Table I below:
(b) (b) (b)
or subject to : tr{Vi (Vi )H } ≤ PT
i=1 a=1 Table I: Gradient projection WSMMSE minimisation
(15) algorithm
where: (a)
μi (a) −1
• Wi(a) = ln2 (Ei ) 1) Set the iteration number n = 0.
(a) (a) −1 (a) (a) −1
• IiF D = log2 |(EiF D ) | and IiHD = log2 |(EiHD ) | 2) Start with an arbitrary initial complex transmit filter
• Therefore: (precoding matrix)(Vib )0 of size (Ni × di ), i : 1 → K
(a) (a) (a) −1 (a) −1 and b ∈ {1, 2}.
IiF D + IiHD = log2 |(EiF D ) | + log2 |(EiHD ) |
(a) −1 (a) −1 3) Find a search direction Sn as : Sn = −Δfn where
= log2 |(EiF D ) (EiHD ) | (b)
fn = −Δf ((Vi )n )
as depicted in equation (15).
4) Determine the optimal step length λ∗n in the search
The change from WSR maximisation to WMMSE minimisa- direction Sn and get the next iteration optimal transmit
tion is justified by the fact the two problems share the same filter (Vib )(n+1) as:
KarushKuhnTucker (KKT) conditions.Thus a proper choice of
(Vib )(n+1) = (Vib )n + (λ∗n × Δfn )
optimum WMMSE-weights will result in the maximum WSR
5) Update n = n + 1. Update the corresponding receive
and therefore generates maximum spectrum efficiency. (a)
filters rF Di using (16).
The corresponding optimum receive filters at a node i(a) for
6) Repeat steps 3 and 4 until convergence, or else up until
both the Full-duplex and Half-duplex modes can be obtained
a predefined number of iterations is reached.
by analogy to equation (18) in [1] as follows:
(a)∗ √ H H
IV. R ESULTS AND DISCUSSION
rF Di = ρi (Vib ) (Hiiab ) × The performance evaluation of this study has been
H H (a) −1
conducted in MATLAB discrete simulator. The key simulation
(ρi Hiiab Vib (Vib ) (Hiab ) + )
F Di set-up parameter values are captured in Table II below:
(a)∗ √ H H
rHDi = ρi (Vib ) (Hiiab ) ×
H H (a) −1 Table II: Simulation parameters
(ρi Hiiab Vib (Vib ) (Hiab ) + ) (16)
HDi Parameter value Description
B. The Rosen’s gradient projection algorithm di 800 Transmission data streams size
K 5 Number of MIMO links
The Rosen’s gradient projection method is based on pro- Ni = Mi 2, 3, 4 Number of transmitters/Receivers per node
jecting the search direction into the subspace tangent to the γ=δ -40 dB Receiver and transmitter distortion parameters
active constraints as illustrated in the figure 4 below. μ 1 Positive WSR weight coefficient
Dichotomous method step length optimisation

By using the above set up, the different conducted simulations

(Vi b ) n (Vi b ) n 1
yields the following results:
gi 0 From the analysis of the obtained results, the following
_ b
observations can be drawn:
(V i ) n - From figure 5, it can be noticed that at low power, the Full-
duplex model outperforms the Mixed strategy model in terms
of spectral efficiency but as power is increased, the Mixed
Fig. 4. Rosen’s gradient projection and restoration concepts strategy model exhibits better performance. However, as power
Where: reaches higher values, the performance of two models seem
b b ieme
• (Vi )n and (Vi )(n+1) are the n and (n + 1)ieme to converge towards a maximum Weighted Sum Rate of round
optimal transmit filters on link i node b. (V̄ib )n is the about 100bits/s/Hz.
projection of (Vib )n on the search direction axis. - From figure 6, it can be noticed that the mixed-strategy
 Full duplex (FD) versus Mixed Strategy (FD-HD) spectral efficiency comparison

10
2
been investigated. After a tedious modeling of the proposed
strategy in terms of the different involved channel impairments
such as interference (both self-interference and inter-user
Sum-rate (bps/Hz)

interference) as well as distortion at the transmitter and the

10
1
receiver; the obtained results prove that the proposed Mixed
strategy (Interleaved Full and Half duplex) model in general
provides better results in terms of spectral efficiency than the
FD
Mixed strategy (FD-HD) fully Full-duplex model. The impact as well as the limitations
10
-10
0

0 10 20 30 40 50
related to the increase in the number of antennas are clearly
SNR (dB) exhibited. This work therefore contributes towards the current
Fig. 5. Spectral efficiency performance comparison between FD and Mixed- research in terms of MIMO systems designs with the aim
strategy (FD-HD) with power level variation
of achieving higher spectral efficiency. Further work could
Full-duplex (FD) versus Mixed Strategy (FD-HD) convergence behavior
consist of an investigation about applying the mixed-strategy
30
model proposed in this paper in a cellular system scenario for
29 example.
28
R EFERENCES
Sum-rate (bps/Hz)

27 [1] A.C Cagatay, R. Wang, Y. Hua and M. Latva-aho, March 2015,

26 Weighted Sum-Rate Maximization for Full-duplex MIMO Interference
25 Channels, In the IEEE transactions on Communications, Vol. 63, No.
24
3, pp. 801–815.
[2] D. Bandyopadhyay and J. Sen, May 2011, Internet of Things:
23
Applications and Challenges in Technology and Standardization, In
22 FD
Mixed-strategy (FD-HD)
the springer transactions on Wireless Personal Communications, Vol.
21 58, Issue 1, pp. 49–69.
20
0 5 10 15 20
[3] F. Negro, S.P. Shenoy, I. Ghauri and D.T.M. Slock, 2010, On the
Iterations (n) MIMO Interference Channel, In the Proc. of the Information Theory
Fig. 6. Convergence comparison between FD and mixed-strategy(HD-FD) and Applications Workshop (ITA), San Diego, CA, USA, pp. 1–9.
WSR maximisation [4] H.J. Choi, C. Song, H. Park, and I. Lee, January 2014, Utility
Maximization In The Half-duplex Two-way MIMO Relay Channel,
In the IEEE transacations on Wireless Communications, vol. 13, No.
Spectral efficiency variation with increased spatial diversity
200 1, pp. 198 – 209.
180
[5] Z. Chen et al, April 2012 Interference Mitigation for Cognitive
Radio MIMO systems based on Practical Precodingl, In the Elsevier
160 2X2 Mixed FD-HD MIMO Physical Communication Journal , vol. 9, pp. 308 – 315.
140 3X3 Mixed FD-HD MIMO
[6] H. Degenhardt and A. Klein, April 2012, Self-inference aware MIMO
Sum-rate (bps/Hz)

4X4 Mixed FD-HD MIMO

120
100
80
60
40
20
0 [8] M. Duarte and A. Sabharwal, 2010, Full-Duplex Wireless Communi-
-10 0 10
SNR (dB)
Fig. 7. Spectral efficiency variation with increased power and spatial diversity
(Full-Half duplex) model converges slightly faster to higher
spectral efficiency values than the Full duplex one. Another
interesting observation is that after the 6th iteration of the
Rosen’s gradient projection algorithm both models’ spectral
efficiency performance remain quite steady and there is no
more significant improvement observed. Figure 7 clearly
shows that as power increases, spectral efficiency increases
too but the key observation here is the effect of increasing the
number of both the transmit and receive antennas on spectral
efficiency. As the number of antennas increases from 2 to 3
to 4, the spectral efficiency of the Mixed-strategy model is
considerably improved to point of doubling at a SNR of 40
dB.
V. C ONCLUSION
In this paper, the possibility to use a Full and Half duplex
Mixed strategy in a K links MIMO interference channels has
Filter Design for Non-regenerative Multi-pair Two-way Relaying, In
the IEEE international Conference on Wireless Communication and
Networking ,Shanghai, China , pp. 272 – 276.
[7] E. Everett and A. Sabharwal, July 2014 A Signal-Space Analysis
of Spatial Self-interference Isolation for Full-Duplex Wireless, In
the IEEE international Conference on Information Theory, Honolulu,
Hawaii , pp. 661 – 665.
20 30 40
cations Using Off-The-Shellf Radios: Feasibility and First Results, In
the IEEE international Conference on Signal Systems and Comput-
ers,Pacific Grove, CA , pp. 1558 – 1562.
[9] A.C. Cirik, R. Wang and Y. Hua, 2013, Weighted-Sum-Rate Maximi-
sation for Bi-directional Full-Duplex MIMO Systems, In the Proceed-
ing to Asilomar Conference Signals, pp. 1632 – 1636.
[10] T.M. Kim et al, June 2013, Distributed Sum- rate Optimization for
Full-Duplex MIMO systems Under Limited Dynamic Range, In the
IEEE Signal Processing Letters, vol. 20, No. 6, pp. 555 – 558.
[11] B.P. Day, A.R. Margetts, D.W. Bliss, and P. Schniter, July 2012,
Full-duplex bidirectional MIMO system,” In the IEEE Transaction
on Signal Processing, vol. 60, No. 7, pp. 3702 – 3713.
[12] H.Suzuki et al, August 2008, Transmitter Noise Effect on The Per-
formance of a MIMOA-OFDM Hardware Implementation Achieving
Improved Coverage,” In the IEEE Transaction on Communication,
vol. 26, No. 6, pp. 867 – 876.
[13] J. Shin and J. Moon, , October 2012, Weighted-Sum-Rate Maximi-
sation Linear Transceiver Filters for the K-User MIMO Interference
Channel,” In the IEEE Transaction on Communication, vol. 60, No.
10, pp. 2776 – 2783.
[14] J.B. Rosen, April 2003, The Gradient Projection Method for Nonlin-
ear Programming. Part II. Nonlinear constraints,” In the Journal of
the Society for Industrial and Applied Mathematics, vol. 9, Issue. 4,
pp. 514 – 532.
 Design and Development of Interval Arithmetic Library
using Q Format Data Representation: A FPGA Approach
Vidya Mansur
Department of EC,
BVBCET,Hubli,Karnataka,India
vidyamansur@gmail.com
Abstract—Basic digital signal operations are addition,
subtraction, multiplication, division, multiply accumulate etc.
Performance consideration depends on speed, area, power,
precision, correctness and reliability of computed results.
Digital signal processing is basically operating on the input
signals to improve the output quality of the signal. To
introduce the novel approach of interval arithmetic hardware
modelling of basic building blocks using Q format in digital
signal processing. Develop the interval arithmetic adder,
subtractor, multiplier. Design and implement the novel
interval arithmetic FIR filter using Q format data operations.
Keywords—Digital signal operations, Interval arithmetic, Q
format.
I. INTRODUCTION
Interval arithmetic is the mathematics used in various
scientific applications. It has been extensively used in the
digital computers during the past three decades. Various
implementation designs on fast computing systems has
revealed that the enhanced computing ability of interval
arithmetic to solve problems in various application domain
has improved the computing accuracy. Result metric is used
in the analysis of the algorithm to observe its performance
against variation in the chosen design parameter. The
performance of the interval arithmetic is evaluated using a
result metric obtained in the range of lower bound and upper
bound. When interval arithmetic is utilized in algorithm
convergence to a desired solution, the optimum value of the
convergence factor, number of times the function is
evaluated, failure to reach a desired solution, execution
time, speed are some of the performance measures.
The primary interval operators namely addition,
subtraction, multiplication, division, comparison design
affect the performance of the interval arithmetic
performance when implemented on system software
platform and VLSI compute efficient interval arithmetic
system modules to solve various scientific domain
applications.
As interval arithmetic computing arithmetic domain is
predominantly used to improve the output signal
interpretation and analysis. Besides the primary interval
operators, the method used in secondary design sub modules
like multiply accumulate unit, squarer sub module etc.
contribute in changing the performance behaviour of the
designs based on interval arithmetic. Further the interval
arithmetic computations especially the usage in application
specific algorithm contribute dominantly in the performance
Dr.R.B.Shettar
Department.of EC,
BVBCET,Hubli,Karnataka,India
raj@bvb.edu
criteria of the designs. The performance of the interval
arithmetic platform primarily depends on the nature of the
problem to be solved.
In this paper the basic building blocks for the interval
arithmetic operators are designed and modelled on FPGA.
These basic building blocks are integrated to form the
subsystem design. The interval arithmetic is a computing
paradigm, where the signal quality is enhanced in terms of
accuracy and precision as shown in [9], [10], [11]. There are
various data representation used in digital signal processing.
In normal IEEE 754 adder using interval arithmetic the extra
overhead of the exponent operation is eliminated using Q
format 32 bit data signals. The described architecture is
optimized for area and power. The design model is applied
for LMS algorithm. Design supports compute intensive
operations in the adaptive filter domain. Design is more
compact using Q format, than the IEEE 754 format design
module.
The proposed interval arithmetic platform enables func-
tional verification of the system level design versus the
hardware model. This work investigates the reconfiguration
criteria and presents the digital solution in communication
systems. Numerous experimental results are illustrated to
justify the performance of the proposed FPGA solution.
During the last few years the data accuracy, fast
computation has been the focus of many researchers and
significant achievement has been observed to provide digital
solution for various real time applications to enhance the
computation efficiency.
In interval arithmetic the lower bound and upper bound of
the data is considered for mathematical processing in the
design model. Future research directions in interval
arithmetic focus on mapping the computational model in
digital computers to make the system more efficient for
compute intensive application specific designs. In [7], [8]
the role of data representation in interval arithmetic is
presented. The emphasis is on those capabilities of interval
arithmetic that guarantee the performance of the infinite
precision result in the computed interval. The basic
principles of interval arithmetic are developed for various
mathematical calculations in [9], [10]. In the earlier stages
of digital numerical computation accurate results were
obtained with singe floating point number as an
approximation. Interval arithmetic consists of a set of real
values. Various arithmetic operations are applied to the set
of numbers. A narrow interval indicates better correctness of
results and is exact. A wide range of numbers in the interval
will definitely contain the result value but it is dangerous to
completely rely on this. The different properties of the
system of interval arithmetic are correctness, totality,
closeness, optimality, efficiency. The proofs involved in the
operations on an interval are discussed extensively in [11].
Matlab provides wide platform for varied mathematical
operations. New and efficient concepts based on interval
arithmetic are developed as a toolbox INTLAB in Matlab.
One of the feature of Matlab is that it doesnt have the
concept of type declaration of variables. They are
automatically declared as per the interpretation principle.
Diverse ways of representing the set of numbers in the form
of an interval and operations on them are presented in [12],
[13]. Interval arithmetic calculations requires proper set of
rules to be followed to compute in interval form. The
interval consists of a lower bound and a higher bound which
when used for performing operations give an enclosure of
correct solution.
In [14] the author Rump presented that by expressing the
set of real values by the midpoint and radius, good results
were obtained. Later the author incorporated all the concepts
in developing the Matlab toolbox Intlab. The properties of
interval arithmetic are presented with the usage of Intlab
tool. Various Intlab functions are evolved to enhance the
performance of mathematical operations. In the next section
the Design methodology used for this research work is
presented.
II. SYSTEM MODEL
Q format Interval Subtractor The inputs to the interval
added are all in the Q format data representation. Interval
representation is represented by
If x = [a, b] y = [c, d]
z = x - y = [a - d, b - c] = [e, f]
When two intervals are subtracted the lower bound of the
resultant output is the difference of the lower bound of X
and upper bound of Y, the upper bound of the output Z is
the difference of the upper bound of input interval X and
lower bound of input interval Y.
Q Format Interval Multiplier
The data representation of the inputs are in Q (1,30) format.
To compute the interval multiplication four products have to
be calculated. The minimum and the maximum of the four
obtained products are computed to get the lower bound and
upper bound of the resultant interval respectively.
If x = [a, b] y = [c, d]
z = x * y = [min (ac, ad, bc, bd), max (ac, ad, bc, bd)] In
Fig. 1 and 2 the data flow representation for Q format
interval multiplier is presented. The interval multiplication
is computed by finding the minimum and maximum of the
four products obtained.

Q format Interval Adder The inputs to the interval adder

are in the form of intervals. The resulting value is also in the
form of an interval. The inputs to the interval added are all
in the Q format data representation. In the previous section
the Q format data representation is presented.

Interval addition is represented by:

If x = [a, b] y = [c, d]

z = x + y = [a + c, b + d] = [e, f]

When two intervals are added the lower bound of the

resultant output is the addition of the lower bound of X and
lower bound of, the upper bound of the output Z is the sum
of the upper bound of input interval X and upper bound of
input interval Y. Our design objective is to model the design
in hardware using FPGA. The interval addition steps
described above should be now mapped to a data flow
graph. The nodes of the data flow graph indicate the
operation to be performed and the arrows to this node
indicate the inputs, arrow from the node indicate the output
result. The resources used in interval arithmetic adder can
be identified by the nodes in the data flow graph. The
interval artihmetic adder needs four registers to store the
input operands a, b, c, d two registers to store the output
values e, f. The arithmetic modules required are two Q
format adders to compute the lower bound and the upper
bound. With this information there is a smooth transition of
the system level design process to RTL design process to
model in Verilog.
Figure 1. Data flow graph for interval multiplier (a)
III. Q FORMAT INTERVAL ARITHMETIC
BUILDING
BLOCKS
A Q format 32 bit binary interval adder is shown in fig. 5.
It is a combinational system. The n bit inputs a = (an-1,
. . . . a0), b = (bn-1, . . . . b0), c = (cn-1, . . . . c0), d = (dn-1, .
. . . d0) representing the input operands. x = (a, b), y = (c, d)
where x and y are two operands. z = (e, f) is the result of
addition. The behavioural description of the interval
arithmetic adder relating inputs and outputs is e = a +c, f = b
+ d A high level specification of Q(30) format inputs is :
a = a31, a30, ..a0 aj belongs to (0, 1) b
= b31, b30, ..b0 bj belongs to (0, 1) c =
c31, c30,..c0 cj belongs to (0, 1)
d = d31, d30, ....d0 dj belongs to (0, 1) e
= e31, e30, ..e0 aj belongs to (0, 1) f =
f31, f30, ..f0 fj belongs to (0, 1)
a31, b31, c31, d31, e31, f31 are the sign bits. If a31 = 0
then the number is a positive number If a31 = 1 then the
number is a negative number
The interval arithmetic adder is a module capable of
realizing addition on lower bound input operands and upper
bound input operands. To achieve this objective, an interval
arithmetic adder has the data inputs and outputs along with
the control inputs. These control inputs are used to select the
specific function, performed at a given time. The reason for
developing the interval arithmetic adder module as a library
Figure 2. Data flow graph for interval multiplier (b)
component is to use the design in many different
applications. The developed module can be effectively used
in the addition can be effectively used in the adaptive filter
LMS operation to perform interval arithmetic adder
operation with the control signal selection. The Q format is
used in the interval arithmetic addition operation.
The input data representation to the adder is in Q (1, 30)
format. The input values are in signed 32 bit representation.
To store the input data four registers are required. To store
the output bounded data two registers are essential, one for
the lower bound and one for the upper bound. The control
signal enadd is used to commence the addition operations.
Another control signal adddone = 1 indicates the completion
of the addition computation.
Consider the input data sample a = 0.54
Following steps indicate the conversion from floating
point real number to Q format.
Normalization step to get the data in the range of 0 <
a<1.
a = 0.54/4 = 0.1350
0.1350 *(230) = 144955146.2
Round this number a = (144955146)d
a = Hex of 144955146 = (08A3D70A)h
a = (00001000101000111101011100001010) b
The inputs to the interval arithmetic adder are x = (a, b) y =
(c, d). TheRega; Regb; Regc; Regd are used to store the
input operands in format adder units consists of addition
functional unit. For the interval addition algorithm the two RESULTS: INTERVAL ADDER ON VERTEX 5
results obtained namely the lower bound and higher bound PLATFORMS
inRegacandRegbd respectively. Whenever the addition of
any two binary data has to be calculated the enable signal of Table III
Q- format interval adder unit has to be activated, since the XC4VfX12
input data is converted to Q- format and given there is no No of slices 1913/5472
need of any conversion unit in the architecture No of 4 input
LUTs 3640/10944
Delay 15.414ns

RESULTS: FLOATING POINT INTERVAL ADDER ON

VERTEX 4 PLATFORM

Subtractor is a module capable of realizing subtraction on

lower bound input operands and upper bound input
Figure 3. Architecture of interval Q format adder operands. To achieve this objective, an interval arithmetic
Table I subtractor has the data inputs and outputs along with the
XC4VFX60 XC4VFX20 control inputs. These control inputs are used to select the
No of slices 32/2580 32/8544 specific function, performed at a given time. The reason for
No of 4input developing the interval arithmetic subtractor module as a
LUTs 64/50560 64/17088 library component is to use the design in many different
Delay 7.211ns 8.123ns applications. The developed module can be effectively used
in the addition can be effectively used in the adaptive filter
LMS operation to perform interval arithmetic subtractor
operation with the control signal selection. The Q format is
RESULTS: INTERVAL ADDER ON VERTEX 4 used in the interval arithmetic subtractor operation.
PLATFORMS
The data flow graph presented in the previous section is
Area and delay estimates mapped to the hardware module. The architecture of the
interval arithmetic Q format subtractor is indicated in the
The architecture of Q format interval adder described in Fig. 8.
the previous section has been synthesized using Verilog
design flow. The design is simulated to ensure the
arithmetic correctness of the design, exhaustive test data
seta are given and design tested, the FPGA ISA tool suite is
used to simulate the data. Xilinx 14.2 version is used to
model the design. Since this a novel approach, there are no
designs for comparison in the literature available.

Interval arithmetic subtractor Architecture It is a

combinational system. The n bit inputs a = (an-1, . . . .

a0), b = (bn-1, . . . . b0),c = (cn-1, . . . . c0),d = (dn-1, .

Figure 4. Architecture of Q format Interval arithmetic
subtractor
. . . d0) representing the input operands. x = (a, b),y = (c,
d)where x and y are two operands. z = (e, f) is the result of
The inputs to the interval arithmetic subtractor are x = (a,
subtraction. The behavioural description of the interval
b) y = (c, d). The Rega; Regb; Regc; Regd are used to store
arithmetic subtractor relating inputs and outputs is
the input operands in format adder units consists of addition
functional unit. For the interval subtraction algorithm the
e = a - d, f = b - c
two results obtained namely the lower bound and higher
bound in Regad and Regbc respectively.
A high level specification of Q(30) format inputs is same
as for the interval arithmetic adder. The interval arithmetic
Whenever the subtraction of any two binary data has to
be calculated the enable signal of Q- format interval
Table II
subtractor unit has to be activated, since the input data is
XC5VLX50 XC5VFX130 converted to Q- format and given there is no need of any
No of slice LUTs 64/28800 64/81920 conversion unit in the architecture.
Delay 5.785ns 4.984ns
 In this work the design of hardware model of interval
arithmetic multiplier is presented as a novel approach using
Q format data representation. Although Q format data rep-
resentation is popularly used in digital signal processing, to
our knowledge this is the first design presented in Q format
interval arithmetic multiplier. The Q format interval
multiplier uses the popular Booth multiplier.
Figure 5. Performance analysis of interval subtractor

Interval arithmetic multiplier Architecture

Audio processing, video processing, communication

systems, biomedical signal processing, multimedia
applications are notably data intense and resource
consuming. These compute intensive applications involve
many digital signal processing tasks using the arithmetic
operations like addition, subtraction, multiplication, division
to execute the algorithmic steps involved in the design
model. The processors evolved over the years are
implemented using state of art submission technology.
Multiplication operation has a processing number of steps in
various types of data representation. In this section the
implementation of interval arithmetic multiplier design
using Q format is illustrated.
Overview
As digital signal processing evolve, the volume of data to be
processed increases significantly may be due to the large
input data sets available or due to the complex system
designs in present days. During the manipulation of the data
information a large cost is involved in terms of execution
time and hardware resources. Processing may require many
operations, hence more time and more computing modules.
Other design space parameters that are of interest are the
accuracy of the result, reliable computing, bit-width opti-
mization, precision to greater extent.Over the years schol-
arly work has been reported considering the performance
issues started above. In this context, optimum designs to
address these parameters have been developed by various
researchers. There is a demand for designs that reduce the
computational cost and also give closeness solution giving
guarantee that the solution is correctly obtained. The
following points are relevant to be thought during the
discussion and these are stated as below
Figure 6. Interval arithmetic multiplier architecture
A Q format 32 bit binary interval multiplier is shown in
Fig 10. It is a combinational system. The n bit inputs a =
(an-1, . . . . a0), b = (bn-1, . . . . b0), c = (cn-1, . . . . c0), d =
(dn-1, . . . . d0)
representing the input operands. x = (a, b), y = (c, d) where
x and y are two operands. z = (e, f) is the result of
multiplication. The behavioural description of the interval
arithmetic multiplication relating inputs and outputs is e =
min (ac, ad, bc, bd), f = max (ac, ad, bc, bd)
From the data flow graph the following architecture is
designed. However despite being a very appealing solution
for processing large computational tasks, there can be
number of alternate solution to the comparator steps
handling the various cases for interval multiplication. The
architecture presented is developed using Verilog
modelling.
The input registers for the two operands x = (a, b) and y =
(c, d) are a, b, c and d. These operands are stored in
registerRega,Regb, Regc and Regd:

Do we need to move towards better reliable computing
systems?
In critical mission control system application, is it
sufficient to obtain a single solution or is it necessary to
obtain a bound of solutions?
There are four 32 bit Q format booth multipliers. The input
operand to the first multiplier are values a and c, to the
second multiplier a and d, to the third multiplier b and c, to
the fourth multiplier b and d.

Based on the importance of applications and tolerance of

the output data obtained after processing various arithmetic
domains are used. Some of them are cordic arithmetic,
distributed arithmetic, normal floating point arithmetic
where effort has been made to decrease the computation
cost. To increase the guarantee of solution domain of
processing may be shifted to interval arithmetic.
The Q format 32 bit multiplier unit computes the product of
two signed integers using booth recorded multiplier. The
booth recoded multiplier design steps are illustrated in the
previous section. The Q format receives the ap-propriate 32
bit input from the repetitive registers in 25 complement
form. The output of each multiplier is stored in Regac;
Regad; Regbc; Regbd respectively.

The Q format 32 bit interval multiplier operates n two

phases.

The first phase is the multiplication phase and

The second phase is the interval bound computation phase.

Whenever the multiplier has to compute the product of two
operands the enable signal of the multiplication unit is to be
activated.
Figure 8. Min comparator for multiplier lower bound

Fig. 8 shows the architecture of lower bound com-putation

unit. The inputs to these are the outputs of the multiplier
namely ac, ad, bc, bd which are stored in Regac; Regad;
Regbc; Regbd. The minimum of these four values is the lower
bound value of the interval multiplier.

To obtain result following steps are performed.

If ac <ad then min = ac Else Min = ad If min< bc then min =

min Else min = bc If min<bd then min = min Else min = bd
Multlow = min

Experimental results

Figure 7. Max comparator for multiplier upper bound
Fig. 7 illustrates the multiplier upper bound computa-tion
unit. The architecture is similar to the lower bound
computing unit, except for now the comparison is to obtain
maximum of two operands.
The interval arithmetic Q format 32 bit multiplier is
implemented using Xilinx ISE 14.2 tool suite. Xilinx is a
powerful simulation tool developed by model technologies
for Xilinx devices. The area estimate of the and delay
estimate of the design is in the following Fig. 9.

Min-max computing unit

Figure 9. Performance analysis(Area and timing) of 32 bit

Q format interval multiplier units using Booth multiplier for
XCV4f and XCV 5f

The design is synthesized in XCV4F and XCV5F. Since

there is no architecture available for comparing the results
of Q format interval multiplier there is no avenue for
comparing our result
IV. RESULT ANALYSIS
d) Partial product addition
In this section the Computational step differences between Low power solution. No of transistors required are less.
floating point multiplier and Q format multiplier are Area optimized design.
discussed first.
Computational time is less.
Interval Floating point multiplier
V. CONCLUSION
– IEEE 754 single precision 32 bit representation.
– MSB is sign bit. The above discussion presents the innovative archi-tecture
– Booth multiplier CSD, Wallace tree any design be design for Q format interval multiplier to establish the
used. competitive advantage of fixed point Q format
– Four multiplication operation to obtain a bounded multiplication. Our proposed architec-ture has formed
solution. approach in interval arithmetic paradigm in term of
– Three comparisons to obtain lower bound. despising basic building block of interval arithmetic
– Three comparisons to obtain upper bound. hardware library. Various examples for data conversion has
– FP multiplier one operation been illustrated with appropiate logical steps. The Q format
inter-val adder, subtractor, multiplier are presented with the
Add exponents a) er = ex + ey, er1 = er - 127 b) mantissa data flow graph and the architecture to map the deign to
result = mx * my hardware. These building blocks are implemented on the
hardware platform to obtain the results shown above.
24 Bit Booth multiplier c) Normalise the result
REFERENCES
Total number of operations = one addition, one subtraction [1] Simon Haykin, “Adaptive filter theory,” Pearson Education
and 24 Bit Booth multiplier and one normalization step. Publication, 4th Edition, May 2012, pp. 231-278.
[2] Emmanuel Ifeachor and Barrie W. Jervis, “ Digital Signal Processing
– 4 FP multipliers ,” Pearson Education Publica-tion, 2nd Edition, Jan. 2013, pp.342-440.
– 4 additions [3] Paulo S.R. Diniz, “ Adaptive Filtering Algorithms and Practical
– 4 subtraction Implementation ,” Kluwer Academic Publications, 3rd Edition, pp. 77-189.
– 4 24 bit multiplications [4] Angelo Bosco, Massimo Mancirso, “ Adaptive filtering for image
– 1 normalization/ shift denoising ,” ICCE. International Conference on Consumer Electronics, Jun.
2001, pp. 208-211.
d) Partial product addition [5] Peter Strobach, “ Low rank adaptive filters,” IEEE Transactions on
Signal Processing, Vol. 44, No. 12, Dec. 1996, pp. 2932-2947.
Increased power solution. No of transistors re-quired are [6] John Hakon Husoy, Mohammad Shams Esfand Abadi, “ A Family of
more. Flexible NLMS-type Adaptive Filter Algorithms ,” 6th International
Area is more Conference on Information, Communication and Signal pro-cessing,
Singapore, Jun. 2007, pp. 1-5.

Computational time is more. Hence the compu-tational cost
is more.
Interval Q format multiplier
32 bit Q format representation. MSB is sign bit.
Booth multiplier CSD, Wallace tree any design be used.
Four multiplication operation to obtain a bounded solution.
Three comparisons to obtain lower bound. Three
comparisons to obtain upper bound.
One 32 bit Q multiplier using booth algorithm
a) Use signed integers b) Only integer multiplica-tions c)
One 32 Bit Booth multiplier
[7] R. E. Moore,“ Interval Arithmetic and Automatic Error analysis in
Digital Computing ,” Technical Report No. 25, Applied Mathematics and
Statistics Laboratory, Stanford University, Stanford, Califr-nia, Nov. 1962,
pp. 1-3.
[8] R. E. Moore, “ Automatic Error Analysis in Dig-ital Computation ,”
Technical Report No. 48421, Lockheed Missiles and Space Co.,
Sunnyvale, Cal-ifornia, Jan. 1959, pp. 1-5.
[9] R. E. Moore, “Methods and Applications of Interval Analysis ,”
SIAM, Philadelphia, Jun. 1979, pp. 4-27.
[10] John Pryce, “ IEEE Working Group P1788 , A standalone for
Interval Arithmetic ,” Dagastuhl Seminar 09471, Cranfield University,
Nov. 2009, pp. 1-20.
[11] T. Hickey, Q. Ju, M.H. Van Emden, “ Interval Arithmetic: from
Principles to Implementation ,” Journal of the ACM (JACM), Vol. 48 No.
5, Sep. 2001, pp. 1038-1068.
[12] S.M. Rump, “ INTLAB - INTerval LABoratory ,” Tibor Csendes,
Editor, Developements in Re-liable Computing, Kluwer Academic
Publishers, Jun. 1999, pp. 77-104.

No addition and subtraction for exponents No [13] S.M. Rump, Hamburg, “ INTERVAL COMPUTA-TIONS WITH
INTLAB ,” Developments in Reliable Computing, Springer Publication,
normalization. Sep. 1999, pp. 77-104.
[14] G. I. Hargreaves,“ Interval Analysis in MATLAB,” Numerical
4 32 bit Booth multipliers No exponent additions Analysis Report No. 416, Dec. 2002, pp.1-4.
[15] Erick L Obester,“ Fixed-point Representation and fractional Math ,”
No exponent subtraction Obester Consulting, Aug. 2007, pp. 1-19.
 Computation Beyond Moore’s Law:
Adaptive Field-Effect Devices for Reconfigurable
Logic and Hardware-Based Neural Networks
Udo Schwalke
Institute for Semiconductor Technology and Nanoelectronics
Technische Universität Darmstadt
Darmstadt, Germany
schwalke@iht.tu-darmstadt.de
Abstract—The success of integrated silicon technology is
based on the down-scaling of minimum feature sizes of silicon
field-effect devices (MOSFETs) in a complementary circuit
configuration (CMOS) according to Moore’s Law. Reducing the
feature size provides more components per chip and higher
speed. However, this continuous miniaturization of MOSFETs
will come to an end as CMOS scaling will soon approach atomic
dimensions. To take computation beyond Moore’s Law requires
breaking at least with two major paradigms: (1) High computing
performance is directly related to high switching speeds of the
single device and (2), the separation of memory and computing.
In this work we report on a novel adaptive nanowire field-effect
transistor (a-NWFET) architecture which provides a release
from paradigms (1) as well as (2). The fabricated a-NWFETs are
originally ambipolar nanowire devices, using midgap Schottky-
barrier contacts as source and drain (S/D) electrodes. The final
unipolar a-NWFET device type (i.e. NMOS or PMOS) can be
created by applying an electric bias at the back-gate. The ability
to select the transistor type by the application of an electrical
signal to the back-gate adds to the versatility of the device
concept, where the two complementary device types are
interchangeable on the fly. A simple and versatile device
structure for logic and intrinsic memory applications with the
potential to realize novel reconfigurable logic architectures and
hardware-based neural networks will be presented.
Keywords—Adaptive Field-Effect Transistor; Moore’s Law;
Reconfigurable Logic; Neural Network.
I. INTRODUCTION
Like no other technology, integrated electronics has
changed our daily life and silicon has been the ultimate
semiconductor material in micro- and nanoelectronics for more
than 50 years. The continuous down-scaling of silicon field-
effect devices (MOSFETs) [1] in a complementary
configuration (CMOS) [2] according to Moore’s Law [3]
provides the basis of the tremendous progress in information
technology. Thanks to silicon CMOS, we can enjoy, for
example, the rich multimedia experience when using the
internet, mobile phones or tablet PCs. In fact, without the
present hardware technology platform, many (if not all) areas
of computing, communication, networking and information
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
processing like big data analysis as we know today would
simply not exist.
The success of integrated silicon technology is based on the
scaling of minimum feature sizes. Within the past 50 years,
microelectronics has completed the transition into
nanoelectronics, i.e. 100µm (micrometer) “big” devices have
been downscaled into the sub-100 nanometer range. For
example, today state-of-the-art advanced silicon CMOS
technologies are utilizing feature sizes of 22 - 14 nm which is
five times below the size of a virus! Reducing the feature size
provides more components per chip and higher speed.
However, this continuous top-down miniaturization of
MOSFETs will come to an end [4] as CMOS downscaling will
soon approach atomic dimensions.
To take computation beyond Moore’s Law requires to
break at least with two major paradigms: (1) High computing
performance is directly related to high switching speeds of the
single device and (2), the separation of memory (e.g. DRAM)
and computing (e.g. CPU). On the other hand, it would be
highly desirable if any new approach would be fully
compatible with the established silicon-CMOS integrated
circuit technology in order to make use of most of the existing
semiconductor industry.
In conventional CMOS technology, NMOS- and PMOS-
FETs are hardware defined by choosing the appropriate doping
of source (S) and drain (D) junctions with respect to the
substrate. However, in this work we report on a novel adaptive
nanowire field-effect transistor (a-NWFET) architecture which
provides a release from paradigms (1) as well as (2). The
fabricated a-NWFETs are originally ambipolar silicon
nanowire devices, using midgap Schottky-barrier contacts as
source and drain (S/D) electrodes. The final unipolar a-
NWFET device type (i.e. NMOS or PMOS) can be selected by
the electric bias at the back-gate. The ability to select the
transistor type by the application of an electrical signal to the
back-gate adds to the versatility of the device concept, where
the two complementary device types are interchangeable on the
fly. This simple and versatile device structure for logic and
intrinsic memory applications has the potential to realize novel

Figure 1. Drawings and SEM images of the fabricated NWFET
structure. a) Overview with cross-sectional positions as indicated. b)
Cross-section A’ along the nanowire. c) Cross-sectional view of front-
gate section B’. d) Cross-sectional view of source/drain region C’. e)
Scanning electron microscopy image of NWFET structure on SOI
across the front-gate section B’. f) The same but across the S/D region
C’.
reconfigurable circuit architectures and hardware-based neural
networks.
II. DEVICE FABRICATION
For device fabrication, a CMOS-compatible top-down
fabrication technology has been used, forming the nanowire by
means of well-established lithography and subsequent reactive
ion etching. Fig. 1 gives a comprehensive view of the device
that has been realized to demonstrate the process technology
and device functionality as well. The devices are fabricated on
commercially available ultrathin-body PremiumSOITM wafers
from SOITEC with a buried oxide (i.e. back-gate oxide) of
145nm thickness [5]. The top- silicon thickness is 70nm and
the lowest background doping level available of 1015
boron/cm3 is used. The 95nm wide silicon fin is patterned into
negative resist using a top-down electron-beam-lithography
(EBL) technology followed by state-of-the art hydrogen-
bromide silicon plasma etching. A gate oxide of 8nm is formed
by dry thermal oxidation at 1000°C in a tube furnace, resulting
in a final silicon width of approximately 80nm. The channel
height is measured to be 60nm by means of atomic force
microscopy. Contact holes are patterned via EBL and opened
by wet chemical etching in buffered hydrofluoric acid. All
metal contacts are structured via a dedicated double-layer lift-
off resist system [6]. The metallization of the S/D contacts and
gate electrode is realized by evaporating 70nm of nickel
capped with 180nm of aluminum. The S/D-contacts are
salicided in a horizontal tube furnace at 500 °C for 5 minutes in
a forming gas ambient consisting of 90% N2 and 10% H2.
During this forming gas treatment the metallic nickel reacts
with the silicon surface forming a nickel-silicide mid-gap
Schottky-barrier contact to the silicon nanowire [7]. Note, that
during the entire fabrication process no doping process step is
required, making this process independent of issues as random
dopant fluctuation, complicated drain constructions which
simplifies processing.
III. DEVICE CONCEPT
A. Basic operation principles – the back-gate transistor
The fabricated NWFET can be thought of a superposition
of two interlaced field-effect-transistors. The global transistor
is formed by the back-gate electrode (or handle wafer) over the
full nanowire length (cf. Fig. 1b), including the Schottky-
barrier S/D regions (cf. Fig. 1d and Fig. 1f). This device
operates as an enhancement-mode transistor and provides the
charge carriers (i.e. holes or electrons) directly at the interface
between top-silicon and back-gate oxide (BG-OX). Fig. 2
shows a schematic drawing of the band-diagram at different
back-gate potentials. In Fig. 2a, a negative voltage is applied to
the back-gate electrode. Provided that the back-gate is biased
sufficiently negative, holes are drawn from the Schottky-barrier
contacts into the silicon channel and accumulate near the top-
Si/BG-OX interface forming a P-type channel. In Fig. 2b, a
sufficiently positive voltage is applied to accumulate electrons
near the top-Si/BG-OX interface to form an N-type channel.
Accordingly, the nanowire channel type can be selected
electrically to be either P-type or N-type by applying negative
or positive back-gate bias. Note that the low background boron
doping of the SOI substrate is neglected for all following
considerations as it is assumed to be of no relevance for the
principal device operation. This assumption has been
Figure 2. Schematic band-diagrams illustrating the influence of the
back-gate on the NWFET channel region. a) Accumulation of P-type
carriers (holes) at the top-Si/BG-OX interface: VBG << 0V, VDS = 0V. b)
Accumulation of N-type carriers (electrons) at the interface: VBG >> 0V,
VDS = 0V.

Figure 3. Simulated charge carrier densities within the NWFET-
structure for positive and negative back-gate polarity along the x-
direction. a) VBG = -20V resulting in P-channel operation and b) VBG =
+20V resulting in N-channel operation. The front-gate voltage was set to
VFG = 0V.
|IDS| (A) @ VDS= |2V| and floating FG
1x10 channel which is shown for inverted polarities in Fig. 4b. The
1x10
1x10 junctions. This has been confirmed by measurements of the
1x10
1x10 contrast to conventional MOSFETs with n+/p+ S/D regions, the
Figure 4. Simulated band-diagrams displaying back-gate control within
1x10 charge carrier injection at the Schottky junction explicitly
the a-NWFET channel region for different source/drain potentials.
a) Accumulation of P-type carriers (holes) at the top-Si/BG-OX
-20
interface: VBG = -20V, VS = +2V.
b) Accumulation of N-type carriers (electrons) at the interface: V BG =
+20V, VS = -2V
confirmed by 2D simulations of the charge carrier densities for
positive and negative back-gate polarities. Figure 3 shows the
2D simulation results of the cross-section along the a-NWFET
channel which has been described in Fig. 2. The simulation
indicates that a negative back-gate voltage of VBG = -20V is
sufficient to accumulate large hole densities at the top-Si/BG-
OX interface (Fig. 3a). The same holds for a positive back-gate
voltage of VBG = +20V where electron-accumulation occurs
(Fig. 3b). The density of charge carriers in both cases is in the
range of several 1019 cm-3.
The origin of these charge carriers is directly related to
the nickel-silicide midgap Schottky-barrier contacts. For a
perfect midgap Schottky-barrier source (or drain) contact, the
barrier heights for holes and electrons are identical, i.e. φB,p ≈
φB,n (Fig. 4). Without the electric field provided from the back-
gate of the device, neither electrons nor holes can overcome the
barriers φB,p or φB,n at room temperature. These assumptions
have been confirmed by 2D simulations of the band structure
of the adaptive NWFETs (Fig. 4). The charge carriers
accumulating in the channel are originating from the
corresponding drain or source contact via tunneling through the
reverse biased and back-gate modulated Schottky-barriers.
Additionally, an increase of band bending at the metal-
semiconductor interface can be induced by the application of a
source/drain bias VDS ≠ 0V. In Fig. 4a the a-NWFET is
operating as P-type (VBG = -20V)
Holes Electrons
Figure 5. Measured back-gate characteristic of fabricated NWFET with
-6
1x10 midgap Schottky-barrier source/drain contacts. The front-gate electrode
-7
was intentionally left floating.
1x10
and a hole current can flow by applying a source/drain bias of
-8
1x10
VDS = |2V|. The same considerations apply for the N-type
-9
insets in Fig. 4 emphasize on the field-emission process of the
-10
charge carriers through the source/drain Schottky-barrier
-11
temperature dependence of the drain current, pointing towards
-12
tunneling-emission (field-emission) of charge carriers [6, 8]. In
-13
-14
avoids
-15 the-10
need -5for dopants.
0 5 10 15 20
BG
V (V)
The assumptions made for the back-gate transistor to
operate as a selector for the nanowire channel-type (i.e. N-type
or P-type) are further supported by measurements of back-gate
voltage sweeps presented in Fig. 5. In Fig. 5 the back-gate
voltage is swept from -20V < VBG < +20V at VDS = ±2V
resulting in a drain-source current IDS, which is plotted on
logarithmic scale. Around VBG = 0V no significant current flow
is observed from source to drain. However, with decreasing (or
increasing) in back-gate bias above |3V| a current flow occurs
between source and drain. In the left branch of the graph hole-
conduction is dominant while the right branch indicates
electron-conduction.
B. Basic operation principles – the front-gate transistor
Taking advantage of the electrically selectable device-type,
fully functional CMOS transistors can be obtained by means of
the front-gate control. The front-gate electrode forms the
second transistor of the previously described two transistor
structure. In contrast to the back-gate, the front-gate transistor
operates like a depletion-mode transistor, i.e. the front-gate
controls the amount of charge carriers traveling through the
channel. The front-gate is set-up in a π-gate-configuration (cf.
Fig. 1c and 1e) permitting excellent electrostatic control over
the whole channel cross-section to completely turn-off the
device [9] as illustrated in Fig. 6. By means of 3D simulation
the impact of the front-gate control on the charge carrier
concentration across the a-NWFET structure at the front-gate
along the z-direction is shown for both, p-type (Fig. 6a) and n-

Figure 6. Simulated charge carrier densities across the a-NWFET-
structure at the front-gate along the z-direction as a function of front-
gate bias for a fixed back-gate voltage.
a) For VBG = -20V resulting in P-channel operation and VFG swept from
-5V to +5V.
b) For VBG = +20V resulting in N-channel operation and VFG swept from
+5V to -5V.
type (Fig. 6b) is shown. For the PMOS device, the hole
concentration is reduced by more than 13 orders of magnitude
when sweeping the front-gate voltage from VFG = -5V to VFG =
+5V. Basically the same behavior is observed for the NMOS
device when sweeping the front-gate voltage from VFG = +5V
to VFG = -5V. Note, that at VFG = 0V the PMOS NWFET
device exhibits a higher carrier concentration than the NMOS
NWFET. This is due to the low boron background doping
mentioned earlier, since truly intrinsic SOI material is not
available.
By performing front-gate voltage sweeps at constant
positive or negative back-gate bias, the sub-threshold
characteristics of both P-type and N-type devices using one and
the same NWFET are measured as shown in Fig. 7. When
plotting the drain current IDS on a linear scale (cf. Fig. 7a) it is
obvious that the respective PMOS and NMOS currents differ
by a factor of approximately 2 which is due to the difference in
carrier mobility for electrons and holes, respectively. In fact
this result is consistent with the presence of field-induced P-
type and N-type NWFETs as discussed above. From Fig. 7b it
can be found, that the sub-threshold slope of the a-NWFET is
influenced by the applied back-gate voltage. Depending on
back-bias, the sub-threshold slope of the P-type NWFET is in
the range of 90 to 120 mV/dec. The values for the N-type
NWFET are slightly higher between 150 to 250 mV/dec, again
depending on back-gate bias conditions. The slightly
asymmetric threshold voltage and the higher sub-threshold
slope of the N-type NWFET are attributed to the low p-type
background doping of the commercial SOI substrate and
interface charges.
IV. RECONFIGURABLE CIRCUIT: A-NWFET INVERTER
Starting from individual devices, simple circuits have been
realized. E.g. a CMOS inverter represents the simplest form of
a logic gate where significant power consumption only occurs
during the switching transition, when PMOS and NMOS
1.2
VBG =+20V
1.0
|IDS| (µA) @ VDS= |2V|
P-Type NWFET
0.8
VBG=-20V
0.6
N-Type NWFET
0.4
VBG=-7,5V
0.2
VBG=+7,5V
0.0
-5 -4 -3 -2 -1 0 1 2 3 4 5
VFG (V)
a) b)
Figure 7. a) Linear plot of the input characteristics of one a-NWFET. P-
type and N-type are selected via back-gate polarity. The back-gate
voltage is stepped in |2.5V| steps from |7.5V| up to VBG = |20V|
b) Semi-log plot of sub-threshold characteristics of the a-NWFET.
Again, the back-gate voltage is stepped in |2.5V| steps up to VBG = |20V|
a) b)
Figure 8. a) Transfer characteristics (solid squares) of the CMOS a-
NWFE inverter. Output currents of individual devices (open circles and
triangles) and characteristic cross-current (open squares).
b) Reconfigurable CMOS-NWFET-inverter demonstration: Transfer
characteristics (solid squares) of the inverter with device (NW A) as P-
type NWFET and device (NW B) as N-type NWFET. Same structure
after back-gate bias and supply voltage polarities are interchanged:
Transfer characteristics (open squares) of the inverter with device (NW
A) as N-type NWFET and device (NW B) as P-type NWFET.
devices are partially turned on [2]. Using the adaptive-NWFET
technology, a CMOS inverter circuit has been realized as
shown in Fig. 8a. By applying a back-gate voltage of -20V to
the a-NWFET (NW A) the PMOS is defined and using a back-
gate bias of +18V the NMOS a-NWFET (NW B) is defined.
The back-gate voltage of the N-type a-NWFET has to be
slightly reduced to adjust the maximum output current values
and threshold voltage (open circles and triangles). These
matching a-NWFETs are used to demonstrate transfer
characteristic (solid squares) of the resulting adaptive CMOS
NWFET-inverter. The measured inverter transfer characteristic
confirms the functionality of the proposed dopant-free a-
NWFET device architecture for CMOS logic applications. The
current peak at the transition point indicates that the circuit is in
fact a true CMOS inverter and not a resistive load inverter.
Furthermore, by changing back-gate and supply voltage
polarities, NMOS and PMOS devices are interchanged but the
same inverter functionality can be obtained. First, in Fig. 8b the
resulting transfer characteristic (filled symbols) is shown with
device (NW A) working as P-type a-NWFET and device (NW
B) is used as a N-type a-NWFET. As before, a decent inverter
transfer characteristic is observed. After interchanging back-
gate bias and supply voltages, i.e. device (NW A) is selected to
work as N-type a-NWFET and device (NW B) as P-type.
Again a well-established inverter behavior is observed (open
symbols), as shown in Fig. 8b. However, a shift of the inverter
characteristic is observed. This deviation is related to process
variations of our university-level prototype technology.
 -7
1x10 -6
P-type channel formation 1x10
-8
before write cycle -7
1x10 after write cycle 1x10
-8
1x10
-9
1x10 -9
memory
1x10 window

IDS (A)
memory

|IDS| (A)
-10
P-NWFET
-10
1x10 1x10 window
 VBG = 3V -11
N-NWFET
-11
1x10
1x10
-12
1x10 P-NWFET
-12 before write cycle
-13
1x10 1x10 after write cycle
N-NWFET
-14
-13 1x10 before write cycle
1x10 after write cycle
-15
1x10
-10 -8 -6 -4 -2 0 -3 -2 -1 0 1 2 3
VBG (V) VFG (V)

(a) (b)

Figure 10. a) Measured back-gate characteristic of a P-NWFET device

Figure 9. Illustration of the proposed adaptive intrnsic NWFET-
memory device on MultiSOITM substrate. The a-NWFET memory
write and readout voltages are applied at the back-gate terminal.
Nevertheless, these results confirm the realization of a novel
reconfigurable adaptive CMOS-NWFET technology for logic
applications which is freely programmable via a simple change
of back-gate voltage polarity. The reconfigurable logic gives
the option to reduce the device count as was demonstrated
recently using a NAND/NOR circuits [10]. In other words:
performance increase by means of reducing circuit complexity
instead of increasing device speed (cf. paradigm (1) in the
introduction).
V. INTRINSIC MEMORY CELL OPERATION
Besides logic circuits, the adaptive-NWFET structure can
also be used for memory applications to store information.
Charge carrier injection and trapping in the back-gate oxide
(BG-OX) can be utilized to open the possibility for creating
non-volatile charge-trapping memory devices with the same
adaptive NWFET structure as illustrated in Fig 9. After charge
injection during the write cycle via Fowler-Nordheim (FN)
tunneling in combination with impact ionization into the thick
back-gate oxide [11], the trapped charge in the buried oxide is
expected to cause a shift in the sub-threshold characteristics of
the a-NWFET which can be used to define the logical state, i.e.
“0” or “1”. In Fig. 10a the hole-conduction branch (P-type) of
the measured back-gate characteristic of a NWFET is shown
before and after a memory write cycle VBG_cycle = 25V has been
applied to the back-gate. Open symbols in Fig. 10a represent
the initial device state when no programming of the memory
has been carried out. The cycle voltage is applied in such a
manner, that source and drain are grounded and a pulse of
VBG_cycle = 25V is applied for a time of tcycle = 276 ms. As a
result of the electron injection and charge trapping, the back-
gate turn-on voltage is shifted from -6V to -3V. By means of
capacitance-voltage measurements of pre- and post-write
cycles the amount of trapped charge carriers inside the oxide
has been estimated to Qtrap ≈ 5•1011cm-2 for the injection
conditions above. The electric field generated by the injected
electrons is now added to the external electric field provided by
the back-gate, shifting the turn-on voltage of the back-gate
transistor. As a result, the front-gate characteristics is also
affected and a shift in the threshold is observed as shown in
Fig. 10b. The same considerations apply for NMOS NWFET
operation, with interchanged voltage polarities. For the given
reading voltage, the difference in the current ratios of ≈ 104 can
be used to assign the stored logical “0” or “1” levels. The
prior and after programming. Filled symbols represent the measurement
before a memory write cycle. Open symbols show the measurement
after a memory write cycle has been performed. The back-gate turn-on
voltage is shifted by ΔVBG = 3V.
b) Measured front-gate characteristics to confirm memory effect of one
CMOS-NWFET, i.e. at both, P-NWFET and N-NWFET operation. A
back-gate write-voltage of VBG_cycle = |25V| is used and the readout back-
gate voltage is VBG_read = |15V|.
adaptive NWFET-memory cell can be erased by FN-tunneling
using the opposite polarity than for programming as usually
practiced in charge trapping memory cells [12]. In view of this
early stage of prototype fabrication, the relatively large and
stable memory window of 104 is very promising for CMOS-
NWFET memory applications. Further optimization is
possible, e.g. by reducing the relatively thick back-gate oxide
thickness down to levels of ultrathin buried oxide SOI wafers
(e.g. 10nm vs. 145nm), to allow for today’s low supply voltage
levels. Moreover, alternative memory device concepts with
optimized SONOS dielectrics [12] or floating back-gate
configuration, similar to floating gate memories [13] can be
imagined as embedded non-volatile memory application in this
novel adaptive nanowire CMOS technology.
When combining these intrinsic memory properties with
the adaptive logic device, paradigm (2) is obsolete (cf.
introduction). This will open a path for a “learning” transistor
to implement hardware-based neural network functionality [14]
for novel and efficient neural computing schemes.
VI. CONCLUSION
A simple and versatile device structure for logic and
intrinsic memory applications has been presented. The
fabricated adaptive nanowire field-effect transistor represents a
universal device which can be reversely configured as PMOS-
or NMOS-FET simply by the application of an electric signal
at the back-gate. Towards computing beyond Moore’s Law,
novel reconfigurable circuits can be envisaged where
performance is increased by means of reducing circuit
complexity instead of increasing device speed. Furthermore,
non-volatile memory function can be easily implemented in
this adaptive NWFET technology by means of buried oxide
charge injection and trapping. The intrinsic combination with
the adaptive logic device will open a path for a “learning”
transistor to implement hardware-based neural network
functionality for novel and efficient neural computing schemes.
Best of all, this approach is fully compatible with the
established silicon-CMOS integrated circuit technology so that
one can make use of the existing semiconductor industry.
 ACKNOWLEDGMENT [6] F. Wessely, T. Krauss, U. Schwalke, Proceedings of the ESSDERC, pp.
356-358 (2010).
The author would like to thank Frank Wessely and [7] E. Buchner et. al., Applied Physics A, 40, pp. 71-77, 1986.
Tillmann Krauss for their support. [8] S.M. Sze, Physics of Semiconductor devices, Wiley Interscience, 2007.
[9] J.-T. Park, J.-P. Colinge, C. H. Diaz, IEEE Electron Device Letters 22,
REFERENCES pp. 405-406 (2001).
[1] Kahng and Atalla, "Electric Field Controlled Semiconductor Device," U. [10] J. Trommer et al., IEEE Electron Device Letters, 35, 141, 2014.
S. Patent No. 3,102,230 (Filed 31 May 31, 1960, issued August 27, [11] U. Schwalke, M. Pölzl, T. Sekinger, M. Kerber, Micoelectronics
1963). Reliability, 41, 1007-1010 (2001).
[2] F. Wanlass and C. Sah, International Solid State Circuits Conference [12] Barbara De Salvo et al., IEEE TRANSACTIONS ON DEVICE AND
Digest of Technical Papers, pp. 32-33 (1963). MATERIALS RELIABILITY, VOL. 4, NO. 3, 377- 389, 2004
[3] Gordon E. Moore, “Cramming more components onto integrated [13] P. Cappelletti, C. Golla, P. Olivo, E. Zanomi, Flash Memories, Kluwer
circuits”, Electronics, 38 (8), p.114, 1965. Academic Publishers (2001).
[4] J. D. J. Frank, R. H. Dennard, E. Nowak, P. M. Solomon, Y. Taur, and [14] I. Aleksander, and H. Morton; An Introduction to Neural Computing,
H.-S. P. Wong, “Device scaling limits of Si MOSFET’s and their Chapman and Hall, London (1990).
application dependencies,” Proc. IEEE, vol. 89, pp. 259–288, 2001.
[5] SOITEC,http://www.soitec.com/en/products-and-services/
microelectronics; available online, 2013.
 Initial Proposal of a Framework in the Context of
Cyberdefense to Articulate Controls and Metrics
Associated
Pablo G. Sack, Jorge S. Ierache
Facultad de Informática Ciencias de la Comunicación y Técnicas Especiales
Universidad de Morón
Cabildo 134 Morón, Argentina
sackpablo@gmail.com
Abstract—This paper provides a group of definitions,
documents, metrics and framework from different authors and
sources in order to state a model of framework linking the
security controls upon the basis of the referenced documents as a
result of the research. It is a proposal to attain a series of
progressive goals starting from a basic goal of introducing the
cyberspace topic, cyber warfare and cyber defense. Satisfied the
first goal we move into the universe of cyber defense, specifically
in security controls where the most important and complete
documents referring this topic were selected. This research will
be used for making a large database containing the security
controls specified in each document. As a next goal, the security
controls are complemented by an additional research of security
metrics that will provide the necessary tools to measure
performance of each security control. These metrics will allow
the assignment of a score which will be used in a table of
equivalence proposed by the authors in order to standardize all
units of measurement into a single measure. The ultimate goal
will be to use this unique measure obtained of each metric to
generate an initial proposal for a framework that facilitates the
management and diagnosis of security in the context of cyber
defense.
Keywords—Cyberspace; Cyber warfare; Cyber defense;
Informatic Security; Metrics; Security Controls
I. INTRODUCTION
The constant human evolution has led to the creation of a
new artificial space that is transverse to natural areas (land,
sea, air, space) in which humanity develops naturally. The
new environment or space called cyberspace is fully integrated
in human activities, does not recognize physical boundaries or
nation-states, and allows the development of operations in
terms of interoperability between systems in different natural
environments. From this new artificial space a battleground
emerges which is present at the door of every computer
infected by malicious software (malware) and may be an
element of large-scale attack to a country / agency or company
without the awareness of its owner. Internet is not a safe place
as there are people ready to commit cybercrimes either for fun,
for money, for political reasons, etc.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
A. Cyberspace
The word cyberspace comes from the combination of the
word "cibernao" - from the Greek meaning "to pilot a ship"
and is commonly used in the field of networks - and space
giving the idea of being piloted or sailing on a virtual world.
Today cyberspace is given a broader meaning as it connotes
all information available digitally, how the information
transferred from one hand to another and the communities that
uses that kind of information. [1].
Cyberspace in the context of the battlefield has been
growing and becoming more difficult to define and defend [2],
unlike our historical understanding of the war, cyberspace as a
battlefield has asymmetrical advantages to the attacking. In
this order it is appropriate to consider that:
 The intelligence and deception are the critical
principles of war in cyberspace.
 Cyberspace is vast and it is easy to stay undercover or
undetected.
 The attackers have the cost advantage against the
defenders.
 The use of cyberspace does not require intensive
materials or capital.
 It may launch attacks from anywhere in the world,
even the same country or attack network providing
anonymity. This point represents an extra difficulty
because it is not easy to locate the aggressor and may
even be independent groups (nongovernmental) or
citizens themselves.
 The attacker can take all the time necessary for the
acquisition of resources or opportunities to ensure the
achievement of its objective.
Different definitions of Cyberspace may occur. For the
National Institute of Standards and Technology (NIST) [3]
Cyberspace is defined as global domain combined of
interdependent networks like internet, telecommunications
networks, controllers, and similar.
The European Union defined as: virtual space along which
the electronic data of the world's computers. For the
International Telecommunication Union is defined as: a place
created through the interconnection of computer systems via
the Internet. United States (National Military Strategy for
Cyberspace Operations) defines it as [4]: The domain
characterized by the use of technology and different
techniques to store, modify and share data via networked
systems. The authors´particular vision of cyberspace proposes
a definition of cyberspace as: an artificial field transverse to
the natural environments (land, sea, air and space) that makes
up the virtual space of interaction in which human activities
themselves are developed and machines related to the creation,
processing, publishing, storage, modification and use of data,
information and digital knowledge, in a distributed context
(cloud computing) through interdependent and interconnected
global networks, public, private, hybrid, software and
firmware machines, whose distinctive character is given by the
use of information and communications technologies.
B. Cyberwar
Cyberwar is defined by Richard Clarke as the actions taken
by a nation that penetrates computers or networks of other
nations with the purpose of causing damage or rupture of the
same. United States (DoD) Cyber warfare (CyW) — Any
action aimed to impose our will on rival systems. [5]
C. Cyber defense
The Cyber defense is defined by NATO as: Developing the
ability to minimize the odds that a cyberattack has success [6].
The defense means [7] the ability to stand in the way of
penetration, such an attempt to identify and thwart through the
interruption and suspension of the work. For this reason, the
computer systems are configured in such way to limit access,
permissions, perform backups and data recovery. For the
United States (Comprehensive National Cybersecurity
Initiative (CNCI)): The defense of all types of threats through
the improve of US counterintelligence capabilities and
increasing the security of critical information.[8]
II. SECURITY CONTROLS
In the context of cyberdefense for its study it is proposed
to build a strong and solid database of security controls taking
as backbone and structure the document “The Critical Security
Controls for Effective Cyber Defense” [9]. Complementing
the database of security controls with the statements in
“Strategies to Mitigate Targeted Cyber Intrusions” [10] and in
the document of “Security and Privacy Controls for Federal
Information Systems and Organizations (800-53 Rev.4)” [11].
It takes as a starting point and reference the document
“Framework for Improving Critical Infrastructure
Cybersecurity” [12]. In this order those documents are
described below.
The document of Critical Security Controls for Effective
Cyber Defense [9], was prepared by a community of more
than a hundred government agencies, private companies and
experts. These controls act as a solid platform to build other
standards; they can be used to create a roadmap or starting
point for those who do not know where to start. The main
objective of this document is to reduce the initial attacks by
increasing levels of security to configure the devices,
identifying compromised devices, interruption of attackers
control over the implanted malicious code and find a way that
everything is integrated in a feedback loop (continuous
evolution). The critical principles on which this document [9],
is based are:
1) The defense receive a feedback from the offense: Use
the expertise of attacks to learn from these events and build
more effective defenses.
2) Prioritize: Investing in controls that provide a greater
degree of risk reduction.
3) Metrics: Use known metrics that let us communicate
the results with management, auditors, computer group and
the security group of the organization.
4) Diagnosis: continuous measurements carried out to
assess and check the efficiency of the measures.
5) Automation: To provide reliable and consistent
measurements is requiered the automation of defenses.
In addition, this document is divided into twenty critical
security controls. These subcontrols are categorized in order to
implement controls in a progressive manner and staggered as
follows:
a) Quick Wins: These types of security controls are used
to prevent the most common attacks and provide the first layer
of defense;
b) Visibility and attribution: The objective of these
types of security controls is to make measurements to improve
the security controls already implemented.
c) Improved information security configuration and
hygiene: These subcontrols have the objective to reduce poor
security practices and their impact.
d) Advanced subcontrols: These subcontrols generally
are more expensive or use techniques that are hard to install
and apply.
The document Strategies to Mitigate Targeted Cyber
Intrusions” [10] is divided into three documents by the
Australia Department of Defense and consists of as follows:
1) Mitigation Strategies 2014: It contains a brief
introduction and a poster with the summary of thirty five
strategies to mitigate cyber-attacks.
2) Mitigation Strategies 2014 Details: Describes each of
the thirty five strategies to mitigate cyber-attacks. A control
code recommended is mentioned.
3) Information Security Manual 2014 Control: describes
the controls codes mentioned in the previous document. While
this document presents thirty five strategies to mitigate cyber-
attacks and remarks that the primary step is to develop and
implement the first four strategies called as essential. Those
strategies consist of the generation of application whitelisting,
application patches, vulnerabilities patches of operating
systems and restricting administrator privileges.
The document “Security and Privacy Controls for Federal
Information Systems and Organizations (800-53 Rev.4)” [11],
is part of a “life cycle of security” proposed by NIST forming
a framework of risk management. The “life cycle of security”
proposed consists of six stages with the corresponding
document:
 Categorize Information Systems: 800-60.
 Select Security Controls: 800-53 Rev.4.
 Implement Security Controls: 800-160.
 Assess Security Controls: 800-53A.
 Authorize Information Systems: 800-37.
 Monitor Security Controls: 800-137.
The “Framework for Improving Critical Infrastructure
Cybersecurity” [12], is classified into three parts:
1) Core: Clasified all security controls on five concurrent
and continuous functions: Identify, Protect, Detect, Respond
and Recover. Each of these functions are divided into
categories, subcategories and Informative References.
2) Implementation levels: Provide a context for how an
organization sees cyber security risks and processes to manage
that risk. It consists of four levels: Partial, Risk Informed,
Repeatable and Adaptive.
3) Profile: Each organization will have a particular
objective or policy that will decide which categories and
subcategories take from the framework.
III. METRICS
This section describes the most significant features of the
most recent documents of metrics that apply to the
development of the Framework, in this order the following
documents are highlighted: The CIS Security metrics” [13],
complementing with documents “Measurement Frameworks
and Metrics for Resilient Networks and Services: Technical
report” [14], “Cyber Resiliency Metrics V 1.0 Rev.1” [15] and
“Measurement, Identification And Calculation Of Cyber
Defense Metrics” [16].
In this order The CIS Security Metrics [13], presents an
initial set of metrics proposed by a group of experts in the
matter that gives their contribution from different
perspectives. The metrics presented on this paper as
considered as a starting point and some metrics mentioned are:
Mean-Time to Incident Discovery, Number of Incidents,
Mean-Time between Security Incidents, Mean-Time to
Incident Recovery.
The Document de Measurement Frameworks and Metrics
for Resilient Networks and Services: Technical report [14] has
the objective to create a technical source of information on
resilience metrics. Some metrics mentioned are: Tolerance
(expresses the degree to which a system can continue to
operate for faults or challenges), Operational mean time
between failures, Fault report rate, Packet loss, Mean time to
repair.
The document Cyber Resiliency Metrics V 1.0 Rev.1 [15],
this white paper describes the initial representative set of cyber
resiliency metrics identified by the Assessment task of the
Resilient Architectures for Mission Assurance and Business
Objectives (RAMBO) project under the FY11 MITRE
Innovation Program. This set of metrics is expected to evolve
in response to practical experience as well as to the ongoing
refinement of the cyber resiliency engineering framework.
This document contains a list of 272 metric techniques that
will nourish the database to be used in security controls
involved in each grid of the framework proposed. Some
metrics from this document are: Percentage of cyber resources
that are properly configured, Number of attempted intrusions
stopped at a network perimeter, Number of attempted
intrusions deflected to a honeypot. From the document
Measurement, Identification and Calculation of Cyber Defense
Metrics some metrics mentioned [16] are: Defensive
Efficiency and Defense Factor, Percent of Successful Attacks,
Percent of Partially Successful Attacks, Mean Number of
Attack Disruptions, Time Spent per Phase,.
IV. FRAMEWORK PROPOSED MODEL
Based on the controls and the framework presented in the
previous section, it presents the proposal of a Framework
Model to facilitate the gradually controls implementation and
track the progress through metrics in each control, providing
reference documents for each control involved and suggested
links. The framework model is based on the security controls
division proposed for Improving Critical Infrastructure
Cybersecurity [12]. This document proposes a security controls
division in five phases. These phases agglutinate security
controls that comply the objective of each phase (see Fig. 1).
Fig. 1. Phases of “Framework for Improving Critical Infrastructure
Cybersecurity
A. Iterations within the framework
In addition to security controls division in each phase, the
framework consists of four iterations that cross all phases of it
(see Fig. 2). These iterations will aim as described in the
categorization of subcontrols discussed in the previous section
of the document The Critical Security Controls for Effective
Cyber Defense [9], thereby achieving formation of a grid
which will place the controls according to the phase (Identify,
protect, detect, respond to, recover) and iteration. The four
iterations are classified as:
 a) Quick Wins (First iteration): Controls, Metrics and
recommendations to provide immediate protection without
increased cost or infrastructure change.
b) Visibility and Attribution measures (second
iteration): greater controls, metrics and recommendations for
monitoring and controlling the effectiveness of such controls.
c) Improved Setup Information Security (third
iteration): Reduce weak security practices, vulnerabilities and
their impacts, proper maintenance and control settings.
d) Advanced subcontrols (fourth iteration): Minimize
chances of success of determinedly attackers, installation of
high cost controls.
PHASE 1 - IDENTIFY
ITERATION 4 ITERATION 3 ITERATION 2 ITERATION 1
CONTROLS, METRICS AND RECOMMENDATIONS TO PROVIDE IMMEDIATE PROTECTION WITHOUT FURTHER COST OR CHANGE OF
TIGHTER CONTROLS, METRICS AND RECOMMENDATIONS FOR MONITORING AND CONTROLLING THE EFFECTIVENESS OF SUCH
REDUCE WEAK SAFETY PRACTICES, VULNERABILITY AND ITS IMPACT. MAINTENANCE AND APPROPRIATE CONTROL SETTINGS
REDUCE CHANCES OF SUCCESS OF DETERMINATION ATTACKERS. INSTALLATION OF HIGH COST CONTROLS.
Fig. 2. Sequence to follow for the proposed framework
B. Framework Function
Each grid formed between a phase and iteration will
contain an "X" number of controls to be implemented and
comply. The sum of compliance with each of the controls
gives the level of compliance of each grid. The framework
will also have the possibility to calculate the degree of
compliance not only of each grid but also by phase and
iteration. Fig. 3 shows an example of the proposed interface,
where you can see a classification by colors indicating
compliance levels.
Fig. 3. Interface Framework proposal exploitation
C. Function of charge interface
As explained above, each grid will consist of a list of
controls in the charging interface. This interface will show for
each security control a link to supporting documentation and
reference to give a value of compliance of the control as well
as links to recommended sites for more information. The
charging interface facilitates parameterize each of the controls
to the needs of each user by selecting the column Mandatory /
Recommended, the weighting that the user assigned or deems
appropriate, as well as assign a score of compliance. On Fig.
4 it can be seen an example of the grid (phase 1 and iteration
1). Remember that the purpose of Phase 1 is to identify
"allowed company resources (SW, HW, PC’s, etc.) and
iteration 1 corresponds to controls, metrics and
recommendations to provide immediate protection without
SEQUENCE TO FOLLOW FOR THE PROPOSED FRAMEWORK
increased cost or infrastructure change.
PHASE 2 - PROTECT PHASE 3 - DETECT PHASE 4 - RESPOND PHASE 5 - RECOVER
When the user accesses the charging interface he will find
INFRASTRUCTURE the list of controls for this grid with recommended standard
settings that can be modified according to the particular need
for it. The user’s first step will be to determine which controls
CONTROLS
are mandatory o recommended. Later he must assign a
weighting of each control (the value will be contained between
0 and 1). With regard to the score the user will complete the
degree of compliance of each control and the input value will
be contained between 0 and 100.
The total of the grid is determined as follows:
1) It shall be calculated the subtotal (maximum) of each
security control: Subtotal (Maximum) = Entered weighting X
100.
2) It shall be calculated the subtotal (real) of each
security control: Subtotal (real) = Entered weighting X score
obtained.
3) The total of the grid: equal to the ratio between the sum
of the subtotals and the sum of the maximum subtotals,
multiplied by 100 to get the percentage.
Following the example of Fig. 4 the total of the grid is
equal to: (407,60/462)* 100= 88,22%.
Fig. 4. Charge of Interface
 D. Example of how to get the score of a security control.
Due to the large amount of security controls that the
framework database will contain, it is necessary to create a
metrics database that allows to translate the information
recovered by the security controls to put a fulfillment score.
The metrics presented in the third section can provide the
expected result and full security control or provide the
necessary information to help you get the final result expected.
For example, the security controls that have the objective
of reducing the exposition time of a vulnerable system before
the installation of security patch, will involve a metric where
the unit of measurement is time. This unit of measure is
inconsistent with the desired weight for the proposed
framework, which should be assigned to our consideration a
scale of weights determined representing 100% to less than 1
hour waiting period and 0% when the time exceeding 96
hours.
The critical control that will use on the example is the
Critical Security Control 4, subcontrol 5 (CSC4-5) [9] and its
equivalent control “0303”, corresponding to the application of
security patches [10]. This security control said: “Implement
automated patch management tools and updating software
tools for Operating Systems as well as for all software /
application tools”. This subcontrol will be on the second
iteration and second phase (protect) of the proposed
framework (I2F2).
The way to determine if the automated tool implemented
works within the established parameters will measure the
average installation time. For this Security control, the metric
mentioned in [13] called MTTP (Mean Time To Patch) will be
used. This metric said that meanwhile the patches are installed
faster, the average patch time will be shorter and the
organization will spend less time with vulnerable systems.
This metric determines the amount of hours between the patch
availability date and the installation date for each full patch.
The last results are averaged with the quantity of patches
completed within a certain time (hours per patch).

As mentioned above, each of the security controls
contained within a grid have a weight of 0-100. Because of
this, it is necessary to convert measurements and collections of
information obtained from the software and hardware
applications under a single unit of measurement. Three
different documents that mention metric proposals that can be
used to match the expected result in the context of securit are
presented in section three. On the example case, the score (%)
is obtained with a conversion table (See table 1) to translate
the number of hours per patch in percentage numbers.
TABLE I. CONVERSION TABLE
1% 96 hours per patch
72 hours per patch
25%
50% 48 hours per patch
75% 24 hours per patch
100% < 1 hours per patch
E. Conclusions and future work lines
An exploratory process is done in terms of the definitions
of cyberspace, cyber warfare and cyber defense.
The various proposals for standards and safety checks
were presented, but the need is highlighted with tools to obtain
a safety rating systems.
The initial proposal provides the basis for the development
of a tool that benefits the process of evaluation of critical
security controls considered according to published reports
and rules, which are listed in this article. At the same time it
generates an initial proposal for exploitation framework that
facilities the diagnosis and management of security in the
context of cyber defense.
In future work lines it is to articulate the assessment of the
controls and analysis with the use of knowledge-based
systems to assist the determination of diagnosis up a unit
within the framework of cyber defense.
References
[1] David A Umphress, T.C. “El Ciberespacio”: ¿Un Aire y un espacio
Nuevo?, Air&Space Power Journal, august 2007.
[2] Geers, K. A, “Brief Introduction to Cyber Warfare”. Common Defense
Quarterly, pp16-17, Spring 2010.
[3] Eric A. Fisher “Federal Laws Relating to Cybersecurity: Overview and
Discussion of Proposed Revisions” June 2013.
[4] Air Force Cyber Command Strategic Vision (Available:
http://www.dtic.mil/cgi-bin/GetTRDoc?AD=ADA479060)
[5] Alford Jr., Lionel D, CYBER WARFARE: PROTECTING MILITARY
SYSTEMS., 2000. (Available:
https://turnitin.com/viewGale.asp?r=53.8117643938158&svr=06&lang=
en_us&oid=25740519&key=427186604ab0a08b8bbb71d05c16f6f4)
[6] C. Czosseck, R. Ottis, K. Ziolkowski, 4TH INTERNATIONAL
CONFERENCE ON CYBER CONFLICT (CYCON 2012), C. , pp162,
(Available:https://ccdcoe.org/sites/default/files/multimedia/pdf/CyCon_
2012_Proceedings_0.pdf)
[7] Tabansky, L. “Basic Concepts in Cyber Warfare”, Military and Strategic
Affairs , vol. 3. N° 1 , pp. 75-92 , May 2011.
[8] The Comprehensive National Cybersecurity Initiative,
(Available:https://www.whitehouse.gov/issues/foreign-
policy/cybersecurity/national-initiative)
[9] The Critical Security Controls for Effective Cyber Defense. Version 5
(CSC-5),Council on CyberSecurity, 2013. (Available:
https://www.sans.org/media/critical-security-controls/CSC-5.pdf)
[10] Strategies to Mitigate Targeted Cyber Intrusions, Department of Defense
Intelligence and Security of Australian Government, february 2014.
(Available: http://www.asd.gov.au/infosec/mitigationstrategies.htm)
[11] Security and Privacy Controls for Federal Information Systems and
Organizations (800-53 Rev.4) de National Institute of Standards and
Technology,april 2013.
(Available:http://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP
.800-53r4.pdf)
[12] Framework for Improving Critical Infrastructure Cybersecurity de
National Institute of Standards and Technology, february 2014.
(Available: http://www.nist.gov/cyberframework/upload/cybersecurity-
framework-021214-final.pdf)
[13] The CIS Security Metrics. The Center for Internet Security, november
2010.(Available:https://benchmarks.cisecurity.org/tools2/metrics/CIS_S
ecurity_Metrics_v1.1.0.pdf)
[14] Measurement Frameworks and Metrics for Resilient Networks and
Services, Technical report ENISA, february 2011.
[15] Deb Bodeau, Rich Graubart, Len LaPadula, Peter Kertzner, Arnie
Rosenthal, Jay Brennan “Cyber Resiliency Metrics V 1.0 Rev.1”.,
MITRE , april 2012.
[16] Juan E. Sandoval, Suzanne P. Hassell. “Measurement, Identification
And Calculation Of Cyber Defense Metrics”, MILCOM, pp. 2174-
2179,october 2010.
Teaching Climate Change Through Blended Learning
A case study in a Private Secondary School in Mauritius
Teena Gomes
University of Technology, Mauritius
La Tour Koenig, Pte. aux Sables, Mauritius
Abstract—Teaching Climate Change’s concepts are hard and
challenging to impress in the minds of adolescent learners in a
traditional classroom environment. Designing appropriate
instructional strategies with the best visual experiences enable
learners to grasp the complex principle behind the concept and
stimulate their interest. Hence, technology can be seen as a
panacea. In this regard, a study was conducted to explore Form
II learners’ view on learning Climate Change with blended
learning approach. The theoretical context of this research is
underpinned by socio-constructivism. As methodology, an action
research was carried out by implementing an interactive
educational software, designed based on learners’ requirements.
A questionnaire was administered to 30 mixed ability learners
from a gender-mixed private secondary school with the aim of
collecting their learning experience. An active participation of the
learners during the face-to-face session demonstrated that
through blended learning, they had a clearer understanding and
they appreciated the concept with a larger improvement in
learners performance. Hence, this active and constructive
learning strategy encouraged collaboration and cooperation, and
celebrated the autonomy of the learners. Moreover, learners were
motivated to take action and address problems in the community,
developing their civic responsibility for sustainable development.
Keywords—Climate Change; Blended learning; Socio-
constructivism
I. INTRODUCTION
Environmental education cuts across multiple disciplines
such as the Sciences, Arts and Humanities. The study of
Climate Change has been under-emphasised in most secondary
education curricula and the capacity to support the integration
of climate change science is still relatively limited [1].
Imparting the science of climate change is not always easy as
the content itself is difficult to comprehend. As such,
technology helps learners to visualise far field Climate Change
impacts and improve their global perspectives. Also, since
Climate Change consists of theoretically complex models and
hypotheses involving interdisciplinary connections, the
traditional classroom is unable to accommodate for the varying
learning styles of learners, to provide adequate explanation,
and suitable and timely feedback in this digital era [2]. The
traditional didactic strategies used to educate learners about
issues of global climate change, are inappropriate. Traditional
ways of teaching, which are largely based on the infusion of
knowledge, are inapt in helping learners to apply knowledge in
understanding the real issues of everyday life [3]. Hence,
innovative instructional approaches and techniques must be
designed. The science of Climate Change demands a variety of
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Shireen Panchoo
University of Technology, Mauritius
La Tour Koenig, Pte. aux Sables, Mauritius
s.panchoo@umail.utm.ac.mu
strategies, techniques and aids to teach because the content
itself is quite intricate to comprehend. Learners demonstrate
difficulty grasping the complex systems involved such as the
greenhouse effect [4]. Learners, in fact, face scepticism, as they
build mental models of Climate Change concept which are not
aligned with models taught in class. They enter the class with
preconceived notion on Climate Change and unfortunately,
educators are only able to probe the accuracy of learners‟ prior
knowledge after summative assessment [5].
Scholars contend that with blended learning, learners are
better able to clarify any lesson at any time and enrich
socialisation opportunities while, simultaneously, enables
educators to share content, lesson plans, and other curriculum
components, thus shifting the role of the educator from content
deliverer to guide [6, 7]. In this sense, educators are moving
away from a dogmatic approach to a more engaging and
interesting approach that encourages critical thinking rather
than just fact accumulation. Moreover, to ensure that learning
takes place through blended learning, an appropriate pedagogy
must be adopted. Evidence shows that socio-constructivism is a
good fit for blended learning because it stimulates active
learning whereby learners are guided through a
conceptualisation of the implications of environmental change
by exploring the effects of Climate Change and analysing real-
world problems [8, 9]. The paper aims at using an appropriate
instructional design to guide learners in reconstructing their
knowledge on the concept of Climate Change. The objectives
of this paper are to adopt blended learning approach to teach
Climate Change science and to celebrate the autonomy of the
learners.
II. LITERATURE REVIEW
A. Learning with technology
The traditional learning approaches have long been used
and proven effective in the past. Nevertheless, they have
recently been questioned in their ability to provide the learner
with "rich" environments and "authentic" experiences of
learning [10]. Traditional classrooms are space bound,
feedback given to learners is delayed and the educator makes
limited use of visual aids and materials. With the advent of
Internet and prevalence of computer technology cutting across
demographic boundaries, educators have turned to online
learning [11]. Visual materials such as graphics, photographs,
concept maps, films, computer and television images support
written texts in all disciplines and make learning fun, motivate
learners and enrich their imagination [12]. However, the

separation of action of teaching and the reaction of learning
causes learners to feel isolated, confused, frustrated and at
times, learners‟ interest in the subject and learning
effectiveness are reduced[13]. As such, to get the best of both
learning methods, blended learning is considered, as it provides
a viable option for learners who seek the flexibility of online
courses but also request for personal contact with the educator
and other learners in classroom setting. Indeed, blended
learning makes pedagogically significant use of the Internet
and other technological tools while reducing seat time (time
spent in the classroom) [14]. Per se, blended learning improves
pedagogy, increases access to knowledge, fosters social
interaction, increases the amount of educator‟s presence during
learning, enhances ease of revision and provides learners with
greater control over their pace of learning, instructional flow,
selection of resources, and time management[15]. Proponents
of this approach believe that this system includes a committed,
sustained, and well thought-out implementation plan, which
combines appropriate technology with traditional classroom
interaction leading to better outcomes for learners [16, 17].
Consequently, due to difficulties in understanding the complex
and dynamic principle behind Climate Change, educators have
to design appropriate instructional strategies and methods [18].
It is vital that learners develop a holistic understanding of the
concept. So, Khan‟s framework as illustrated in Fig. 2.1 can
serve as a guide to plan, develop, deliver, manage, and evaluate
blended learning programs for Climate Change.
Fig. 2.1: Khan's Blended Learning Framework [19]
B. Socio-Constructivism
In seeking to improve learners‟ performance and
satisfaction, educators may consider adopting a socio-
constructivist approach to teaching and learning. Accordingly,
it is believed that learning is an active and constructive process;
learners not only construct knowledge, but the knowledge they
already possess affects their ability to gain new knowledge
[20]. Learners are believed to be enculturated into their
learning community, based on their existent understanding,
through their interaction with the immediate learning
environment. In a school setting, three elements are always
active: the environment, the learner, and the educator [21]. This
conception of the social and the individual being closely
interconnected, functionally unified, constantly interacting, and
the change and development in one relentlessly influencing the
other, provides a valid explanation for both social and
individual change. The socio-constructivist pedagogy is
2
obviously crucial for helping learners understand Climate
Change concepts, that is, in meaning-making, and guides
learners to build new mental models [22]. Educational
researchers have advanced that socio-constructivism is the
most appropriate pedagogy to be adopted for blended learning
[5, 8, 9].
III. METHODOLOGY
The purpose of this study is to adopt a blended learning
approach to Climate Change science by designing and
implementing an interactive educational software to mixed
ability learners so as to cater for the complexities of the
concept faced in traditional classroom. Khan‟s blended
learning framework guided the design, implementation and
evaluation processes of the software which is further discussed
in the Research Technique section further below. The research
was supported by socio-constructivist learning theory.
A. Research Design
Researchers and practitioners in fields of blended learning
regularly use quantitative methods to measure the effectiveness
of this model among learners. So, based on recent literature
reviews, a questionnaire with questions similar to those as in
[23] was administered to 30 mixed ability learners of Form II
from a mixed-gender private secondary school. The
questionnaire used included a five-point Likert-type scale
anchored at Strongly Disagree=1 and Strongly Agree =5,
indicating learners‟ disagreement or agreement with each item.
The questions were worded based on Khan‟s Blended Learning
Framework in relation to the features of the pedagogical tool
and the satisfaction of learners. The questions were set to
evaluate the effectiveness of learning in an online environment
and the interface design, the appropriateness of the resources
adopted for learning the science of Climate Change, the
accuracy of the instructions and the clarity of the evaluation
criteria.
B. Research Technique
An action research was carried out for the study as it
seemed more appropriate. “Educators collaborate in evaluating
their practice and try out new strategies to render their practice
more consistent with educational values they espouse and thus
develop a shared theory of teaching by research practice” [22]
Fig 3.1 illustrates the processes undertaken for the action
research for designing, implementing and evaluating the
educational software for learning Climate Change science.
Fig. 3.1: Action Research [24]

Step 1: Making the commitment
A requirement analysis was done before designing the
software. The lesson was performed in the traditional
classroom and “en amont” (prior) a diagnosis evaluation was
carried out to measure the performance of the learners and
simultaneously to identify weaknesses and the needs level
(Maslow‟s level of needs) of the learners. Following the
evaluation, it was observed that learners demonstrated
difficulty in grasping the concept and its related aspects such as
the greenhouse gasses and their relationship to global warming.
The ambiguities of the cumulative effects of hundreds of
environmental factor blurred the concept of Climate Change.
So, it was motioned that the traditional teaching method of
Climate Change might be inefficient for the digital natives to
understand such complexities.
Step 2: Designing the study
Based on the results of diagnosis evaluation (performance
of learners), the requirements and needs‟ level of the learners -
identified through their weaknesses - , a prototype of an
interactive educational software was designed. The educational
tool consisted of inter-connected pages and links to ease
navigation and interaction. The software followed an
interdisciplinary approach and pages were enriched by video
clips from YouTube, pictures, simple texts and interactive web-
based content to allow learners to have a closer appreciation of
the concept. Instructions were given on each page so as the tool
becomes user-friendly. Alongside, Thorndike‟s 3 laws of
learning were considered in the design process. The interface
design aimed at making the learner, physically, mentally and
emotionally ready to learn (Law of Readiness). Animated as
well as static pictures both in the foreground and the
background of the pages projected the learners in an
environment conducive to learning Climate Change science
with sounds of water and natural hazards. Links were also
inserted to direct learners to specific Climate Change concepts
such as global Warming and its consequences or causes of
climatic changes Video Clips on YouTube. In addition, the
simple texts on the pages allowed learners to further enhance
what they saw from the Video Clips. Navigating icons allowed
the learners to access one page to another.
Likewise, at the end of each the sub topic, there was a
recapitulation of what has been covered and, only at this point,
that the learner was able to engage in the evaluation tasks. The
evaluation tasks comprised of MCQs, Cloze text, gap-fill,
matching and crossword which were created through Hot
Potatoes 6 and linked to the educational tool. Hence, to enable
learners to construct new knowledge and master the concept,
the formative evaluation at the end of each sub topic was set
with immediate feedback (Law of Exercise). Learners were
provided with appropriate feedback immediately for both
correct and wrong answer. They were allowed to work out the
evaluation task thrice. Then, at the end of the topic on Climate
Change, a summative evaluation was set to reinforce learning
through an on-task and field activity (Law of Effect).The
educational software was uploaded on Google drive.
Step 3: Making Sense of Experience
3
In trying to solve problems of traditional teaching, “en
aval” (after) the diagnosis evaluation, the educational software
was administered to 30 mixed ability learners of Form II from
a mixed-gender private secondary school. The learners were
asked to download the educational software and worked out the
concept at home at their own pace. Prior to this, in class, the 30
learners had to create their Gmail account in the Computer Lab
and were given a demonstration of how to download and use
the educational tool. They had a period of one week to
complete the task. During the class session (35 minutes), after
the deadline, the educator was asked to clarify and enhance
aspects of Climate Change which seemed ambiguous for the
learners. Learners responded positively in class and shared
their own research work to their classmates. At the end of the
classroom session, the learners were administered a
questionnaire to fill in to measure their appreciation and views
on the educational software. The questionnaire was collected
one day after its administration. Lastly, learners‟ degree of
understanding of the concept of Climate Change was measured
through a written evaluation ranging from lower order to
higher order questions (Bloom‟s Taxonomy). The performance
from the second evaluation was compared to that of the
diagnosis evaluation.
Step 4: Improving your practice
Results from the questionnaire, highlighted the new needs
of the learners such as synchronous online communication with
educators and among learners. Learners wished to
communicate with their peers and educator while they were on
task so that they could ask questions and share their views. The
needs were reviewed and rethought of.
Step 5: Beginning again
At this point, the new needs of learners were clearly
identified and an innovative strategy was re-designed. A forum
for online synchronous discussion was integrated. Eventually,
the change was implemented for better learner performance
and satisfaction.
IV. RESULTS
Prior to designing the educational software, the needs of the
learners were measured through a diagnosis evaluation on the
concept of Climate Change to identify the needs, weaknesses
and performance of learners. Hence, based on the results
obtained, an interactive educational tool was designed and
implemented to cater for the needs of the learners. The
educational tool provided the learners with animated pictures,
simple texts, interactive web-based content and Video Clips as
well as evaluation tasks to test the knowledge of the learners at
the end of each sub topics. After the implementation of the
educational tool, the learners‟ views on the software were
captured through a questionnaire and consequently, their
performance was measured through a summative evaluation
(class test).
Data collected from the learners during the survey were
compiled, analysed and discussed. The results provided
insights on the learners‟ view on learning Climate Change
through blended learning. The effectiveness of blended

learning was assessed by measures of learners‟ success in
terms of performance and satisfaction [25].
Results demonstrated that 80% of the learners responded
that blended learning helped them in learning the concept. It
was observed that both the interactive educational tool used
and the face-to-face session, were helpful to gain understanding
of the concept. Nevertheless, the remaining 20% found the
learning model unsuitable. The percentage structure is shown
in Fig 4.1.
10%
10% Extremely
Very
50%
Not Enough
30% Does Not
Fig. 4.1: "Blended learning helps in the learning of the concept"
Indeed, learners usually react positively towards blended
learning; however, they use traditional approach as the simplest
form of discussion on the content and for receiving feedback
from the educator. A similar result was obtained as in [26].
Nonetheless, the insufficient level of two-way interaction
among learners and the educator during their online session
resulted into 10% of the learners not finding the learning model
appropriate and helpful. The remaining 10% responded
negatively to the blended learning approach as they favour
mostly the face-to-face interaction. Since they have always
been cradled by the traditional learning model, they needed the
constant personal contact with the educator.
Undeniably, though most learners reacted positively to the
educational software in terms of clarity and
straightforwardness, a mean of 4.80 reflected that statistically
the majority of the learners agreed that there should be
synchronous online communication so as to enable discussion
of the concept among themselves and to ask for clarification to
the educator while doing the work. It is to be pointed out that
learners meet their educator only twice a week for a period of
35 minutes each. Also, results from the survey (mean=4.50)
demonstrated that the instructions were precisely given to
facilitate navigation from page to page, to view the video Clips
from YouTube and to perform the evaluation activities.
Learners agreed that the design of the educational software;
video clips, audio, animated pictures, simple texts and
interactive web based content were well selected for enhancing
the learning of Climate Change science, making it clear and
simple to understand (mean=4.05) and as a results the
resources met their needs (mean=4.04). Learning in the online
environment was a new experience for the learners and
eventually they were more eager to evolve in this virtual space
being digital natives.
Moreover, for both right and wrong answers, learners were
provided with immediate feedback. In case of a wrong answer,
they were given the opportunity to work out the activity again.
Research advanced that in most circumstances feedback that is
immediate and specific, results in better learning [27, 28].The
concept was made closer to reality and increased their interest
4
on the subject. A majority of learners (mean=4.04) advanced
that they were able to work anytime and anywhere at their own
pace. So, they were eager to search for more information on
Internet (“I‟m bored when I‟m learning online”, mean=1.73).
They, besides, became more aware of the climatic changes in
their own environment by building on what they see and what
they learn. Actually, the summative activity was a real
challenge which they were eager to tackle. Learners initially
had a low level of comprehension on the concept however after
the online lessons which were further sustained during
classroom interaction enabled them to better appreciate the
concept. In fact, learners are motivated by activities that reflect
those in „real world‟ [29]. Thus the learners agreed that that
blended learning model is an efficient one (mean=4.00)
In terms of personal contact during face-to-face session, the
mean of 3.73 demonstrated that learners have divergent views.
Some learners agreed that personal contact with the educator is
imperative during the learning process while others do not. As
contended in [29], personal contact with the educator is
essential to provide help and support. Educators play a major
role in social presence. Indeed, the results of the summative
evaluation were compared to the diagnosis evaluation. There
was significant improvement in all learners‟ performance. In
fact, the 10% of the learners who did not favour the blended
learning approach did perform better.
This conclusion is analogue to the result of [30] conducted
over a much larger sample in their study on the use of blended
learning.
V. CONCLUSION
The paper aimed at designing and adopting a blended
learning model to guide and reconstruct the knowledge of
learners on the science of Climate Change and to celebrate the
autonomy of the learners. The results from this action research
suggested that the combination of well-designed and supported
blended model for learning the Climate Change science enables
learners to appreciate the concept and mounted their interest to
investigate more on the concept on their own; the climatic
happenings in their environment and worldwide. Learners
developed an overarching philosophy that will help them set
goals for sustainable futures and develop a world view. Hence,
it is observed that the autonomy of the learners was celebrated.
Likewise, it was concluded that learners‟ performance
improved with this learning model compared to that of the
traditional classroom. As advanced in [31] “there is convincing
evidence that online learners do just as well if not better than
learners in face-to-face courses” Indeed, the findings which
emanated from the action research also provided a platform for
educators to enhance learners‟ knowledge on Climate Change
through blended learning. The role of the educator in blended
learning environment is essential and should be considered for
the success of blended learning programs. Educator‟s role in
blended Learning environment includes management,
guidance, providing feedback, and evaluation [19]. The
greatest value of an effective blended learning model is that it
fosters a more interactive, collaborative and engaging learning.
Nevertheless, this study needs to be supported by larger-
scale studies to reveal the impact and to have a broader

appreciation of using Blended Learning approach in
conceptualising Climate Change. Also, learners could be
encouraged along with the objectives of the curricula to
participate actively in their community life to make the change
for a sustainable world.
ACKNOWLEDGMENT
I would like to thank the learners at St Helena‟s College for
their participation. I would like to extend my gratitude to Ms.
V. Bundhoo and Mrs. N. Domun for their assistance during the
research. Last but not the least, a special thanks to Mrs. M.
Dhuny for her help and support.
REFERENCES
[1] D.K. Bardsley, and A.M. Bardsley, A.M. “A Constructivist Approach to
Climate Change Teaching and Learning Geographical Research”
Geographical Research, Volume 45, Issue 4, 2007
[2] E. Aladag, and B. Ugurlu, “Global Climate Change Education in
Turkey”, 2008. Unpublished paper
[3] V. Papadimitriou, “Prospective Primary Teachers‟ Understanding of
Climate Change, Greenhouse Effect, and Ozone Layer Depletion”,
Journal of Science Education and Technology, Vol. 13, No. 2 , 2004
[4] National Research Council, “Climate Change Education in Formal
Settings, K-14: A Workshop Summary”. A. Beatty, Rapporteur. Steering
Committee on Climate Change Education in Formal Settings, K-14.
Board on Science Education, Division of Behavioral and Social Sciences
and Education, 2012, Washington, DC: The National Academies Press.
[5] F. Movahedzadeh, “Improving students‟ attitude toward Science through
Blended Learning”, Science education and civic engagement 3:2, 2011
[6] E. Aboukhatwa, “Blended Learning as a Pedagogical Approach to
Improve the Traditional Learning and E-Learning Environments”, The
Second International Arab Conference on Quality Assurance in Higher
Education, 2012.
[7] E. Banados, “A Blended-learning Pedagogical Model for Teaching and
Learning EFL Successfully Through an Online Interactive Multimedia
Environment”, CALICO Journal, Vol. 23, No. 3, 2006
[8] A. Koohang, and K. Harman, Open source: “A metaphor for e-learning.
Informing Science”, The International Journal of an Emerging
Transdiscipline, 8, 75-86, 2005, Retrieved from
http://inform.nu/Articles/Vol8/v8p075-086Kooh.pdf
[9] A. Koohang, “A learner-centered model for blended learning design”,
International Journal of Innovation and Learning, 6(1), 76-9, 2009.
[10] A. Relan, and B.J. Gillani, “Web-based instruction and the traditional
classroom: Similarities and differences”. In Khan, B. (Ed.), Web-based
instruction (pp. 25–37), 1997, New Jersey: Educational Technology
Publications.
[11] M.T. Stone, and S. Perumean-Chaney, “The Benefits of Online
Teaching for Traditional Classroom Pedagogy: A Case Study for
Improving Face-to-Face Instruction”, MERLOT Journal of Online
Learning and Teaching Vol. 7, No. 3, September 2011
[12] A.E. Bozdogan, “The effects of instruction with visual materials on the
development of preservice elementary teachers‟ knowledge and attitude
towards global warming” The Turkish Online Journal of Educational
Technology, volume 10 Issue 2 , 2011
5
[13] J. Humphries, Gauging Faculty Attitudes Toward Online and Hybrid
Learning. Journal of Applied Computing 5 (1): 28–32., 2009
[14] A. deNoyelles, C. Cobb, and D. Lowe, “Influence of Reduced Seat Time
on Satisfaction and Perception of Course Development Goals: A Case
Study in Faculty Development”, Journal Of Asynchronous Learning
Networks, 16(2), 85-98., 2012
[15] Y. Blieck, M. de Jong, and L. Vandeput, “Blended Learning for Lifelong
Learners in a Multicampus Context (MuLLLti)”, L. Leuven University
College, Belgium, 2012
[16] C.J. Bonk, and C.R Graham, eds, “The Handbook of Blended Learning:
Global Perspectives”, Local Designs, 2006, San Francisco: Pfeiffer.
[17] Garrison, D.R., and Vaughan, N.D., “Bended Learning in Higher
Education: Framework, Principles, and Guideline”, 2008, San Francisco,
CA: John Wiley and Son.
[18] C.M. DiEnno, and S.C. Hilton, “High school students‟ knowledge,
attitudes, and levels of enjoyment of an environmental education unit on
non-native plants”. The Journal of Environmental Education. 37(1), 13-
25., 2005
[19] B. Khan, “Managing ELearning Strategies: Design, Delivery,
Implementation and Evaluation”. Information Science Publishing, 2005,
London.
[20] E. Etkina, and J.P Mestre, “Implications of Learning Research for
Teaching Science to Non-Science Majors.” SENCER Backgrounder
PRESENTED AT at SSI 2004
[21] V.V Davydov, translated by S.T Kerr, The influence of L. S. Vygotsky
on educational theory, research, and practice. Educational Researcher,
24 (3), 12-21., 1995
[22] M. Turuk, “The relevance and implications of Vygotsky sociocultural
theory in the second language classroom”, ARECLS, Vol.5, 244-262,
2008
[23] V. Aleksić, and M. Ivanović, “Blended Learning in Tertiary Education:
A Case Study”, Sun SITE Central Europe, Vol. 1036., 2013
[24] F. Rust, and C. Clark, HOW TO DO ACTION RESEARCH IN YOUR
CLASSROOM, Taking Action with Teacher Research, Heinemann
PressG. Eason, B. Noble, and I.N. Sneddon, “On certain integrals of
Lipschitz-Hankel type involving products of Bessel functions,” Phil.
Trans. Roy. Soc. London, vol. A247, pp. 529-551, April 1955.
[25] C.C. Wai, and E. Lim Kok Seng, “Measuring the effectiveness of
blended learning environment: A case study in Malaysia”, Educ Inf
Technol (2015) 20:429-443, 2013
[26] I. Мiliszewska, “Transnational Education Programs: Student Reflections
on a Fully-Online Versus a Hybrid Model”, Victoria University,
Australia, 2007.
[27] L.M. Jeffrey, J. Milne, G. Suddaby, and A. Higgins, “Blended learning:
How teachers balance the blend of online and classroom components”,
Journal of Information Technology Education: Research, 13, 121-140,
2014.
[28] S.D. Miller, “How high- and low-challenge tasks affect motivation and
learning: Implications for struggling learners”, Reading & Writing
Quarterly, 19(1), 39-57, 2010.
[29] A.R. Artino, and J.M. Stephens, “Academic motivation and self-
regulation: A comparative analysis of undergraduate and graduate
students learning online”, Internet & Higher Education, 12(3/4), 146-
151., 2009, doi: 10.1016/j.iheduc.2009.02.001
[30] S. Reasons, K. Valadares, M. Slavkin, “ Questioning the Hybrid Model:
Student Outcomes In Different Course Formats”. JALN, Vol. 9, No. 1,
2009
[31] A.W.T. Bates, and A. Sangra, “Classroom assessment techniques”,
2011, San Francisco: Josey Bass.
 Performance Analysis of Parallel CBAR in
MapReduce Environment
Sayantan Singha Roy Chandan Garai
Department of C.S.E Department of C.S.E
NITTTR, Kolkata NITTTR, Kolkata
Kolkata, India Kolkata, India
sayantansingharoy@outlook.com chandangarai@hotmail.com
Abstract—Clustering of data set is a very contemporary
problem for handling big data and parallelizing the process of
clustering helps in improving efficiency for applications which
involve frequent searching. Various clustering techniques are
used for grouping of data set and CBAR is one such very
frequently used technique used for different applications.
Parallelization of CBAR is very necessary for handling Bigdata
and Hadoop MapReduce platform provides a suitable
environment to improve efficiency for any problem dealing with
huge volume of data by using appropriate segmentation
technique. In this work, we designed and developed a few
algorithms for implementing CBAR using MapReduce technique
and tested the results in different clusters of up to 4 nodes.
Significant improvement has been observed and analysis and
explanation on these results have also been presented in our work
with suitable example.
Keywords—Association Rule Mining; BigData; CBAR;
Clustering; Hadoop; MapReduce
I. INTRODUCTION
Volume of data increases day by day. In last two years alone
90% of data in the world has been generated [1]. To analyze
this kind of huge decentralized data, engineers need to process
these data in distributed way and this can be achieved by
parallel algorithms which can handle big data very efficiently.
Some tools and techniques (i.e. Hadoop, MapReduce) are
available which help to achieve above goal.
Clustering of data in general helps in searching data
elements in an efficient way. This preprocessing of huge data
into several clusters is an accepted way to improve search
related activities and being used in various applications [2][3]
for quite some time. Clustering is a tool which also helps in
data miningby acting as a tool to gain insight into the
organization of data. Data mining and clustering become
challenging due to high volume of input data.
In this paper, we design and develop parallel algorithm to
implement CBAR mining into Hadoop MapReduce
framework. The proposed algorithm is organized in
<key:value> pair in order to execute in parallel manner on
above mentioned framework effectively and the result also
demonstrate that our algorithm can handle large data sets.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Ranjan Dasgupta
Department of C.S.E
NITTTR, Kolkata
Kolkata, India
ranjandasgupta@ieee.org
II. REVIEW WORK
A. Clustering and Data Mining
We do consider both the topics of clustering and data mining
in our review work as our proposed parallel CBAR algorithm,
(the serial version of which is a prominent clustering
technique) can be used efficiently in data mining applications
particularly where the volume of data increases very rapidly.
We first discuss the major clustering technique and in the
second part of our review work we emphasis on data mining
issues.
Clustering is a task of grouping some data elements,
objects which has similar properties or attributes. These
groups containing similar types of elements are called clusters.
The main goal of clustering is to reduce the distance between
the elements reside in same cluster and maximize the distance
between two separate clusters. Clustering helps us to identify
the density of data from that we can identify a pattern. There
are many clustering algorithm or clustering model available.
The well known clustering algorithms include: i) K-means
algorithm, ii) Agglomerative hierarchical clustering, iii)
Support vector machine, iv) Fuzzy clustering. In K-means
algorithm [5][17], K denotes number of clusters the algorithm
will divide a particular data set. In this algorithm, K(number
of clusters) must be provided in advance. It clusters a data set
by calculating centroid of the data. Initially K numbers of
centroid is selected. Then we assign closest points to each
centroid to form a cluster and re compute the centroid of each
cluster. We repeat the above step until centroid position does
not change. The main drawback of K-means algorithm is if we
change the initial centroid positions, the final cluster will
change. K-means clustering is applied in various applications
like general word recognition system without using general
parameter, automatically detection of roads from GPS data
[15][16]. Agglomerative hierarchical clustering[6] process
starts with individual points represented as individual cluster
and iteratively merging to its closest clusters. This clustering
uses bottom up approach. This clustering is graphically
represented using a tree like structure where we can easily see
the cluster and sub cluster relationship. Whether a cluster will
be merged with another cluster or not is determined using
dissimilarity metric. The complexity of agglomerative
clustering is O(n3). Support vector machine (SVM) [7] is a
supervised learning model. For a given set of data it tries to
separate two clusters by determining the maximum gap
between them. In case of linearly separable clusters, it
generates a decision line. When a new point arrives, it mapped
the point into a cluster based on the distance line. In this way,
SVM maximize the inter cluster distance. For linearly non-
separable data, SVM generates hyper plane to separate two
clusters. This hyper plane acts as a decision plane. The
maximum distance between training data set and hyper plane
symbolize a good classification. Based on some data set we
often see that some data can belong to more than one cluster.
In this scenario, fuzzy clustering is used.
With the increase volume of data, we need parallel
clustering algorithms. The above mentioned algorithms have
some drawbacks in terms of processing large scale data. They
assume all data set are stored in some particular area of
memory and data do not arrive continuously. In order to
cluster the data which arrive continuously, K. Govindarajan et
al proposed an algorithm [3] on which they tried to cluster the
student based on performances in terms of accuracy,
efficiency, quality. Generally learning efficiency is calculated
based on some pre defined questions. But it depends on many
factors like learning style, prerequisites to learn the topics,
peer interaction etc. Big data is used to handle the continuous
data which collected during the learning process. Continuous
data consists of video analysis, activity monitoring, social
networking activities. This data analysis helps to identify the
growth of learner‟s competency.Authors proposed a clustering
algorithm based on unsupervised learning which can handle
continuous arrival of data. Learning efficiency is calculated by
four attributes of a student i.e. student id, accuracy, efficiency,
quality. The algorithm calculates the local best particle, local
position of the particles and global best particle and global
best position of the particle in order to analysis the learning
efficiency.
W. Zhao et al proposed parallel K-means algorithm which
executed on map-reduce programming model[8]. Map-Reduce
consists of two main functions: i) Map function, ii) Reduce
function. The input data set on which cluster will be done is
distributed to all node where all map function can access it. So
that distance can be calculated in parallel manner. Input data
set is also organized in key-value pair. Key is the byte of
record and value is a string of content of the record. Input of
map functions are: i) an array which contains centers of the
clusters, ii) key and value from the input file. Map function
generated a key-value pair where key is the index of the array
position where closest center position belongs and value is the
string of all information. This [8] introduced a combiner
function between map and reduce which calculated the partial
sum of the values assign to the same cluster. It is done in the
local memory of each node. So this information is not been
distributed over communication channel. In Reduce function,
the partial sum of combine function‟s output is taken as a
value from each node and key is the index value of the array
position. Reduce function calculates the total number of values
associated in same cluster by adding the partial sum collected
from each node. The number of records reside in same cluster
is also stored in an array from it calculated the new centers for
the next iteration.
Now we discuss some related research work on data
mining [13][19] and association rule mining. R. Agarwal et al
proposed apriori association rule algorithm[1] for generation
of association rule in items sold bysupermarket. This
algorithm is applied on transactional data. It is widely used
algorithm in data mining. The problem is divided in two
phases. In phase I, the item sets which are above minimum
support are selected. Minimum support is used to discard the
less frequent candidate item sets from frequent item set.
Itemsets are categorized according to their length i.e. the
number of item present in the itemset. Let, Ip is an itemset
where p is number of items present in the itemset. Let DB is a
database which contains multiple transactions of items. Each
transaction hasa unique ID. Firstly all possible combination of
itemsets of length 1 (I1) is generated. This is called candidate
key (C1). The itemsets which are above user-defined minimum
support are called frequent 1-itemsets. In next step, the
candidate 2-itemsets are generated using the element of
frequent 1-itemset. The itemsets which are above minimum
support are called frequent 2-itemset. This process continues
until candidate set is null. In phase II, rule among frequent
itemsets has been defined. It determines the confidence of an
item in the database i.e. how frequent an item is bought with
other set of items. The main drawback of above mentioned
algorithm is it scans the whole database every time to
calculate frequent item sets.
Yuh-Jiuan Tsay et al proposed cluster-based association
rule mining [2] technique to generate frequent itemsets in
reduced database scan. It improves the performance of
association rule mining. In order to generate the frequent
itemset, the proposed algorithm creates cluster tables first.
Clusters are created based on the length transactions i.e. if
length of transaction or items bought on the particular
transaction is „i‟, then the transaction would belong in the ith
cluster. So, the maximum number of cluster will be the longest
transaction present in the database. Cluster table contain
transaction ID along with item sold on that particular
transaction. Then it generates candidate 1-itemsets. Candidate
n -itemsets means all possible combination of itemsets contain
exactly n number of items. Each candidate itemset is cross
checked with the equivalent cluster table for minimum
support. If a particular itemset support is greater than
minimum support then the itemset is included in large itemset.
If support of an itemset is lower than minimum support on the
same cluster table, then it is checked with the upper level of
cluster i.e. itemset In is first checked with nth cluster, if the
support of In is lower than minimum support then it‟s support
is checked with (n+1)th cluster. If the aggregated support is
greater than minimum support then it is added in frequent
itemset. This process continues until frequent itemset is null or
end of cluster table is reached. This process breaks the whole
database into some cluster. Clusters are mutually exclusive.
Let frequent I -itemset is FI , initial cluster table is N and end
cluster table is M . When the FI is generated the corresponding
candidate I -itemset (CI) is compared with ith and (i+1)th to Mth
cluster tables. But it is not compared with Nth to (N - i)th
cluster. In this way, this algorithm reduces the database scan.
III. SCOPE OF WORK
MapReduce [12][18] is a programming model which
process large amount data in parallel way on large clusters of
machines [4][14]. MapReduce program mainly consists of two
functions i.e. map function and reduce function as described in
Fig. 1.Map function takes value as input and generates
key:value pair. When all the values get a key, this
programming model groups all the values together according to
their keys.This is the job of combiner module. The output of
the combiner module becomes the input of reduce function.
Reduce function takes a key and list of values as an input.
Reduce function processes on the input and generates output as
per requirement. Users define map and reduce function and the
runtime architecture automatically distribute the task, take care
of machine failures, handles communication among different
nodes, balance the load among different nodes. Hadoop
provides MapReduce runtime system along with a distributed
file system which provides high fault tolerance and scalability.
Hadoop distributed file system replicate the data across the
node which increases availability of data. The file system uses
TCP/IP for communication. There are five kinds of server
available in hadoop as shown in Fig. 2. Name node, data node,
secondary name node handle data storage, retrieval and fault
tolerance. Job tracker and task tracker handle map reduce
computational part[9][10].
Fig. 1 MapReduce Programming Model
Fig. 2 Different Kinds of nodes in hadoop
In case of applying existing association rule algorithm, the
performance is degrading when it is applied in very large data
sets [2]. The performance is degrading because the whole
database is scanned repeatedly to generate candidate keys. In
cluster based association rule (CBAR) mining[3], the database
is scanned only once to generate cluster table. The main
advantage of clustering in CBAR is the less number of cluster
scan is required in order to generate a frequent item set. To
generate a frequent N-item set, Nth cluster is searched directly.
Lower level of clusters (i.e. N-1 cluster) scanning is not
required. This reduces the search space. In case of sequential
clustering as mentioned in CBAR, the time complexity to
cluster a data set is quadratic in nature. Let n is number of
records presents in the transactional data set and m is the data
element present in a record. For a large data set where n
crosses peta threshold, m become very minimal compare to n.
The processing time increases along with n due to thequadratic
nature.
The existing clustering algorithms consider the input data
set is static in nature with time (i.e. the volume of input data set
would not change with time).If the input data set size is n unit
and q unit size is increased over t1time. The existing algorithm
would process the n amount of data firstly to create cluster.
After t1 time, it process (n + q) amount data to create output
cluster. It process n amount of data repeatedly though it is
processed beforehand and n is also very high in terms of
volume. Thus the processing time decreased dramatically. In
case of K – means algorithm the centroid positions will change
if new data points arrives after processing. So, to entry of new
set of data, the entire set of data is reprocessed. On the other
hand, CBAR processes the newly arrived data separately and
append the new data with the previously created clusters
according to the matching. So it can handle dynamic data.
In this paper, we parallelized cluster table generation and
also apply association rule in parallel manner in order to
increase performance to some degree by arresting the growth
of the quadratic function. We generate the cluster table by
scanning the input data set one time, and then we
accommodate the user defined query or association rule
multiple times. In this way, we need not to scan the whole
database multiple times which is time consuming. We can go
straight to the cluster based on the query itself. All clusters
scanning are not required. To incorporate the parallelism and
manipulate very large data set, we adept MapReduce
framework developed by Hadoop.
IV. PROPOSED SOLUTION
Dealing with terabytes of data in normal computing
scenario is not only time consuming but difficult too. To deal
with such high volume of data Cloud computing is a possible
solution. In Cloud computing environment, MapReduce is a
parallel processing framework which provides the architecture
to process Bigdata in parallel fashion. When the volume of data
is very large, clustering in CBAR become very time
consuming. So, MapReduce provides the technology by which
this large amount of data for clustering can be handled by
doing parallel processing.We incorporate only the cluster table
structure from CBAR [2] in this paper and user defined query
is transformed into association rule. In this way, user can get
desired output based on the query itself.
A. Algorithm
In our proposed algorithm, the input data set are stored in
Hadoop distributed file system (HDFS) [10] [11]. Input data
set is a comma separated value file containing series of
recordswhich has a unique id and data elements separated by
comma. Cluster creation phase generates clusters i.e. produces
different files which contain similar kinds of records generated
from input data set. Final output phase takes user defined query
as input and produces results with the help of clusters
generated by cluster creation phase.Our proposed solution has
two phases. Phases are: 1) Cluster generation phase, ii) Final
output phase.Details of the two proposed phases are presented
below.
The cluster creation phase has three functions i.e. Map
function, Reduce function, Multiple output file writing function
as shown in Fig 3.
Fig. 3 Cluster generation block diagram
Map Function: The main purpose of the function is to
determine which record will go in which cluster. The input
parameters of Map function are key and value. This function
reads records line by line from input set stored in HDFS. The
input key and value of Map function are line number of the file
to be read and a string of record of input file respectively. The
input file is available to all map functions. So, the cluster
determination can be executed in parallel way. A key is
generated from input value which determines the cluster
number. Value is the concatenated form of record id and
comma separated data elements. Output of map function is
above mentioned <key:value> pair. The algorithm is given
below:
map (Key, Value)
Input: Line number of file as Key, a record consists of record
id and comma separated data elements as Value
Output: _key:_value pair
1. Generate _key from Value;
2. Construct _temp as concatenate form of record id and
data elements;
3. Assign _temp into _value;
4. Output _key:_value pair;
5. End
Reduce Function: When all the records are processed by
map function, output of map function is feed into reduce
function. It aggregates the values according to common key
(i.e. Cluster number). Reduce function produces a list of values
for same key. The algorithm of reduce function is:
reduce (_key, _value)
Input: _key is Map function generated key, _value is Map
function generated value
Output: Key: _key:List_of_values pair
1. While (_value != NULL) {
2. List_of_values := Aggregate the values for same _key;
3. }
4. Output the _key:List_of_values pair;
5. End
Multiple output files writer function:This function takes
input a list of values against a key which is the output of above
mentioned reduce function. The list of values are written into
different files (i.e. cluster) based on the key. The algorithm is:
multiple_output_file_writer (_key, List_of_values)
Input: _key is the reduce function generated key,
List_of_values is the list of values generated by reduce
function
Output: Multiple files
1. Generate a unique file for _key;
2. List_of_values are written in the above generated file;
3. End
The final output phase has two functions i.e. Map function and
reduce function as shown Fig 4.
Map Function: This function takes user defined query
which is a string of value as an input. Map function generates
key and an intermediate value based on the query. This map
function contain association rule analyzer module which
generates a value by searching the output of cluster generation
block (i.e. collection of clusters) based on the input query. We
append the association rule analyzer generated value with the
intermediate value and generate a <key:value> pair. The
purpose of this function is to incorporate association rule,
search the clusters, and generate tmp_value in parallel manner.
The algorithm is:
map (key,_value,collection_of_clusters) 7. End
Input: key is the line number of the query, user defined query
which is a string of value as a _value, collection_of_clusters is
the clusters generated by cluster generation block B. Step by Step Execution
Output: _key:_Value pair We provide an example to further explain the proposed
algorithm. We applied the algorithm in student database. The
1. Generate _key from _value; input file contains student id and question numbers which are
correctly answered by the students. Here for the case study,
2. Generate intermediate_value from _value;
there are 10 students and each student has to answer 5
3. Search _value in collection_of_clusters; questions. The algorithm reads line by line from the input file.
In cluster generation phase, map function generates a key „3‟
4. Generate a Valasso from search result;
as number of correct answers done by the student is three for
5. _Value = Append Valasso with intermediate_value; the first line and the value is the concatenated form of student
6. Output _key:_Value pair; id and a series of zeros‟ and ones‟. Zero and one are placed in
such a way where zero represents wrong answer and one
7. End represents correct answer and the series length is five as
number of total questions is five. Question number one, three
Reduce Function: The output of map function i.e. _key and and five are correctly answered by Student S1. So, first, third
_Value are feed into reduce function. The input _Value is and fifth position in a set of five positions are represented as
split to fetch the association rule analyzer generated value one and others positions are zero. From this rule we get the
(i.e. Valasso) which is aggregated based on the same key. value S1 10101. The output <key:value> pair of the first line
Output of this function is the <key:value> pair where key is is <3 : S1 10101>. In this way, rest of the records in the file
the user defined query and value is the above mentioned are evaluated and <key:value> pair is generated as shown in
aggregated value of same key.
Table 1. The output of the map function is fed into reduce
function as input. Reduce function aggregates the values based
on the same key. As two values (i.e. S3 01000 and S10 00100)
are registered under key 1, these values are aggregated in
reduce function. Others aggregated values are also shown in
Table 1. In multiple file output writer phase, files are created
with a unique name. The name of a file is concatenated forms
of two things i.e. cluster and key, which is fed from reduce
function. For key 1, the file name is cluster 1. In this file the
list of aggregated values of key 1 is written. For every key a
file is created and list of values for the key are written on the
file.In final output phase, a user defined query and clusters
(i.e. generated from the above phase) are fed as input map
function. In this case, the query is „how many students have
correctly answered question number one and three?‟ Query is
fed into the function in the form of 1, 3. Map function
Fig. 4 System block diagram
generates the key, same as the query and an intermediate value
is generated as per above mentioned rule of map function
reduce (_key, _Value) incluster generation phase. Then, cluster is searched for
finding the match of the user defined query. In the query, the
Input: _key is Map function generated key, _Value is Map number of questions which are done correctly is two.So, the
function generated value algorithm starts searching from cluster 2. Here, numbers of
Output: key: V pair matches found are 0,2,1,1 in cluster 2, cluster 3, cluster 4 and
cluster 5 respectively and Valasso is generated for a particular
1. Initialize an array arr[]; cluster. Valasso mainly keeps track of the students
2. Split _Value and store in arr[]; whoanswered the questions mentioned in the query, correctly.
Here, Valasso is the concatenated form of the student id and
3. Fetch the association rule generated value (Valasso) who answered question number 1 and 3 correctly. In cluster 3,
from arr[] and store; the Valasso is S1 1 and S4 1.Rest of the Valasso„s are shown
4. _tmp = Valasso ; in Table 1. In the reduce function, the key is kept same as map
function (i.e. 1, 3). For the same key, Valasso is split based on
5. V = Aggregated value of _tmp based on same _key
the tab and stored in an array. Then, the match found count is
6. Output the key:V pair; fetched from the array and summed together. The total sum is
the final value of the reduce function. The output of the reduce
function is the key (i.e. the query itself) and value pair.

V. RESULT
We evaluate our proposed algorithm‟s performance in Fig.
5. Performance experiments were executed on a cluster of 4
computers. Each computer has 3.6 GHz quad core i7
processors and 4 GB RAM.Hadoop version 1.0.3 and java
version 1.8.0_31 were used to run map reduce job.
In Fig. 5, we have increased the number of records and
number of computers or nodes also increased for each set of
records. Here, we can see that time taken to process the set of
records is decreasing along with increase number of nodes.
This proves parallelism nature of our proposed algorithm. It
also shows when number records becomes high, single node
takes longer time but node 4 takes relatively same time for all Fig. 5 Speed up graph
the set of records. Time complexity to create cluster is
quadratic in nature as mentioned earlier, we arrest the growth In Table 2, we analyse the growth of execution time
of the function by increasing number of nodes. As the number between single node and four nodes. For 10 million records
of records increases, time spent by the algorithm is relatively with respect to 1 million records, execution time growth for
moderate.So we can say that our proposed algorithm can single node is 129%. Whereas execution time growth for four
handle large data. nodes is 60% for same condition.

Table 2: Execution time growth comparison between single node and cluster
Table 1: Case study
of 4 nodes

VI. CONCLUSION
More than 50% improvement has been observed in some
cases of applications which are quite impressive from
computational perspective. It has also been observed that, the
time for clustering becomes almost stationary for higher
number of nodes even the input volume of data has been
increased from 7 million to 10 million. Thus, CBAR being
very useful clustering technique, using it in cloud environment
(Hadoop MapReduce) for processing BigData has some
inherent advantages and may be used for various applications
as discussed in our work.
 REFERENCES
[1] Rakesh Agrawal, Ramakrishnan Srikant, “Fast algorithms for mining
association rules”. Proceedings of the 20th International Conference on
Very Large Data Bases, VLDB, pp 487-499, Santiago, Chile, September
1994.
[2] Yuh-Jiuan Tsay, Jiunn-Yann Chiang, “CBAR: an efficient method for
mining association rules”,Elsevier,Knowledge-Based Systems, pp. 99–
105, 2005
[3] Kannan Govindarajan , Thamarai Selvi Somasundara, Vivekanandan S
Kumar , Kinshuk, “Continuous Clustering in Big Data Learning
Analytics”, IEEE Fifth International Conference on Technology for
Education, pp. 61-64,2013
[4] Jeffrey Dean, Sanjay Ghemawat “MapReduce: Simplified Data
Processing on Large Clusters”, 6th symposium on operating systems
design and implementation,2004
[5] Hartigan, J. A.; Wong, M. A., "Algorithm AS 136: A K-Means
Clustering Algorithm". Journal of the Royal Statistical Society, Series C
28 (1): 100–108. JSTOR 2346830, 1979
[6] Rokach, Lior, and Oded Maimon. "Clustering methods." Data mining
and knowledge discovery handbook. Springer US, pp. 321-352, 2005
[7] Cortes, C.; Vapnik, V. ,"Support-vector networks". Machine Learning ,
Vol.20, Springer, 1995
[8] Weizhong Zhao, Huifang Ma, Qing He, “Parallel K-Means Clustering
Based on MapReduce”, First International Conference, CloudCom 2009,
Beijing, China, pp 674-679, Springer Berlin Heidelberg, ISBN 978-3-
642-10664-4, December 1-4, 2009
[9] "Running Hadoop on Ubuntu Linux (Multi-Node
Cluster)"http://www.michael-noll.com/tutorials/running-hadoop-on-
ubuntu-linux-multi-node-cluster/
[10] http://en.wikipedia.org/wiki/Apache_Hadoop
[11] Borthakur, D.: The Hadoop Distributed File System: Architecture and
Design, 2007
[12] Chandan Garai, Ranjan Dasgupta, "An Improved Job Scheduling
Algorithm by Utilizing Released Resources for MapReduce," Fourth
International Conference of Emerging Applications of Information
Technology (EAIT), doi: 10.1109/EAIT.2014.55,pp. 9-14, 2014
[13] Singh, A.; Chaudhary, M.; Rana, A.; Dubey, G, “Online Mining of data
to generate association rule mining in large databases”, Recent Trends in
Information Systems (ReTIS), IEEE, pp. 126 – 131, 2011
[14] Jing Zhang; Gongqing Wu; Haiguang Li; Xuegang Hu; Xindong Wu, "A
2-Tier Clustering Algorithm with Map-Reduce",ChinaGrid Conference
(ChinaGrid), pp.160-166,doi: 10.1109/ChinaGrid.2010.14, July2010
[15] Wilpon, J.G.; Rabiner, L., "A modified K-means clustering algorithm
for use in isolated work recognition" in Acoustics, Speech and Signal
Processing, IEEE, vol.33, no.3, pp.587-594, Jun 1985
[16] Wagstaff, Kiri, Claire Cardie, Seth Rogers, and Stefan Schrödl.
"Constrained k-means clustering with background knowledge."
In ICML, vol. 1, pp. 577-584. 2001.
[17] Kanungo, Tapas, David M. Mount, Nathan S. Netanyahu, Christine D.
Piatko, Ruth Silverman, and Angela Y. Wu. "An efficient k-means
clustering algorithm: Analysis and implementation." Pattern Analysis
and Machine Intelligence, IEEE, vol. 24 (7), pp. 881-892 , 2002
[18] Chu, Cheng, Sang Kyun Kim, Yi-An Lin, YuanYuan Yu, Gary Bradski,
Andrew Y. Ng, and Kunle Olukotun. "Map-reduce for machine learning
on multicore."Advances in neural information processing systems,
vol. 19,pp. 281, 2007
[19] Rajaraman, Anand, and Jeffrey D. Ullman. Mining of massive datasets.
Vol. 77. Cambridge: Cambridge University Press, 2012.
 Using a Location-Based Mobile Application to
Locate a Traditional Health Practitioner in South
Africa
Johannes, M. Kekana1 and Robert, T. Hans2
Department of Computer Science
Tshwane University of Technology
Soshanguve, Pretoria, South Africa
1
joe_kekana@yahoo.com
2
hansr@tut.ac.za
Abstract—Mobile technology has permeated many industries,
including the health industry. Different applications are used for
various things in the health industry, from checking blood
pressure to monitoring patients’ health remotely. Following a
legislation which was passed in South Africa to recognize
traditional health practitioners as practicing doctors many bogus
traditional health practitioners have emerged and are robbing
unsuspecting patients their money. To curb this problem the
South African government established a traditional health
council whose mandate is to ensure that all legally practicing
health practitioners register with it. However, prospective
patients still find it difficult to locate these registered health
practitioners. This paper proposes and presents a location-based
mobile application to enable prospective patients to search for
the nearest traditional health practitioner anytime and anywhere.
Keywords—mobile technology; location-based; mobile
application; traditional health practitioner; e-health.
I. INTRODUCTION
The advent of mobile technology has transformed how we
lead our lives. The number of mobile phones has risen to
being more than the number of personal computers in
developed and developing countries [1]. Mobile phones are
portable and enable people to perform internet based functions
that would have demanded use of personal computers [2]. For
some people mobile phones have become expression of their
personality [3].
Mobile technology has been adopted in many different
industries. Mobile applications are used on social networking
to enable people to interact [4]. Advertisement remains the
major source of revenue for mobile applications [5]. The smart
phones are also being used to print documents that they might
have been reading online and want to refer to later in hard
copy [6].
The use of mobile applications has also permeated the
health industry in a big way. Users are able to use their mobile
phones to establish the levels of blood pressure, check as to
whether they are not suffering from cancer, diabetes, etc. [7].
The South African government recently passed a legislation
aimed at recognizing traditional health practitioners as
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
practicing doctors, just like medically trained doctors. This
legislation is in line with what has been observed on the
ground - people have shown to have trust in consulting
traditional health practitioners who are based in their
communities [8]. As a result of these developments, dubious
traditional health practitioners have cropped up all over the
country conning people their hard earned money. In an
attempt to curb this problem the government has established a
council which is aimed at ensuring that all legally practicing
traditional health practitioners register with the council.
However, people who intend consulting traditional health
practitioners are still unable to find them as easily as possible
and also on the move. This paper therefore is aimed at closing
this gap by proposing a location-based mobile application
which enables a user to find the nearest traditional health
practitioner.
The remainder of this paper is structured in the following
manner. Section II discusses related work in mobile
applications and their usage. Section III provides insight on
the functionality of the proposed mobile application. Section
IV discusses the experimental results. Section V discusses the
benefits which will be realized by the proposed system.
Section VI presents the conclusion and future work.
II. RELATED WORK
In the 21st century mobile phones have become integral
part of almost everything we do. Mobile phones are found in
entertainment, education, health [7],[9] and so on. It is just
difficult to imagine one going through a day without having
used a mobile phone in one way or another. Mobile phones
owe their popularity to their ability to put information at the
users’ finger tips anytime and anywhere [10].
In recent years the health industry has seen a boom of
mobile applications which are meant to address various issues
including stress related matters [11], maternal mortality
related issues [12] and heart related illnesses [7], and many
more. In support of the abovementioned assertion [13] states
that mobile technology, has been leveraged to provide
healthcare services in many countries, with some of the
examples being emergency response systems for road traffic
accidents, disease surveillance and control etc. Mobile
technology and ubiquitous computing present enormous
opportunities in as far as e-health is concerned [14]. In
developing countries mobile technology can assist greatly in
addressing health related challenges, such as distance
monitoring of diabetes patients in Africa [15]. However, poor
infrastructure, cultural and ethical barriers, limited technical
capacity, lack of awareness and uncertainty on e-health costs
may be a stumbling block on the adoption of mobile
applications in the health industry in the developing countries
[15]. In spite of all these challenges the statistics indicate that
the adoption of mobile devices is in the rise [16] and no
studies which the authors are aware of have been done to
establish the drive behind this growth. The increase in mobile
technology usage in South Africa as a developing country
stands the proposed system in good stead.
In some instances users of mobile applications tend to
require that such applications be intelligent location-based
ones. According to [17] location-based applications are
expected to be in high demand in the 21 st century. There are
already mobile location-based systems which are used in the
health industry, for example, an application which determines
the change in ones sleeping patterns [17]. Mobile location-
based applications provide location-specific, context-based,
user-focused and personalized information. According to [18]
location-based applications provide services which are either
push- or pull-oriented. The pull location-based services are
user initiated, for example navigation services, while push
location-based services are automatically delivered to the user
even though his/her consent is required [18]. The mobile
location-based system presented in this research paper
provides a pull-oriented service which is dependent on user
preferences.
III. FUNCTIONALITY OF THE SYSTEM
A. System Architecture
The proposed traditional health practitioner mobile
application is created with an intention of assisting patients to
locate registered traditional health practitioners much easier in
South Africa. The application was developed using HTML5,
Cascading Style Sheets (CSS), JavaScript, Jquery, PHP and
MySQL. Moreover, the application takes advantage of the
Google Maps APIs to find the nearest location of the
traditional health practitioner. The system is a web-based
mobile application. The choice of going with a web-based
application as opposed to either a native- or hybrid-based
mobile application is premised on that a web-based mobile
application does not require a user to download the application
[10]. Fig. 1 presents the main components of the system.
Mobile Traditional Database
Phone Health Practitioner
Patient
System
Workstation
Council
Administrator
Figure 1. Traditional Health Practitioner System Architecture
The system has two sets of users - the traditional heath
council administrator and a patient. Each of these users can
perform certain functions as discussed in the next sub-section.
B. Users of the system
1) Council administrator
The council administrator is able to add new registered
traditional health practitioners on the registered traditional
health practitioners’ database using the traditional health
practitioner system. For every new traditional health
practitioner that is registered on the system a unique
system auto generated practice number is assigned to
him/her. The council administrator can modify the details
of a traditional heath practitioner and can also remove a
traditional health practitioner from the system.
2) Patient
A patient is able to search for a registered traditional
health practitioner using the name, suburb or distance
search parameters. If a patient searches for a traditional
health practitioner and does not specify the distance, by
default the system will search the database using the given
search parameters with the radius of up to 100 kilometres
from the location of the mobile phone user. That is, the
search will be limited to the radius of 100 kilometres from
location of mobile phone user. In the event where no
traditional health practitioner is found within the specified
or default radius, the system notifies the patient of this and
then displays the closest traditional health practitioner, if
there is one found. The search results will be arranged in
an ascending order based on the distance from the location
of the mobile phone user. Table 1 presents a list of all
different possible permutations of search results based on
the search parameters.
 Table 1 - Different Permutations on Search for a
Traditional Health Practitioner
Name Suburb Distance Results
Results based on the name, suburb and
Y Y Y distance taking into account the user’s
present location.
Results based on the suburb and
N Y Y distance based on the user’s present
location.
Results based on the distance from the
N N Y user’s present location.

Results based on the default

N N N distance/radius of 100 km taking into
account the user’s present location.
Results based on the name, suburb and
default distance/radius of 100km
Y Y N
taking into account the user’s present
location.
Results based on the name and the
Y N Y distance taking into account the user’s
present location.
Results based on the suburb and
default distance/radius of 100km
N Y N Figure 2. Traditional Health Practitioner Registration Interface
taking into account the user’s present
location.

IV. EXPERIMENTAL TEST RESULTS

The next sub-sections present experimental test results of
the system as performed by the two types of users, which were
mentioned above. The results presented and discussed below
are sample field test results that were carried out by the
authors.

A. Registration of a Traditional Health Practitioner

Once an application for registration as a traditional health
practitioner has been approved by the traditional health council
then the council administrator can register the new applicant
on the database of the Traditional Health Practitioners System
using the desktop interface of the application. Fig. 2 shows
how an applicant’s details were captured on the system’s
database by the council administrator.

B. Search for a Traditional Health Practitioner

A user of this mobile application can search for registered
traditional health practitioner whose details have been
captured on the system. The sample test search results which
are based on the search parameters that were provided (see
Fig. 3) are displayed in Fig. 4. The closest traditional health
practitioner from where the patient sent the search request is
the first one in the search results. Again, different search
results were discussed in Table 1. A patient may click the
navigation button to see the map directions to the traditional
health practitioner’s location. In the case of the search results Figure 3. Traditional Health Practitioner Search Interface
as displayed in Fig. 4, the user pressed the navigation button
associated with the first name and the map directions were
given as shown in Fig. 5.
 V. EXPECTED BENEFITS OF THE SYSTEM
Based on the purpose and the experimental test results of
the system the following are the expected derived benefits of
the system:
 Easy location of traditional health practitioner – The
system enables a patient to locate a traditional health
practitioner who is closest to his/her location. The system
therefore allows the user to locate a traditional health
practitioner on the move, anytime and anywhere.
 Protects potential patients from bogus traditional
health practitioners – The use of the system protects patients
from unscrupulous traditional health practitioners because
patients can only locate registered health practitioners.
Through this system prospective patients of traditional health
practitioners are saved from being robbed their money by
unregistered bogus traditional health practitioners.
 Promotes the increase of traditional health
practitioners’ clientele – The convenience that the system
provides for a patient to locate a traditional health practitioner
will encourage many prospective patients to be willing to
consult registered traditional health practitioners. In turn this
will mean fewer people will consult bogus traditional health
practitioners and thus less people will lose their money to the
unscrupulous traditional health practitioners.

VI. CONCLUSION AND FUTURE WORK

Figure 4. Search Results This paper proposed and presented a location-based mobile
application which enables a prospective patient to search for
the nearest traditional health practitioner. Furthermore, a
council administrator can use the desktop interface of the
application to load the information of the new traditional
health practitioner applicant on the system’s database.
Experimental results on the system usage have also been
presented and discussed above. These experimental results
have shown that a council administrator is able to register a
traditional health practitioner on the system. Furthermore, the
test results also show that a patient is able search and locate a
traditional health practitioner closer to his/her current location.
The test results therefore are indications that the system has
met user requirements. This paper also discussed expected
benefits which accrue from the system usage, and these
include easy location of registered traditional health
practitioners, protection of prospective patients against bogus
traditional health practitioners and promotion of clientele for
traditional health practitioners.
Finally, future work may include a survey on the benefits
of the system – establishing if the expected benefits as
mentioned in this paper are indeed realized.

REFERENCES
[1] V.C.E Bahamondez, J. Häkkilä and A. Schmidt, “Towards better UIs for
mobile learning: experiences in using mobile phones as multimedia tools
at schools in rural Panama,” MUM '12: Proceedings of the 11th
International Conference on Mobile and Ubiquitous Multimedia. 2012.
[2] Y. Bang, D.J. Lee and K. Han, “The Impact of Mobile Channel
Figure 5. Map directions
Adoption on Purchase Time Dispersion in e-Commerce,” ICEC '15
 Proceedings of the 17th International Conference on Electronic
Commerce 2015. Article No. 34.
[3] A,Meschtscherjakov, D.Wilfinger and M,Tscheligi, “Mobile attachment
causes and consequences for emotional bonding with mobile phones,”
2014. CHI '14 Proceedings of the SIGCHI Conference on Human
Factors in Computing Systems , pp 2317-2326.
[4] D.Ferreira, J.Goncalves. V.Kostakos, L.Barkhuus K.Anind and
A.K.Dey, “Contextual experience sampling of mobile application micro-
usage,” 2014.MobileHCI '14 Proceedings of the 16th international
conference on Human-computer interaction with mobile devices &
services, New York, United States of America, pp. 91-100.
[5] J.P. Rula, B.Jun and F.Bustamante, “Mobile AD(D): Estimating Mobile
App Session Times for Better Ads,” 2015.HotMobile '15 Proceedings of
the 16th International Workshop on Mobile Computing Systems and
Applications, New York, United States of America, pp 123-128.
[6] B.Leeladevi, C.P.R Raj, and A.Tolety, “A study on smartphone printing
approaches,” Information & Communication Technologies (ICT), 2013
IEEE Conference on, JeJu Island.
[7] N. Nirwal, N. Sardana, and A.J. Bhatt, “Hopeful Hearts: A Mobile
Health Care Application,”. 2014 Seventh International Conference on
Contemporary Computing (IC3), pp. 351 – 356.
[8] J.P. van Niekerk, (2012),Traditional healers formalised? [Online]
Available From: http://www.scielo.org.za/scielo.php?pid=S0256-
95742012000300001&script=sci_arttext [Accessed: 16 August 2015].
[9] J. Pirker, P. Weiner, C. Gütl, and V.M. Garcia-Barrios, “Location-based
Mobile Application Creator: Creating Educational Mobile Scavenger
Hunts,” 2014 International Conference on Interactive Mobile
Communication Technologies and Learning (IMCL), pp. 160-164.
[10] D.M. Mahmud and N.A.S. Abdullah, “Mobile Application
Development Feasibility Studies,” 2014 IEEE Conference on Open
Systems (ICOS), pp. 30-35.
[11] J.Sakata, M,Zhang, S.Pu, J,Xing, and K.Versha, “Beam: a mobile
application to improve happiness and mental health,” 2014 CHI EA '14
CHI '14 Extended Abstracts on Human Factors in Computing Systems,
New York, United States of America, pp. 221-226 .
[12] S.R.Shinde, R.Shinde, S.Shanbhag, M.Solanki and P.Sable,
“mHEALTH-PHC - Application design for rural health care,” 2014
Humanitarian Technology Conference - (IHTC), pp. 1 – 5.
[13] A.Maitra and N.Kuntagod, “A novel mobile application to assist
maternal health workers in rural India”, 2013 5th International Workshop
on Software Engineering in Health Care (SEHC), pp. 75-78.
[14] R.M.González, J.Francisco, F.J. Álvarez, A.Rodríguez, J.M. Arteaga
and A.M. González, “Guidelines for designing graphical user interfaces
of mobile e-health communities,” INTERACCION '2012 Proceedings of
the 13th International Conference on Interacción Persona-Ordenador,
Article No. 3 , ACM New York, NY, United States of America.
[15] S.Brown and T.X. Brown, “Value of mobile monitoring for diabetes in
developing countries,” ICTD '2013 Proceedings of the Sixth
International Conference on Information and Communication
Technologies and Development: Full Papers - Volume, ACM New
York, United States of America, pp 267-273.
[16] R. De Bruin and S.H. von Solms, “Securing mobile applications in
hostile rural Environments,” IST-Africa 2014 Conference Proceedings
Paul Cunningham and Miriam Cunningham (Eds), pp 1-9.
[17] S. Sathe, R. Melamed, P. Bak, S. Kalyanaraman, “Enabling Location-
Based Service 2.0: Challenges and Opportunies,” 2014 IEEE 15th
International Conference on Mobile Data Management, pp 313-316.
[18] T.J. Gerpott and S. Berg, “Adoption of Location-Based Service Offers
of Mobile Network Operators”, Mobile Business and 2010 Ninth Global
Mobility Roundtable (ICMB-GMR), pp 154-160.
 A New Efficient Algorithm for Executing Queries
over Encrypted Data
Rasha Refaie1, A. A. Abd El-Aziz2, Nermin Hamza3, Mahmood A. Mahmood4, Hesham Hefny5
Dept. of Computer & Information Sciences
Institute of Statistical Studies and Research
Cairo University
e-mail: {1rosha2030@gmail.com, 2a.ahmed@cu.edu.eg, 3nermin_hamza@yahoo.com, 4mahmoodissr@cu.edu.eg,
5
hehefny@ieee.org}

Abstract—Outsourcing databases into cloud increases the
need of data security. The user of cloud must be sure that his
data will be safe and will not be stolen or reused even if the data-
centers were attacked. The service provider is not trustworthy so
the data must be invisible to him. Executing queries over
encrypted data preserves a certain degree of confidentiality. In
this paper, we propose an efficient algorithm to run
computations on data encrypted for different principals. The
proposed algorithm allows users to run queries over encrypted
columns directly without decrypting all records.
Keywords—Database security; query processing,
Homomorphic encryption; CryptDB; MONOMI; and Secure
indexes
I. INTRODUCTION
When a database is provided as a service, a service provider
may be not trustworthy and data needs to be protected from
the database service provider. The problem in using
encryption is that no way to execute queries over encrypted
data without decrypting it[1]. Hence the data will not be
protected, which present a problem for users. There for we
propose an efficient algorithm to run queries over encrypted
data without decrypting it. So the data will remain
confidential and invisible even to the cloud provider.
II. RELATED WORK
Different techniques have been suggested to maintain a certain
degree of privacy on database outsourcing scenarios. In
addition they allow to execute some SQL queries efficiently.
Fully homomorphism encryption (FHE), is a new concept of
security. This system can calculate any type of function on
encrypted data[2]. Using fully homomorphic encryption still
very expensive and slow[3].
CryptDB’s approach is to execute queries over encrypted data
on the DBMS server as it would on an Plaintext database[4].
The proxy manages all the communication to or from the
database and uses secret keys to encrypt all data which
included in queries issued to the DBMS[5,6]. In this approach
data is encrypted in a layered way which is called as
onion[7,8,9,10,11]. CryptDB is much more efficient but
cannot support most analytical queries over encrypted data
and cannot run computations on values which encrypted using
different keys at DBMS[3].
MONOMI is a system to execute analytical workloads over
encrypted data on not trusted database server. In MONOMI a
part of the query is run on the not trusted server over
encrypted data and the remaining of the query on the plain
database at the user server[12]. MONOMI send the encrypted
result for the user which decrypt it and run the final
computation which is more efficiently to compute at the user.
Installing new database design still a big problem in
MONOMI and any security constraints is not taken into
account [3].
[13] suggest a good algorithm to search over encrypted data.
The limitations of the existing techniques for fuzzy match and
range queries are efficiently eliminated using this algorithm.
This algorithm is efficient for searching when the result of
query is less than 40%of the total data.
[14] proposes a new architecture to support data's
confidentiality, integrity and availability. This architecture
illustrate a secure and robust cloud storage by combining
cryptography and access control with two layered encryption.
Using pre-computed index make the execution of queries in
databases faster. But using standard indexes become
Ineffective when the data is encrypted[15]. Moreover, if
several users with different access rights use the same index,
each one of them needs access to the entire index, possibly
including indexed elements that are beyond his access rights.
A simple but elegant solution to this problem is suggested by
split the index into several sub-indexes where each sub-index
relates to values in the column encrypted using the same
key[3].
Security issues and related costs are the most strategic issues
which related to outsourcing database. [16] proposes a model
which includes the variability of database workload and cloud
prices and the related cost of encryption schemes. By applying
the model to actual cloud provider prices, the encryption and
adaptive encryption costs for data privacy can be determined.

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

 III. PROPOSED TECHNIQUE
Our proposed technique builds on CryptDB’s design.
CryptDB’s design consisting of two main parts: a database
proxy and DBMS. CryptDB makes the DBMS run the queries
over an encrypted database as if it was plaintext. The proxy is
a trusted server which stores the database schema and a secret
master key. It works as a middle layer, that encrypts and
decrypts all data and changes some query operators, while
keep the semantics of the query. The proxy manages all the
communication to or from the database and uses secret keys to
encrypt all data that included in queries issued which send to
the DBMS[5,6]. CryptDB can chain encryption keys to user
passwords. In CryptDB the user can decrypt his authenticate
data item only by using his password. Consequently, if all
TABLE1
servers are compromised an attacker cannot decrypt the data
of any user who is not logged in and even a database
administrator cannot access to decrypted data[3].
CryptDB has a problem to run computations on values which
encrypted using different keys at DBMS , even if the
application has the authority of all principals in question[3].
The traditional way is to perform the computation after
decrypting the data. It may be practical for some
computations, but for large-scale aggregates computations this
approach is too expensive. For example if we have Table1 and
its encrypted form Table 2 which has column (salary)
containing values encrypted by different keys( encrypted
using K1, encrypted using K2). Table 2 is saved at DBMS
server.
TABLE2: Encrypted_Table

ID Emp_Name Salary Dept_Num ID Emp_Name Salary Dept_Num

1 Bob 60 1 1 x9ak x934bc1 1
2 Alice 100 1 2 98wu x5a8c34 1
3 Eve 800 1 3 u8sb x84a21c 1
4 McDonald 100 1 4 3781e2 x5a8c34 1
5 Bob Marley 60 2 5 ed34ef x1eab81 2
6 Alice 100 2 6 98wu x638e54 2
7 McDonald 60 2 7 3781e2 x1eab81 2

Suppose a user issues the following query:
SELECT Emp_Name, Salary, Dept_Num
FROM TABLE2
WHERE Salary ≥ 100
In CryptDB data is encrypted in a layered way which is called
as onion. The term "onion" refers to layers of encryption,
Overlapping like the layers of an onion. These onions have
different layers each encrypted by using different algorithms.
The outer layer of an onion is the most secure layer. A value
has only one current layer in each onion. ORD onion is used to
adjust the order leakage for the queries including comparison.
Order-preserving encryption (OPE), is used to encrypt values
and keep the order of these values. If x < y, then OPEK(x) <
OPEK(y), for any secret key K[4]. Therefore, if a column is
encrypted with OPE, the server can perform range queries and
queries including comparison. In our example salary is
encrypted by different key. The problem is that when the same
value encrypted by different keys it gives different
values[EK1(100) ≠ EK2(100)]. So, server can not perform range
queries and queries including comparison on a column
containing values encrypted by different keys.
To execute this query using CryptDB. The proxy will encrypt
the query and send it to DBMS server to run it. The server
cannot check if total ≥ 100 because salary encrypted by
different keys. Thus, CryptDB executes the following query at
the DBMS server:

SELECT Emp_Name, Salary, Dept_Num

FROM TABLE2

Once the proxy receives the results, it decrypts them, and

executes the WHERE Salary ≥ 100 condation. Any matching
results are sent to the user [9]. Here we had to decrypt the
whole two columns Emp_Name and Salary.
Fig. 1. Different Layers of ORD Onion[4].
Our proposed technique try to overcomes these limitation by
encrypted the query by different keys and run every query on
the corresponding data which encrypted by the same key. Our
proposed technique builds on CryptDB’s design which
consisting of two main parts: a database proxy and DBMS as
previously explained.

Q1
User1
Encrypted Q2
Password P1 Database
Proxy server Q3

(Trusted server) Q...

User2
DBMS server
Password P2 (Untrusted server)

Figure 2: Proposed security technique

The following algorithm will describe the outlines of the suggested technique:

Input: user query

Output: query result
Begin
1. User issues query
2. The proxy check the query condition If (query condition is not on column containing data encrypted by
different keys) Go to step 3 Else if (authorized user) Go to step 4 Else Go to step 6
3. i. The proxy intercepts and rewrites the query which forwards to the DBMS server ii. DBMS retrieve data from
Encrypted_Table and returns it to the proxy Go to step 5
4. i. Proxy encrypt the query using different keys which forwards to the DBMS server ii. DBMS run every query on
the corresponding data which encrypted by the same key and returns the results to the proxy.
5. The proxy decrypts the query results and returns it to the user .
6. Exit
End

When any authorized user wants to search some records and
query condition is on the column which encrypted using
different keys, so the proxy will encrypt the query using
different keys (K1, K2, K3, ............) to (Q1, Q2, Q3,
................). The proxy send these queries to DBMS to run
every query on the corresponding data which encrypted by the
same key. These queries will return these records which
exactly the user wants from Encrypted_Table. This approach
does not need to decrypt all the values of entire encrypted
column; rather it decrypts only those values which satisfy the
user query. The following example can describe the searching
operation of the suggested algorithm:
SELECT Emp_Name, Salary, Dept_Num
The proxy will encrypt this query using (K1, K2) to (Q1, Q2).
The proxy send Q1 and Q2 to the DBMS. The DBMS run Q1
on the data which encrypted by K1and Q2 on the data which
ecrypted by K2 to return exactly these records which the user
wants. The DBMS server returns the encrypted query result
(TABLE 3,4).
Q1: SELECT Emp_Name, Salary, Dept_Num
FROM TABLE2
WHERE Salary ≥ x5a8c34

FROM TABLE2 WHERE Salary ≥ 100

TABLE3: RESULT of Q1 REFERENCES

Emp_Name Salary Dept_Num [1] (2011) The Forbes website. [Online]. Available:http://
http://www.forbes.com/sites/andygreenberg/2011/12/19/an-mit-magic
98wu x5a8c34 1 trick-computing-on-encrypted databases-without-ever-decrypting them/
u8sb x84a21c 1 [2] TEBAA.M, EL HAJJI.S & EL GHAZI.A "Homomorphic Encryption
3781e2 x5a8c34 1 Applied to the Cloud Computing Security". In Proc. of the World
Congress on Engineering 2012, Vol I WCE 2012, London, U.K, July4-
6, 2012.
[3] R. Refaie, A. A. Abd El-Aziz, N. Hamza, M. A. Mahmood and H. Hefny
Q2:SELECT Emp_Name, Salary, Dept_Num "A Survey on Executing Query on Encrypted", The International
Conference on Intelligent Information Technologies (ICIIT 2014),
FROM TABLE2 Chennai, India, 12/2014.
[4] R. A. Popa, C. M. S. Redfield, N. Zeldovich, and H. Balakrishnan
WHERE Salary ≥ x638e54 "CryptDB: Protecting confidentiality with encrypted query processing".
In Proc. of the 23rd
TABLE4: RESULT of Q2 [5] R. A. Popa, C. M. S. Redfield, N. Zeldovich, and H. Balakrishnan. "
Review of \CryptDB: Protecting Confidentiality with Encrypted Query
Processing". May 16, 2012 SOSP, pages 85100, Cascais, Portugal, Oct.
Emp_Name Salary Dept_Num 2011.
98wu x638e54 2 [6] R. A. Popa, C. M. S. Redfield, N. Zeldovich, and H. Balakrishnan. "
CryptDB: A Practical Encrypted Relational DBMS". Available:http:
//lib4shared.com/doc-file/cryptdb-a-practical-encrypted-relational-dbms
The proxy decrypts these results and returns it to the user [7] Dayioglu.Z, Kiraz.M, Birinci.F, and Akin.I. "Secure Database in Cloud
Computing: CryptDB Revisited". 6t INTERNATIONAL
(TABLE5). INFORMATION SECURITY & CRYPTOLOGY CONFERENCE 20-
21 September/Eylül 2013 | Ankara / TURKEY
TABLE5: THE RETURNED RESULT FROM PROXY SERVER’S SIDE
[8] http://en.wikipedia.org/wiki/Tor_(anonymity_network)
Emp_Name Salary Dept_Num [9] Ferretti.L, Pierazzi.F, Colajanni.M, and Marchetti.M."Security and
Confidentality Solutions for Public Cloud Database Services "
Alice 100 1 SECURWARE 2013 : The Seventh International Conference on
Eve 800 1 Emerging Security Information, Systems and Technologies.
[10] P. SRIVANI, S. RAMACHANDRAM and R. SRIDEVI. " A SURVEY
McDonald 100 1 REPORT ON CRYPTDB " Computer Science , Communication &
Alice 100 2 Instrumentation Devices Editors: Janahanlal Stephen,Harish Rohilk and
SVasavi Copyright © 2015 AET-2014k6 Organisers. ISBN: 978-981-
In this technique the proxy does not need to decrypt all the 09-5247-1.
values of entire encrypted column; rather it decrypts only [11] C. Curino, E.P. C. Jones, R. Popa, N. Malviya, E. Wu, S. Madden, H.
these values which match the user query. Balakrishnan and N. Zeldovich. " Relational Cloud: A Database-as-a-
Service for the Cloud " available at
https://people.csail.mit.edu/nickolai/papers/curino-relcloud.pdf.
[12] Stephen Tu, M. Frans Kaashoek, Madden.S and Zeldovich.N. "
IV. CONCLUSION Processing Analytical Queries over Encrypted Data". In Proc. of the
39th International Conference on Very Large Data Bases (VLDB), Riva
In this paper, we outlined Various techniques which offer del Garda, Italy, August 2013.
privacy on database outsourcing scenarios. A secure algorithm [13] M. Sharma, A. Chaudhary, S. Kumar. " Query Processing Performance
for searching over encrypted data was suggested. Our and Searching over Encrypted Data by using an Efficient Algorithm"
proposed technique builds on CryptDB’s design. CryptDB’s International Journal of Computer Applications (0975 – 8887) Volume
62– No.10, January 2013.
approach is to make a secure communication between user and
[14] R. Kalaichelvi and L. Arockiam " Secure and Robust Cloud Storage
the encrypted database by executing queries over encrypted with Cryptography and Access Control " R. Kalaichelvi et al./ Elixir
data on the DBMS server as it would on an plaintext database. Comp. Sci. & Engg. 56 (2013) 13481-13484. Available online at
Our proposed algorithm will efficiently eliminate the www.elixirpublishers.com (Elixir International Journal)
limitations of computations on values encrypted for different [15] Shmueli.E, Waisenberg.R, Elovici.Y and Ben-Gurion.E. "Designing
principals. We will apply our algorithm to ensure its Secure Indexes for Encrypted Databases. In Proc of Data and
Applications Security". 19th Annual IFIP WG 11.3 Working
effectiveness. Conference, USA, 2005.
[16] Ferretti.L, Pierazzi.F, Colajanni.M, and Marchetti.M." Performance and
Cost Evaluation of an Adaptive Encryption Architecture for Cloud
Databases". IEEE TRANSACTIONS ON CLOUD COMP TING, VOL.
2, NO. 2, APRIL-JUNE 2014.
 Memristor Model For Massively-Parallel
Computations
Dalibor Biolek
Department of Microelectronics / Electrical Engineering
Brno University of Technology / University of Defence
Brno, Czech Republic
dalibor.biolek@unob.cz
Abstract—The model of memristor described in the paper is
designed for building models of large networks for analog
computations. A circuit containing thousands of memristors for
finding the shortest path in a complicated maze is a typical
example. The model is designed to meet the following criteria: 1.
It is a model of HP memristor with linear dopant drift while
respecting the physical bounds of the internal state variable. 2.
Reliable operation in the SPICE environment also when
simulating extremely large networks. 3. Minimization of the
simulation time while computing bias points and during transient
analyses. A benchmark circuit for testing the applications of
various complexities is presented. The results confirm a perfect
operation of the model also in applications containing thousands
of memristors.
Keywords—memristor; model; massively-parallel
computations; SPICE
I. INTRODUCTION
Circuits for massively-parallel computing, when the
collective cooperation of a large number of memristors results
in effects unattainable by digital computers, are important
potential application areas of memristors. Fast searching for
paths in complicated labyrinths, when the computation time
grows exponentially with the maze size, is a typical example.
An interesting method for solving efficiently this problem via
a memristor network is proposed in [1]. By means of analog
switches, this network can be reconfigured to the form of the
analyzed labyrinth. When applying DC voltage across
terminals, which represent the entrance and the exit of the
maze, the current flows only through the memristors which lie
on the sought paths inside the labyrinth. After disconnecting
the source, these paths are memorized in the form of
memristances of the corresponding memristors. The path
lengths can be evaluated via measuring the total resistances of
individual paths. Such a principle is used in [2] for finding
optimal trajectories of the passenger on the London tube
network, which is modeled by a memristive grid.
Experimenting with such networks requires hundreds if not
thousands of memristors. Their samples are not currently
accessible. However, problems also exist with computer
simulations of such large circuits. The numerical problems
This work was supported by the Czech Science Foundation under grant
No 14-19865S – Generalized higher-order elements. Research described in
this paper was financed by the Czech Ministry of Education in the frame of
the National Sustainability Program under grant LO1401. For research, the
infrastructure of the SIX Center was used. The research was also supported by
the Project of Specific Research, K217 Department, UD Brno.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Viera Biolkova and Zdenek Kolka
Dept. of Radio Electronics
Brno University of Technology
Brno, Czech Republic
{biolkova, kolka}@feec.vutbr.cz
burdening simple models of the HP memristor with frequently
used window functions [3] increase with increasing
complexity of the application circuit. The experiments up to
now with these models reveal their malfunction in the SPICE
environment also for primitive circuits containing only one
memristor. The well-known Pickett’s model of TiO2
memristor [4] is rather complicated in itself: in a SPICE
implementation, suggested in [5], it represents 26 additional
equations. When utilizing such a model in the application
network containing 1000 memristors, then it represents
additional 26 thousand rows and 26 thousand columns in the
circuit matrix. In addition, the Pickett’s model labors with
serious convergence problems. A simple circuit can be shown,
containing 8 memristors, each modeled by the Pickett’s
analog model, whose DC solution cannot be found in SPICE due to
numerical problems.
The paper describes a model of the HP memristor with
linear dopant drift, which is mathematically equivalent to the
classical Strukov model [6], but taking into account the
physical limits of the internal state variable x  [0,1], which is
the normalized width of the doped (and thus conductive) TiO 2
layer [7]. This limitation is commonly modeled by a
rectangular window function [7]. However, such a model fails
numerically if the memristor reaches either border of the state
variable [8]. For our purposes we will therefore use the
method of nonlinear transformation [3] of the native state
variable (charge) into the physical state variable (x), which
must be optimized in order to minimize the total count of
equations for the simulation program.
II. PROPOSED MODEL OF HP MEMRISTOR
The classical model of the TiO2 memristor is in the form
[7]
v  Rm x   i , Rm x   Ron x  Roff (1  x)  Roff  Rx , (1)
dx R
 kif w (x) , k  v 2on . (2)
dt D
 Here v, i, and Rm are the memristor voltage, current, and
resistance (memristance), Ron and Roff are the limiting values
of the memristance at the boundary states x = 1 and x = 0, and
R = Roff - Ron. The boundary speed in (2) is directly
proportional to current, where k is a technological constant,
dependent on the dopant mobility V, on the resistance Ron,
and on the total length D of the oxide. The window function
fw(x) models the boundary phenomena, when the boundary
speed in the vicinity of the states x = 1 and x = 0 is slowed
down. The role of the simplest rectangular window (fw(x) = 0
for x  0 and x  1, fw(x) = 1 for other x) is to stop the
boundary movement in order that x does not fall outside its
physical limits.
The differential equation (2) can be integrated in the form
[3]
dx
 f w ( x)
 Fw ( x)  Cw  k  idt  kq  q
where Cw is a constant of integration, q is the normalized
charge, and the primitive function Fw(x) transforms the state
variable x into the normalized charge. The inverse FI w (q ) for
the case of rectangular window and for CW = 0 has a piece-
wise linear form according to Fig. 1.
x(t) = FIw(q(t) + q0) = FIw(kq(t) + q0)
1
x0
0.5
0
-1 - 0.5 0 q 0.5
q0
q(t)
Fig. 1. Nonlinear function for transforming the normalized charge q into the
normalized physical state variable x of TiO2 memristor with rectangular
window.
Instead of the differential equation (2), a simpler equation
between current and charge
dq
i
dt
will be a part of the memristor model. Its numerical solution is
trivial. The computed charge must then be normalized by
multiplying it by the constant k from Eq. (2), and transform it
subsequently into the state variable x via the nonlinear
transformation according to Fig. 1. The memristance can be
evaluated from x by Eq. (1). The initial state of the memristor
can be defined in one of three ways: (A) As initial
memristance, which is related unambiguously to the initial
state x0 via Eq. (1). (B) As initial state x0. Its limiting values 0
and 1 correspond to the memristances Roff and Ron. (C) As
initial normalized charge q0 . According to Fig. 1, its values -
0.5, 0.5, and 0 correspond to the x states 0, 1, and 0.5, and to
the memristances Roff, Ron, and (Ron+Roff)/2. Method (C) is
most general because it enables defining the initial state of the
memristor also outside the region of linearly increasing
characteristic in Fig. 1.
The mechanism of memristor modeling is shown in Fig. 2
(3) (a): The memristor current is sensed and transformed into a
charge via integration. The block FIw transforms the charge
into the state variable x, which controls the memristance.
Figure 2 (b) represents the model for its effective
implementation in SPICE. In addition to the couple of
terminals A and B, only one additional node q appears in the
model, whose voltage is equal to the computed charge. The
memristance is modeled via the controlled current source. To
avoid a potential conflict of ideal current sources in the
application circuit, the A and B terminals are shunted by the
auxiliary resistor Raux1, whose resistance is much greater than
that of Roff. The auxiliary resistor Raux2 is necessary for
defining the DC path from node q to ground.
A
q
I 
Rm=Roff ‒ Rx
x
1 V
FIw( )
B
(a)
(4)
 A into the node. All the memristors have the same initial
memristances. As is obvious from the current direction, all
q vertical/horizontal memristors will decrease/increase their
resistances. The initial resistance of the in node is given by the
I I
resistance of the network with identical 6 k resistors,
whereas the final resistance is given by the network of vertical
1F Ron and horizontal Roff resistors. The corresponding initial and
1 G
final voltages are shown in Fig. 4. The non-monotonic
1 G transient can be achieved via selecting a proper value of Rini.
Raux1 Gres Gq C int Raux 2
Figure 5 provides the simulation results of the same circuit
for sinusoidal excitation. It should be noted that the pinched
V hysteresis loop can exhibit crossing points also outside the v–i
V
Roff  R FI w (kV (q)  q0 ) origin. This status can be adjusted via a proper selection of the
B initial memristance.
(b)
The computation times for transient analyses from Figs 4
Fig. 2. Model of HP memristor, (a) ideological, (b) for SPICE and 5 were approximately the same, ca 4 seconds. Similar
implementation. analyses of more complicated circuits with M=N=50, thus
containing 5100 memristors, and M = N = 100 with 20200
III. BENCHMARK CIRCUIT memristors take ca 9 and 75 seconds, respectively.
The benchmark circuit in Fig. 3 consists of 2MN+M+N 10
identical memristors. Its complexity can be arbitrarily varied Vin
by selecting the parameters M and N. The signal source is 8.2585 V
applied between the in and ground terminals. The current [V] 8
source has been selected for experiments described below,
providing either DC current or sinusoidal waveform with
adjustable amplitude, frequency and initial phase, with the 6
voltage response being evaluated in both cases. For sinusoidal 4.3898 V
excitation, the fingerprints of v–i pinched hysteresis loops can
be tested. The experiments were done in Micro-Cap 11 and 4
LTspice on a PC with Intel(R) Core(TM) i7-3770 CPU @
3.4GHz and 16GB RAM, Windows 7 Ultimate 64 bit.
2
in 1 2 3 M

0
1 0 5 10 time 15 20
[sec]

2
Fig. 4. Results of transient analysis of the circuit from Fig. 3 for M=50,
N=30: voltage response to a constant current of 250 A flowing into the in
node; Rini = 6 k, Ron = 100, Roff = 10k, step ceiling = 0.2 sec.
3

Fig. 3. Benchmark circuit compounded of 2MN+M+N identical memristors.

Figures 4 and 5 summarize the Micro-Cap simulations of

an extremely large circuit with M=50 and N=30, thus
containing 3080 memristors. Figure 4 shows the evolution of
voltage at the in node driven by DC 250A current flowing
 1 5 It is obvious from the MxN benchmark circuit in Fig. 3 that
I Vin it contains (M+1)(N+1) nodes and 2MN+M+N memristors. If
[mA] [V] the memristor submodel contains a number of additional
0.5 2.5 internal nodes, then the total number of the nodes n, which
results in the (n x n) circuit matrix processed by the simulation
program, is
0 0 n = (M+1)(N+1)+a(2MN+M+N) =
(1+2a)MN+(1+a)(M+N)+1.
For the case M=N=100 with 20200 memristors, mentioned
-0.5 -2.5 in Section III, n = 10201+20200a. Since the proposed
memristor macro circuit in Appendix 1 contains only one
internal node (q), the solver of the simulation program
-1 -5 manipulates with the circuit matrix of (n, n) size, where n =
0 0.5 1 time 1.5 2 3MN+2M+2N+1. For the case M=N=100 with 20200
[sec] memristors, mentioned in Section III, n = 30401. However, the
model from [4] with 26 internal nodes leads to n = 535401
5
Vin equations, and the simulation task crashes just for much
[V] smaller-sized benchmark circuits.
2.5 In case it is necessary to model nonlinear boundary effects
in the TiO2 memristor, a sigmoidal function according to [3]
can be used instead of the piece-wise-linear transforming
0 function between native and physical state variables, with the
principle of the model and its advantageous parameters
remaining preserved. It should be noted that this increase of
-2.5 the model complexity is not accompanied by increasing the
number of its internal nodes.

-5
-1 -0.5 0 I 0.5 1
[mA]

Fig. 5. Transient analysis of the circuit from Fig. 3 for M=50, N=30, driven
by sinusoidal 1 mA/ 1 Hz current; Rini = 1.3 k, Ron = 100 , Roff = 10 k;
step ceiling = 0.02 sec.

IV. CONCLUSIONS
Computer experiments with extremely large circuits,
containing the proposed model of the HP memristor, confirm
the robustness of this model. It enables a fast and reliable
analysis of circuits containing thousands of memristors.
Comparable parameters are absolutely unavailable for
classical memristor models, irrespective of whether they
utilize the window functions or the Pickett – type physical
models.
It follows from Appendix 1 that variable resistance can be
modeled in Micro-Cap directly via a resistor with formula-
dependent resistance. This way modeled memristive port
speeds-up the analysis of large circuits containing memristors
in comparison to the classical SPICE modeling via controlled
sources as applied to PSpice and LTspice (see Appendix 2). In
some special cases, the latter approach can provide more
accurate results: the accuracy of the first model is guaranteed
with smaller step ceiling. Since it results in reducing the speed
of the transient analysis, the advantage of the direct modeling
of memristance is rather questionable.
 APPENDIX 1 – MEMRISTOR MACRO FOR MICRO-CAP 11 GM IN1 IN2 VALUE =
+ {v(IN1,IN2)/(ROFF-DELTAR*FI(XINI-0.5+K*V(Q)))}
R3 IN1 IN2 1G
.PARAMETERS(Ron=100,Roff=10k,Rini=1k) .ENDS MEMRISTOR_IDEAL
Plus
I(RM)
IN1 q
Cq R2
RM {Rmem} Gq 1 1G References
Minus

IN2
[1] Y. V. Pershin, and M. Di Ventra, “Solving mazes with memristors: A
.define uv 1e-14 massively parallel approach,” Phys. Rev., vol. E 84, p. 046703, 2011.
[2] Z. Ye, S. H. M. Wu, and T. Prodromakis, “Computing shortest paths in
.define k (uv*Ron/D**2)
2D and 3D memristive networks,” arXiv:1303.3927v1 [physics.comp-
.define D 10n ph], 15 Mar 2013.
.define deltaR (Roff-Ron) [3] D. Biolek, Z. Biolek, V. Biolkova, and Z. Kolka, “Reliable Modeling of
Ideal Generic Memristors via State-Space Transformation,”
.define xini ((Roff-Rini)/deltaR)
Radioengineering, vol. 24, no. 2, pp. 393–407, 2015.
.define FI(x) TABLE(x,-0.5,0,0.5,1) [4] M. D. Pickett, D. B. Strukov, J. L. Borghetti, J. J. Yang, G. S. Snider, D.
.define Rmem {Roff-deltaR*FI(xini-0.5+k*v(q))} R. Stewart, and R. S. Williams, “Switching dynamics in titanium dioxide
memristive devices,” Journal of Applied Physics, vol. 106, p. 074508,
2009.
APPENDIX 2 –SUBCIRCUIT FOR PSPICE AND LTSPICE [5] H. Abdalla, and M. D. Pickett, “SPICE modeling of memristors,” In
Proc. IEEE Int. Symp. Circuits and Systems, Rio de Janeiro (Brazil), pp.
1832–1835, 2011.
.SUBCKT MEMRISTOR_IDEAL In1 In2 [6] D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, “The
+ PARAMS: RON=100 ROFF=10K RINI=1K missing memristor found,” Nature, vol. 453, pp. 80–83, 2008.
*
[7] Z. Biolek, D. Biolek, and V. Biolkova, “SPICE model of memristor with
.FUNC FI(X)={TABLE(X,-0.5,0,0.5,1)}
nonlinear dopant drift,” Radioengineering, vol. 18, no. 2, pp. 210–214,
.PARAM K={(UV*RON/D**2)} D=10N UV=1E-14 2009.
+ XINI={(ROFF-RINI)/DELTAR} DELTAR={ROFF-RON}
* [8] D. Biolek, Z. Biolek, V. Biolkova, and Z. Kolka, “Modeling of TiO2
memristor: from analytic to numerical analyses,” Semiconductor Science
GQ 0 q VALUE = {I(GM)}
and Technology, vol. 29, no. 12, p. 125008, 2014.
CQ q 0 1
R2 q 0 1G
*
 Stability of Digitally Emulated Mem-Elements
Zdenek Kolka, Viera Biolkova
Department of Radio Electronics
Brno University of Technology
Brno, Czech Republic
{kolka | biolkova}@feec.vutbr.cz
Abstract—The paper analyzes the stability issue of a special
emulator of memristive, memcapacitive and meminductive
systems as well as higher-order elements from Chua’s periodical
table, and other two-terminal devices. It is demonstrated that, in
order to provide stable behavior, the analog port of the emulator
must be implemented by the controlled voltage or current source
depending on the type of the emulated two-terminal device. As a
practical consequence, the emulator must contain both of the
above analog interfaces in order to be universal.
Keywords— memristor; memcapacitor; meminductor; emulator
I. INTRODUCTION
Memristors [1] and memristive systems [2] belong to
promising building blocks for constructing future computer
systems [3] and bio-inspired electronics [4]. In addition to this
potential application area, there are also many studies dealing
with analog applications of memristors [5], drawing on the
analog nature of the memristor. These novel circuit ideas,
utilizing the unique features of memristors, cover various areas
of analog signal processing commonly used for communication
circuits, such as oscillators [6] and relaxation generators [7],
rectifiers [8], electronically controlled filters [9], controllers
[10] and amplifiers [11], modulators and demodulators [12],
A/D and D/A converters [13], and systems for image [14] and
audio [15] signal processing.
Re-designing the above systems from the conventional to
the memristive platform is inconceivable without the
corresponding experimental work. However, there is a lack of
real mem-devices on the market which could be useful for such
experimenting. As a result, usually the only feedback for the
designers is the computer simulation [16], with all its obvious
and well-known limitations in comparison to life experiments.
In addition to memristors and memristive systems, also
memcapacitors, meminductors, memcapacitive and
meminductive systems [17], and general higher-order elements
(HOEs) [18] from Chua’s periodical table [19] come on the
scene. The lack of useful aids for experimenting is yet worse
than for memristive systems.
In comparison with the computer simulations, the hardware
emulation of memristive systems [20] can be an important step
towards laboratory experiments with newly designed
memristor-based applications. Such emulators can concurrently
stimulate the interest of students and designers who are not
familiar with unconventional electronic components. With
This work was supported by the Czech Science Foundation under grant
No 14-19865S – Generalized higher-order elements. The research is a part of
the COST Action IC1401 and is financially supported by the Czech Ministry
of Education under grant No LD15033. Research described in this paper was
financed by the Czech Ministry of Education in the frame of the National
Sustainability Program under grant LO1401. For research, the infrastructure
of the SIX Center was used. The research was also supported by the Project of
Specific Research, K217 Department, UD Brno.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Dalibor Biolek
Department of Microelectronics / Electrical Engineering
Brno University of Technology / University of Defence
Brno, Czech Republic
dalibor.biolek@unob.cz
regard to the variety of potentially useful unconventional
circuit elements, there is a need for mimicking the behavior of
arbitrary devices, the memcapacitor, meminductor and HOEs
being a small subset of them.
Many circuit ideas of the emulators of the above systems
have been published so far. Their survey is described, for
example, in [20], [21], [22]. All of them can be classified as
analog or hybrid (i.e. with digitally-programmed analog
interface), and as single-purpose or universal. The ambition of
emulating flexibly an arbitrary component can only be
accomplished by using a hybrid universal emulator. The first
step towards its implementation was described in [11]: the
resistance of a digital potentiometer is controlled by a
microcontroller depending on the program which evaluates the
differential state equation of a considered memristive system.
However, the use of the potentiometer for emulating the analog
port limits the applications to memristive devices alone. In
order to mimic also memcapacitive, meminductive and other
systems, the so-called memulator was proposed in [21]. The
analog port is emulated via a controlled voltage or current
source. The way of the control depends on the mathematical
model between the port voltage and current of the emulated
two-terminal device.
With regard to the duality of the voltage and the current
sources, it would seem that such a memulator could be of an
economical construction, thus providing only one type of the
above controlled sources. However, practical experiments with
the memulator set up serious problems with the stability of the
emulated devices. It turns out, for example, that the
memcapacitive/meminductive devices cannot be satisfactorily
emulated via the current/voltage sources. This stability issue is
analyzed in depth in Section II of this paper. It is clearly shown
that the universal hybrid emulator must use a controlled
voltage and also current source for mimicking devices of
arbitrary nature. Section III presents experimental results with a
real hardware demonstrating the usefulness of this approach.
II. STABILITY ANALYSIS OF EMULATED ELEMENTS
A. Hybrid Emulator
The hybrid emulator can be configured either as a voltage
controlled current source or as a current controlled voltage
source. Modern integrated DC-DC converters, such as
ADuM5000, allow designing the device as fully floating with a
parasitic capacitance in the order of several picofarads. Fig. 1
shows the block diagram of a designed emulator, which is
powered and programmed via USB interface. In addition to
controlled voltage and current sources there is also an IO
module with digitally controlled resistor, which is shown here
for completeness.
of the continuous-time network, which is sampled as xn+1 after
the sampling period ts has elapsed.
B. Emulator Testbed
Let us suppose a voltage controlled emulator embedded in a
purely resistive circuit, which can be considered as the simplest
application scenario. Results for a current-controlled emulator
can be obtained analogously. The whole chain from the D/A
converter output to the A/D converter input will be linear. To
make the analysis trackable we will consider only a 1st order
low-pass filter combining the functions of antialiasing and
reconstruction filters from Fig. 2 (a) and omit the quantization
effects. Resistor R in Fig. 3 represents the Thévenin equivalent
of any resistive circuitry. The equivalent DC source is omitted
as it does not influence stability.
memulator i

A/D MCU D/A R v

Fig. 3. Linear resistive testbed.

The output quantity of the D/A converter, which is

Fig. 1. Block diagram of fully-floating hybrid emulator.
The microcontroller unit (MCU) measures the controlling
quantity by means of the A/D converter, computes the
response, and sets the value of the controlled quantity by the
equivalent to the emulator current (y  i), passes through a 1st-
order low-pass filter and controls the current source.
Let in denote the converter output at the time instant nts.
Then the voltage vn+1 sampled at the instant (n+1)ts is

 
D/A converter. The MCU runs a discretized system of
differential-algebraic equations of the emulated element vn 1  vnet s /  F  in R 1  et s /  F , (1)
between the terminals A and B.
where τF is the filter time constant. If we denote
Let us consider an emulator embedded in a continuous-time
electrical network, which closes a feedback loop. Generally,   e t s /  F , (2)
the loop also contains an antialiasing linear filter at the input of
the A/D converter and a reconstruction filter at the output of then the z-transform of (1) is
the D/A converter. Vz  V  I R1    , (3)

electrical and finally

network
V 1
 R (4)
I z 
anti- x y recon.
alias. A/D MCU D/A filter represents a complex transimpedance of the continuous-time
filter linear circuit providing a feedback to the discrete-time MCU.
(a)
C. Memristive System
computation
time tc The memristive system modeling is based on the Port
analog Equation (5) and the State Equation (6), which in the case of
transient
yn-1 xn yn yn+1
voltage control are as follows [16]
(n-1)ts n ts (n+1)ts i  g ( x, v, t )v , (5)
(b) x  f (x, v, t ) , (6)
Fig. 2. Emulator in application circuit (a); Timing diagram (b).
where i and v are port quantities, g() and f() are generally
Let us denote the emulator controlling quantity by x and the nonlinear functions, and xRn represents the state variable(s).
controlled quantity by y. The computation time tc of MCU will
be considered negligible in comparison with the sampling Let us assume that the dynamics of (6) is sufficiently slow
period ts. MCU computes the output sample, say yn, which in comparison with the sampling process and the bandwidth of
appears at the output of the D/A converter. It starts a transient the reconstruction filter. To get a qualitative insight, it will be
sufficient to analyze the local behavior of the memristive ts
system with frozen x, i.e. an emulated resistor. in  in 1  vn (14)
LM
Then the discretized Port Equation (5) will be simply
and in the z-domain
in  vn / RM , (7)
ts z
I V . (15)
where RM is a slowly varying memristance. In the z-domain we LM z  1
obtain
Combining (15) and (4) we obtain the characteristic
1 equation
I V. (8)
RM
 (1   ) z 
Combining (8) and (4) the characteristic equation of the I 1    0, (16)
 ( z   )( z  1) 
system is
where
 R 1 
I 1  0
 (9)
M z  
ts
 R  (17)
LM / R
with a single root
is the ratio of the sampling period and the time constant of the
R
z1   (1   )   . (10) emulated inductor and the equivalent external resistor R.
RM
The condition of stability in the z-domain |z1| < 1 leads to
R 1 
 . (11)
RM 1  
Table I summarizes the stability limit (11) for three filter
scenarios.

TABLE I. STABILITY CONDITION FOR EMULATED RESISTOR

τF α stability condition
0 (no filtering) 0 R < RM

ts exp(-1) R < 2.16 RM

10 ts exp(-0.1) R < 20 RM Fig. 4. Magnitude of roots of (16) as a function of β.

Let us note that using τF = 10 ts significantly decreases the Taking into consideration the sampling process of the
emulator bandwidth. emulated system dynamics, it makes sense to analyze the
stability for β < 1. Fig. 4 shows the magnitudes of both roots of
The same analysis performed for a current-controlled (16) for the same filter scenario as for the resistive system (see
emulator leads to an inverse condition to (11) Table I). In all cases we obtain a stable behavior for β < 1.
R 1 E. Memcapacitive System
 . (12)
RM 1   Similarly to both previous systems we will analyze an
emulated capacitor CM
D. Meminductive System
The analysis of the meminductive system is based on the dv
i  CM . (18)
same assumption which was used for the slow dynamics of the dt
memristive system (5), (6). We will analyze an emulated
inductor LM Applying the simple forward rule for differentiating, the
discretized version of (18) is
t
1 vn  vn 1
i  i (0) 
LM  v( )d ,
0
(13) in  CM
ts
. (19)

where i(0) is the initial condition. Analogously to the inductive system we obtain the
characteristic equation combining (19) and (4) as
Applying the simple rectangle integration rule the
discretized version of (13) will be in the time domain
  1 z 1  b(i  I t ) for i  I t
I 1  (1   ) 0, (20) 
  z( z   )  f t (i )  b(i  I t ) for i   I t , (24)
 0
where  otherwise

ts introduces the current threshold into the behavior of the

 (21)
R CM
is the ratio of the sampling period and the time constant of the
emulated capacitor and the equivalent external resistor R.
Fig. 5 shows the magnitudes of both roots of (20) as a
function of γ for three filter scenarios.
element to provide the offset and drift immunity. For |i| < It the
memristor state is frozen.
The window function wR() confines the memristance
change to the interval between Ron and Roff. It uses a special
form of the Biolek window function [23]
wR ( RM , i)   (i) ( Roff  RM )   (i) ( RM  Ron ) , (25)

where Ɵ is the step function.

Fig. 6. Test circuit for memristor and meminductor and experiment setup.

The memristor was connected into a divider structure in

Fig. 5. Magnitude of roots of (20) as a function of γ.
Fig. 6. The circuit was driven by a sinusoidal generator with an
amplitude of 1V and a frequency of 100 Hz. The memristor
It is obvious that for all three filter settings there exists a parameters were set as follows: It = 0.2 mA, Roff = 2 kΩ,
lower limit of γ below which the emulation is unstable. It is an Ron = 800 Ω, b = 8109 Ω s-1 A-1. The sampling period of the
impractical behavior because if RCM increases as a result of the emulator was 10 μs and the reconstruction filter was disabled
memcapacitive action the system becomes unstable. In (τF = 0). According to (11) the emulation is stable for R < RM.
addition, to reach a usable sampling ratio γ  0.1 it is necessary The resistance limit Ron was chosen below the value of R.
to significantly decrease the bandwidth of emulation by using a Therefore whenever the state variable RM goes below 1 kΩ, the
low-bandwidth reconstruction filter. “fast” emulation process becomes unstable, which can be
A comparison of the results for inductive and capacitive clearly seen on the oscillogram in Fig. 7. Such unstable
systems clearly shows that the meminductive systems should behavior would not occur when the memristor was emulated
be emulated by a voltage-controlled current source. A similar using always stable digitally controlled resistor.
analysis with a current-controlled voltage source would show
that current control, on the contrary, is suitable for
memcapacitive systems.
RM < R
III. EXPERIMENTAL RESULTS
It was shown in [5] that the so-called bipolar mem-system
with threshold can be useful in a number of various linear and
nonlinear applications. Its simplest memristive model can be
described by the following port and state equations
1
i  RM v , (22)
d
RM  ft (i ) wR ( RM , i) , (23)
dt
where the state variable is the memristance itself. The
activation function Fig. 7. Memristor waveforms of v (channel 1) and vM (channel 2).

Analogously to the memristive system (22) – (25), a bipolar

meminductive system with threshold can be defined with the
meminductance LM as internal state variable, via the following
equations [16]
  LM i , (26)
d [4]
LM  f t (i) wL ( LM , i ) , (27)
dt
where  is the generalized flux (time integral of voltage) and
the function ft is defined in (24). The window function
wL ( LM , i)   (i) ( Lhigh  LM )   (i) ( LM  Llow) (28)
restricts the meminductance change to the interval between Llow
and Lhigh.
According to the results of analysis in Section II.D the
meminductive system should be emulated by a voltage-
controlled current source.
Fig. 8. Meminductor waveforms of v (channel 1) and vM (channel 2).
The meminductor was connected in the same circuit as in
Fig. 6 and driven by a sinusoidal generator 1V, 250 Hz. The
meminductor parameters were set as follows: It = 0.5 mA,
Lhigh = 400 mH, Llow = 100 mH, b = 106 H s-1 A-1 with the same
emulator setting as in the case of memristor. The oscillogram
in Fig. 8 confirms always stable emulation.
IV. CONCLUSIONS
The paper analyzed the stability conditions for emulated
resistive, inductive, and capacitive systems. It was shown that
always stable emulation of meminductive systems is possible
with a voltage-controlled current source and the emulation of
memcapacitive systems is possible with a current-controlled
voltage source. The memristive systems can be emulated by
both sources with a certain limitation or with a digitally
controlled resistor, which is always stable. Therefore a
practically usable hybrid emulator should have all three
interface modules like the one in Fig. 1.
REFERENCES
[1] L. O. Chua, “Memristor – The missing circuit element,” IEEE Trans.
Circuit Theory, vol. CT-18, no. 5, pp. 507–519, Sept. 1971.
[2] L. O. Chua and S. M. Kang, “Memristive Devices and Systems,” Proc.
of the IEEE, vol. 64, no. 2, pp. 209-223, Feb. 1976.
[3] J. J. Yang, D. B. Strukov, and D. R. Stewart, “Memristive devices for
computing,” Nature Nanotechnology, vol. 8, pp. 13-24, 2013.
Y. V. Pershin and M. Di Ventra, “Experimental demonstration of
associative memory with memristive neural networks,” Neural
Networks, vol. 23, no. 7, pp. 881-886, 2010.
[5] D. Biolek, V. Biolkova, and Z. Kolka, “Memristive systems for analog
signal processing,” in ISCAS 2014, Melbourne, Australia, pp. 2588-
2591, 2014.
L. O. Chua and M. Itoh, “Memristor oscillators,” Int. Journal of
[6]
Bifurcation and Chaos, vol.18, no.11, pp. 3183-3206, 2008.
[7] M. A. Zidan, H. Omran, C. Smith, A. Sayed, A.G. Radwan, and K.N.
Salama, “A family of memristor-based reactance-less oscillators,” Int. J.
of Circ Theory and Applications, vol. 42, no. 11, pp. 1103-1122, 2014.
[8] J. Vavra and D. Biolek, “A full-wave rectifier based on memristive
systems,” in 2nd Int. Conf. on Modelling, Identification and Control
(MIC 2015), pp. 91-94, 2015.
[9] A. Ascoli, R. Tetzlaff, F. Corinto, M. Mirchev, and M. Gilli,
“Memristor-based filtering applications,” 14th Latin American Test
Workshop (LATW), 2013, pp. 1-6, 2013.
[10] G. Saha, “Theory of Memristive Controllers: Design and Stability
Analysis for Linear Plants,” Int. Journal of Computer Applications, vol.
34, no. 10, 2011.
[11] Y. V. Pershin and M. Di Ventra, “Practical Approach to Programmable
Analog Circuits With Memristors,” IEEE Trans. on Circuits and
Systems I: Regular Papers, vol. 57, no. 8, pp. 1857-1864, 2010.
[12] I. C. Goknar, F. Öncül, and E. Minayi, “New Memristor Applications:
AM, ASK, FSK, and BPSK Modulators,” Antennas and Propagation
Magazine, IEEE, vol. 55, no. 2, pp. 304-313, 2013.
[13] L. Gao, F. Merrikh-Bayat, F. Alibart, X. Guo, B. D. Hoskins, K.-T.
Cheng, and D. B. Strukov, “Digital-to-Analog and Analog-to-Digital
Conversion with Metal Oxide Memristors for Ultra-Low Power
Computing,” Nanoarch'13 Transactions, pp. 4, 2013.
[14] C.-R. Han, S.-J. Lee, S.-K. Oh, and K. Cho, “Memristor-MOS Analog
Correlator for Pattern Recognition System,” Journal of Nanoscience and
Nanotechnology, vol. 13, no. 5, pp. 3365-3370, 2013.
[15] S. K. Duan, X. F. Hu, L. D. Wang, and Ch. D. Li, “Analog memristive
memory with applications in audio signal processing,” SCIENCE
CHINA Information Sciences, SP Science China Press , pp. 1-15, 2013.
[16] D. Biolek, M. Di Ventra, and Y. V. Pershin, “Reliable SPICE
Simulations of Memristors, Memcapacitors and Meminductors,”
Radioengineering, vol. 22, no. 4, pp. 945-968, 2013.
[17] M. Di Ventra, Y. V. Pershin, and L O. Chua, “Circuit elements with
memory: memristors, memcapacitors, and meminductors,” Proc. of
IEEE, vol. 97, no. 10, pp. 1717-1724, 2009.
[18] L. O. Chua, “Device modeling via basic nonlinear circuit elements,”
IEEE Trans. Circ Syst., vol. CAS-27, no. 11, pp. 1014–1044, 1980.
[19] L. O.Chua, “Nonlinear circuit foundations for nanodevices, Part I: The
four-element torus,” Proc. IEEE, vol. 91, no. 11, pp. 1830-1859, 2003.
[20] D. Biolek, “Memristor emulators,” In Memristive networks. Springer
book, New York, pp. 487–504, 2014.
[21] Z. Kolka, D. Biolek, and V. Biolkova, “Hybrid modeling and emulation
of mem-systems,” Int J Numer Modell Electron Networks Devices
Fields, vol. 25, no. 3, pp. 216–225, 2012.
[22] D. Biolek, V. Biolkova, Z. Kolka, and J. Dobes, “Analog emulator of
genuinely floating memcapacitor with piece-wise-linear constitutive
relation,” Circuits Syst Signal Process., published online: 05 May 2015.
[23] Z.Biolek, D.Biolek, and V.Biolkova, “SPICE model of memristor with
nonlinear dopant drift,” Radioengineering vol. 18, no. 2, pp. 210–214,
2009.
 Depiction of Experimental Findings for a Cloud
Enterprise Architecture
Promise Mvelase, Nomusa Dlodlo, Happy Sithole
Centre for High Performance Computing
CSIR Meraka Institute
Pretoria, South Africa
{pmvelase, ndlodlo, hsithole}@csir.co.za
Abstract—Research and commercial efforts are currently
addressing challenges and providing solutions in cloud
computing. Business models are emerging to address different
use scenarios of cloud computing. In this paper, we present a
virtual enterprise (VE) model of cloud computing to enable
Small, Medium and Micro Enterprises (SMMEs) to respond
quickly to customers’ demands and market opportunities,
therefore Enabling Small, Medium, and Micro Enterprises
through Cloud utility Infrastructure: gaining agility and
flexibility needed for business success. In our virtual enterprise
model, temporary co-operations are created to realize the value
of a short term business opportunity that the partner SMMEs
cannot (or can, but only to a lesser extent) capture on their own.
This model is based on the realization that it is not economically
viable for SMMEs to acquire their own private cloud
infrastructure or even subscribe to public cloud services as a
single entity. The pricing model obtained from our proposed
business model shows the benefits that are derived from using the
VE cloud model over subscription to public cloud as a single
business enterprise. The pricing structure of our VE cloud model
is up to 17.82 times economical compared with equivalent
Amazon EC2 instance type pricing model.
Keywords—Cloud computing; service computing; SMMEs,
Virtual Enterprise
I. INTRODUCTION
Cloud computing is a computing paradigm in which every
layer of computing from infrastructure to application is a
service. It enables usage of computing hardware and
hardware belonging to a „third-party‟ thus lowering cost of
ownership of computing and enabling mobile computing [1],
[2], [3], [4], [5], [6], [7]. Adoption of cloud computing as a
utility infrastructure lowers total cost of doing business for
the Small, Micro and Medium Enterprises (SMMEs). The
SMMEs contribute to economic growth and promotion of
equitable development. The employment potential of SMMEs
at low capital cost has been the major advantage of the sector.
Employment intensity of the SMMEs sector is much higher
than that of the large enterprises. The SMMEs constitute over
90% of total enterprises in most of the economies and are
credited with generating the highest rates of employment
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Matthew Adigun
Centre of Excellence for Mobile e-Services
Department of Computer Science
University of Zululand
Kwadlangezwa, South Africa
profmatthew@gmail.com
growth and account for a major share of industrial production
and exports [31, 32, 50].
What appear to be the major constraints to the development
of SMMEs in many developing countries are limited access
to finance, technology, markets and management skills.
Access to and awareness of business information is also the
main constraint to the development and growth of SMMEs in
developing economies. Limited access to information
resources to start, survive and grow is one of the challenges
faced by SMMEs in enterprise development [33].
In the current e-business environment, individual enterprises,
including SMMEs cannot survive on their own. It is crucial
that SMMEs engage effectively with their partners and
customers. These enterprises require a certain way of e-
business interaction with their partners. The virtual enterprise
(VE) business concept, also known as the networked
enterprise, consists of distributed business functions and
utilities, outsourced to partners that work with the firm to
deliver the product to end customers. The VE model is one
such business environment that can facilitate cloud
computing for SMMEs. Emerging technologies, including
cloud computing, have the potential to transform and
automate the business processes of SMMEs and enable them
to engage with trading partners and customers in global
networks [34]. Our VE model of cloud utility infrastructure is
attempts to enable SMMEs take advantage of cloud
computing in a VE business model.
The network enterprise model is identical to the Grid-based
Utility Infrastructure for SMMEs-enabled Technology
(GUISET) ([8], [9]) project. GUISET is modelled to provide
computing utility infrastructure for small and medium
enterprises as well as the average rural dweller of African
communities from a cooperative/networked enterprise
viewpoint [8]. Cost is a key constraint for African users of
technology, thus our business model of VE for cloud
computing addressed pays specific attention to cost of
computing utility [2], [10]. Also, because a larger percentage
of users of the proposed VE model use mobile devices its
design addressed challenges associated with generic cloud
and mobile cloud [3], [4]. These are some of the ways in
which the design of the proposed VE model is peculiar.
A good number of cloud computing solutions exist in
commercial and research efforts, however, many issues are
largely open to research. They include „security, availability,
scalability, interoperability, service level agreement, data
integration, data governance, trust pyramid, user-centric
privacy, transparency, political and legal issues, business
service management‟ [11] among others [12], [13], [14].
Business service management is the main challenge
addressed in this research.
II. ADVANTAGES OF CLOUD COMPUTING
The major benefits of cloud computing must be understood
by architects in order to provide a future-state architecture –
an architecture that considers cloud computing potential.
Many major benefits emerge when cloud computing
technology starts to take hold:
 Separation and decoupling of business functions from
the framework required to run it.
 Adaptability to select various vendors that offer
scalable and reliable business services, development
environments and a framework that can be leveraged
out of the box. This is charged on the basis of
metering with no permanent contracts.
 Adaptable nature of the framework in order to
speedily allocate and de-allocate extremely scalable
assets to business services on the basis of demand.
The cloud ensures that the costs of attaining, delivering and
managing computing power are cut down by allowing
agencies to buy only the required computing services, rather
than investing in expensive and complex IT infrastructures.
The development costs can be borne by the agencies. In
addition to this, agencies can cut down the costs of
maintaining and testing existing and new systems.
Organizations can also drive down the costs of administration
and licensing of software as they can use online service with
the help of the cloud.
Within a network access device, the cloud guarantees to
provide users access to high-powered computing and storage
resources. By providing these services, cloud computing
technology helps to facilitate telework initiatives. In addition
to this, it also helps to strengthen the continuity of
organizational operations. The cloud is an on-going
computing resource that helps users to get used to the
consumption of resources to satisfy their needs. Significantly
scalable, cloud computing technology facilitates its
infrastructure to expand in an expedient and efficient manner,
without the need for major capital investments. Since the
resources are required and ended within a very short time
interval, capacity should be added. Therefore enterprises can
avoid the expense, latency and risk of buying software and
hardware that absorbs the data space center, and cuts down
the conventional time needed for scaling up the application to
maintain a job.
Cloud computing facilitates enterprises to easily shift
direction, remove capacity and therefore expenses as required
[35]. Cloud computing can provide the advantage of resource
maximization. This technology significantly eases the load on
already thin IT resources which are important for an
enterprise lacking IT professionals.
The advantage of collaboration is also provided by cloud
computing. The cloud provides an atmosphere in which users
are allowed to develop the software-based services that
improve collaboration and promote greater sharing of
information, not just inside the organization but also among
other private and government departments [36].
Cloud computing also gives the benefit of customization.
Cloud computing technology is a platform with extraordinary
potential for customizing and creating applications in order to
address a variety of challenges and tasks. Its agility means
that particular processes can be altered easily in order to meet
the varying agency needs, as the processes are mainly
changeable by making an organizational change, and not by
forcing redevelopment of the back-end systems.
A. Research goal and objectives
The main goal and objectives for this research are:
1) Research goal
The goal of this study is to develop a VE-enabled cloud
enterprise architecture that will enable SMMEs to participate
in a virtual operating environment.
2) Research objectives
In order to achieve the main goal of this research, we employ
the following objectives:
 To identify issues affecting SMMEs
 To identify existing VE models, that can be adapted
to meet the acute infrastructural deficiency of
SMMEs.
 To model a VE-driven cloud computing architecture
custom-made or custom-tested for SMMEs and
deploying a cloud for evaluation purposes.
a) Research Objective 1
Existing VE technology models were reviewed, which helped
us to identify useful VE models that can be adapted to meet
SMMEs‟ needs at both business and IT levels at optimized
cost.
b) Research Objective 2
This was achieved by identifying relevant literature on
existing VE architectures and cloud computing enterprise
architectures to enable us to design VE-driven cloud enterprise
architecture. A linkage between VE and cloud computing
enterprise architecture was discovered. The transformation of
cloud enterprise from virtual enterprise is facilitated by means
of the Internet, SOA, web services, B2B and BPMS
technologies. A conceptual analysis of existing VE
architectures and cloud enterprise architectures relating to our
study was conducted.
 c) Research Objective 3
An experimental set-up of a VE-driven cloud enterprise
architecture customized for SMMEs was deployed and
evaluated.
3) Resaerch Protocol
a) Objective 1
To identify issues affecting SMME productivity
 A case study conducted in KwaNongoma on
issues hindering SMMEs‟ productivity and
growth was useful in assessing the above
issue. Additional related literature was
reviewed.
b) Objective 2
To identify existing VE models that can be adapted to meet
the infrastructural needs of SMMEs
 The literature on VE models was useful as we were
able to determine how to scale down to an
advantageous model that could be adopted for our
proposed architecture.
 The existing literature on VE architectures was also
reviewed to help us find VE architecture components
that relate to cloud enterprise architecture to enable
us to design and deploy our customised cloud
architecture for SMMEs.
 A conceptual analysis was done of existing VE
architectures and cloud enterprise architectures
relating to our study.
c) Objective 3
To model a VE-driven cloud computing architecture custom
made or custom tested for SMMEs and deploy a cloud for
evaluation purposes
 We designed a VE-enabled cloud computing
architecture for SMMEs by combining the
capabilities of both existing VE
architectures and cloud enterprise
architecture.
 Keeping in mind the financial state and the
needs of SMMEs, we evaluated the cloud
computing deployment models to choose a
suitable one for SMMEs.
 We performed system design to be able to
define the architecture, components,
modules, interfaces and data for our
prototype that satisfied specified
requirements.
 We implemented and evaluated an
experimental set-up of a VE-driven cloud
computing architecture for SMMEs.
III. RESEARCH METHODOLOGY
A. Data Collection
A background study on rural communities and e-
government models was conducted. This study spoke into a
questionnaire to gather the information from both the
community and SMMEs on their needs. The questionnaire was
administered to three communities in KwaNongoma rural.
These were KwaKhangela, KwaMeme and KwaSomkhele.
Respondents were asked to give their opinions on how far
their needs were being met and comment on how the current
situation could be improved. The survey covered several
issues related to health, energy, water, education, IT usage in
the running of small businesses and transport, to name but a
few. The questionnaires were arranged in sections to find out:
 The government services that were available in the
deep rural area of KwaNongoma.
 How the community viewed these services.
 The levels of literacy and ICT literacy in the
community.
B. Conceptual analysis
1) Data Analysis
The data were analysed according to the research questions
parallel with prototyping. The findings of the survey showed
that a higher percentage of SMMEs do not use ICTs in
running their businesses. This is due to the lack of knowledge
and affordability of such resources.
The conceptual analysis was done to find existing virtual
enterprise and cloud enterprise architectures that can be used
to develop a prototype of VE-driven cloud computing
architecture custom made or custom tested for SMMEs.
2) Primary and Secondary Methods
The main research method is conceptual design, where a
VE-enabled cloud enterprise architecture custom made for
SMMEs is designed to enable SMMEs to share IT resources,
distribute responsibilities and capabilities hence become
responsive to market changes and customer demand. The
proposed architecture should provide SMMEs with the
flexibility, agility and adaptability required for them to be able
to cope with the rapidly changing market environment. We
carried out a literature review and conceptual analysis in
designing our architecture.
The secondary methods were: a literature review, and
prototype. The literature review provided a theoretical
background to cloud computing, virtual enterprise and other
underlying technologies that make cloud computing possible.
A software prototype was used in evaluating our proposed
cloud pricing model against the equivalent amazon EC2
instance type.
IV. BUSINESS MODEL OF THE PROPOSED VE MODEL
Figure 1 is the conceptual view of the business model of our
proposed VE-enabled cloud. The business model is pay-as-
you-use just like any other cloud infrastructure [1]. The goals
of cloud computing, among others, are to lower total cost of
ownership (TCO) and provide “business flexibility” for
acquisition and use of computing resources either for
business, governance, research, personal or any other use [1],
[2], [15], [7]. This goal is more critical for the proposed VE
model because the target users are the small and medium
enterprises in Africa. It is also intended for personal
computing needs of the common African citizens in E-Health,
Emergency response, E-learning, M-Commerce etc.
Obviously, cost is a critical factor for these set of target users.
In achieving this goal, the proposed VE model is designed to
use licensed open source infrastructures wherever possible
without compromising the high agility, resilience,
availability, reliability and scalability that a high performing
computing infrastructure like cloud requires [16], [17]. This
will ensure that services are provided to consumers at the
most affordable cost possible at optimum quality. We have a
situation in which small SMMEs in rural settings cannot
afford the services of a dedicated lawyer, human resource
practitioner, accountant, etc. Since these are services that are
required once-off as and when necessary the architecture
proposes that these SMMEs get those services from a private
cloud of service providers, which is a network of medium-
size enterprises (SMMEs) participating in a virtual enterprise
(VE) enabled cloud business structure. The private network is
made up of the virtual enterprise setting which in itself is the
business aspect and the technology.
A VE-enabled cloud enterprise structure and operation
consists of the medium-size enterprises in VE alliance and all
its business process utility and IT clouds computing services
and their providers. The cloud enterprise architecture for
medium-size enterprises participating in a VE setting is made
up of the business context, business services, business
processes and IT services. The business context layer is
responsible for the definition of business goals, strategies,
structure, policies and performance metrics and indicators.
The main users of services at this level are business owners
and executives who are hardly ever IT experts. The main
functions of a business such as human resources, payroll,
accounting, etc. are defined as coarse-grained services, called
“business services” in the business services layers. Users such
as business or IT architects may define or select the required
business services from out-of-box business services
blueprints. The IT services layer represents the services that
are obtainable in the cloud. Finally, the business processes
layer is the illustration of selection, design, integration and
composition of IT services in the form of workflows that
fulfil the needs of outlined business services. In this
architecture, the medium-size enterprises in the virtual setting
share business context, business services, and business
processes to improve competitive advantage, and quickly
respond to market opportunities. Hence the VE enabled cloud
enterprise architecture comes down to a value system, which
involves a number of companies‟ value chain that is
collaborating to deliver the end product to the customer. The
aspect of value chain is not covered in this paper.
Rather than relying on established organizations (e.g., Google
cloud, Amazon EC), SMMEs in the VE setting form their
own private cloud, where they collaborate their existing IT
infrastructure, skills, processes, organizational models and
core-competencies. There should be in place, strategies on
sharing competencies. We should remember that the SMMEs
also compete with one another; hence they cannot expose all
their competencies. The reason that they do not host their IT
services on a third party cloud service provider is because of
the advantages that come from the collaboration of resources
within the VE. Hosting their entire IT infrastructure to the
cloud service provider could cost them even more, and the
resources underutilized. Assuming that the SMMEs in the
alliance understand each other better, there can be no one
MSE dominating. In case of insufficient resources the
SMMEs can then tap into external public cloud. This comes
down to a hybrid deployment model. Therefore, cloud
computing capabilities provide the VE alliance with agility,
flexibility, and adaptability required due to its‟ highly flexible
ICT infrastructure.
Figure 1: Business model of VE-enabled Cloud Enterprise Architecture for
SMMEs
V. VE-ENABLED CLOUD ENTERPRISE PROTOTYPE DESIGN
The VE enabled cloud enterprise architecture for SMMEs
prototype shown in figure 1 illustrates the basic functions of
our customized architecture. We now present the prototype
design of our model with the aspect of the architecture use
case scenario described in this paper. The prototype design
supporting the SMMEs needs is explained using UML design
diagrams.
A. Use case scenario
The use case scenario is illustrated in Figure 2. The
technical management team in this VE model operates as an
independent business entity from the collaborating SMMEs.
This independent technical operator is regarded as cloud
service provider in this case. The cloud service provider
customizes the images from the desktop before uploading it to
the cloud. For Example, we assume that a cloud service
provider would have done his task, of finding out about the
users that he will be providing services to. Therefore the
images are customised according to individual user groups.
There can be a windows image which is loaded with HR,
inventory, ERP, CRM, legal services, accounting applications.
To cater for users who might not need all these services, other
images may have one or two of these services. In this case we
are avoiding a situation where a user might own accessing a
virtual machine with services that he does not need, which will
make the bill to be unnecessarily high.
The customer requests IT services from the cloud, the
system machine image provides a hard disk snapshot used to
launch a virtual machine (VM) instance. The VM instance has
RAM, number of CPUs and the network interface attached.
Depending on individual‟s customers‟ requirements the VM
instance may contain SaaS (HR system, inventory
management, legal services, etc) and a specified RAM and
number of CPUs. The cloud service provider creates a
customer‟s account and instantiate a VM instance dedicated to
that customer. The requester logs in to access the VM
instance, as and when needed. The VM instance appears as if
it runs on the user‟s machine when in actual sense it is seated
in the cloud server because of the abstraction that the cloud
provides. The customer only uses the cloud services when
needed and billed according to usage (i.e Utility business
model). Customers are recognised by the IP addresses
assigned to their VM instances. A customer is billed when the
VM instance is on the running mode. For a customer to
completely end the session the VM instance must be shut
down. Logging off keeps the VM instance running and the
customer is billed even though it is not using the cloud
services. The cloud automatically scales up and down
depending to the users‟ needs. Now the customer can enjoy
the benefits of the cloud by having access to services that he
could not afford had it not been for the cloud business model.
Once the session is closed by the customer, the administrator
sends the bill to the customer.
Figure 2: Use Case Scenario of VE-enabled Cloud Enterprise Architecture for
1) Customer SMMEs
In this case the customer is a small SMME user who
cannot afford to purchase ICT infrastructure due to financial B. Workflows of stepwise activities
constraints. The customer utilises the services of the cloud on Figure 3 shows the activity diagram of our VE-enabled
a pay per use basis which is cheaper than buying the ICT cloud enterprise architecture. The flow of events starts by
infrastructure. The cloud service provider does not grant the service request action initiated by the customer, where the
customer permission to create an instance, instances are customer logs into the system. The user management object
created by the cloud SP. Utilizing the services of the cloud authenticates the user; this is where a decision has to be made.
enables small SMME to become agile, grow in their If the user is a registered customer, the request is attended to.
businesses, and compete with advanced SMMEs. At this stage the request is fulfilled by creating the VM
instance. The image management object uploads VM instance
2) Administrator else the cloud server is scaled to cater for the request. The VM
The administrator is an organization or department which instance containing application requested by the user now can
uses the Nimbula director who will be responsible for the cost be provided to the requestor of services. It then uploads to the
incurred from using the system. In actual sense, this is the desktop of the requestor through a web interface. It must be
cloud SP site. From the definition found in Nimbula director noted that for services to be accessed from the cloud, the
this is the customer, which has been explained above. customer must have internet access, therefore any device with
internet connectivity will do. When the requestor finishes
utilizing the cloud services, the administrator bills the
requestor per utilised services. The requester closes the session
by paying for the utilised services.
 Figure 3: Activity Diagram of VE-enabled Cloud Enterprise Architecture for SMMEs
C. Resource management in the proposed VE model
One of the key features of resource management in
GUIEST is energy-awareness [19], [20]. Power constrain is a
serious concern for technology implementation and usage in
Africa. Another aspect of resource management concern is
flexible information processing [1], [21], [22], [23], [24], [25].
This will include dynamic service evolution to meet constantly
changing service consumers‟ needs.
VI. 9.1 PROGRAMMING, APPLICATION AND SERVICES
COMPOSITION MODEL OF THE PROPOSED VE MODEL
A simple programming for cloud computing platforms is
proposed in [26]. In [27] a workflow-oriented cloud
computing programming model was developed. The
applications developed on GUISET-Cloud are service oriented
applications [28]. As a software infrastructure, therefore a
major concern in the application and services model is
interoperability of application services among heterogeneous
platforms. Quality of services (QoS) is also a prime concern as
the services are user-centric. Applications are developed as
web services that are dynamically re-versioned to meet
B. Model prototyping
In this section we present simulated prototype of the
proposed model. The cloud deployment is illustrated in terms
of automation, identity, permissions and delegation, openness
and choice.
1) Basic assumption of the simulation
In developing our simulation, the following assumptions
are considered, bearing in mind the duration of the project.
● The cloud infrastructure is running, services
are deployed and service consumers request services.
constant changes in users demand. Applications services are
also multi-tier [29]. The application and services composition
model of the proposed VE model must therefore, address
challenges associated with these features as well as related
challenges. This is done by addressing the end-user
perspective with the goal of providing simplicity of access to
and use of services. The architectural perspective of the
proposed VE model is a large evolutionary system dealing
with ontologies and metadata; adaptation models and context-
awareness and reasoning under uncertainty and increasing
complexity. The infrastructure is made up of interconnected
nodes of any type providing resources ranging from
computational to knowledge resources [30].
A. Innovative security in the proposed VE model
Security issues in cloud computing is still an open research
topic just as many other cloud related issues [12], [13], [14],
[11]. The proposed VE model addressed security challenge in
cloud computing by finding the appropriate identity
management model that supports single-sign on, selective
authentication and authorisation of infrastructure entities and
agents [30].
● Services are consumed by users on a pay-
per-use basis, hence maintenance, scalability,
infrastructure is of third party‟s concern
● VM are created and customized per users‟
request and requirements
2) Description of the simulation
The scenario described in chapter four is considered in
simulating our model. End-consumer requirements,
infrastructure capabilities and cloud service catalogue that
operate in fulfilling a particular request were considered. In
order to fulfil a request, end-user requirements are the
deciding factor. The cloud is scaled according to the number
of users.
3) Simulation environment
The simulation of our model was done by installing
Nimbula Director. “Nimbula Director is an automated cloud
management system which allows customers to easily
repurpose their existing infrastructure and build a private
computing cloud in the safety of their own data centre.” This
cloud deployment model is appropriate for our architecture
since the medium sized enterprises will be forming a cloud
using their existing infrastructure to become cloud SPs for
very small enterprises. To install a Nimbula director site, we
needed a minimum of 3 machines setup, and a seed node
machine with a DVD. The machines comprising the Nimbula
Director cluster(s) need to comply with the following
hardware and software requirements, therefore the machines
we used meet the standard.
Nimbula director UI is divided to the top pain and the
bottom pain, this allows you to create, modify and destroy
objects.
The five main tasks at the top pain perform the following:
● User management create and manage
groups and their permission
● Image list contains a persistence list of
machine images that could be used to keep track of
different versions of a machine image.
● A machine image is a VM template that you
can launch into a running machine instance
● Virtual network allows you to create and
manage VEthernets and VDHCP servers. VE are
virtual layer 2 networks that provide isolation and in
the implement of using VLANs
● VDTP servers can be created for each
VEthernet to dynamically assign IP addresses to VM
instances running in that Vethernet.
● Network security list let you configure a
built in distributed firewall for isolating instances and
regulating traffic in and out of the cloud that is
dynamically configured and independent of the
underlying network.
● Instance management allow you to view and
launch machine images into running machine
instances.
A running VM instance created by the SMME cloud
administrator was used for our demonstration. The VM
instance is created using quick launch. The VM instance is
created via instance management task, here an instance can be
launched and viewed. Instances are customizable according to
user requirements. The administrator verifies that the new VM
instance has been launched and is running. The Web interface
displays details like the image list, state, placement
requirements, etc. for the new VM instance. An IP address to
use to connect to the VM instance is also provided (i.e. VM:
172.18.1.18). Before creating a VM instance the SMME
administrator creates the user SMME under the SMME
customer.
VII. COST SAVING EVALUATION
In evaluating the performance of our model, we have
evaluated our model according to utility evaluation of an
SMME to see if the utility requirements are fulfilled. Cost
saving is the basic requirement of our VE-enabled cloud
model. We compared our model with the Amazon Elastic
Compute Cloud (EC2) pricing. We used Standard on Demand
EC2 cost model in evaluating our model. EC2 has a number of
pricing models (Amazon.com
http://aws.amazon.com/ec2/pricing/). The Standard on
Demand model is the pricing model equivalent to the cloud
infrastructure in this work. In the Standard on Demand model,
the user pays for compute capacity by the hour with no long-
term commitments or upfront payments. Our pricing starts at
$300 per year per processor core - including support and
maintenance. This is equivalent to our model. Our model is
based on Nimbula Director, where software price is only
based on number of physical processor cores on which it runs
(i.e. the bigger the physical infrastructure, the more you pay
because you have more cores). This is the same criteria for
similar EC2 pricing models. However, for the configuration
and proposed model in this research, the Standard on Demand
pricing model is the ideal comparable model. The comparison
of pricing model of the proposed VE-enabled Cloud
Enterprise Architecture for SMMEs is therefore based on the
EC2 Standard on Demand pricing model. This does not
suggest rigidity in the pricing model of the proposed
architecture, but the analysis is done to show the cost saving
capability of the model.
Cloud providers provide four basic cost models 1 – 4 as
stated in [36]:
 Cost model for data storage
(size(total)×tsub×cost(storage) where tsub is the
subscription time),
 Cost model for computational machine
(cost(machine)),
 Cost model for data transfer into the cloud (cost(trans
f erin)), and
 Cost model for data transfer out to the cloud
(cost(trans f erout))
Amazon EC2 provides the flexibility to choose from a
number of different instance types to meet flexible computing
needs (see TABLE I.). Each instance provides a predictable
amount of dedicated compute capacity and is charged per
instance-hour consumed. The standard instance type has
memory-to-CPU ratios suitable for most general-purpose
applications.
 TABLE I. THE EC2 STANDARD INSTANCE TYPES

Standard On-Demand Instances Linux/Unix Usage Windows Usage

Small (Default) $0.080 per Hour $0.115 per Hour

Medium $0.160 per Hour $0.230 per Hour

Large $0.320 per Hour $0.460 per Hour

Extra Large $0.640 per Hour $0.920 per Hour

The Nimbula configuration equivalent to the highest EC2 standard instance was used in our prototype. This is shown in
TABLE II
TABLE II. NIMBULA DIRECTOR INSTANCE TYPE USED FOR OUR VE-ENABLED CLOUD

Instance Virtual Cores Memory Instance Store Platform Platform Price

Volumes

8 GB (3 x 8 250 GiB (3 x 250

Standard On-
4 (4x 4 virtual RAM) & 4 GB HD) & 500 GiB
Demand 64-bit High $0.034 per hour
cores)= 16 cores (1 x 4 RAM)= (1 x 500 GiB)=
Instances
32GB 1250 GiB

To illustrate the cost estimation, we examined the case of
VE-Enabled Cloud Enterprise Architecture using the Nimbula
Director Instance type in Table 3 and the amazon instance
type in Table 4. TABLE III and Table 4 show the estimated
costs based on instance type obtained in our private cloud and
Amazon EC2 respectively. TABLE III shows a Linux-based
saving for the SMMEs who are the target users of the
proposed architecture.
TABLE III. AMAZON EC2 INSTANCE TYPE WITH HIGH-MEMORY DOUBLE EXTRA LARGE
double extra-large instance in Amazon EC2. The
configuration we used is equivalent to a double extra-large
EC2 machine instance-type. The price of the EC2 instant type
configuration is 17.82 times more expensive than the
equivalent VE-cloud configuration in the proposed
architecture. This is a huge

EC2 Instance
Virtual Plat-
Instance Type Name Compute Memory store I/O Price
cores form
units (ECU) volumes

Standard on- High- M2.2x 13 4 (with 3.25 34.2 GiB 840 GiB (1 64-bit High $0.640
demand memory large ECUs each) x 840 GiB) per hour
instances double

in applicability. For instance the business model presented can

VIII. CONCLUSION AND FUTURE RESEARCH
In this paper, we presented the architectural design of a
cloud infrastructure targeted at providing service computing to
SMMEs in economically disadvantaged business context. The
various perspectives presented are specific in focus but global
be applied to developing economy in any part of the world.
The research issues highlighted such as security model,
software and application model and resource management is
informative for researchers, designers and managers of cloud
and service computing.
 Our on-going research is intended to build an evolutionary
system that demonstrates dynamic adaptation and
personalisation of functionality so that consumer and provider
of services can add and withdraw business resources without
needing to redesign the system. In this description, we
presented a high-level view of the architecture. Decomposition
of the building blocks is assigned to specific research topics
for implementation and prototyping as a proof of concept.
REFERENCES
[1] Foster I, Yong Z, Raicu I, Lu S. Cloud computing and grid computing
360-degree compared. Grid Computing Environments Workshop, 2008.
GCE '08. 2008; (DOI: 10.1109/GCE.2008.4738445): 1 – 10.
[2] Xinhui L, Ying L, Tiancheng L, Jie Q, Fengchun W. . The method and
tool of cost analysis for cloud computing. IEEE International Conference
on Cloud Computing .2009;CLOUD '09(DOI:
10.1109/CLOUD.2009.84.): 93 - 100
[3] Fernando N, Loke SW, Rahayu W. Mobile cloud computing: A survey. In
Press, Future Generation Computer Systems.2013;29: 84–106
[4] Seneviratne A, Thilakarathna K, Petander H, Wasalthilake D. Moving
from clouds to mobile clouds to satisfy the demand of mobile user
generated content. IEEE 5th International Conference on Advanced
Networks and Telecommunication Systems (ANTS).2011; (DOI:
10.1109/ANTS.2011.6163645): 1 – 4
[5] Buyya R, Yeo CS, Venugopal S, Broberg J, Brandic I. Cloud computing
and emerging IT platforms: Vision, hype, and reality for delivering
computing as the 5th utility. Future Generation Computer Systems.
2009;25: 599 - 616
[6] Angeli D, Masala E. A cost-effective cloud computing framework for
accelerating multimedia. In press, J. Parallel Distrib. Comput. 2012;
(doi:10.1016/j.jpdc.2012.06.005)
[7] Wang X, Wang B, Huang J. Cloud computing and its key techniques.
IEEE International Conference on Computer Science and Automation
Engineering (CSAE). 2011 ; 2: 404 - 410
[8] Adigun MO, Emuoyibofarhe OJ, Migiro SO . Challenges to access and
opportunity to use SMME enabling technologies in Africa. 1st all Africa
Technology Diffusion Conference, Johannesburg, South Africa. June 12-
14, 2006; 1: 34-43
[9] Ekabua OO. Change impact analysis model-based framework for service
provisioning in a grid environment. South Africa. A PhD thesis in the
department of Computer Science, University of Zululand; 2009
[10] Rodero-Merino L, Caron E, Muresan A, Desprez F . Using clouds to
scale grid resources: An economic model. Future Generation Computer
Systems. 2012; 28 :633–646
[11] Rimal BP, Jukan A, Katsaros D, Goeleven Y. Architectural requirements
for cloud computing systems: An enterprise cloud approach . J Grid
Computing .2011;9(DOI 10.1007/s10723-010-9171-y):3–26
[12] Sun D, Chang G, Sun L, Wang X. Surveying and analyzing security,
privacy and trust issues in cloud computing environments. Procedia
Engineering .2011;15(doi:10.1016/j.proeng.2011.08.537 ):2852 - 2856
[13] Tom T-h, Mohammed S. Toward designing a secure biosurveillance
cloud. J Supercomput.2011;Springer Science+Business Media, LLC(DOI
10.1007/s11227-011-0709-y):
[14] Mohamed EM, Abdelkader HS, EI-Etriby S. Enhanced Data Security
Model for Cloud Computing. The 8th International Conference on
INFOrmatics and Systems (INFOS2012).2012;14-16 May():CC12 -CC17
[15] Lin Y-K, Chang P-C. Performance indicator evaluation for a cloud
computing system from QoS viewpoint. Qual Quant.2011;Springer
Science+Business Media B.V. 2011(DOI 10.1007/s11135-011-9613-z ):
[16] Mauch V, Kunze M, Hillenbrand M . High performance cloud computing
. Elsevier Future Generation Computer Systems.;In
press(doi:10.1016/j.future.2012.03.011 ):
[17] Vecchiola C, Pandey S, Buyya R . High-Performance Cloud Computing:
A View of Scientific Applications . 10th International Symposium on
Pervasive Systems, Algorithms, and Networks .2010;IEEE Computer
Society(DOI 10.1109/I-SPAN.2009.150 ):4 - 16
[18] Alvaro J, Barros B . A new cloud computing architecture for music
composition. Elsevier Journal of Network and Computer Applications.
2012; In press (http://dx.doi.org/10.1016/j.jnca.2012.04.015)
[19] Beloglazov A, Abawajy J, Buyya R. Energy-aware resource allocation
heuristics for efficient management of data centers for Cloud computing.
Elsevier Future Generation Computer Systems .2012;28():755–768
[20] Cui Y, Ma X, Wang H. A survey of energy efficient wireless
transmission and modeling in mobile cloud computing. Springer Mobile
Netw Appl.Published online 19 April 2012;(DOI 10.1007/s11036-012-
0370-6):
[21]Lin C. A novel college network resource management method using cloud
computing. Elsevier Physics Procedia.2012;24 ():2293 – 2297
[22] Maurer M, Brandic I, Sakellariou R. Adaptive resource configuration for
Cloud infrastructure management. Future Generation Computer
Systems.2012;Accepted Manuscript(doi:10.1016/j.future.2012.07.004):
[24] Espadas J, Molina A, Jiménez G, Molina M, Ramírez R, Concha D. A
tenant-based resource allocation model for scaling Software-as-a-Service
applications over cloud computing infrastructures. Future Generation
Computer Systems.2012;In press(doi:10.1016/j.future.2011.10.013):
[25] Stanoevska-Slabeva K, Wozniak T, Hoyer V. Practical guidelines for
evolving IT infrastructure towards Grids and Clouds. In: Stanoevska-
Slabeva K (eds.)Grid and Cloud Computing: A Business Perspective on
Technology and Applications. 1st ed. Berlin Heidelberg: Springer-Verlag;
2010. p225 - 243
[26] Ranjan R, Zhao L. Peer-to-peer service provisioning in cloud computing
environments. J Supercomput.2011;Springer Science+Business Media,
LLC 2011(DOI 10.1007/s11227-011-0710-5):
[27] Martins J, Pereira J, Fernandes SM, Cachopo J. Towards a simple
programming model in cloud computing platform. In Proceedings of
First International Symposium on Network Cloud Computing and
Applications.2011;IEEE Computer Society(DOI
10.1109/NCCA.2011.21):83 - 90
[28] Pang J, Cui L, Zheng Y, Wang H. A workflow-oriented cloud computing
framework and programming model for data intensive application.
Proceedings of the 15th International Conference on Computer
Supported Cooperative Work in Design.2011;():356 - 361
[29] Tang K, Zhang JM, Feng CH. Application centric lifecycle framework in
cloud. Eighth IEEE International Conference on e-Business
Engineering.2011;IEEE Computer Society(DOI
10.1109/ICEBE.2011.32):329 - 334
[30] Iqbal W, Dailey MN, Carrera D, Janecek P. Adaptive resource
provisioning for read intensive multi-tier applications in the cloud. Future
Generation Computer
Systems.2011;27(doi:10.1016/j.future.2010.10.016):871–879
[31] Mobileeservices.org. Research themes. Centre for Mobile e-Services,
University of Zululand. http://mobileeservices.org/themain/node/10
(accessed 07 August, 2012).
[32] Aremu, MA and Adeyemi SL (2011) Small and Medium Scale
Enterprises as A Survival Strategy for Employment Generation in
Nigeria. Journal of Sustainable Development, 4(1), 200 – 206. Accessible
at http://ccsenet.org/journal/index.php/jsd/article/view/9240. Accessed
last on Tuesday, 27 November 2012.
[33] Sharma, M, Mehra, A, Jola, H, Kumar, A, Misra,M, Tiwari, V (2011)
Scope of cloud computing for SMEs in India. Journal of Computing 2(5):
144 – 149. Accessible at
http://arxiv.org/ftp/arxiv/papers/1005/1005.4030.pdf. Last accessed
Tuesday, 27 November 2012
[34] Abor, PA, Abekah-Nkrumah, G, Sakyi, K, Adjasi, CKD, Abor, J (2011)
The socio-economic determinants of maternal health care utilization in
Ghana. International Journal of Social Economics, Vol. 38 Iss: 7, pp.628
– 648. DOI: 10.1108/03068291111139258.
[35]Dai, W. (2009) The Impact of Emerging Technologies on small and
medium Enterprises (SMEs). Journal of business Systems, Governance
and Ethics 4(4): 53 – 60. Accessible at
http://www.jbsge.vu.edu.au/issues/vol04no4/Dai.pdf. Last accessed on
Tuesday, 27 November 2012
[36] Mantri, S., Nandi, G., Kumar, S., Kumar (Eds.), High perfomance
architecture and grid computing (pp. 4-10). Proceedings of the
International Conference, HPAGC 2011, Chandigarh, India. Series:
Communications in Computer and Information Science, Vol. 169.
[37] GTSI solutions. (2009). Building a framework for successful transition.
White Papers/Cloud Computing. Retrieved from
http://www.gtsi.com/cms/documents/White-Papers/Cloud-
Computing.pdf.
[38] Bloomberg, J., Schmelzer, R. (2006). The Business Inflexibility Trap.
Service Orient Or Be Doomed: How Service Orientation will Change
your Business How Service Orientation will change your Business (pp. 1-
11). (Original work published 2006). Retrieved from
Amazon/http://www.amazon.com/Service-Orient-Be-Doomed-
Orientation/dp/0471768588.
[39] Arteta, B. M., Giachetti, RE. (2004). A measure of agility as the
complexity of the enterprise system. Robotics and Computer Integrated
Manufacturing: 13th International Conference on Flexible Automation
and Intelligent Manufacturing, 20(6), 495-503. Retrieved from
Elsevier/http://www.sciencedirect.com/science/article/pii/S07365845040
00717.
[40] Camarinha-Matos, L.M (2005). Brief Historical Perspective for Virtual
Organizations, Virtual Organizations Systems and Practices
[41] Sobotta, A. (2008). Enhancing agility promoting benefits of service-
orientation with utility computing. Masters Dissertation, Copenhagen,
Denmark: IT University of Copenhagen (ITU). Retrieved from
http://gotze.eu/wp-content/uploads/2009/03/adrian-sobotta-master-
thesis.pdf.
[42] Davis, E. (2007.) What is on American minds? Management Review,
April, 14-20.
[43] Bloomberg, J., Schmelzer, R. (2006). The Business Inflexibility Trap.
Service Orient Or Be Doomed: How Service Orientation will Change
your Business How Service Orientation will change your Business (pp. 1-
11). (Original work published 2006). Retrieved from
Amazon/http://www.amazon.com/Service-Orient-Be-Doomed-
Orientation/dp/0471768588.
[44] Bharadwaj, A. S (2000). A resource-based perspective on information
technology capability and firm performance: An empirical investigation.
MIS Quarterly Executive, 24(1), 169-196.
[45] Leonard-Barton, D. (1992). Core Capabilities and Core Rigidities: A
Paradox in Managing New Product Development. Strategic Management
Journal, 13(5), 112-125.
[46] Overby, E., Bharadwaj, A., Sambamurthy, V. (2006). Enterprise agility
and the enabling role of information technology. European Journal of
Information Systems, 15(2), 120.
[47] Zhang, L., Zhou, Q. (2009). Cloud computing open architecture.
Proceedings of the IEEE International Conference on Web Services,
ICWS 2009, 607-616. Retrieved from IEEE Digital
library/http://ieeexplore.ieee.org/xpl/login.jsp?tp=&arnumber=5175875&
url=http%3A
[48] Weil, P., Broadbent, M. (1998). Leveraging the new infrastructure: How
market leaders capitalize on information technology. Boston: Harvard
Business School Press.
[49] Conboy, K., Fitzgerald, B. (2004). a study of agility in different
disciplines. Proccedings of the 2004 ACM Workshop in Interdisciplinary
Software Engineering Research, 37-44.
[50] Katua, N. T. (2014). The Role of SMEs in Employment Creation and
Economic Growth in Selected Countries. International Journal of
Education and Research Vol. 12 (2), 46-472
 Agent Development Platforms for Bioinformatics
Ajit Kumar Gopee, Geerish Suddul, A. Chikudo
University of Technology, Mauritius
Pointes-aux-Sables, Mauritius
agopee, g.suddul@umail.utm.ac.mu
atchikudo@gmail.com
Abstract—During the past decade, Molecular Biology
wet labs and sequencing activities have been generating a
vast amount of data, causing an increase in the adoption of
new methods and tools for identification and analysis. In
particular, the use of software agent technology in
bioinformatics applications has been on the rise, due some
of its unique features as autonomy and remote execution
capabilities. However, nowadays there exists multiple
software agent development platforms, and their potential
application to bioinformatics require an assessment based
on their inherent features. We therefore present a
comparative analysis of the main available platforms
focusing on standard compatibility, communication,
mobility, security policy, availability, usability and
development issues. Our result, based on a scoring system,
demonstrate that the JADE platform is the most
appropriate and promising one to be used in tackling
bioinformatics problems.
Keywords—Software Agent; Development Platform;
Bioinformatics.
I. INTRODUCTION
Bioinformatics is an interdisciplinary subject involving
Molecular Biology, Computing and Statistics to collect and
analyze biochemical and biological data. The major areas in
this field are: sequence analysis, Gene expression, and Protein
expression, Structural Bioinformatics, Network and System
Biology, Data mining, Gene prediction among others.
Bioinformatics has in the recent years, generated an avalanche
of data. Development and use of unsupervised and automated
analysis tools for these data became imperative.
All living organisms are made up of cells. Most of these
cells contains hereditary material, the chromosomes (DNA
and RNA).
Fig 1: Major parts of a DNA: genes (exon), non-coding
regions; noncoding regions called introns, within exons.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Y. Jaufeerally-Fakim
University of Mauritius
Reduit, Mauritius
yasmina@uom.ac.mu.ac
To be able to extract, sequence and analyze the DNA,
appropriate tools, hardware and software are essential in
Bioinformatics. During wet lab experimentation, DNA are
extracted and purified though a number of steps and then
sequenced.
Fig 2: Major steps in DNA Extraction and purification
DNA is composed of four major bases (Adenine, Guanine,
Cytosine and Thymine). The exact order of these bases
determine the gene and the latter determines the
characteristics, structure and function of proteins.
Determining the exact sequence of these bases is one of
the most fundamental task in Molecular Biology and
Bioinformatics. The next logical step is to be able to determine
its function from its underlying sequence. New generation
sequence technologies generate „Big data‟ which need to be
analyzed. „Big data‟ possess four essential characteristics, the
four popular V‟s: volume (amount) of data, data processing
velocity, data sources inter- and intra-variability, and veracity
of the data quality. Usually these data are very diverse,
complex (network interaction and dynamics): data can no
more be analyzed in isolation), and often of several scales. For
example, the Human Genome Project required the know-how,
technology, and human resource from twenty different
institutions spread over 13 years of work with a budget of
more than three billion dollars to determine the whole genome
structure which consists of approximately three billion
nucleotides. However the same tasks now can be done within
three days. This is due to high-throughput genomics. Specific
bioinformatics activities are employed for specific purposes.
According to existing literature, System integration seems
to be the most appropriate solution to deal with these
avalanche of data. Three essential aspects of system
integration are: distribution, autonomy and heterogeneity.
The development of software agents as well as their wide
usage requires good underlying infrastructure. Thus, simply
having an infrastructure is usually not enough. As user-
acceptance of the infrastructure depends on ease of application
development, versatility of applications, support for various
standards and most importantly the security of communication
among agents, security of agents on remote platforms and
security of mobile agent hosting platforms. Software Agents
are suitable for applications that involve repetitive, autonomy,
complex and time-consuming activities for examples merging
heterogeneous knowledge sources which are scattered in the
world; composite changing systems simulation. All these
features are very inherent to Bioinformatics. They strongly
favored the onset of Agent Technology in Bioinformatics.
Software agents seems to be the most appropriate and useful
technology for deploying distributed applications [12]. They
possess features like autonomy, flexibility, optimized use of
network resources, frequent entity interactions for various
tasks among others. Moreover, they have also been accepted
as boosting technology for bioinformatics applications
development. Agent technology suits applications that imply:
iterative and tedious activities; data and knowhow handling,
such as combination of different data provenance and
modelling of complicated, varying systems.
The analysis of Bioinformatics „Big data‟ can be carried out
using a system integration approach. Most system integration
model provides a three-tier architecture: a „user‟ level, a
„mediator‟ level, and a „wrapper‟ level. Below this layer are
the different physical heterogeneous databases.
Agent technologies can be used at different level in a system
integration model for e.g. at the user level, the „Mediator‟
level, at the „Wrapper‟ level. Each agent can specialize in a
specific task or class of tasks. They are programmed with self-
organizing features to deal with dynamic changes in data,
tasks, agent scalability, and to wisely avoid potential
bottleneck in the system. The agent has the capability of
decomposing (task decomposition) a complex task into
smaller and simpler ones which can be easily tackled by other
agents (task allocation). Moreover the three-tier model for
message passing among agents presents a huge achievement in
resolving heterogeneity. Agent systems can easily be adapted
to numerous semantic ontologies which is very typical of
Bioinformatics
Software Agents help to distribute the computing load, while
limiting the demand for transferring huge amounts of data. It
can also support the automation of information-gathering.
In this current research, we assess the characteristics of
existing agent IDEs that can be potentially applied in
bioinformatics. Each agent IDE assessment is based on:
● Standard compliance and compatibilities: for e.g.
FIPA, MASIF (OMG), among others.
● Support for inter-platform communications.
● Agent mobile capabilities: strong (code migration
and actual state of executing module), minimal
(shifting of code only).
● Policy: for secure intra- and inter-platform
communication.
● Available support and updates from developers
● Updates on documentations and usability: user and
developer level of acceptance.
● Updates on projects involving practical application
development.
II. LITERATURE REVIEW
Michael Luck et al. defines the core contribution to the
Open network for agent-based framework: AgentLink and
AgentLink II [16]. In the AgentLink III specialized meeting
in 2004, the Working Group on Agents in Bioinformatics
(BIOAGENTS) was created and this provided avenues for
merging the agent and the bioinformatics areas to unveil a new
and different paradigm suitable for computational frameworks.
Bioinformatics was identified as one of the applications where
agent technologies could be used. Since then there have been
many research groups around the world working in this field.
Below is a summary of some research groups and their major
research works:
Some preliminary work involving agents: genome
investigation initiated within Gene Weaver project[13];
myGrid eScience project, [22], DECAF project [14,15], and
intracellular signaling pathways [23].
Luciano Milanesi gave the presentations on Grid.it. Nicola
Cannata and Andrew Martin, outlined several situations where
Software agents could be used. Flavio Corradini, an expert on
formalization of composite systems, shared his knowhow on
applying formal and semi-formal approaches for specifying
such systems in Bioinformatics. [24]
Luciano Milanesi presented on the Italian Grid.it, a
platform targeting intensive computing, in continuously
growing institutions. Within the project, a dedicated team
group was mandated to develop grid-based solution for life
sciences research; one example is the relationship between
protein surfaces and their explicit functions. [24]
Giuliano Armano expose was on the prediction of protein
secondary structure by software agents (and which is still a hot
topic nowadays [12]) to support the development of open-
source system, and to merge heterogeneous databases.
Some previous research has also been done revolving
around Software Agent Development platforms. Some
examples are: Perdikeas, Chatzipapadopoulos, and Venieris
studied Aglets, Odyssey, Concordia, Voyager, and
Grasshopper, and also showed some preliminary results, [25];
Silva, Soares, Martins, Batista and Santos unveil the outcome
of some experimental research on some existing
platforms[26]; Dikaiakos, and Samaras assess the features of
Concordia and Aglets, [1], [27]; Silva, Romao and Diego
Silva studied mobile agent IDEs such as AgentSpace ,
Telescript, and Aglets, [28]; Burbeck, Garpe, and Tehrani
evaluate experimentally Tryllian , Aglets, and Grasshopper,
[29]; (Leszczyna compares agent platforms such as
Grasshopper ,Tryllian, and JADE, although not with mobile
agents,[30]. Shakshuki presented results with Voyager and
Aglets, by using agent toolkits such as JADE and Aglets, and
provided some promising results, but mobile agents were not
considered [31]. Consequently this emerged as an interesting
avenue that can be explored in the future.
All these previous work done is based on general qualitative
comparisons of various development toolkits. In the next
chapter, a detailed study is carried out on some of the most
popular platforms.
III. POPULAR SOFTWARE AGENT DEVELOPMENT PLATFORMS
A. JADE
JADE (Java Agent Development Environment) [10], a
Java based framework, was created by Tilab for the
deployment of multi agent applications (MAS) based on the
sharing paradigm architecture. The latest version, 4.3.3 was
developed in December 2014. It facilitates the development
of MAS through a middleware that complies with the latest
FIPA specifications. The IDE supports the debugging and
deployment phases of agent development by the use of
integrated graphical tools. The communication environment
evolves gradually according to peers‟ needs and requirements.
In JADE an instance of runtime environment is known as a
container, and it manages all the agents created in it.
Collection of such containers is called a platform and it
incorporates a uniform layer which masks the ramification of
concealed hardware and software. Finally, it is essential to
unveil its support for semantic representation of knowledge
about agents. Every container in JADE comprises of Directory
Facilitator (DF) agent, Remote Monitoring Interface (RMI)
agent & Agent Monitoring System (AMS) agent. In JADE,
agents communicate using asynchronous message passing
technique which is most widely accepted model for distributed
and loosely coupled communications.
JADE maintains the security of agents through strong
authentication mechanisms to verify any agent‟s right..
Messages exchanged among agents comply with Agent
Communication Language (ACL) defined by FIPA. JADE
supports execution of multiple parallel tasks within the same
java thread. This feature ensures scalability as well as meets
resource constraints of environment. In the absence of Proxies
an agent locates its destination by interacting with the AMS
(Agent Management System).
B. Aglet Software Development Kit(ASDK)
ASDK [32] is a JAVA-Based mobile agent development
platform, latest version being Aglet 2.5 alpha. It is an open
source freely available toolkit, which provides a good
graphical user interface for agent development. It consists of
two packages, The Aglet Building Environment (ABE) and
the Aglet Workbench. Aglet workbench aims at developing
stand-alone mobile agents. The ABE (SDK) comprises of
Aglet API, the Aglet Server known as Tahiti and the Agent
Web Launcher called Fiji along with documentation and
sample Aglets. Aglets are basically java objects which
constitutes two major components i.e. Aglet Core and Aglet
Proxy. Aglet server Tahiti is an application program that
works as agent server for aglets. It provides users with a good
GUI and allows users to create and dispatch an agent, monitor
it, dispose it off when required. It gives user the ability to set
agent‟s access privileges on the server.
Aglets support both agent mobility as well as predefined
movement of the agent on the network also called as Itinerary.
One of its strength is that it follows the OMG Mobile Agent
System Interoperability Facility (MASIF) specification. Aglets
integrates a single-thread processing with a message-flow
based communication infrastructure. Features to support both
concurrent and asynchronous calls are available. Aglets use
brokers as a suitable mechanism to reference distant agents.
Agent Communication is carried out using synchronous and
asynchronous message flow. Agent Transfer Protocol (ATP)
along with Java Agent Transfer and communication Interface
(J-ATCI) also facilitates this flow of messages.
A major drawback of the platform is that the proxies are
ineffective once the referenced agent change location. Thus,
the programmer must request for an updated surrogate, if
required, before using it. Since single-thread processing is
used, the programmer must avoid executing time-consuming
instructions , else this would prevent other agent events (for
example incoming messages) from being processed. Remote
calls to agents or user-friendly identifiers assignment are not
supported. Its security is knitted in the concept of restricting
transfer of aglets only to its own servers. Scalability is another
problem, since aglets are not inter operable with other
platforms or their agents, due to their restriction of working
with their own server.
C. Anchor
Anchor agent toolkit [33] was developed by Lawrence
Berkeley National Laboratory, in U.S.A. It facilitates the
conveyance and control of secure mobile agents in disparate
distributed environments. Its architecture includes an Agent
API, an agent viewer graphical User interface, an Anchor
server, Anchor security manager), Anchor class loader ,secure
agent transfer protocol handler, Anchor Java Naming and
Directory Interface and Anchor Java Native Interface
components.
The agent server in Anchor is the operating environment
which acts as foundation of this IDE. A precise port is
allocated and runs on a host. It performs all system related
functions. Anchor server supports the agent migration through
SATP. Secure socket layer (SSL) is used to provide agent
authentication. Akenti, an essential component of Anchor and
which is an access control system, enforces controlled access
of distributed resources. This latter uses Public Key
infrastructure. Access control decisions are made using
digitally signed certificated based on X.509 standard.
Integration of Akenti component in Anchor provides a strong
security system. AJANDI component in Anchor provides a
naming protocol through which every agent can be activated
and display its actual status.
D. Grasshopper
Grasshopper [34] was initially proposed by IKV++ in
1999 before being taken up by Enago Mobile. It is a user-
friendly IDE for mobile agents. A Grasshopper system
 consists of several „regions‟ and provides nice features,
including a GUI to handle agents and regions. The
different regions provide a mechanism to simulate
changing brokers. The major weakness of Grasshopper is
that there are no current maintenance and updates. Since it
has to update every surrogate right before using it, this can
create an impediment on the region server. Another
problem with Grasshopper is that a call to a moving agent
can end up abruptly at origin Finally, as in Aglets, the
same preset protocol is always observed till the end of an
agent track.
E. Vogager
Voyager [35] is an agent development tool developed by
Object Space, in mid-1996. The latter has been taken over by
Recursion Software Inc. since 2001. The latest version
available is Voyager 8.0. It‟s a simple but powerful Java-based
technology for creating mobile agents. It provides features
such as mobile codes, agents, and dynamic CORBA proxies
generation. Through remote call protocols, and by using
proxies Agents communicate among themselves. By hopping
in between the proxies, Voyager hides the underlying location.
Voyager is an easy-to-use platform, with a plethora of
functionalities, which simplifies the development of
distributed systems. It allows message flow among agents,
even while traversing the network, irrespective of its position
in the network. It supports synchronous communications.
Whenever an agent moves, it leaves behind a forwarder object
which forwards the message to its new location. Though the
tool provides flexible life spans for its agents, it is not
available for free, which could discourage many researchers
from using it. Another weakness is that to find an agent the
whole chain must be traversed.
F. SPRINGS
SPRINGS [36] developed recently at the University of
Zaragoza in Spain, deals with scalability and reliability in in
situations where moderate and high number of mobile agents
prevails. It has a region-based hierarchical architecture. It
offers location transparency: 1) there is no need to specify
network addresses but just the name of the destination; and 2)
by using varying surrogates for call management. Moreover, it
tries to reduce the live-lock problem that can arise when
agents migrate quickly. SPRINGS does not comply with FIPA
standard agent communication. Though it is easy to use, it
does not provide an appropriate GUI. Lastly, documentation is
sparingly available about this new platform.
G. Tracy
Tracy, [37] developed at the University of Jena in
Germany has a plugin-oriented architecture. Therefore, Tracy
agencies are lightweight and scalable (new addons can be
easily included to manage new services). The platform has
several agent migration protocols. However Tracy does not
allow remote agent-agent communications. An agent must
move down to an active agent in order to communicate with it.
The platform was essentially used as a test environment where
different migration and class loading protocols could be
assesses, and in so doing, the authors came up with laudable
and stimulating research work. Though it is a new platform,
there has been very few interactions in the blog and the
mailing list among the agent communities.
H. \Zeus
Zeus [38] is a comprehensive IDE for fast collaborative
agent applications deployment and it was proposed by
Advanced Applications & Technology Department of British
Telecommunication labs. It is open source freely available
toolkit. It is compatible with most hardware platforms since it
is entirely implemented in Java. It also complies with FIPA
standards. Most generic agent functionalities are available
under Zeus. The planning and scheduling of agent's actions are
easily managed by employing sophisticated protocols. A
comprehensive set of software components and tools are
available to implement and manage agent-based systems.
Agent Communication Language (ACL) or Knowledge Query
Manipulation language (KQML) is used to establish agent-
agent communication while the security aspect is handled
through public and/or private key cryptography, and digital
signature technologies.[5].
IV. TOOLKIT COMPARISON
This section compares the above toolkits major features
which can affect their applicability.
● Nature of product: whether a toolkit is open source
or commercially available influences end users a
great deal, as if a product is freely available some
compromises can be made with features provided as
compared to the one which involves some cost.
● Standard Implemented: Foundation for Intelligent
Physical Agents (FIPA) is a non-profit organization
involved in Agent protocols standardization and
specification. Its goal is to establish internationally
accepted specifications so as to maximize
interoperability across agent based applications,
services and equipment. Thus if a toolkit comply with
FIPA standards it increases its utility and scalability.
● Communication Technique:
Asynchronous communication is preferred over
synchronous communication since it does not require
simultaneous presence of all parties.
● Security Mechanism: Considering the nature of
applications where agents are employed these days,
strong security features embedded in a toolkit can
make it more appealing especially that they operate
in heterogeneous distributed environments.
● Agent mobility: Network load and traffic can be
reduced if agents are mobile in nature, and at the
same time the efficiency of agents are increased. It‟s
a desired feature when developing agent platforms.
● Migration Mechanism: RMI mechanism consumes
more time and resources in agent migration compared
to socket mechanism.
Table 1 (on page 6) shows a qualitative Comparison among
existing Agent Development Platforms. The different
platforms‟ specific feature were given specific scores
depending on the importance and relevance of each feature.
Some scores were allocated based on the availability or in-
availability of a feature. Overall, the JADE platform had the
highest score, followed by the Grasshopper, Anchor and
Aglets platforms respectively.
Table 2 (below) shows a comparison among four most popular
Agent Platforms and their Task Suitability in relation to three
major bioinformatics tasks.
Table 2: Agent Platforms and Bioinformatics Task Suitability
V. DISCUSSION
After comparing the various platforms scores were
allocated based on the various factors and characteristics of
each platform. According to the comparisons, the JADE
platform had the highest score. This is mainly due to the fact
that the platform uses the behavior model and is FIPA
compliant. In addition to that, the JADE platform is very
secure and has a user friendly GUI. The Grasshopper platform
was second from JADE, but it uses a procedural model and
only implements the basic mechanisms, despite being FIPA
compliant.

Agent Data Sequence Gene VI. CONCLUSION & FUTURE WORK

Platform Mining Alignment Prediction By now the Bioinformatics researchers have realized that
JADE Yes (8) Yes (10) Yes (10) the integration of agents into bioinformatics will unleash
Grasshopper Yes (10) Yes (6) Yes (6) untapped future potential. This is mainly due to the fact that
Anchor No (0) Yes (7) Yes (7) the Bioinformatics domain is generating an avalanche of data
every day. Conventional programming paradigms (for e.g.
Aglets No (0) Yes (7) Yes (7) procedural or Object Oriented) can no more effectively cope
with such situation. Software agents seems to be the promising
avenues in terms of autonomy, flexibility, self-
A. Data Mining learning/adapting, optimal use of resources including network,
One of the trends of modern genomic research is the among others. In this research we assess the features of current
generation of avalanche of raw sequence data. Two highly software agent platforms used in bioinformatics and apply a
desirable and practical objectives of data mining, are simple scoring scheme in an attempt to identify the most
prediction and descriptive annotation. The Grasshopper and appropriate platform(s) to be used in Bioinformatics at large.
JADE platforms support the remote call functionality and are In this context, we are considering carrying out further
therefore suitable for data mining. The JADE platform is evaluation, comparison and implementation through the
suited to this task but not as much as grasshopper since the eventual development of a cross platform bioinformatics tool
later supports proxies and dynamic proxies. The Anchor and to be deployed for the collection and analysis of information
Aglets platforms on the other hand cannot implement remote on insulin related genes using the JADE framework.
calls.

B. Sequence Alignment
Sequence comparison is essential in bioinformatics
analysis. It is of paramount importance toward functional and
structural analysis of newly discovered sequences. The JADE
platform is the most suited for this task due to its use of the
behavior model which greatly simplifies the creation of the
sequence alignment algorithms. Aglets and Anchor use the
event model which is more suitable than the procedural model
implemented by Grasshopper.

C. Gene Prediction
Gene prediction is a prerequisite for detailed functional
annotation of genes and genomes. The current gene prediction
methods can be categorized into two major classes, abinitio–
based and homology-based approaches. Both approaches
adopt the sequence alignment technique as a step in the
prediction procedure, therefore the suitability ratings are
similar to the sequence alignment column.
According to the evaluation, the characteristics of the
JADE platform gave the highest score amongst the evaluated
platforms, making it the most suitable platform for the three
main bioinformatics tasks namely data mining, sequence
alignment and gene prediction.
 Table 1: Qualitative Comparison among Agent Development Platforms
REFERENCES
[1] M. D. Dikaiakos and G. Samaras, “Performance evaluation of mobile
agents: Issues and approaches, lecture notes in computer science,” in
Performance Engineering, vol. 2047. Springer, May 2008, pp. 148–166.
[2] P. Braun and W. R. Rossak, Mobile Agents-Basic Concept, Mobility
Models, and the Tracy Toolkit. Morgan Kaufmann Publishers,
December 2005.
[3] C. Spyrou, G. Samaras, E. Pitoura, and P. Evripidou, “Mobile agents for
wireless computing: the convergence of wireless computational models
with mobile agent technologies,” Mobile Networks and
Applications(MONET), vol. 9, no. 5, pp. 517–528, October 2004.
[4] D. Milojicic, F. Douglis, and R. Wheeler, Mobility: processes,
computers, and agents. Addison-Wesley Professional, April 2009.
[5] Bellifemine, F. L., Caire, G., and Greenwood, D. (2007). Developing
Multi-Agent Systems with JADE. Wiley.
[6] Foundation for Intelligent Physical Agents (FIPA). Foundation for
Intelligent Physical Agents (FIPA). Website. Available at
http://www.fipa.org/ (last access: August 07, 2008).
[7] Aglet community http://aglets.sourceforge.net/ [Last date accessed:
27/07/15]
[8] FIPA-OS http://fipa-os.sourceforge.net/ [Last date accessed: 27/07/15]
[9] Grasshopper – the agent platform http://www.grasshopper.de/ [Last date
accessed: 27/07/15]
[10] JADE http://jade.cselt.it/ [Last date accessed: 27/07/15]
[11] Merelli et al, Agents in Bioinformatics, Computational and System
Biology, Briefings in Bioinformatics, VOL 8. NO 1. 45-59,
doi:10.1093/bib/bbl014, Advance Access publication, May 26, 2006
[12] AgentLink III , http://www.agentlink.org [Last date accessed: 29/07/15]
[13] Bryson K, Luck M, Joy M, Jones D. Cooperative Information Agents
IV, Lecture Notes in Artificial Intelligence. Berlin/Heidelberg: Springer-
Verlag; 2000. Applying agents to bioinformatics in Geneweaver; p. 60-
71.
[14] Graham JRA, Decker K, Mersic Decaf – a flexible multi agent system
architecture. Auton Agents Multi Agent Syst 2003;7:7-27.
[15] Agent system for genomic annotation. Int J Cooperative Inf Systems
2002;11:265-92.
[16] M. Luck, P. McBurney, and C. Preist, Agent Technology: Enabling Next
Generation Computing (A Roadmap for Agent Based Computing),
AgentLink, 2003.
[17] E. Merelli, R. Culmone, and L. Mariani ,Bioagent: A mobile agent
system for bioscientists, in NETTAB - Agents in Bioinformatics,
Bologna, 2002.
[18] K. Bryson, M. Luck, M. Joy, and D. Jones, Applying agents to
bioinformatics in geneweaver, in Cooperative Information Agents IV,
Lecture Notes in Artificial Intelligence, Springer-Verlag, 2000, pp. 60–
71
[19] Decker K, Khan S, Schmidt C, et al .Biomas: a multi-
[20] R. A. S. Moeller, M. Schroeder,Conflict resolution for the automated
annotation of transmembrane proteins, Computational Chemestry., 26
(2001), pp. 41–46.
[21] K. Webb and T. White,Cell modeling using agent-based formalisms, in
Autonomus Agent and Multi-Agent
[22] http://www.mygrid.org.uk [Last accessed: 29/07/15]
[23] González PP, Cárdenas M, Camacho D, Franyuti A, Rosas O, Lagúnez-
Otero J.-Cellulat: an agent-based intracellular signalling model,
Biosystems. 2003 Feb-Mar;68(2-3):171-85.
[24] http://www.grid.it [Last accessed: 29/07/15]
[25] Chatzipapadopoulos, FG; Perdikeas, MK; Venieris, IS
Mobile agent and CORBA technologies in the broadband intelligent
network, IEEE COMMUNICATIONS MAGAZINE, issue: 6, volume: 38,
year: 2000, pages: 116 – 124
[26] LM Silva · G Soares · P Martins · V Batista · L Santos - Comparing the
performance of mobile agent systems: a study of benchmarking,
Computer Communications 04/2000; 23(8). DOI:10.1016/S0140-
3664(99)00237-6
[27] M. Dikaiakos, Melinos k., G. Samaras, "A Performance Analysis
Framework for Mobile-Agent Systems, " 5th International Conference
on Mobile Agents. Atlanta, Georgia, USA, December 2 - 4, 2001.
Published by Springer-Verlag in the Lecture Notes on Computer Science
series.
[28] A.R. Silva, A. Romao, D. Deugo and M. Silva, "Towards a reference
model for surveying mobile agent systems", Autonomous Agents and
Multi-Agent Systems, vol. 4, no. 3, pp. 187-231, September 2001.
[29] Burbeck, K. Garpe, D. Nadjm-Tehrani, S.: Scale-up and performance
studies of three agent platforms. In: Performance, Computing, and
Communications, 2004 IEEE International Conference on, pages: 857-
864. April 15-17, 2004, Linkoping University.
[30] R. Leszczyna, 2004, “Evaluation of agent platforms,” European
Commission, Joint Research Centre, Institute for the Protection and
Security of the Citizen, Ispra, Italy, Tech. Rep.
[31] E. Shakshuki, “A methodology for evaluating agent toolkits,” in
International Conference on Information Technology: Coding and
Computing (ITCC‟05), Las Vegas, Nevada, USA, vol. 1. IEEE
Computer Society, April 2005, pp. 391–396
[32] Aglets Workbench, by IBM Japan Research Group.
http://aglets.trl.ibm.co.jp [Last date accessed: 27/07/15]
[33] Mudumbai S.S., William J. and Abdelliah E., „Anchor Toolkit-A
Secure Mobile Agent System‟. Published in eScholarship. Available at
http://escholarship.org/uc/item/2594j56c [Last date accessed: 28/07/15]
[34] C. Bäumer, Thomas Magedanz - Grasshopper - A Mobile Agent
Platform for Active Telecommunication, , IATA '99 Proceedings of the
Third International Workshop on Intelligent Agents for
Telecommunication Applications, Pages 19-32, Springer-Verlag
London, UK ©1999
[35] „Voyager and Agent Platforms Comparison‟published by ObjectSpace
Inc. for Voyager 1.0, available online at
http://www.cis.upenn.edu/~bcpierce/courses/629/papers/unfiled/AgentPl
atformsW97.PDF, September 1997.
[36] Sergio Ilarri, Raquel Trillo, Eduardo Mena - SPRINGS: A Scalable
Platform for Highly Mobile Agents in Distributed Computing
Environments. Conference Paper · January 2006, DOI:
10.1109/WOWMOM.2006.103 ·Source: DBLP Conference: 2006
International Symposium on a World of Wireless, Mobile and
Multimedia Networks (WoWMoM 2006), 26-29 June 2006, Buffalo,
New York, USA, Proceedings
[37] Braun, P., Müller, I., Schlegel, T., Kern, S., Schau, V., & Rossak, W.
(2005). Tracy: An extensible plugin-oriented software architecture for
mobile agent toolkits. In Software Agent-Based Applications, Platforms
and Development Kits (pp. 357-381). Birkhäuser Basel.
[38] J C Collis, D T Ndumu, H S Nwana and L C Lee - The ZEUS agent
building tool-kit , BT Technology Journal, July 1998, Volume 16, Issue
3, pp 60-68
[39] K. Shunmuganathan et al, Agents Based Bioinformatics Integeration
using RETSINA, The International Arab Journal of Information
Technology, Vol 5, No 3, July 2008.
[40] Rusli Abdullah, Hamidah Ibrahim, Rodziah Atan, Suhaimi Napis, Azmi
Jaafar, Mohd Hasan Selamat, Nurul Haslina Hairudin and Siti Hernazura
Hamidon, M. Jamil, Applying Agent Technology to
FacilitateKnowledge Sharing Among Bioinformatics, IJCSNS
International Journal of Computer Science and Network 310 Security,
VOL.8 No.4, April 2008
[41] G.S. Bhamra et al, Intelligent Software Agents: An Overview,
International Journal of Computer Applications, March 2014, DOI:
10.5120/15474-4160
[42] G.S. Bhamra et al, Agent enabled mining of Distributed Protein Data
banks, International Journal in Foundations of Computer Science and
Technology, Vol 5, No 3, May 2015
 Confused which Educational Video to Choose?
Appropriateness of YouTube Videos for Instructional Purposes- Making the Right Choice
Perienen Appavoo Meera Gungea
p.appavoo@open.ac.mu m.gungea@open.ac.mu
Open University of Mauritius
Réduit, Mauritius
Abstract—The ubiquity of computers and increasing Internet
connectivity is leading to the emergence of innovative ways of
learning such that terms like mixed-ability teaching, independent
and flexible learning and interactive lessons are becoming
buzzwords for the learning society. The availability of the
enormous variety of online free resources poses problems when it
comes to picking out the relevant, updated and appropriate ones
be it for general use or for educational purposes. YouTube Edu
offers a learning platform with hundreds of thousands of
instructional materials, only a mouse click away. In an attempt
to spot the best videos, many are guided by the number of views
and likes. This study investigated the appropriateness of this
selection approach. Media specialists and video producers
brainstormed, discussed and agreed on a list of 16 criteria which
were then applied to 30 videos dealing with three subjects of
Mathematics, namely Fractions, Sets and Integers. Each video
was attributed a suitability video index (SVI) on a scale of 0-48.
The study revealed that if in the first instance, the number of
views and likes can be a first step towards selecting appropriate
videos, yet a more scientific approach needs to be adopted and
this study has compiled a set of carefully selected criteria to
assess the suitability of videos for instructional purposes. These
criteria can be used as important guidelines for video makers
wishing to produce sound and pedagogical instructional videos.
Keywords—YouTube Education, innovative video technology,
learning mathematics
I. INTRODUCTION
Created in February 2005 by Chad Hurley, Steve Chen, and
Jawed Karim, three former PayPal employees, YouTube
provides a public-access Web-based platform that allows
people to upload, view and download videos, and the
possibility to share video clips on the Internet through other
web sites, mobile devices, blogs, and emails [1]. With more
than 1 billion users everyday, people spend hundreds of
millions of hours on YouTube and generate billions of views.
YouTube introduced YouTube EDU which brings learners and
educators together in a global video classroom where viewers
can have access to a broad set of educational videos.
YouTube created two programs to help schools and teachers
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Teenah Jutton Preetamsingh Dookhun
t.jutton@open.ac.mu p.dookhun@open.ac.mu
utilize YouTube EDU most effectively: YouTube for Schools
and YouTube for Teachers [2].
II. LITERATURE REVIEW
A. History
The advent of technology has revolutionized learning in many
innovative ways. Non-print multimedia-based technologies,
such as instructional videos, are now regarded as cost-
effective, interactive, effective teaching and learning tools [3].
Since the mid-20th century, instructional technologies and
educational media have undergone major development.
Paving the way from learning and teaching through audio in
the 1950s, the very first instructional videos appeared in the
late 1960s. Parallel with the introduction and nearly universal
use of personal computers, an evolution of instructional
technology was witnessed. This continued to expand with the
appearance of the Internet for interactive educational lessons,
webinars, podcasts, live distance video interactions, and Web-
based courses and programs [4].
If before the millennium people had to buy or rent films,
YouTube has been able to position itself as a giant library with
millions of free videos to choose from. YouTube emerges as
the third most popular website behind Google and Facebook
[5], being the go-to site for video on the web.
B. Contribution of ICT, including YouTube videos to learning
Videos are just one constituent of the complex classroom
activity system. Thus the learning outcomes depend largely on
the way video viewing is integrated with other learning
resources and tasks [6].
Anderson [7] defines ICT in education as ''the digitizing of
human knowledge, cloud computing, social networking,
touch-screen technology, and the convergence of mobile and
PC technologies''
Incorporating technology into the classroom benefits not only
the students but the teachers as well [8]. In addition, all
students do not have the same learning abilities, and
technology posits itself as a tool that allows teachers to reach
out to them [9].
Snelson [10] argues that there have been various studies in the
past to analyse the potential of YouTube in the arena of online
education. He further stipulates that playlist lessons have been
created to meet learning objectives across the cognitive,
affective, or psychomotor learning domains in real-world
online classrooms[11], [12]. Hence, when selecting videos, the
process should be according to the objectives the teacher
wishes to achieve [15].
Findings have also revealed that videos offer students the
opportunity to gain insight into events that they usually cannot
experience in real life [13]. In a study by Badr A. Al-Jandan
et al [14] students considered the inclusion of videos in the
lectures to be useful. The usefulness of video in teaching could
be attributed to the fact that it leads to enhanced visualization
and better understanding of the procedures/facts being
explained during teaching, thereby leading towards a better
clarification of those facts, increased interest in the lecture and
easy reproduction of knowledge gained. For example, videos
can better illustrate the eruption of a volcano or the habitat of
sea animals.
Over and above improving academic level of students,
educators also widely use educational videos for their own
professional development. Today, the biggest challenge is to
provide both teachers and learners with videos that can meet
their needs. Niken & Van der Voort [16] identified nine
quality criteria in evaluating children‟s television programmes.
These are comprehensibility (to understand the main messages
in the program easily), credibility (to decide if the program has
a significant value), entertainment (to find the program
joyful), aesthetic quality (to evaluate the program as
artistically well designed), involvement (programme‟s getting
attention of the children easily), presence of role models (to
find positive role models for children in the programme),
innocuousness (to find the programme not harmful for
children), restfulness (to find the programme making children
relaxed) and thought provocation (to evaluate the programme
to stimulate creative thinking of children). However the two
quality standards children considered most important were
comprehensibility and aesthetic quality.
Melissa et al. [3] developed a series of criteria to select
educational videos which constituted of four groups, namely
quality, content, engagement and delivery. In another study by
Morain & Swarts [17], a rubric comprising three levels was
chosen to evaluate the appropriateness of software tutorials
namely, physical design, cognitive design and affective
design.
These studies highlight the importance of certain
characteristics which have to be present in videos so that they
meet the objectives of the viewer.
C. ICT in Mauritius
The outburst of technology has been witnessed as prominent
in Mauritius, as well with an explosive growth of its
telecommunications sector since the late twentieth century.
As at 2014, 58.3% of inhabitants had access to the Internet
[18]. Mauritius has an ICT Development Index of 5.67 in a
scale of 0 (lowest ICT development) to 10 (highest ICT
development) [19]. In addition, 97% of secondary schools in
Mauritius have access to the Internet with 20 students per
computer. Moreover, 95.8% people in the age group 12 to 19
make use of the computer for learning [18].
Mauritian is thus well poised to benefit from the affordances
of ICT for pedagogical gain. This research study will therefore
investigate the level and extent at which freely available
learning resources on YouTube can be beneficial to our
learners.
III. SIGNIFICANCE OF THE STUDY
The Internet is flooded with millions of videos, the majority
being produced and uploaded by amateurs, such that learning
using these videos can be problematic. Haphazard selection
can be disappointing despite the fact that there is a lot of good
worth to be tapped. This study has a two-fold benefit.
1. Firstly, to empower educators make the right choice of
instructional videos for teaching and learning of
Mathematics. They can thus build a recommended list of
videos to be used by learners. The use of online learning
materials has often been cautioned by teachers with
regards to accuracy, completeness and relevance [19].
2. Secondly, this set of criteria can be used as guidelines by
teachers and video producers who wish to produce
instructional videos to support classroom lessons.
IV. AIM AND OBJECTIVES
This paper is aimed at assessing the appropriateness of
educational programmes on YouTube for use by educators.
The objectives are as follows:
1. To establish a list of criteria which can be used as
guidelines when selecting educational videos.
2. To analyse a set of selected educational videos for three
topics of Mathematics namely Fractions, Sets and Integers
based on these criteria.
3. To propose a framework to educators to enable them to
select the best instructional videos for classroom
integration.

V. METHODOLOGY
The methodological approach was both qualitative and
quantitative in nature. Based on literature review, some of the
guidelines characterising a good educational video have been
extracted.
Using these guidelines as a starting point, we conducted 2
focus group discussions as the first phase of our study as the
first phase of our study.
A focus group discussion can be defined as an in-depth, open- VI. DATA ANALYSIS, FINDINGS AND DISCUSSIONS
ended group discussion of 1 to 2 hours' duration during which
a specific set of issues on a predefined and limited topic is
explored. Such group usually consists of 5 to 8 participants
and are convened under the guidance of a facilitator [21].
Through such group discussions, people‟s knowledge and
experiences can be explored along with what people think,
how they think and why they think that way [22].
Our focus group consisted of a total of 8 educational media
producers and 3 professional technicians involved in the
production of educational audio-visual materials.
Some of the leading questions that fuelled our discussion
included:
1. What are the main stages in video production?
2. How is the raw script worked out?
3. What guides the translation of a raw script to audio
visual scripts?
4. Which elements are considered?
5. What is the input of TV Directors in the process?
6. Which are the components which add value to an
educational video?
We also discussed about the primary objectives of producing
educational videos and how they can be successfully
integrated in the classroom.
After in depth discussion, a list of criteria emerged. The
results are discussed in the next section. Based on these
results, we developed a marking rubric based on a 4-point
Likert scale, again discussed in depth in the next section.
Mathematics learning disabilities is rampant among Mauritian
learners and ICT is being viewed as a channel to bridge the
learning gap. YouTube resources are often accessed by both
teachers and students, especially to complement learning and
teaching. We therefore decided to analyse a number of
mathematical videos as per the established set of criteria. Ten
videos for each of the following 3 topics Fractions, Sets and
Integers were analyzed against the criteria. These topics were
based upon various MES reports [23] which mentioned the
three most difficult topics of the syllabus which are Fractions,
Sets and Integers.
Hence, videos for each of the three topics were analysed. A
search for each topic resulted in thousands of pages and only
the most viewed videos were selected for analysis. This
approach was adopted based on the fact that most individuals
rarely venture beyond the first page of their search results
[24].
Phase I
The results of the first phase of this study are provided in
Table 1 and they show the 16 criteria discussed and agreed by
the media specialists. Furthermore, the participants of the
focus group pointed out that not all criteria have the same
degree of importance such that it would be inappropriate to
discard a video based on some of these criteria only. For
instance, the group opined that the video cannot be discarded
if there was no objectives. On the other hand, criteria such as
language and accuracy were deemed to be mandatory.
Based on this argument, we agreed upon further categorizing
the 16 criteria into three groups, namely Mandatory, Essential
and Desirable. The explanatory note for each criterion was
also provided:
Mandatory: The two groups stressed on the importance of this
criteria. Unless these criteria are present to some extent, the
video should be rejected. For instance, if the content is not
according to the syllabus, the video can be rejected. The same
rigour applies to language which must be clear and
understandable by the students.
Essential: Over and above the mandatory criteria, the groups
drew a list of these essential criteria. One such criteria
includes the duration. Research shows that the majority of
people watch only the first 4 minutes of videos on YouTube
[25]. This means the longer the duration of a video, less
people will watch it till the end. Hence, there is a need to
select videos that are not too long.
Desirable: Finally the last six criteria were considered as
desirable and included the following: aesthetics, presenter, use
of prop/captions/jingles/music and addressing mixed ability
students amongst others.
Despite this categorization, the groups agreed that all three
categories have a certain degree of importance, making the
whole set of criteria a necessary reference guide for
identification of good instructional videos.
TABLE 1: Set of criteria to determine appropriateness of online educational B the higher order skills of the high scorer.
videos
L
Syllabus/Curriculum E
Illustrations involving real-life examples
Content of the video must be aligned with
Real life examples should be used to explain
school syllabus/curriculum.
M difficult concepts especially for slow learners
A Language to have a better understanding of the
N Language must be easily understood by the application of the concept.
D students. Use of props/captions/jingles/music
A Instructional design The use of props, captions and music are
T Content must be properly structured so that desirable features of a video that help to keep
O understanding and assimilation are viewers‟ interest while aiding them in better
R facilitated. understanding difficult terms and concepts.
Y Accuracy/Adequacy Objectives and recap
The content of the video must be accurate The objectives of the lesson should be well
with regards to all concepts and thesis to be defined right at the beginning of the video
mastered. Explanation should be exhaustive. and a recap is desirable at the end to allow
viewers to summarize main learning points
of the video.
Clarity of image/sound Encourages independent and self-learning
The quality of the images and the sound The video should promote independent and
should be of high standard. self-learning among students where
Engagement minimum assistance is required to
The video should catch and retain the understand the lesson.
attention of the viewers and keep them Based on this framework at Table 1, a marking rubric scale
motivated throughout. was developed to allocate a score for each criterion. Table 2
Presenter/Narrator proposes a sample for the criterion „objectives and recap‟.
Presenter should be well groomed and have TABLE 2: Objectives and Recap criterion
the right presentation skills to keep the
attention of the audience. 3 2 1 0
E
Duration Both Only Objectiv No
S
The duration of the video should not be too objectives objective es and objective
S
long but just enough to meet the objectives and recap s or recap not s and
E
of the conceptual content. Short duration is are recap well- recap
N
recommended. included included defined
T
I Supports/enhances classroom lessons
A Phase II
L Video should support lessons taught in class
and should add value to them. Thirty videos, 10 for each of the three topics, namely
Pace of delivery (Accent, tone, fractions, sets and integers, have been previewed and a
punctuation, pitch) number of observations were made. The results of the search
The content should be delivered at a displayed hundreds of thousands of suggestions. Like Google,
reasonable pace, with the appropriate that is YouTube uses ranking factors to determine which videos end
the accent, tone and pitch of the narrator so up at the top of each search results page (SERP). YouTube
that explanation are very clear. looks at a video‟s number of views, how long users watch it
Aesthetics and how many positive ratings and comments it has. For good
The visual appearance, judicious use of measure, they also throw in the number of channel
colors and format of the video should elicit subscribers, how many times the video appears in a user‟s
D and maintain viewing interest. playlist, how often it‟s added to a favourite list or playlist and
E how many times it‟s been embedded on a website.
Addressing mixed ability students
S
The video should be designed in a way that it Videos with the highest number of views from the first two
I
caters for mixed ability students making use pages were chosen and assessed as per the 16 criteria. The
R
of prior knowledge and illustrations to scores were recorded and analysed. The average duration of
A
introduce new concepts while also triggering the videos was 7 minutes and yearly views ranged from 5000
to 340,000. We noted that for Fractions, there were many
more videos and the yearly views were higher than for Sets
and Integers.
For analysis purposes, we considered an average yearly views
as videos on YouTube are posted at different points in time.
As an example, a 3 year old video might be there with 20,000
views while another one might be there for less than one year
and have around 10 000 views.
Suitability Video Index (SVI)
Each video was attributed a SVI value based on the sum of the
scores. For each of the 16 criteria [max= 48 (16*3), min=0].
The range of SVI values was 22 to 46, with most of the videos
scoring above 28, which is a good threshold for videos of
acceptable pedagogical worth.
that the higher the yearly likes, the higher the SVI value tends
to be. But again, the chart indicates cases of diversion from
this rule, such that it could be a mistake to select an
instructional video solely on account of the yearly likes,
though this might be a first step towards earmarking a set of
possible good instructional videos.
No. of likes per year against SVI
50
40
30

SVI
20
10
Finding 1: 0
0 500 1000 1500
In general, videos with high number of yearly views scored
well on the SVI scale and are thus appropriate for sound
instructional purposes. However, as shown in Figure 1, higher No. of likes per year
yearly views did not necessarily mean higher SVI, such that
Fig 2. Variation of SVI with yearly likes
there were videos with an average of 100,000 yearly views
and with SVI values ranging from 25 to 48. So, while high
yearly views can be indicative of the popularity of a video, yet
the ultimate choice cannot simply be made on this parameter. Finding 3:
This corroborates with research carried out by Duncan et al
[22] suggesting that the overall 100 most easily accessible or Figure 3 reveals that yearly dislikes is also far from being
watched YouTube videos may not necessarily be robust or indicative of the inappropriateness of a video for instructional
invested with high quality production values. purposes. Data shows that on one hand there are videos with
low yearly dislikes but with low SVI values, and videos with
Variation of SVI with number high yearly dislikes but with high SVI values too.
of views per year
50
40 No. of dislikes per year against
30 SVI
SVI

20 50
10 40
30
SVI

0
20
0 100000 200000 300000 400000
10
No. of views/year 0
0 50 100 150 200
Fig 1. Variation of SVI with yearly views
No. of dislikes per year
Finding 2:
Fig 3. SVI versus number of dislikes
The analysis regarding yearly likes revealed a weak but
significant association with the SVI value. Figure 2 shows
Finding 4:
Figure 4 shows the comparison of SVI with percentage of VII. CONCLUSIONS AND RECOMMENDATIONS
likes per view. The scatter diagram once again shows that If some of the parameters on the YouTube page like number
there is no relationship between these two factors, such that a of views and likes can be a first step towards choosing a video
high percentage of likes/view did not correlate with high SVI for viewing purposes, yet none of them is a strong, reliable
values. metric. Therefore, teachers should not rely on either of these
parameters to select appropriate content on YouTube for
classroom integration. Based on the findings of this study, we
SVI v/s Percentage of likes per recommend that the best approach to selecting YouTube
views videos is for teachers to view the whole content and to assess
50 it as per the 16 criteria mentioned in this research paper.
40 Furthermore, teachers need to build a database of videos for a
30 particular topic as no one video caters for a whole topic. In
SVI

20 some cases, teachers may even wish to select appropriate and

10 relevant excerpts of a video that add value to what has already
been taught in class. Moreover, the children may require an
0
introduction and some basic guidelines to the video. YouTube,
0.00% 0.20% 0.40% 0.60% 0.80% 1.00% by nature of the medium, only offers a didactic form of
Percentage of likes per views teaching where concepts and basic principles are transmitted
to the learner. The interactivity element of YouTube videos is
limited to play, pause, forward and backward. However, the
Fig 4. SVI versus percentage of likes per views concepts of Mathematics are more easily understood,
remembered and applied when learners tackle problems,
Finding 5: exercises and activities which are unfortunately not possible
on the video format. To complement these videos, it is
Analysis of each of the 16 criteria for the 30 videos revealed therefore imperative that appropriate interactive software are
that those with the 5 highest scores were among the mandatory identified or designed so that learners can practice acquired
and essential ones, namely language, presenter, image/sound, concepts.
syllabus and duration. Three mandatory criteria on the low
end and which should draw attention of video makers are In order to maximize on the integration of YouTube videos in
accuracy, instructional design and engagement. Mathematics learning, teachers are advised to prepare pre-
viewing and post-viewing activities. Such activities must be
Table 3 shows that videos posted on YouTube are lacking in prepared in advance after thoroughly viewing the video. The
their potential to address mixed ability students, have very pre-viewing activities will prepare students before viewing the
poor objectives and recap sections and use few real-life video, while post-viewing activities will help students
examples to illustrate mathematical concepts. consolidate the concepts they have understood and mastered
during the viewing session. This is in line with what
TABLE 3. Sample score per category Karppinen [6] advocates.
Score Criteria Category The number of views can be a very tricky way of selecting a
83 Duration Essential video as videos with high number of views are likely to be
80 Syllabus Mandatory clicked more often, hence, generating more views, without
80 Image/Sound Mandatory necessarily being viewed throughout. Thus, for educational
72 Presenter Essential programs, a more appropriate indicator on the YouTube Edu
69 Language Mandatory page would be the average percentage a video has been
67 Delivery Essential viewed. This would be more appropriate as it would show how
62 Support lessons Essential much time on average people spent watching the video and
61 Aesthetics Desirable not just whether the video was “clicked”.
60 Independent learning Desirable Videos posted on YouTube are produced mostly by amateurs
59 Engagement Mandatory
and very often lack the basics of good instructional video
56 Instructional design Mandatory
design. As confirmed in this study, the content can be good
52 Captions Desirable
but the packaging of it, including first and foremost the
51 Accuracy Mandatory
instructional design must be professionally looked into. The
48 Objective/Recap Desirable
YouTube Edu platform has much to offer to our young
43 Mixed ability Desirable
Examples Desirable students. Content is only a mouse click away. It is therefore
22
 imperative that quality videos that meet instructional gh_Policy_Development/resources/worldbank/countries/mauritius/co
a.htm [Accessed on 03 September 2015]
objectives be designed, posted and indexed to enable our
[3] N. L. Y. A. Melissa, S. F. Fong, and S. L. Ong, “An Evaluation of
emerging young independent learners make the right choice. Instructional Videos in EDUWEBTV: Technical Qualities,
Finally, the authors proposed a video selection framework Pedagogical Aspects, Engagement and Perceived Impact on
(Figure 5) that illustrates the steps which can help in the Learning,” Malaysian Journal of Educational Technology, vol. 10,
issue 2, pp. 101-113, 2010.
process of identifying and selecting appropriate video lessons
[4] T.C. Shorkey and M. Uebel, “History and Development of
to support teaching and learning. Instructional Technology and Media in Social Work Education,”
Journal of Social Work Education, vol. 50, pp. 26-247, 2014.
[5] The Top 500 Sites on the web, Alexa 2013. [Online]. Available from
http://www.alexa.com/topsites [Accessed 3 September 2015]
[6] P. Karppinen, “Meaningful learning with digital and online videos:
Theoretical perspectives,” AACE Journal, vol. 13, issue 3, pp. 233-
250, 2005.
[7] J. Anderson, “The Role of ICT in Teaching and Learning: Past,
Present and Future Perfect,” Proceeding of the UNESCO 13th Apeid
International Conference 2013. [Online]. Available from:
www.unescobkk.org/education/apeid/apeid-international-
conference/13-th-apeid-international-conference/speakers-and-
speeches/jonathan-anderson/the-role-of-ict-in-teaching-and-learning-
past-present-and-future-perfect/
[8] N. Alias, S. Hajar, A. Razak, G. Elhodad, N. Mat Noh, K. Kanjambu
and P. Muniandy, “A content analysis in the studies of YouTube in
selected journals,” Social and Behavioral Sciences, vol. 103, pp. 10 –
18, 2013.
[9] L. M. Gorder, “A study of teacher perceptions of instructional
technology integration in the classroom,” The Delta Pi Epsilon
Journal, vol. 2, pp. 63-76, 2008.

[10] C. Snelson, “Web-based video for e -Learning: Tapping into the

YouTube phenomenon,” in Collective Intelligence and E-Learning
2.0: Implications of Web-Based Communities and Networking, H.
Yang and S. Yuen, Eds. Hershey, PA: IGI Global, 2009, pp.147-166.
[11] C. Snelson, “Mapping YouTube video playlist lessons to the learning
domains,” in Planning for cognitive, affective, and psychomotor
learning, C. Crawford et al.,Eds. Chesapeake, VA: Association for
the Advancement of Computing in Education, 2010a, pp. 1193-1198.
[Online]. Available from: http://www.editlib.org/p/33518
[Proceedings of Society for Information Technology & Teacher
Education International Conference]
[12] C. Snelson, “Virtual movie sets and branching video: Developing
interactive educational machinima with Second Life and
YouTube,” Proceedings of the Technology, Colleges, and
Community Worldwide Online Conference, 2010b. [Online].
Available from: http://etec.hawaii.edu/proceedings/2010/Snelson.pdf
[13] M. Merkt, S. Weigand, A. Heier and S. Schwan, “Learning with
videos vs. learning with print: The role of interactive features,”
Learning and Instruction, vol. 21, pp. 687-704, 2011.
[14] A. A. Badr, F. Imran and S. Q. Khan, “Students‟ perspectives on the
relevance of internet-based educational videos in dental
Fig. 5. Video Selection framework education,” Journal of Taibah University Medical Sciences, vol. 10,
issue 3, pp. 288-292, 2015.
[15] C. Gaudin and S. Chalies, “Video viewing in teacher education and
professional development: a literature review,” Educational Research
Review, doi: 10.1016/j.edurev.2015.06.001, 2015.
[16] P. Nikken and T. H. A. Van Der Voort, “Children's Views on Quality
Standards for Children's Television Programs,” Journal of
Educational Media, vol. 23, issue 2/3, p. 169, 1997.
VIII. REFERENCES [17] M. Morain and J. Swarts, “YouTutorial: A framework for assessing
instructional online video,” Technical communication quarterly, vol.
[1] S. C. Burke and S. L. Snyder, “YouTube: an innovative learning 21, issue 1, pp. 6-24, 2012.
resource for college health education courses,” International [18] Economic and social indicators, ICT statistics. Central Statistics
Electronic Journal of Health Education, 11, pp.39–46, 2008. Office, 2014. [Online] Available from:
[2] YouTube Statistics 2014. [Online]. Available from: http://statsmauritius.gov.mu/English/Publications/Pages/ICT-Stats-
http://colfinder.net/materials/Supporting_Distance_Education_Throu Yr-2012.aspx. [2 Feb 2015]
[19] International Telecommunications Union (ITU) ICT facts and
figures: The world in 2014. [Online]. Available from:
http://www.itu.int/en/ITUD/Statistics/Documents/facts/ICTFactsFigu
res2013-e.pdf [Accessed 8 Aug 2015].
[20] P. Appavoo, K. M. Soyjaudah and V. Armoogum, “Readiness of
mathematics‟ educators to incorporate ICT as a teaching tool in
Mauritius,” in Proc. of the IEEE Africon 2013 conference., 2013,
doi: 10.1109/AFRCON.2013.6757764
[21] N. Robinson, “The use of focus group methodology with selected
examples from sexual health research,” Journal of Advanced
Nursing, vol. 29, issue 4, pp. 905-913, 1999.
[22] J. Kitzinger, “Introducing focus groups,” in Qualitative Research in
Health Care, N. Mays and C. Pope, Eds. London: BMJ Publishing
Group, 1996, pp. 36-45.
[23] Mauritius Examinations Syndicate), Annual Statistics Report, Réduit,
Mauritius, 2002-2011.
[24] I. Duncan, L. Yarwood-Ross and C. Haigh, “YouTube as a source of
clinical skills education. Nurse Education Today,” vol. 33, pp. 1576–
1580, 2013.
[25] BBB Active (2010) Available from:
http://www.bbcactive.com/BBCActiveIdeasandResources/Education
videosTenwaystousethemwell.aspx accessed 10.09.15
 Analysis of IEEE 802.11n Network Access
Categories in EDCA Non-Saturated Networks
Prosper Sotenga, Peter Chuku and Thomas Olwal
Department of Electrical Engineering
FSATIE/Tshwane University of Technology
Pretoria 001, South Africa
Abstract—In recent years demand for high data rate amongst
internet users has increased due to the use of real time data
services such as video and voice conferencing. The IEEE 802.11n
standard is an advancement from the previous a,b and g versions.
Its design goal is to increase data rate, and maintain compat-
ibility with the previous versions. The IEEE 802.11n standard
introduces improvements at both the PHY and MAC layers. This
standard still adopts the EDCA and HCCA mechanism used for
enhancing QoS in the network by assigning priorities to different
classes of data called Access Categories (ACs). In this paper we
compute a theoretical throughput model for 4 different stations
transmitting separate ACs in non saturated conditions. Networks
with non saturated conditions find real applications in WSN and
WiFi enterprises. We then perform a practical experimentation
with 4 different nodes and compare the throughput results with
the theoretical one. Based on the comparison, we determine
the fairness linked with the QoS mechanism. The comparative
throughput performance indicates that the analytical results
compare well with the practical experimental results considering
non saturated ACs.
I. I NTRODUCTION
Over the years, we have experienced increased usage of
WLAN for several applications. In fact, WLAN has become
very flexible and ubiquitous. For that matter WLAN is now
being used for a wide range of services such as, video stream-
ing, audio, web, printing, gaming etc. The first development
of Wireless LAN (WLAN) saw several improvements into
different versions which were all intended to satisfy different
needs. The IEEE 802.11a standard uses Orthogonal Frequency
Division Multiplexing (OFDM) to provide data rate of upto 54
Mbps in a 5 GHz band [1]. This standard yields more data rate
than the IEEE 802.1l and IEEE 802.1lb versions equivalent to
2 Mbps and 11 Mbps respectively. The IEEE 802.1la was very
expensive to integrate into chips due to the operating band
of 5 GHz. Its variant, the IEEE 802.1lb was more popular
even though it achieved relatively less data rate. The quest
for improved cost effective data rate at 2.4 GHz prompted the
development of a new version called IEEE 802.1lg, which
provided the same data rate at the same frequency as its
variant. The most recent of all these standards is the IEEE
802.1ln which boasts of data rate of up to 600 Mbps.
Improvement of data rates have therefore provided the
chance for heavy traffic to be accommodated in the WLAN
networks. This comes at a cost as the channel is capacity
limited which therefore causes network constraints and com-

ª
*&&&
promises Quality of Service (QoS) [2]. Efforts were made
to provide QoS for the increasing traffic load of WLANs.
This led to the amendments in the IEEE 802.1ln standard.
In the IEEE 802.1ln amendments, new MAC techniques have
been introduced to enhance QoS and throughput. The MAC
layer improvement techniques envisaged in the latest IEEE
802.1ln standard amendment includes [3]: Enhanced Dis-
tributed Channel Access (EDCA), Frame Aggregation, Block
Acknowledgement and Reduced Inter-frame Space. For the
purpose of our work we will be making use of EDCA.
EDCA, is basically an extension of the Distributed Coordi-
nation Function (DCF) which was part of the amendment of
the IEEE 802.1le to provide enhanced QoS of service based on
priority classes. Information to be transmitted by each station
are classified into priority classes called Access Categories
(ACs). Information such as Video and Audio are delay, jitter
and packet loss sensitive and hence requires more attention to
QoS. Other applications like internet surfing, sending emails
etc. do not really have real time effects. EDCA tends to
address this by giving higher priorities to real time data and
lower priorities to other forms of data. Even though one
may say QoS is being enhanced, users of lower priority may
still experience low QoS especially in instances where High
Priority ACs are enormous within the network. This paper
seeks to evaluate the fairness of the IEEE 802.11n network
in relation to categorising different data types into ACs. We
achieve this by computing the theoretical throughput of 4
stations. Each station transmits one ACs at several random
instances and analysis on the fairness can be made. We also
evaluate this in real time using Observer Expert as a test-bed.
Observer Expert provides real time VoIP and Video Expert
Analysis, Stream Reconstruction, Multi-Hop Analysis, and
Connection Dynamics [14]. This paper provides a comparative
analysis of IEEE 802.11n ACs, between analytical and real
life measurements using observer expert. The analytical results
show some conformance level with the real life performance.
The remainder of the paper is outlined as follows: Section
II will elaborate on the new MAC layer enhancements of
the IEEE 802.11n standard. Section III will highlight the
Analytical Model used for throughput evaluation. Section IV
will present the simulation and experimental set up. In section
V results obtained from the computations and experiments
 Fig. 1. Basic DCF Mechanism [9]
are presented and discussed. Finally conclusion is provided
in section VI.
II. IEEE 802.11 N MAC L AYER E NHANCEMENTS
A. Enhanced Distributed Channel Access (EDCA)
Since EDCA is the application of DCF we will high-
light on DCF before proceeding to EDCA. DCF is
a Carrier Sense Multiple Access with Collision Avoid-
ance(CSMA/CA)mechanism that is used in IEEE 802.11 net-
works. Each station contesting for channel utilisation, first
senses the channel for fixed duration of time called the DCF
Inter-Frame Space(DIFS). When the channel is sensed to be
idle, the station then utilises the channel. However when the
channel is sensed to be busy after the DIFS, the station defers
for a DIFS time and waits for a random countdown counter
to reach 0. After 0, if the channel is not occupied, the station
may then transmit. The random countdown is called the back-
off counter which is essential for collision avoidance. The
width of the random number generated for countdown is called
the Contention Window (CW). When the channel is not idle,
stations may be waiting to access the channel as soon as
the channel is idle, which may cause collision. Therefore the
random back-off time provides a stochastic medium access
hence reducing the probability of collisions just after the
channel has been released. The DCF mechanism is clearly
illustrated in figure.2. EDCA maintains this channel access
mechanism but in turn separates the data arriving at the MAC
layer into ACs [4]. ACs are differentiated according to priority
and assigned parameters. Table I shows how differentiation is
done for ACs. The parameters assigned to each AC statistically
prioritise the channel access for one AC over another. In
EDCA, a new parameter called the Arbitration inter-frame
space (AIFS) is defined. The AIFS varies from one AC to
another hence often denoted as AIFS[n], where n is the nth
AC. A higher priority AC has a shorter AIFS thereby giving
it a higher precedence over the other ACs [3]. Similarly, each
AC has its own pair of CWmin and CWmax [5].
TABLE I
ACCESS C ATEGORIES FROM H IGHER P RIORITY (ROW 2) TO L OWER
P RIORITY (ROW 5)
PRIORITY AC AIFSN CWmin CWmax DESIGNATION
Highest VO 2 7 15 Voice
VI 2 15 31 Video
BE 3 31 1023 Best Effort
Lowest BK 7 31 1023 Background
Fig. 2. Two level frame aggregation mechanism [8]
B. Frame Aggregation & Block Acknowledgement Protocol
(BA)
The earlier 802.11 MAC layer experienced inefficiency due
to the numerous transmission of acknowledgement packets for
every disaggregated data unit sent. This was however resolved
through the development of the 802.11e standard by using
the Block frame and the Block Acknowledgement transfers.
Further on, the former was then enhanced by the 802.11n using
frame aggregation. Thus two aggregation schemes are defined:
Aggregated MAC Service Data Unit (A-MSDU) and the Ag-
gregated MAC Protocol Data Unit (A-MPDU) [6]. In simple
terms, instead of sending a single data frame, the transmitting
party combines several frames together at the MAC layer for
transmission [7]. In A-MSDU, multiple MSDUs are combined
together to form an single MPDU. An A-MSDU comprises
of a number of sub-frames. Every sub-frame has a header
that consists of a Destination address,Source Address, Length,
MSDU, and padding bytes. The A-MSDUs also comprises of
a MAC header and frame check sequence (FCS). This means
an A-MSDU frame cannot be sent to multiple receiver [8].
The maximum length of an A-MSDU frame can be 3839 or
7955 bytes [8].
In an A-MPDU, multiple MPDUs are combined together.
These MPDUs may contain MSDUs and or A-MSDUs. When
an A-MPDU contains multiples A-MSDUs, this is called two
level frame aggregation as illustrated in figure 2. Each MPDU
is prefixed with a delimiter and postfixed with a padding.
As mentioned earlier on BA was envisaged in the design
goals of the 802.11n. BA works by acknowledging multiple
data frames with just a single acknowledgement frame. The
acknowledgement frame is a bitmap used to confirm the
successful or unsuccessful delivery of each transmitted frames.
A typical example of the Block Acknowledgement mechanism
is illustrated in figure 3.
III. A NALYTICAL M ODEL
When analytically modelling the MAC layer of the IEEE
802.11n standard, one can consider two scenarios namely
saturated and unsaturated network conditions.
A. Saturated Network Condition
In a saturated network condition, all stations within the
network always have packets in queue to be transmitted as
soon the the medium is available. In such a scenario the
probability for collisions to occur is high since there is constant
contest between stations to access the medium. This type of
network is prone to packet losses as there is a possibility
of buffer overflow at the MAC layer during the entire MAC
servicing time. According to [11], EDCA analytical models
for saturated cases provides performance measure, however
the downfall of saturated network models is that, it is not
applicable for real WLAN networks. The author in [11]
proposes an analytical model for EDCA saturated network
based on markov chain. This model is regarded as high in
complexity by [11].
B. Unsaturated Network Condition
In an unsaturated case only a portion of the stations contend
for the medium. This means some stations have empty buffers
while others have occupied buffers. With that regard packets
are transmitted with a low probability of collision. Authors in
[11] points out that the unsaturated condition lacks proposed
analytical models. In the work conducted by [4] a proposed
model based on non saturated condition is presented. The fact
that this model is based on non-saturated makes it mimic
realistic applications whereby not all ACs have packets in a
queue to be transmitted. Thus amongst all 4 ACs, Best Effort
(BE) and Background (BK) will not always have packets in a
queue. Hence our experimental set up is based on this model.
C. Throughput computations
This paper is mainly focused on throughput analysis. We
therefore adopt the model presented by [4]. The throughput is
denote by Si . The expression for throughput is defined in (1).
pi,s T data B
Si = (1)
(1 − pb )Te + ps Tsi + (pb − ps )Tci
pi,s is the probability that a packet from any of the ni back-
off instances contending for channel access of the ith AC is
transmitted successfully. This parameter is computed in (2).
τi represents the probability that the ith AC transmits during
a back off instance irrespective of a collision or not. This
parameter is computed in [2]. B however is the nominal bit
rate.
Fig. 3. Block Acknowledgement Mechanism [10]
pi,s = ni τi (1 − pi ) (2)
Moving further, pi is the probability of an unsuccessful
transmission. However, the probability of unsuccessful trans-
mission can be estimated from two different angles. The first is
by ignoring virtual collisions (VCs) that may occur during the
queuing of frames that arrive at the MAC layer. The second
is by considering virtual collisions that may occur during the
queuing of frames. Typically when a station is transmitting
more that one ACs, colliding station may have higher priority
ACs contending for channel access while the colliding lower
priority ACs may go into back-off. In this work we will
not consider VCs because our experimental set-up will only
consist of stations each transmitting separate dedicated ACs.
Without considering VCs, pi is given by (3) as in [2].
1 − pb
pi = 1 − (3)
1 − τi
pb is the probability that the channel is busy and is also
presented in (4).
N
 −1
pb = 1 − (1 − τi )ni (4)
i=0
In equation (4), N is the entire number of ACs. Now, referring
back to (1), ps is the probability of successfully transmitting a
packet from the ith AC in a given time slot. This is calculated
in (5).
N
 −1
ps = pi,s (5)
i=0
T data is the average duration required to successfully trans-
mit a packet. Te , Tsi and Tci respectively denote the real
time duration of an empty time slot, duration of successful
transmission of a packet and duration of collision of the AC
attempting to transmit. Tdata , Tsi and Tci are computed as in
(6), (7) and (8).
Tsi = TH + T data + ζ + SIF S + TBA + ζ + AIF Si (6)
Tci = TH + T ∗data + SIF S + TBA + ζ + AIF Si (7)
T data = αTA−M P DU (8)
TH is the transmission duration of the PHY and MAC
header whereas ζ is the propagation delay. SIF S is the period
between which a packet is successfully transmitted and an
acknowledgement frame is sent back. The Short Inter-Frame
(SIFS) is a predefined parameter. TBA is the time duration
for a Block Acknowledgement. In this work we will assume
that all packets have the same length. Therefore T ∗data which
represents the time duration for the longest packet payload TABLE II
involved in a collision is equal to the average duration of a C OMPUTATIONAL PARAMETERS
successful transmission of a packet T data . As indicated in Parameter Value
section II, AIF Si is the time duration whereby the channel B 600Mbps
has to be idle before the ith AC can transmit or start a random Time Slot 9μs
back-off countdown. The AIF Si parameter is given by (9). ζ 1μs
TP HY 20μs
LBA 320 bytes
AIF Si = SIF S + AIF SNi T imeslot (9) Length of MSDU header 14 bytes
LP AD 4 bytes
LM D 4 bytes
With reference to (8), TA−M P DU is the duration of a single SIFS 16μs
MPDU which is defined in (10). LA−M SDU which denotes N 4
the length of A-MSDU is scaled by a parameter θ which
tells the total number of A-MSDUs in a single MPDU. The
length of a single A-MSDU denoted as LA−M SDU in (11) is
is used we do not compromise the results. Rather we can say
a combination of the length of the destination address (DA),
our results is based on a worst case scenario. The model is
source address (SA) and the length prefixes as well as the
then computed for 100 random transmission instances and the
MSDU and Padding (PAD) postfix.
results are graphed.
Similarly, the experimental set-up is shown in figure 4. We
LM D + θLA−M SDU + LP AD made use of 4 different stations connected to the same 802.11n
TA−M P DU = (10)
B network via an Access Point(AP) in Tshwane University
Of Technology/French South African Institute of Technology
laboratories. One station solely used video streaming to depict
LA−M SDU = LDA +LSA +Llength +LM SDU +LP AD (11) the Video (VI) AC, the second station was used for voice
At this point, we have dealt with all the parameters required traffic over Skype to indicate Voice (VO) AC. The third station
for computing the throughput equation. Meanwhile, we recall was used for intensive web browsing to indicate Best Effort
that stations randomly choose an integer to start with the back- (BE) AC while the fourth station was just used for messaging
off countdown procedure. This random integer is within the in other to depict Background (BK) AC. The traffics of all
range of 0 > Wi,j . Wi,j is the contention window for each these stations were captured independently and consecutively
ith AC and j th retransmission attempt. Wi,j is calculated in using observer expert. After capturing these data we were able
(9). to acquire throughput analysis graphically and numerically in
other to compare with the computed model.

2j (CWi,min + 1) for j = 0, 1, . . . , mi − 1
Wi,j = (12)
CWi,min + 1 for j = mi , . . . , Li

IV. S IMULATION AND E XPERIMENTAL S ETUP

In this section we present the procedures followed in other
to compute the analytical model presented in the previous sec-
tion. The parameters chosen for simulation are also presented
in this section and we use MATLAB for computing the model.
We then move on to discuss the physical experimental set-up
for obtaining real life results. We use observer expert which
is a tool used for analysing networks and traffics.
The parameters indicated in Table I and II are used to
compute the model in MATLAB. To obtain the desired results
we take a different approach compared to what other authors
in literature have done. We assume we only have 4 stations
Fig. 4. Experimental Set-up.
transmitting on the networks. Each station transmits only one
type of AC and is different from the other stations. For each
station we compute the throughput model with its statistical TABLE III
parameters, thus the probability of a successful transmission, C OMPUTATIONAL PARAMETERS OF AC D IFFERENTIATION
collision etc. For simplicity we also assume there is only
Parameter VO VI BK BE
one retransmission opportunity during the back-off stage. θ 22 5 3 2
Moreover since we want to compare this to real life scenario, α 4 4 4 4
by assuming an ideal situation where only one retransmission LA−M SDU 160 660 1000 1600
 V. R ESULTS AND D ISCUSSION

In this section we present the results and discuss them. In

the experimental results, the throughput is given as packets per
transmission times in seconds. Thus we can still relate the out-
come of the physical experiment with the simulated theoretical
computation because bits and packets are proportional. These
results show some correlation with the theoretical model. For
each experimental results, the graph/signal in blue depicts the
captured packets.
The simulated results obtained for voice as in figure 5
clearly indicates that the station transmitting and receiving
voice data has higher throughput at most transmission times
over the network. In the experimental results of voice, it shows
that there is a cluster of packet transmission at the highest
throughput level. The station transmitted voice using SKYPE (a) Simulation (in b/s vs sec) (b) Experimental (in pkt/s vs sec)
audio call and results obtained is similar to the simulation. In Fig. 5. Throughput vs Transmission times for Voice AC.
figure 6, for simulated Video transmission we can see that the
result is fair. Thus there seem to be a balance between the
maximum throughput achieved and the minimum throughput
over all transmission times. With the experimental results,
when we started streaming a video, the signal suddenly shoots
up to the maximum attainable packets per second rate and
with moderate amount of drops in throughput just as in the
simulation results. In figure 7, the simulated results for the
Best effort traffic sees the throughput quite concentrated at
a minimal value. In the experimental results for Best effort,
the station exhibited many low drops in throughput/packets
in several instances of time but still clustered. Meanwhile, in
figure 8, the Background simulation results shows the worst
of them all. Its is highly concentrated at the minimal value
of the throughput axis. This correlates with the experimental
results.
From these observations we can say that there is a correla- (a) Simulation (in b/s vs sec) (b) Experimental (in pkt/s vs sec)
tion between the theoretical model computation and the real Fig. 6. Throughput vs Transmission times for Video AC.
life physical scenario as evident from the results obtained. In
a nut-shell, we can assume that Video and Voice transmissions
dominate within the 802.11n network. This creates unfairness
even though BE and BK ACs are not of higher importance in
terms of time.
The entire result really tell us in theory that, the EDCA
mechanism used for enhancing QoS does not bring about fair-
ness in the network. According to [13], fairness of 802.11n has
not received much research attention. We therefore compute
the fairness for each AC and provide results. To compute the
fairness we apply the common Jain’s index which is computed
as:
Np 2
( i=1 Si,c )
f (x1 , x2 , . . . , xn ) = Np 2 (13)
Np ( i=1 Si,c )
(a) Simulation (in b/s vs sec) (b) Experimental (in pkt/s vs sec)
The results of the fairness spans from N1p (worst case) and Fig. 7. Throughput vs Transmission times for Best Effort AC.
1(best case). Si,c is the throughput of the ith AC during the
cth transmission time. Np represents length of transmission
time. What we are clearly trying to achieve is the fairness of the throughput level across all times for all ACs. The results of

(a) Simulation (in b/s vs sec) (b) Experimental (in pkt/s vs sec)
Fig. 8. Throughput vs Transmission times for Background AC.
Fig. 9. Jains fairness Index for simulation results based on 4 stations with
different ACs.
computation of the fairness are indicated in figure 9. In this
figure it is evident that, AC1 (Highest) obtains the highest
fairness, whilst AC2 receives less fairness compared to AC1 .
The trend decrease until AC4 which has the lowest fairness
index. The results of the fairness index is due to the fact
that there is variation in throughput of all ACs. This impacts
negatively on the QoS in terms of throughput for ACs with
lower priorities in a non saturated networks.
VI. C ONCLUSION
In this paper we have elaborated on the main MAC mech-
anisms beginning with DCF to EDCA. We have presented
the enhancements associated with the EDCA mechanism. We
indicated the purpose of the enhancement which was mainly
for improving QoS as envisaged in the IEEE802.11e standard.
Moving further we used a theoretical model from literature
based on a stochastic approach to compute the throughput of
the network. We then provided results and compared them
with physical experimental results. In that regard we evaluated
the fairness caused by the ECDA mechanism. We conclude
based on the results obtained that EDCA does not provide
fairness amongst stations transmitting different categories of
data. This is evident in both theory and practice. For further
developments, we propose that, a reasonable level of fairness
may be attained by considering an adaptive approach. This
recommendation may improve the network by autonomously
sensing the network dynamics and readjusting the AC param-
eters to provide fairness for all ACs both analytically and
experimentally.
R EFERENCES
[1] C. Kai, Y Chen and N Yu, ”Performance Analysis of DCF under two
access mechanisms in IEEE802.1la WLAN”, 2nd International Conference
on Mobile Technology, Applications and Systems, pp. 1-7,2005.
[2] P. E. Engelstad and O. N. Osterbo, ”Delay and Throughput Analysis of
IEEE 802.11e EDCA with Starvation Prediction”, in LCN, Zurich,pp.647-
655, 2005.
[3] E.Perahia, R. Stacey, ”Next Generation Wireless LANs Throughput,
robustness, and reliability in 802.11n”, Cambridge University Press 2008,
ISBN-13 978-0-521-88584-3 , eBook (EBL).
[4] P.E. Engelstad and O.N. Osterbo, ”Non-Saturation and Saturation Anal-
ysis of IEEE 802.11e EDCA with Starvation Prediction”, in International
symposium on Modeling, analysis and simulation of wireless and mobile
systems, Montral, Quebec, pp. 224-233, 2005.
[5] D He, C.Q. Shen, ”Simulation Study of IEEE 802.1l e EDCF”, Vehicular
Technology Conference, pp. 685-689, 2003.
[6] Y. Daldoul, T. Ahmed and D. Meddour, ”IEEE 802.11n Aggregation
Performance Study for the Multicast”, Wireless Days (WD), pp. 1-6,
2011.
[7] S. Maaroufi, W. Ajib, H. Elbiaze, ”Performance Evaluation of New MAC
Mechanisms for IEEE 802.11n”, , 2007.
[8] J. Kolap, S. Krishnan, N. Shaha, ”Comparison of Frame Aggregation
Mechanism in 802.11n WLAN”, , 2011.
[9] T. Vanhatupa, ”Wi-Fi Capacity Analysis for 802.11ac and 802.11n:
Theory & Practice”, ,Ekahau Wi-Fi Design White Paper, 2013.
[10] R. Leutert, ”WLAN 802.11n MIMO Analysis”, ,SHARKFEST, Stanford
University, 2010.
[11] E. Charfi, L. Chaari, L. Kamoun, ”Fairness of the IEEE 802.11n
aggregation scheme for real time application in unsaturated condition”,
,IFIP WMNC, 2011.
[12] K. Xu, Q. Wang, and H. Hassanein, ”802.11e enhanced distributed
coordination function (EDCF) in WLAN”, ,in Globecom, San Francisco,
pp. 1048- 1053, 2003.
[13] L. Kriara, Mahesh, K. Marina and A. Farshad, ”Characterisation
of 802.11n Wireless LAN Performance via Testbed Measurements and
Statistical Analysis”, ,in SECON, pp. 158- 166, 2013.
[14] Network Instruments,”Observer Expert User Guide”, 2013.
 Security in the Internet of Things through
Obfuscation and Diversification
Shohreh Hosseinzadeh, Sampsa Rauti, Sami Hyrynsalmi and Ville Leppänen
Department of Information Technology
University of Turku, Finland
Emails: {shohos, sjprau, sthyry, ville.leppanen} @utu.fi
Abstract—Internet of Things (IoT) is composed of hetero-
geneous embedded and wearable sensors and devices that
collect and share information over the Internet. This may
contain private information of the users. Thus, securing the
information and preserving the privacy of the users are of
paramount importance.
In this paper we look into the possibility of a p p l y i n g the
two techniques, obfuscation and diversification, in IoT.
Diversification and obfuscation techniques are two outstand-
ing security techniques used for proactively protecting the
software and code. We propose obfuscating and diversifying
the operating systems and APIs on the IoT devices, and also
some communication protocols enabling the external use of
IoT devices. We believe that the proposed ideas mitigate the
risk of unknown zero-day attacks, large-scale attacks, and also
the targeted attacks.
Keywords—Internet of Things, IoT, security, privacy, obfus-
cation, diversification
II.
I. INTRODUCTION
Operating systems and software in IoT
Internet of Things (IoT) or Internet of Everything, was
introduced by MIT Auto-ID Labs in 1999 [1], is a network
of physical devices (objects) connecting to each other (wire-
less) for sending/receiving data. The tiny chips embedded
in the devices enable them to communicate without the
human interaction. This aims to make the human lives more
intelligent, automated and thus more comfortable. IoT is
known as the third revolution in information technology
after the Internet and mobile communication networks, and
today it is being used in multitude public and private
sectors, ranging from the public safety to health care. By
the continuous growing trend, more and more”things” are
getting connected to each other every day, collecting and
transmitting personal and business information back and
forth. Cisco IBSG [2] reports that the number of connected
devices in 2015 is 25 billion, and this number is expected
to grow to 50 billion by 2020. However, the security in IoT
is still a major challenge. According to [3], 70 percent of
the IoT devices are vulnerable to exploits that c o u l d be a
doorway for attackers to the network. On this basis,
researchers and developers are continually seeking effective
techniques that boost the security in this environment, which
is compatible with the limitations of the participating nodes
in IoT. To the best of our knowledge, none of the existing
research works in the field of IoT security have l o o k e d
into the obfuscation and diversification techniques as the
potential techniques for mitigating the risk of malware.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
In this paper, we propose a novel idea that addresses
possible security threats in IoT. We base our idea on two
promising techniques, obfuscation and diversification, that
have been proved to be successful in impeding the malware
in various domains [4]. In this work-in-progress paper, we
propose using these two techniques to protect the operating
systems and APIs of the devices participating in IoT, and
also to introduce an additional level of security at the
network level, by diversifying some protocols used in the
communication. This study is a research proposal, in which
we present our novel ideas. In the future works, we consider
implementing these ideas, and demonstrate the effectiveness
of the proposed approaches.
The remainder of the paper is structured as follows: in
Section 2 we present a background on the characteristics of
IoT, the software and the protocols used in these domains.
Section 3 discusses our proposed idea in detail, and Section
4 concludes the paper.
CHARACTERISTICS OF THE I O T
IoT comprises a wide variety of heterogeneous sensors
and devices, of which some are powered by more potent 32-
bit processors (e.g., smart phones) and some are controlled
by lightweight 8-bit micro-controllers [5]. Therefore, the
chosen software should be applicable to a range of devices,
including the lowest power ones. On one hand, it should be
capable of supporting the functionality of the object; and
on the other hand, should be in line with the limitations of
these devices in memory, computational power, and energy
capacity. The software in IoT should have the following
characteristics [6]:
• Heterogeneous hardware constraints: the IoT soft-
ware should have limited CPU and memory re-
quirements, so that it could support the constrained
hardware platforms.
• Autonomy: It should be energy efficient, reliable,
and adaptive to the network s t a c k .
• Programmability: It should provide a standard Ap-
plication Program Interface (API) for software de-
velopment and support the standard programming
languages.
These factors have leaded the developers to think of the
operating systems that are adaptive to the diverse low-power
objects in IoT. Among all, Contiki [7] and TinyOS [8] are

Figure 1. IoT network stack, and some examples of the corresponding
protocols used in each layer [6].
the most dominant operating systems used on IoT devices.
Contiki [7] is an open-source Operating System (OS) devel-
oped in C with a modular structure. It connects low-power
micro-controllers to the Internet, and supports a wide range
of networking standards, including IPV4, IPV6, CoAP. This
lightweight OS is considered reasonably memory efficient
by using only few kilobytes of memory1. Contiki is de-
signed based on the event-driven multi-threading kernel.
Moreover, Contiki has a dynamic nature, i.e., it allows the
dynamic loading and unloading of applications at run-time.
TinyOS [8] is an open source, application specific OS de-
signed and widely used by low-power devices in WSN, and
ubiquitous environments. It is implemented using the
component-based programming model and supports the
event-driven concurrency. TinyOS is multi-threading and has
monolithic structure.
Access protocols in I o T
IoT encompasses a huge number of devices and appli-
cations connected to each other. Basically, there are three
different types of connections in IoT [9]: a) Device to Device
(D2D) is the connection between the devices, b) Device to
Server (D2S) is the connection for sending collected data
from the devices to the servers, and c) Server to Server
(S2S) is the connection between the servers to share the
collected data. For making the connections feasible there
exist protocols that are designed based on the requirements
and characteristics of the participants in IoT. As an exam-
ple, Constrained Application Protocol (CoAP) [10] is an
application layer protocol which is designed to be used in
resource-constrained devices and WSN nodes. This protocol
is simply translated to HTTP which simplifies the integration
with the web. Figure 1 illustrates the current network stack
used in IoT, and some of the communication protocols that
are used in each la yer .
1. Contiki: The Open Source OS for the Internet of Things:
http://www.contiki-os.org
Security of IoT
IoT is still not a mature technology and accordingly, the
security measures considered for it are still in early stages.
Due to the fact that IoT is based on the Internet, it is
subject to traditional security challenges threatening the
Internet; and in addition to those traditional challenges, the
key characteristics of IoT add new security threats on top
of them.
These characteristics are [11], [12]: Diversity: The par-
ticipating devices in IoT range from low power/cost to high
performance devices. Hence, the security measure designed
needs to be compatible with wide range of devices. Scale:
The number of sensors and devices in IoT is tremendously
growing, which makes it harder to control the devices and
the collected information. Wireless connection: devices are
connecting to the Internet wireless through different links
(Bluetooth, 802.11, ZigBee). The links should be protected
so that the communicated information does not leak. Em-
bedded use: most of the IoT devices are designed for single
purpose with different communication patterns. Mobility:
The devices in IoT are mobile and connected to the Internet
through various providers.
These properties make IoT more prone to the security
threats compared to the Internet and traditional sensor net-
works. Security threats known in IoT could be due to insuf-
ficient security software or firmware, insufficient security in
the Web interface, insecurity in cloud interface, insecurity
in mobile interface, insecurity at authentication or autho-
rization point, lack of proper security measures in network
services, lack of encryption, poor security configuration,
poor physical security, and privacy issues [12]. To overcome
these security challenges, different tactics are suggested to
be taken into account for implementation, including secure
firewall and intrusion prevention system, secure identity
management and access control program, multi-factor au-
thentication, emergency response program, and IoT security
standardizations [13]. However, we believe that they d o not
sufficiently secure the IoT and more concrete security
measures are required. This motivated us to propose a novel
idea to boost the security in IoT.
III. ENHANCING THE SECURITY OF IOT USING
OBFUS- CATION AND DIVERSIFICATION
TECHNIQUES
Background on obfuscation and diversification
By using the obfuscation and diversification techniques
we do not aim at removing the security vulnerabilities, but
making it challenging and costly for the intruder to take
advantage of them.
Obfuscation refers to making the code more compli-
cated and harder to understand. As a result, the attacker
needs to spend more time and energy to comprehend the
code [14]. Figure 2 illustrates how obfuscation affects the
code: a) is a piece of JavaScript code, and b) is obfuscated
version of the same code. As the figure shows, the code is
scrambled/obfuscated in a way that it is harder to understand
the purpose of the code.
a)
function setText(data) {
document.getElementById("myDiv").innerHTML
= data;
} securing the systems in the current era of OS
b)
function ghds3x(n) {
h = "\x69\u006En\u0065r\x48T\u004DL";
a="s c v o v d h e , n i";x=a.split("
");b="gztxleWentBsyf";
r=b.replace("z",x[7]).replace("x","E").replace("s
","").replace("f","I")
["repl" + "ace"]("W","m")+"d";
c="my"+String.fromCharCode(68)+x[10]+"v";
s=x[5]+x[3]+x[1]+"um"+x[7]+x[9]+"t";d=this[s][r](
c);if(+!![])
{ d[h]=n; } else { d[h]=c; }
} system or perform what an attacker desires [18], in order
Figure 2. a) Original version of a piece of JavaScript code, and b)
obfuscated version of the same code
Diversification makes the program instances unique, so
that the attacker is no longer able to exploit a vast number
of devices via a single attack model. The attacker needs
to design various versions of attack models intended for
various program instances. This makes the attack costly and
difficult. Figure 3 illustrates the program diversification, in
which different replicas of program P are generated and
distributed to the users. All these programs function the
same, but have different internal structures. Therefore, even
if the attacker succeeds in attacking one version of the
program (P3), the other versions are safe.
Figure 3. Diversification generates unique instances of the software. A
single attack model works only for one instance.
The idea in diversification is to change all kinds of
internal interfaces appearing inside IoT systems, or between
IoT device and its backend (cloud) system. As diversification
can be done with respect to known reference interfaces,
software developers of IoT systems need not to be aware
of diversification details. In practice, the purpose of diversi-
fying an interface is to create a machine/system-wise unique
secret that is shared only between the legal ”clients” of
the interface. However, a diversified implementation is not
the same as an encrypted implementation – a s the idea
is that a diversified implementation is still executable as
such. Moreover, the idea behind the diversification
techniques is not to remove the vulnerabilities and the
security holes in software, but to prevent the attackers to
take advantage of them, since even if an attacker would
be able to use a security hole to inject malware (code)
into the system, the malware would not work as it would
need to access resources using the diversified (secret)
interfaces instead of the prevailing setting where
accessing only requires knowledge on reference interfaces.
Therefore, diversification is considered very useful for
monoculture and in the settings where it is hard to patch
security problems of OS or system later on.
There exist different diversification techniques that use
Various transformation mechanisms at different levels of
software life cycle [16]. The idea of program diversification
for effectively protecting the operating systems has been
pioneered by Cohen [17]. In our previous works we have
applied the diversification and obfuscation to different OSs,
e.g., Linux to secure it against malware. As malware is
meant to run on the user’s computer to manipulate the
to thwart some malware to access resources via the system
calls, we have to introduce a way to block the system calls
made directly and indirectly from an untrusted application.
For this purpose, we can diversify the system call numbers,
and the library functions that make system calls [19], [20].
In [21] we diversified also the indirect entry points to the
system calls and propagated the new information on entry
points to all legal applications, so that the malware will not
have the information to access the resources. On that
account, we designed a concrete diversification tool which
diversifies the symbols in the Linux ELF binaries and makes
them unique.
The proposed a p p r o a c h
As mentioned, security threats in IoT could potentially
be on the application layer (the software on the IoT devices)
or on the network layer. On this basis, we propose two novel
techniques to provide security in these two layers. We
propose: 1) obfuscating/diversifying the OSs and APIs used
in the IoT devices, and 2) obfuscating/diversifying some of
the access protocols among them.
The sensors and devices in the IoT contain embedded
chips which function with the help of an OS and APIs. This
OS and the APIs are prone to malware. Our first proposed
idea aims at protecting the IoT devices by applying diversi-
fication and obfuscation techniques to the OSs and APIs and
securing them against the malware. As discussed earlier, in
our previous works [19], [20], [21], we demonstrated that
by obfuscating the OS and diversifying the APIs we can
successfully make it harder for the malware to interact with
the interfaces and access the resources.
We believe the same approaches can be applied on the
OSs and APIs of the IoT devices to protect them from
the malware. By making the OS and the APIs unique
through diversification, we manage to thwart the massive-
scale attacks, i.e., the attacker by designing a single attack
model cannot take a large number of devices under control.
Additionally, we mitigate the risk of zero-day attacks.
Secondly, we propose applying diversification at the access
protocol level. In a communication network, an application
level protocol specifies the interfaces and the shared
protocols used by the communication parties. Protocol
identification is the act of identifying what protocol is used
in the communication session [22]. Protocol identification
can be done via static analysis methods and comparing the
protocol used in the communication with the common existing
protocols. The information gained from this analysis could be
used by an intruder which endangers the integrity and
confidentiality of the communication. In order to protect the
protocol from identification, it could be obfuscated so that it
is more difficult to be recognized by the traffic classification
machines. Protocol obfuscation attempts to remove the
properties that make the protocol identifiable, e.g., packet size
and byte sequence (make them look random). The most
common way for protocol obfuscation is using cryptography.
Depending on the need of the network, different levels of
encryption could be applied [22].
We propose obfuscating the communication protocol
among a small set of nodes (e.g., within a home) in a
way that the obfuscation method is kept secret among them
and only the nodes which know the secret are able to
communicate with each others. By changing/complicating
the form of the protocol and making it different from the
default format, we aim at generating a huge number of
unique diversified protocols from a reference protocol.
Applying diversification to the application is feasible, al-
though it has two main challenges: 1) the recognition of
transition (message sending/receiving) from programs, 2)
complication and slowdown of the program. Besides the
protocol obfuscation, we propose protocol diversification,
which considers the protocol as an operation of two state
machines, so that (synchronized) state changes are messages
sent between parties. The original implicit state machine of a
protocol can be diversified by adding/splitting new states and
transitions.
Motivation
For securing IoT and the participants of these envi-
ronments, we propose using diversification and obfuscation
mechanisms. We believe that these two techniques are effec-
tive for this purpose, by mitigating the risk of unknown zero-
day attacks and also preventing the large-scale and targeted
attacks. The motivation behind our idea is as follows:
• Additional Security: There have been various secu-
rity mechanisms designed and applied on IoT at the
network level. We believe that presenting security at
the device level is an orthogonal proactive measure for
security. In this way, the malware is stopped at one
node and not propagated to the whole network.
• Energy-efficiency: The participating nodes (devices) in
IoT are extremely constraint in resources, i.e., they are
limited in terms of memory, energy capacity, and
computational power. Therefore, the security
mechanism designed for them should be rather
lightweight. For instance, the anti-virus programs often
cannot be used in IoT devices (because of their huge
effect on the performance and energy consumption).
We believe that API diversification will not slow down
the execution at all: however, obfuscation and protocol
diversification affect the efficiency to some extent.
• No complexity for the manufacturer: The devices
in IoT are controlled by the tiny chips embedded in
them which may not tolerate a complex design. Con-
sidering this fact, our security mechanism does not add
any complexity overhead for the manufacturer.
• Alleviate the risk of malware: The devices in IoT
contain tiny chips with a lightweight OS on them. The
OS handles the operation of the device by ex-
ecuting the code. Code execution presents a surface
for attacks for unwanted software or malware. This
malicious behavior could range from unauthorized
accessing/reading of the data to altering it towards
the attacker’s intent. Thus, we believe the OS can
be protected using obfuscation and diversification
mechanisms, in order to make the malware ineffec-
tive. In order to make our approach applicable with
the limited capabilities of the devices (in computa-
tional power) we limit the level of obfuscation and
stick to the less complicated diversification methods,
e.g., identifier renaming.
• Alleviate the risk of massive-scale attacks: Currently,
there are billions of devices manufactured and dis-
tributed in a”monoculture” manner, i.e., they are
designed and produced identical which makes them
have the same layout and consequently, the same
security vulnerabilities. By designing one single at-
tack model, an attacker can simply invade a wide
range of devices. We believe that diversification, by
introducing”multiculturalism”, in software produc-
tion is a proper proactive security mechanism for the
widely distributed environments including IoT. The
idea of multiculturalism in the software production
alleviates the risk of massive-scale attacks.
• Amend the update limitation in embedded devices:
One of the drawbacks of the embedded devices is
that they usually cannot be updated, i.e., in case a
security hole is found they cannot update themselves
to receive the security patch. Hence, we need to
think of a solution that does not try to remove the
security holes and vulnerabilities of the software, but
to make it difficult for the attacker to exploit them.
Diversification and obfuscation techniques are
helpful to be prepared for the unknown attacks;
because the general idea behind these techniques is
not to remove the security holes but to avoid the
attacker from taking advantage of them.
Limitations
Despite of the great benefits our proposed approach
brings along, there might be some limitations for it,
including cost and adaptability of the network.
Obfuscation on one hand protects the software and
applications from the malicious reverse engineering by
making them hard to read; and on the other hand, it
introduces costs in terms of execution overhead, memory
consumption, and code size increase. Additionally, when
obfuscation is applied on the protocol, it should be assured
that the both parties in the communication session support
the obfuscated protocol.
V. CONCLUSION AND FUTURE WORKS
In this paper we pointed out that due to the special char-
acteristics of IoT, the security is more challenging compared
to the traditional networks. Therefore, it requires security
measures that are in line with the capacity of IoT devices.
Diversification and obfuscation techniques are the two
promising security techniques that have been presented pre-
viously, and practitioners can already start to adapt these
techniques into use to protect their IoT networks. In this
 work-in-progress we proposed two approaches using the two
techniques, obfuscation and diversification as well as request
further work in this topic.
First, we propose applying the techniques on the operat-
ing systems and APIs of the IoT devices to make it harder
to breach through the devices. This is a traditional way of
utilizing the techniques and it is expected that this has a
significant impact on improving the security of the devices.
Second, we propose applying these techniques on related
application layer protocols. The idea is to increase the
workload of a malicious attacker by using these techniques
together with other techniques such as cryptography to secure
the communication between the devices and servers.
Third, to the best of our knowledge, there are no
thorough study of attack vectors of IoT networks. Further
inquiries should focus on defining different attack vectors in
order to prove security analyses tools to start working with
countermeasures.
Finally, to extend this work-in-progress, our future work
is towards diversifying an IoT operating system of a device, and
obfuscating and diversifying an application level (internal)
access protocol used between the IoT devices and its cloud
based back-end system. For the device/application we have
planned to use a health care sector IoT device based on
Contiki as it is the most common operating system. The target
of protocol obfuscation and diversification is the CoAP used in
the same device.
ACKNOWLEDGEMENT
The authors gratefully acknowledge Tekes, the Finnish
Funding Agency for Innovation, DIGILE Oy and Cyber
Trust research program for their support.
posed security model and threat taxonomy for the Internet of Things
(IoT),” in Recent Trends in Network Security and Applications, ser.
Communications in Computer and Information Science. Springer
Berlin Heidelberg, 2010, vol. 89, pp. 420–429.
[12] R. Roman, J. Zhou, and J. Lopez, “On the features and
challenges of security and privacy in distributed Internet of Things,”
Computer Networks, vol. 57, no. 10, pp. 2266 – 2279, 2013.
[13] S. Sicari, A. Rizzardi, L. Grieco, and A. Coen-Porisini, “Security,
privacy and trust in Internet of Things: The road ahead,” Computer
Networks, vol. 76, no. 0, pp. 146 – 164, 2015.
[14] C. Collberg, C. Thomborson, and D. Low, “A taxonomy of ob-
fuscating transformations,” Department of Computer Science, The
University of Auckland, New Zealand, Tech. Rep., 1997.
[15] “ The International Obfuscated C Code Contest:
http://www.ioccc.org/years.html,” verified 2015-07-16.
[16] P. Larsen, S. Brunthaler, and M. Franz, “Automatic software diver-
sity,” Security Privacy, IEEE, vol. 13, no. 2, pp. 30–37, Mar 2015.
[17] F. B. Cohen, “Operating System Protection through Program Evolu-
tion,” Comput. Secur., vol. 12, no. 6, pp. 565–584, Oct. 1993.
[18] E . Skoudis, Ma lwa re: Fighting m a l i c i o u s c od e. Prentice
Ha ll Professional, 2004.
[19] S. Rauti, J. Holvitie, and V. Leppänen, “Towards a
diversification framework for operating system protection,”
ser. CompSysTech ’14. New York, NY, USA: ACM, 2014, pp.
286–293.
[20] S. Rauti, S. Laur´en, S. Hosseinzadeh, J.-M. M¨akel¨a, S. Hyrynsalmi,
and V. Leppänen, “Diversification of system calls in linux binaries,”
in InTrust2014. LNCS, to appear September 2015, p. 15 pages.
[21] S. Laur´en, P. M¨aki, S. Rauti, S. Hosseinzadeh, S. Hyrynsalmi, and
V. Leppänen, “Symbol diversification of linux binaries,” in
WorldCIS-2014. Infonomics Society, 2014, pp. 74–79.
[22] E. Hjelmvik and W. John, “Breaking and improving protocol
obfuscation,” Chalmers University of Technology, Tech. Rep, vol.
123751, 2010.

REFERENCES
[1] I. Bose and R. Pal, “Auto-id: Managing anything, anywhere, anytime in
the supply chain,” Commun. ACM, 48(8), pp. 100–106, Aug. 2005.
[2] “The internet of things- how the next evolution of the internet is
changing everything,” verified 2015-07-08. [Online]. Available:
https://www.cisco.com/web/about/ac79/docs/innov/IoTIBSG 0411FINAL.pdf
[3] “Internet of things research study - hp report,” verified 2015-07-08.
[Online]. Available: http://www8.hp.com/h20195/V2/GetPDF.aspx/4AA5-
4759ENW.pdf
[4] P. Larsen, A. Homescu, S. Brunthaler, and M. Franz, “SOK: Auto- mated
software diversity,” in Security and Privacy (SP), May 2014, pp. 276–
291.
[5] K. T. Nguyen, M. Laurent, and N. Oualha, “Survey on secure com-
munication protocols for the Internet of Things,” Ad Hoc Networks,
vol. 32, pp. 17 – 31, 2015.
[6] E. Baccelli, O. Hahm, M. Gü nes, M. Wählisch, and T. C. Schmidt,
“Operating systems for the IoT–goals, challenges, and solutions,” in
WISG2013, January 2013.
[7] A. Dunkels, B. Gronvall, and T. Voigt, “Contiki - a lightweight and
flexible operating system for tiny networked sensors,” in Local Com-
puter Networks, 2004. 29th Annual IEEE International Conference
on, Nov 2004, pp. 455–462.
[8] P. Levis S. Madden, J. Polastre, R. Szewczyk, K. Whitehouse, A.
Woo, D. Gay, J. Hill, M. Welsh, E. Brewer, and D. Culler, “Tinyos: An
operating system for sensor networks,” in Ambient Intelligence.
Springer Berlin Heidelberg, 2005, pp. 115–148.
[9] M. Corson, R. Laroia, J. Li, V. Park, T. Richardson, and G. Tsirt-
sis, “Toward proximity-aware internetworking,” Wireless Communi-
cations, IEEE, vol. 17, no. 6, pp. 26–33, December 2010.
[10] Z. Shelby, K. Hartke, and C. Bormann, “The constrained application
protocol (CoAP),” in Internet Engineering Task Force (IETF), 2014.
[11] S. Babar, P. Mahalle, A. Stango, N. Prasad, and R. Prasad, “Pro-
Proactive Digital Forensics in the Cloud using Virtual
Machines
D.J. Ras
Department of Computer Science
University of Pretoria
Pretoria, South Africa
dras@cs.up.ac.za
Abstract—With the advent of cloud computing systems it has
become possible to provision large scale systems in a short time
with little effort. The systems underpinning these cloud systems
have to deal with massive amounts of data in order to function.
Should an indecent occur that requires some form of forensic
investigation it can be very challenging for an investigator to
conduct the investigation. This is due, in large part, to the
volatility of data in cloud systems.
In this paper, a model architecture is proposed to enable
proactive forensics of cloud computing systems. Using a reference
architecture for cloud systems, an add-on system is created to
enable the capture and storage of forensic data. The captured data
is then available to the investigator should the need for an
investigation arise. This must be achieved with minimal alteration
or interruption of existing cloud systems.
The system is described and a theoretical architectural model
is given. An evaluation discusses the possible advantages and
disadvantages of such a system and how it can be implemented as
a proof of concept. It also relates the proposed model to the ISO
27043 standard of forensic investigations.
Keywords—Cloud computing; digital forensics; cloud forensics;
standards; ISO standard 27043
I. INTRODUCTION
In the period from the early 80s to the present day,
computer storage capabilities have massively increased,
beginning low capacity floppy disks to massive clustered
cloud computing system. As these technologies have evolved
and improved, so too has the need to perform forensics on
devices and systems. Where it is relatively simple to
forensically analyze a single disk, performing a forensic
investigation on a cloud computing system is much more
complex, due to the inherent complexity of such a system.
Problems like data location, segregation and recovery, to name
but a few, plague investigators making cloud forensics very
challenging [1]. Another, more general problem, is that up
until recently there has not been a single standard on which a
forensic investigation process could be based. This problem
should hopefully be alleviated with the introduction of the ISO
27043 standard [2]. In combination with a standard
investigative process, the application of forensic readiness
principles and procedures could further aid in making cloud
forensics less challenging [3]–[5].
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
H.S. Venter
Department of Computer Science
University of Pretoria
Pretoria, South Africa
hventer@cs.up.ac.za
In this paper, a model architecture is proposed that
implements forensic readiness as described by the ISO
standard 27043 [2] by means of proactive forensics of cloud
computing systems. The proactive forensic model is used as a
method to address the problems of data volume, data isolation,
data integrity and availability of the cloud system. The
proposed architectural model captures potential forensic
evidence with the aid of a forensic monitor and transmits the
potential evidence via a secure channel, to a cloud forensic
system, where the potential evidence can be analyzed.
The goal of the architecture is to provide a mechanism that
would allow an investigator to rapidly and effectively acquire
data needed for an investigation. It is not the goal of the
system to fully replicate the virtual machines in the cloud
system. This would in effect massively increase the size of the
cloud and exacerbate the problem of large data volumes. The
exact details of what potential digital evidence is captured and
how the transmission channel is secured is not in the scope of
this paper, only the system that could facilitate the collection.
The paper is structured as follows: Section II contain
background information regarding the current state of the
cloud, cloud forensics and the ISO 27043 standard. It also
contains the related work on which this paper is based. Section
II.D contains models for the possible implementation of the
ISO 27043 standard (section II.C). Section III contains the
proposed model for the realization of the ISO 27043 standard.
The model is evaluated in section IV finally, section V
concludes the paper. All instances where the phrase the
standard is used, the phrase refers to the ISO 27043 standard.
II. BACKGROUND
In order to realize an architecture to enable forensic readiness
in cloud computing systems, background information is
needed with regards to the general concepts of cloud
computing, the current state of cloud forensics and work
related to forensic readiness of cloud systems. The ISO 27043
standard outlines the forensic process that must be followed
during an investigation by separating the phases into different
process classes [2]. In this paper the readiness class is used as
a measure as to whether the proposed cloud architecture is
useful. The following sections elaborate on the
aforementioned topics.
A. Cloud Computing
According to the National Institute of Standards and
Technology cloud computing can be defined as a model for
enabling convenient, on-demand network access to a shared
pool of configurable computing resources (e.g. networks,
servers, storage, applications, and services) that can be rapidly
provisioned and released with minimal management effort or
service provider interaction [6].
Cloud computing is the culmination of different computing
technologies converging to form a new technology that
removes the traditional limitations of strict policy driven
information technology infrastructure [7]. These core
technologies are virtualization and clustering of infrastructure
[8].
1) Virtualization
Virtualization is one of the core technologies used in cloud
computing infrastructure [7], [9], [10]. This enables users to
have access to scalable, on demand computing systems [11].
Virtualization works by abstracting computing resources from
their physical counterparts [8] into a resource pool from which
the resources can be drawn by users. A virtual machine can
thus be provisioned by a user as long as the resources required
for the virtual machine do not exceed that of the physical host
machine. This allows multiple virtual machines to run on a
single host [12], [13].
2) Clustering
The clustering of computers occurs when two or more
computing systems work together to perform some function
[14]. The aim of this is to have a scalable solution that gives
flexibility in terms of computing power, redundancy or
availability. As cloud computing relies on large pools of
resources to operate, the practice of clustering many nodes
becomes a very attractive proposition. This is because it
allows many redundant nodes to form the resource pool.
3) Hypervisor
A hypervisor or virtual machine manager is a piece of
software that allows the simultaneous running of one or more
operating systems known as virtual machines (VM) [15], [16].
The virtual machines running on the system is called the guest
operating systems. Each guest is allocated resources by the
hypervisor from the host resource pool. Hypervisors are
divided into Type 1 and Type 2 hypervisors.
 Type 1: This type is known as a bare metal or native
hypervisor, which runs directly on the hardware
without a host OS.
 Type 2: This type is known as a hosted hypervisor,
which runs on a host OS. The host OS is installed on
the host system and the hypervisor is installed on the
host OS.
The architecture proposed in this paper is independent of the
type of hypervisor.
B. Cloud Forensics
As digital forensics of systems become well defined, cloud
forensics must naturally follow suit. However, because of the
nature of cloud systems, cloud forensics have some unique
difficulties associated with it [17]. Cloud computing forensic
science is defined by the US National Institute of Standards
and Technology (NIST) as the application of scientific
principles, technological practices and derived and proven
methods to reconstruct past cloud computing events through
identification, collection, preservation, examination,
interpretation and reporting of digital evidence [1], [18].
The NIST also defines nine major groups of difficulties
associated with cloud forensics. These groups are architecture,
data collection, analysis, anti-forensics, incident first
responders, role management, legal, standards and training. At
the moment cloud architectures are not geared toward digital
forensics as it has not been a requirement.
The current architectures provide rapid provisioning of
cloud services making data volatility an issue that must be
addressed [19]. In addition to the architecture problems, the
problem of data collection is also an issue. As data can be
migrated in cloud systems, either manually or automatically
[20], it can be very difficult to isolate and collect relevant data.
The isolation problem can also translate into anti-forensics
[21], since someone could actively use this to hamper an
investigation. The large volume of captured data can also be
an issue when it comes to analysis. By using the proposed
architecture for proactive forensics, the groups of architecture,
data collection and analysis are addressed.
C. ISO Standard 27043
The ISO 27043 standard describes the idealized structure for
an investigation process across various investigation scenarios
[2]. The process classes are the readiness, initialization,
acquisitive, investigative and concurrent process classes. This
paper focuses on the readiness processes class, as it aids in the
execution of the other classes.
The readiness class includes the processes that enables an
organization to be set up in a way that should the need for the
collection of digital evidence arise, the potential of successful
collection is maximized while the cost and time of the
investigation is minimized. As stated earlier, this class of
processes is optional and dependent on the individual
organizations to implement readiness measures. Four goals are
associated with the readiness class: to have the digital
evidence used to its maximum potential, to minimize the cost
implications of the investigation, both in terms of financial
cost and cost of impact to the affected system, to have
minimal impact on the business processes of the affected
organization and to have the level of information security, of
the organization's systems, be improved or preserved.
 TABLE I SUMMARY OF PROPOSED ARCHITECTURAL MODELS

Operating System Hypervisor Communication Single tenant VM Multi-tenant VM

Embedded Embedded layer
Implementation Built into OS. Can Built into Built into cloud Runs on Runs on
be ran as an OS Hypervisor. OS / Hypervisor as a Hypervisor as a
service, daemon Application on VM tenant. VM tenant.
or program. cloud
OS.
Implementation Medium. High. Must form High. Must be on Low. Nested VMs Low. Nested VMs
Complexity Dependent on the part of the the protocol layer
security required. Hypervisor requiring rewrites
implementation. of said protocols.
Data segregation High. Per OS High (best case). Low to Medium. High. One guest Medium. Multiple
instance, can Managed by Dependent on OS per instance of guest tenants in
easily be Hypervisor, implementation. forensic VM forensic VM.
differentiated. dependent on All data must pass
implementation. via the cloud OS.
Tamper resistance Dependent on High. Cloud user Dependent on High. Cannot be High. Cannot be
implementation of does not have implementation. accessed or accessed
monitor in guest access to High when detected by cloud or detected by
OS. Hypervisor. embedded. Low user. cloud user
when a separate
application.
As a Service Software as a Platform as a Infrastructure- Platform as a Platform as a
model Service Service /Platform Service Service
as a Service,
depending on
implementation

The ISO 27043 standard serves as a baseline reference
point for this paper, in terms of the goals that must be
achieved in order to realize forensic readiness. The aim of the
proactive forensic architecture is realize the goals set forth in
the readiness class of the ISO 27043 standard.
D. Previous work
The work of Ras and Venter [22] describes 5 possible
architectures for enabling proactive cloud forensics. Table I
summarizes the different architectures in terms of their
implementation, implementation complexity, data segregation,
tamper resistance, and As a Service paradigm. The model
architectures are based on the NIST reference architecture for
cloud systems [1] where hardware nodes are virtualized via
clustering and the installation of a cloud operating system
(OS). A hypervisor then allows guest virtual machines (VMs)
to be tenants on the cloud system. Using the information in
Table I, a selection can be made of which model to implement
as a proof of concept. The model in this paper builds on this
previous work and extends the model into a theoretical
architecture that can be implemented to address the problems
of architecture, data collection and data analysis.
III. PROPOSED MODEL
From the previous work, a full system model can be
designed which is shown in Fig. 1. Due to the fact that this
paper deals with a proof of concept and that the model must
still be implemented, the Single tenant VM model is chosen as
a departure point for the full model architecture. This model is
selected because it offers the least complex implementation
and best data segregation. The most prominent drawback is
that the Single tenant VM model has an adverse impact on the
performance of the cloud system.
It is not the intention of the system to capture all the
information contained in all the virtual machines of the cloud
system. Capturing all the information of each VM in the cloud
is not a feasible solution. By capturing all the information,
each guest VM would take up roughly twice the amount of
disk space i.e. the space required for the VM to execute and
the space required for storage of the captured information. The
implementation of the model is intended for a specialized case
of private cloud systems. In this case private cloud systems
can be defined as cloud systems owned by companies or
organizations that are not open for use by the general public.
The model, as shown in Fig. 1, is comprised of four
components namely: the cloud hosting system (block 1), the
forensic controller (block 2), the cloud forensics system (block
3) and the cloud security system (block 4).
A. Cloud hosting system
The single tenant forensic monitor model, uses nested
VMs to enable forensic monitoring and is shown in Fig. 1

Fig. 1 Cloud Forensic Model Architecture
block 1. With this model, a single guest OS is contained
within a single forensic VM. Thus, the forensic VM acts as a
wrapper for the guest OS isolating it from the underlying
cloud structure. In this case the forensic VM runs on top of the
hypervisor. This is done so that the implementation of any
existing cloud system need not be changed.
This model thus requires that all data passes from the guest
OS, to the rest of the cloud stack via the forensic VM.
Data is intercepted and captured by the forensic VM where
some automated forensic analysis can be performed. Hashing,
for instance, can then take place. The function of the forensic
VM is to capture data and transmit it to the cloud forensic
system. This is done to minimize the required computational
overhead in the cloud hosting system.
The operation of the forensic VM is transparent to the user.
From the user point of view, the provisioned guest operating
system run directly on the cloud hosting system.
B. Forensic controller
The forensic controller (Fig. 1 block 2) governs the
forensic VMs in the cloud system. It is connected to the
security system of the cloud infrastructure.
The controller has the function of assigning the level of
forensic data acquisition to the different forensic VMs. In
order to save both space and computational overhead, the
forensic VM can be configured to different levels of data
acquisition. At the lowest level, the VM only captures minimal
data, for example log files. Whereas on the highest level, the
entire guest VM can be captured for forensic analysis. This
level of data acquisition is determined by the configuration set
by the cloud service provider.
The forensic controller receives input from the cloud
security system. Should a specific VM be targeted by some
external threat, the cloud security system detects the threat and
signals the forensic controller. The forensic controller can
escalate the level of data acquisition by the forensic VM for
the targeted guest VM. This can be done either for a single
VM or for a group of VMs, depending on the detected threat.
The captured forensic data is transported via a secure
channel to the cloud forensics system where automated
analysis can be performed.
C. Cloud forensic system
The cloud forensic system (Fig. 1 block 3) is the
destination for data captured by the forensic VMs. It is a
separate system that stands apart from the cloud hosting
system. The cloud forensic system consists of a forensic hash
store, a forensic payload store and an analysis engine.
The forensic hash store contains the hash signatures of that
is generated by the forensic VMs. These hashes are catalogued
and stored in a way that it is easy to differentiate their origin.
The hashes serves the purpose that in the case of an
investigation, the forensic hashes can be compared to the
hashes generated by the investigator in order to determine
where changes to data took place. It can also be used as a
short-cut by negating the need to generate the hashes from
scratch.
The forensic payload store contains the forensic data
collected from the different forensic VMs. When the forensic
VM receives a signal to increase the level of data acquisition,
the acquired data is transmitted to the forensic payload store
via a secure channel. The acquired data is catalogued and
sorted in order to easily determine its origin from within the
cloud hosting system. The data is segregated and secured
should it be needed in an investigation.
The analysis engine performs a predetermined analysis on
the forensic data. This can be configured to such tasks as
cataloguing and analysis of images, analysis of
communications and anomaly detection. The proactive
analysis data can save time for the investigator and speed up
the investigation.
D. Cloud security system
The cloud security system (Fig. 1 block 4) is the standard
security systems that are present in a cloud environment for
example firewalls and anti-virus systems. As these standard
systems are already optimized to their purposes, it stands to
reason that they should not be replaced. Thus the only addition
to this existing security system would be the interface to
forensic controller.
IV. EVALUATION
There are four goals described in readiness class of the
ISO 27043 standard. These goals are: to have the digital
evidence used to its maximum potential, to minimize the cost
implications of the investigation, both in terms of financial
cost and cost of impact to the affected system, to have
minimal impact on the business processes of the affected
organization and to have the level of information security, of
the organization’s systems, be improved or preserved.
In terms of using the forensic data to its maximum
potential, cloud data is volatile making it difficult to use this
data to its maximum potential as it can easily be lost. The
cloud forensic system enables better use if this data due to the
fact that the forensic data is captured in a proactive manner.
Using a proactive approach to forensics reduces the need for
large scale disruption of the cloud system, in that the forensic
data is already captured, catalogued and to an extent analyzed
by the forensic system. This in turn reduces the cost associated
with such an investigation both in system downtime and time
needed by the investigator. As it is not necessary to shut down
the entire cloud system for an investigation, the cloud service
provider can continue its business operations with minimal
disruption. This also maintains the information security of
other users of the cloud service provider as the investigation
can be immediately targeted to the affected systems.
The implementation of the proposed architecture will have
an impact on the performance of the cloud system. This
impact will not be seen by the end user in terms of response
times, but rather but the cloud provider in terms of lost
resources. The lost resources will be in the form of computing
resources, like hard disk space and computation cycles used
by the forensic systems. While these problems will have a cost
implication to the cloud services provider, it can be offset by
passing the costs to the user, which in the case of private cloud
would most likely be internal business units. Also, should the
need arise for an investigation, there would be no need to shut
down the entire cloud for the investigation. This will lead to a
massive saving due to not losing revenue due to system down
time.
Service level agreements forms an integral part of the
forensic system. It must be made clear to the user that data
will be captured and for what purpose it will be used.
Naturally every effort must be made to secure captured data in
order to maintain the privacy of the clients of the cloud service
provider.
It is noted that there will be an impact on the performance
of the cloud system. This impact will have some cost
associated with it due to the computational overhead and
additional hardware requirements for the forensic systems.
The extent of the performance impact is outside the scope of
this paper and will be dealt with in future research.
V. CONCLUSION AND FUTURE WORK
Cloud forensic remains challenging field due to challenges
for example data volume, data isolation, data integrity,
availability of the cloud system and a lack of standardization
of cloud forensics. By applying the processes found in the
readiness class found in ISO 27043 and creating a model
architecture to enable proactive forensics of cloud computing
systems, a possible solution is found. The work of Richter et
al. [23] shows a method of accessing the file system of the
guest VM that could serve as an implementation strategy for
the proposed model.
This solution addresses the some of the problems
associated cloud systems. While this proposed solution might
show a way forward, it still requires further research. The next
steps are to implement the proactive cloud forensic model.
Metrics must be defined in order to measure what the impact
of a forensic system would be in a cloud environment. These
metrics should include, performance in terms of input/output
response, CPU time used for forensic system, hard disk space
used by the forensic system and so forth.
In addition to the impact assessment, the utility of the
forensic system must also be judged. This can be done in
terms of quality of the data captured i.e. would the data be
admissible in a court of law, cloud stem uptime and usefulness
of the captured data.
With these factors taken into account and with the aid of a
proof of concept it should be possible to determine if the
model is suitable to practically address the problem of cloud
forensics.
References
[1] N. US Department of Commerce, “NIST Cloud Computing Forensic
Science Challenges NIST Cloud Computing,” 2014.
[2] ISO 27043, “Incident investigation principles and processes,” Inf.
Technol. - Secur. Tech., vol. 2015, pp. 1–27, 2015.
[3] S. Ballou, J. Carazo, B. Crane, F. Demma, G. Gottfried, S. Guttman, J.
Herig, T. Hutchison, D. Icove, B. Jarzen, T. Johnson, K. Matthews, M.
Pollitt, D. Poole, M. Riley, K. Schmid, H. A. Schmidt, R. Schmidt, C.
Selavka, S. Sepulveda, T. Shipley, C. Stippich, C. M. Whitcomb, and W.
Williams, “Electronic Crime Scene Investigation: A Guide for First
Responders,” Washington, 2001.
[4] G. Pangalos, C. Ilioudis, and I. Pagkalos, “The Importance of Corporate
Forensic Readiness in the Information Security Framework,” 2010 19th
IEEE Int. Work. Enabling Technol. Infrastructures Collab. Enterp., pp.
12–16, 2010.
[5] R. Rowlingson, “A ten step process for forensic readiness,” Int. J. Digit.
Evid., vol. 2, no. 3, pp. 1–28, 2004.
[6] P. Mell and T. Grance, “The NIST Definition of Cloud Computing (
Draft ) Recommendations of the National Institute of Standards and
Technology,” 2011.
[7] T. Lillard, Digital Forensics for Network, Internet, and Cloud
Computing: A Forensic Evidence Guide for Moving Targets and Data,
1st ed. Syngress, 2010.
[8] K. Hess and A. Newman, Practical Virtualization Solutions:
Virtualization from the Trenches. Prentice Hall, 2009.
[9] M. Christodorescu, R. Sailer, and D. Schales, “Cloud security is not
(just) virtualization security: a short paper,” in Proceedings of the 2009
ACM workshop on Cloud computing security, 2009, pp. 97–102.
[10] R. Moreno-Vozmediano, R. S. Montero, and I. M. Llorente, “Elastic
management of cluster-based services in the cloud,” in Proceedings of
the 1st workshop on Automated control for datacenters and clouds -
ACDC ’09, 2009, p. 19.
[11] B. Siddhisena, L. Warusawithana, and M. Mendis, “Next generation
multi-tenant virtualization cloud computing platform,” in Advanced
Communication Technology (ICACT), 2011 13th International
Conference on, 2011, vol. 54, no. 2, pp. 405–410.
[12] D. Barrett, Virtualization and Forensics A Digital Forensic
Investigator’s Guide to Virtual Environments. Syngress, 2010.
[13] H. A. Lagar-Cavilla, J. a. Whitney, R. Bryant, P. Patchin, M. Brudno, E.
de Lara, S. M. Rumble, M. Satyanarayanan, and A. Scannell,
“SnowFlock,” ACM Trans. Comput. Syst., vol. 29, no. 1, pp. 1–45, Feb.
2011.
[14] E. Manoel, C. Carlane, L. Ferreira, S. Hill, D. Leitko, and P. Zutenis,
Linux Clustering with CSM and GPFS. IBM Redbooks, 2002.
[15] P. Dash, “Hypervisor,” in Getting Started with Oracle VM VirtualBox,
Packt Publishing, 2013.
[16] O. Karakok, A. Afrose, H. Fayed, D. Klebanov, and N. Shamsee,
“Server Virtualization,” in CCNA Data Center DCICT 640-916 Official
Cert Guide, Cisco Press, 2015.
[17] K. Ruan, J. Carthy, T. Kechadi, and M. Crosbie, “Cloud Forensics,” in
Advances in Digital Forensics VII, G. Peterson and S. Shenoi, Eds.
Springer Berlin Heidelberg, 2011, pp. 35–46.
[18] P. Mell and T. Grance, “The NIST Definition of Cloud Computing:
Recommendations of the National Institute of Standards and Technology
(2011),” Gaithersburg, 2012.
[19] D. Birk, “Technical Challenges of Forensic Investigations in Cloud
Computing Environments,” in Workshop on cryptography and security
in clouds, 2011, pp. 1–6.
[20] Y. Jadeja and K. Modi, “Cloud computing - Concepts, architecture and
challenges,” in 2012 International Conference on Computing,
Electronics and Electrical Technologies, ICCEET 2012, 2012, pp. 877–
880.
[21] K. Dahbur and B. Mohammad, “The anti-forensics challenge,” in
Proceedings of the 2011 International Conference on Intelligent
Semantic Web-Services and Applications - ISWSA ’11, 2011, pp. 1–7.
[22] D. J. Ras and H. S. Venter, “Models for the Forensic Monitoring of
Cloud Virtual Machines,” in 13th European Conference on Cyber
Warfare and Security ECCWS 2014, 2014, pp. 290–299.
[23] W. Richter, C. Isci, B. Gilbert, J. Harkes, V. Bala, and M.
Satyanarayanan, “Agentless Cloud-Wide Streaming of Guest File
System Updates,” 2014 IEEE Int. Conf. Cloud Eng., pp. 7–16, 2014.
 Fusion of Local and Global Features for
Face Recognition
Nazmeen B. Boodoo-Jahangeer* & Sunilduth Baichoo
Department of Computer Science, University of Mauritius,
Reduit, Mauritius
nazmeen182@yahoo.com*

Abstract—In recent years, there have been lots of research
developments done in face recognition systems. Face recognition
systems are widely used for access control, border control,
surveillance and in law enforcement. Among other biometrics, it
is the most natural and acceptable way of identifying an
individual. Face recognition system does not require physical
interaction with the user. Research is still being done intensively
to produce systems that can cater for several challenges such as
changes in pose, illumination, occlusion and low resolution
images. Algorithms reported in literature use either global
feature extraction or local feature extraction. In this work, a
different technique is proposed that combines both global and
local approaches for face recognition. The Principal Components
Analysis (PCA) and Local Binary Patterns (LBP) have been
employed. Face recognition yields a recognition rate of 90 % with
PCA and 92 % with LBP. However, results show an
improvement in recognition rate to 95 % when both approaches
are fused.
Keywords—Face Recognition; Biometric;, Global Feature;
Local Feature; PCA; LBP
I. INTRODUCTION
The needs for reliable authentication techniques have
increased due to information security, privacy concerns and
rapid advancements in networking, communication and
mobile systems. Increasing demands and deployments of
biometrics are observed in diverse environments such as
airports, banks, law enforcements, secure access controls,
commercial and forensic applications. Biometric-based
authentication has started to gain acceptance as a legitimate
method for determining an individual’s identity in many
countries.
Biometric refers “to identifying an individual based on his
or her distinguishing characteristics” [1]. Biometric traits are
often classified as physiological or behavioral traits.
Physiological characteristics are related to the shape of the
body. Examples include, but are not limited to fingerprint,
face, iris, hand geometry, palm print, retina and DNA.
Behavioural characteristics are related to the behavior of a
person, including, but not limited to: Keystrokes, signature,
gait and voice. A more recent term in this field is soft
biometric. Soft biometric means certain information about the
user, which can be combined with his biometric identity to
provide a stronger authentication method.
Traditional methods of authentication were done using two
approaches. Firstly, the user can possess something that will
help for authentication, such as User ID, keys, smart card and
badge. However, these things can be easily shared,
duplicated, lost or stolen. A second approach to identify a
person is that the user can remember something such as a pin
or password. This approach also is not secure as the pin or
password can be easily obtained by any third party who wants
unauthorized access to valuables and information [2].
A different method of authentication is by using
biometrics. Biometric-based methods easily deal with the
problems of traditional methods since users are identified by
who they are, not by something they have to remember or
carry with them [3]. Biometric traits are more difficult to
forge, copy, share, misplace or guess[4]. A biometric system
requires the person being authenticated to be present at the
time and point of authentication. Biometric has another
advantage over traditional authentication methods by
providing negative identification functionality. These
situations arise when the user tries to avoid successful
identification (in case of thieves) or an imposter trying to
establish a false identity. Biometric recognition has thus
become an indispensable tool for authentication in this
technologically-advanced society.
Though many biometric traits are available nowadays, not
all biometrics are suitable for all types of applications [5].
Each application has its own requirements in terms of
recognition accuracy, resource requirements, reliability and
cost [6]. In this work, the focus is on face recognition.
A facial recognition system is a computer-based
authentication system that can identify users based on their
unique facial characteristics. Unlike other biometrics such as
fingerprint, hand geometry, DNA, face recognition does not
require direct contact with the sensor, since face image can be
taken from a distance, thus hygiene issues do not arise. Also,
face recognition is the most socially and culturally accepted
internationally, since already existing manual facial
recognition is being used for passports. However, face
recognition systems suffer from occlusion of hair and make-
ups. Moreover, changes in facial expressions, illuminations
and pose also affect the system. In order to cater for these
problems, in our research, a fusion of global and local features
has been done to make the recognition system more robust.

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

 II. BACKGROUND STUDY
Face recognition is currently highly researched area of
computer vision and pattern recognition [7]. While many
algorithms are being developed, they are usually compared to
existing ones quite superficially and only simple comparisons
are reported. Given the numerous theories and techniques that
are applicable to face recognition, it is clear that detailed
evaluation and benchmarking of these algorithms is crucial.
Effort done by FERET researchers in their evaluations [8]
pushed face recognition algorithm comparisons to the next
level.
Research in automatic face recognition dates back to the
1960’s [9]. However, the system was not reliable for
recognizing faces in real-time. A survey of face recognition
techniques has been given by Zhao, W. et al. [7]. Face
recognition methods can be classified in three categories based
on feature extraction methods used in the recognition process:
global approach, component-based approach and hybrid
approach. In the global approach, the whole face image, which
is represented by a single feature vector, is used as input to a
classifier. The commonly-used techniques for global approach
are PCA, LDA, ICA support vector machine, hidden markov
model (HMM). The global methods work quite well when
using the frontal face images, however the performance of the
system degrades when there is
The component-based approach analyzes local features
such as nose, eyes and their geometric relationships.
Sometimes this approach is known as only feature-based
approach or geometry-based approach. The advantage of this
technique is that it can compensate for slight changes in pose.
The hybrid approach combines both the local feature and
global feature for recognition to create a robust system. The
following section outlines the commonly-used algorithms for
face recognition.
The Principal Component Analysis (PCA) has been
extensively applied to face recognition. Kirby and Sirovich
were among the first to apply PCA to face images [10]. Turk
and Pentland popularized the use of PCA for face recognition
[11] . This technique is a holistic method that takes as input
the face image directly. The PCA algorithm is used to find a
subspace whose basis vectors correspond to the maximum
variance directions in the original n dimensional space. The
PCA algorithm allows any face image to be reconstructed by
the most discriminative eigenfaces.
The main advantage of the PCA algorithm is that is caters
for dimensionality reduction. It is simple and efficient when
images used are taken under controlled conditions. However,
PCA does not provide good performance when there is a
change in lighting or pose [12].
Researchers began to search for other subspaces that might
improve performance. One alternative is Fisher’s Linear
Discriminant Analysis (LDA) [13]. The goal of LDA is to find
a basis for projection that minimize the intra-class variation
but preserve the inter-class variation. The algorithm makes
use of PCA as the first step [14]. According to Martinez and
Kak [15], the PCA algorithm outperforms LDA when the
training dataset is small. The fisher faces is insensitive to
changes in light intensity and can cater for different facial
expressions. However, the dimension of the fisherface is not
as compact as eigenface, which require larger storage space
and more processing time [16].
In PCA, the basis images depend only on pairwise
relationships between pixels in the image database. In face
recognition, there may be important information contained in
the high-order relationships among pixels. Thus the basis
images may be found by methods sensitive to these high-order
statistics. One method that can be used is the component
analysis (ICA), a generalization of PCA. ICA provides a
representation based on statistically independent variables
[17]. According to Liu and Wechsler [18], ICA has better
performance than PCA algorithm in the context of face
recognition.
Singular Value decomposition (SVD) is another way of
extracting algebraic features from an image. The SVD
technique converts the high-dimensional face image into a
subspace, similar to PCA. Classification is done by comparing
position of test image with the positions of the known faces in
the subspace. The SVD of a face image is said to have a good
stability and is suitable for use when when images are taken
under different noise and view point conditions [19]. When
applied to facial expression recognition, SVD yields
satistactory results [20].
The traditional PCA caters for linear dimensionality
reduction. This method becomes invalid if the data has more
complicated structures which cannot be simplified in a linear
subspace. The Kernal PCA allows for generalization of
traditional PCA to nonlinear dimensionality reduction [21].
Researchers have made use of local approach for feature
extraction such as local binary patterns [22]. Local descriptors
are found to be more robust against occlusion, expression
variation, pose variation and small sample sizes than
traditional holistic algorithms
The following paragraphs provide an overview of a few
among the recent techniques applied by researchers for face
recognition.
Agarwal et al. [23] used principle component analysis for
feature extraction and the feed forward back propagation
Neural Network for recognition. They tested this approach on
400 images. All the variants of feature extraction is considered
for the calculation of the recognition score. The Olivetti and
Oracle Research Laboratory (ORL) face databases were used
for testing. Experimental results gave a recognition rate of
97.018%.
The area of face recognition still attracts lots of interest.
One of the recent works includes that of Sellahewa and Jassim
[24] who proposed a quality-based adaptive approach to face
recognition. An objective measure of illumination quality of a
given face image was derived to decide if the image would be
pre-processed to normalize its illumination. Also, the authors
devised a global quality-based normalization scheme extended
to a regional quality-based approach to adaptive illumination
normalization. Finally, the illumination quality measure was
used as a means to adaptively select the weighting parameters
of the fused wavelet-based multi-stream face recognition
scheme.
Hsieh C. et al. (2010) created a system that is able to
recognize faces with expressions under the restriction of one
single training sample per class. Using a constrained optical
flow algorithm, the authors calculated the expressional
motions from each neutral faces in the database to the input
test image, and estimated the likelihood of such a facial
expression movement. Using the optical flow information,
neutral images in the database could be further warped to
faces with the exact expression of input image. The authors
exploited the two different types of information, i.e., the
computed optical flow and the synthesized image, to improve
the accuracy of face recognition. Experiments conducted on
the Binghamton University 3-D Face Expression Database
showed the improved performance of the proposed face
recognition system.
Though extensive research has been done in face
recognition, there are still several unresolved problems. These
include remote face recognition, video-based face recognition,
face recognition across aging and face recognition in
unconstrained conditions, among others [26]. In fact, Park et
al. [27] proposed a method of 3D aging modelling technique
to compensate for the age variations. The method adapted
view-invariant 3D face models to the given 2D face aging
database and it was evaluated on three different databases
(FG-NET, MORPH, and BROWNS) using FaceVACS, a
state-of-the-art commercial face recognition engine. The
experimental results proved that the accuracy of the face
recognition system was significantly improved after applying
the aging model in the three databases.
Another study [28] emphasize on the use of normalization
to increase the face recognition rate. A U-face concept was
proposed that removes background, hair, and other factors that
are not conducive to face recognition; retain only the focus of
the face region. This method was tested using Kernel Principal
Component Analysis (KPCA) and Principal Component
Analysis (PCA). The recognition rate has increased by 7 % by
both methods when applying such normalization.
According to literature, face recognition systems work
well when the training and test images are acquired in similar
conditions. However, the system’s performance degrades
when there are variations between the test and training images.
The difficulties of face recognition systems remain changes in
pose, expression, age, illumination and occlusion. In this
work, a technique for combining local and global features will
be proposed in order to improve performance of existing face
recognition systems.
III. METHODOLOGY
Proposed algorithms for biometric recognition can fall
into two major categories: Global and Local. The global
approach takes into consideration the whole image and use it
for feature extraction and recognition, whereas local methods
make use of local feature regions in that image as descriptor.
Most shape and texture descriptor falls in global approach
category, they provide a very compact representation of the
images, where each image is a point in a high-dimensional
space. The dimensionality corresponds directly to the size of
the image in terms of pixels. Pattern recognition is very
complex when using high-dimensionality. One way to deal
with it is to apply dimensionality reduction technique. The
Principal Components Analysis (PCA) is commonly-used to
transform an image into a low dimension feature space such
that the major variations are preserved. Since global methods
are sensitive to clutter and occlusion, it is important to have
only one object in the image or to have a good segmentation
of the object from the background.
The advantages of using local features are that they have
low computational complexity and speed up the image
processing [29]. Local methods for feature extraction are more
robust to clutter and occlusion. However, in cases where
images are not of good quality, global method is preferred.
One limitation of local feature is that there may be differing
number of feature points in each image, thus making
comparison more difficult
In order to get the advantages of both local and global
approach of feature extraction, both methods are applied to
face images and the results are fused. The Principal
Components analysis is chosen as global method while the
Local Binary Patterns are chosen as the local method.
A. Principal Components Analysis (PCA)
Principal components analysis (PCA) is a technique most
commonly used for dimensionality reduction in computer
vision [11]. The features extracted are based on Karhunen-
Loeve (KL) expansion. The corresponding algorithm in the
context of face recognition is called eigenface method. In fact,
the eigenface method generates features that capture the
holistic nature of the faces through PCA.
On the other hand, the main disadvantages of KL methods
are its complexity and that the extracted base is data-
dependent: if new images are added to the database the KL
base need to be recomputed. The main idea is to decompose a
face picture as a weighted combination of the orthonormal
base provided by the KL transform. The base corresponds to
the eigenvectors of the covariance matrix of the data, known
as eigenfaces or eigenears.
Thus, the decomposition of a face image into an eigenface
space provides a set of features. The maximum number of
features is restricted to the number of images used to compute
the KL transform, although usually only the more relevant
features are selected, removing the ones associated with the
smallest eigenvalues.
The basic idea of PCA is to find an optimal linear
transformation that maps the original n-dimensional data space
into an m-dimensional feature space (m < n) to achieve
dimensionality reduction. The PCA algorithm chooses a
dimensionality reducing linear projection that maximizes the
scatter of all projected samples. PCA seeks to find the vectors
that best describe the data in terms of reproducibility; however
these vectors may not include the effective information for
classification, and may eliminate discriminative information.
The PCA algorithm is as follows: A 2-D facial image can
be represented as 1-D vector by concatenating each row (or
column) into a long thin vector. Let’s suppose we have M
vectors of size N (= rows of image × columns of image)
representing a set of sampled images. P j’s represent the pixel
values.
The images are mean centered by subtracting the mean
image from each image vector. Let m represent the mean
image.
And let wi be defined as mean centered image
The goal is to find a set of ei’s which have the largest
possible projection onto each of the wi’s. A set of M
orthonormal vectors ei is found for which the quantity
is maximized with the orthonormality constraint
It has been shown that the ei’s and ʎi’s are given by the
eigenvectors and eigenvalues of the covariance matrix
where W is a matrix composed of the column vectors wi
placed side by side. The size of C is N × N which could be
enormous. For example, images of size 64 × 64 create the
covariance matrix of size 4096 × 4096. It is not practical to
solve for the eigenvectors of C directly. A common theorem in
linear algebra states that the vectors ei and scalars ʎi can be
obtained by solving for the eigenvectors and eigenvalues of
the M × M matrix WTW. Let di and µi be the eigenvectors and
eigenvalues of WTW, respectively.
By multiplying left to both sides by W
which means that the first M - 1 eigenvectors ei and
eigenvalues ʎi of WWT are given by Wdi and µi, respectively.
Wdi needs to be normalized in order to be equal to e i. Since a
finite number of image vectors, M can be summed up, the rank
of the covariance matrix cannot exceed M – 1.
The eigenvectors corresponding to nonzero eigenvalues of
the covariance matrix produce an orthonormal basis for the
subspace within which most image data can be represented
with a small amount of error. The eigenvectors are sorted from
high to low according to their corresponding eigenvalues. The
eigenvector associated with the largest eigenvalue is one that
reflects the greatest variance in the image.
A facial image can be projected onto M1 (<< M)
dimensions by computing
where vi = eTi wi . vi is the ith coordinate of the facial image
in the new space, which came to be the principal component.
The vectors ei are also images, so called, eigenimages, or
eigenfaces. They can be viewed as images and indeed look
like faces. Ω describes the contribution of each eigenface in
representing the facial image by treating the eigenfaces as a
basis set for facial images. The simplest method for
determining which face class provides the best description of
an input facial image is to find the face class k that minimizes
the Euclidean distance
where ϵk is a vector describing the kth face class. If ϵk is less
than some predefined threshold ϴϵ, a face is classified as
belonging to the class k.
B. Local Binary patterns
LBP operator, originally introduced by Ojala et al. [22], is
a powerful way of describing the texture of an image. The
original version of the local binary pattern operator works in a
3 × 3 pixel block of an image. The pixels in this block are
threshold by its center pixel value, multiplied by powers of
two and then summed to obtain a label for the center pixel. As
the neighborhood consists of 8 pixels, a total of 28 = 256
different labels can be obtained depending on the relative gray
values of the center and the pixels in the neighborhood.
 The steps for LBP operator are given into more details
below:

The image is first divided into cells, consisting of a

number of 3 × 3 pixel blocks. The center pixel, pxc, in a block
is compared to each of its 8 neighbors, pxi (i=0, 1, …,7). The
pixels are followed along a circle, either clockwise or counter-
clockwise.

Fig. 2: LBP number calculation

Where the center pixel's value is greater than the

neighbor's value, "0" is written. Otherwise, "1" is written. This
gives an 8-digit binary number. Fig. 1 illustrates the basic LBP Fig. 3 shows a sample image and the corresponding
operator. histogram.

Fig. 3: Image and histograms

The PCA and LBP algorithm will be applied in the UOM

database, details of which is given in the next section.

Fig. 1: Basic of LBP,(a) Gray Value, (b) Corresponding Binary Pattern

Then, the values of the pixels in the thresholded
neighborhood are multiplied by the weights given to the
corresponding pixels. Finally, the values of the eight pixels are
summed to obtain the LBP number for the centre pixel:
C. Dataset
The UOM database was created to collect face images
from over 100 volunteers. The latter are students and staff
from the University between 20 to 50 years old, Five CCTV
cameras have been used and positioned at -900, -450, 00, 450
and 900, as shown in the diagram below:

The LBP number derived is 110 ( 2 + 4 + 8 + 32 + 64), as

shown in Fig. 2. The histogram, over the cell, of the frequency
of each "number" occurring is computed. The histograms of
all cells in the image are then concatenated. This gives the
feature vector for the image.

Subject
Fig 4. Setup of cameras
The CCTV cameras took videos of subjects for about one
minute. The frames were then extracted from the videos, 10
frames per subject, resulting in 1000 images. The resolution of
the images is 252 by 288 pixels. Five images per subject were
used for training set and five for the test set.
D. Fusion strategy
Integration of information in a Multimodal biometric
system can occur in three main levels, namely feature level,
matching level or decision level. At feature level, the feature
sets of different modalities are combined. Fusion at this level
provides the highest flexibility but classification problems
may arise due to the large dimension of the combined feature
vectors. Fusion at matching level is the most common one,
whereby the scores of the classifiers are usually normalized
and then they are combined in a consistent manner. At fusion
on decision level each subsystem determines its own
authentication decision and all individual results are combined
to a common decision of the fusion system.
In this study, fusion at the decision level is applied for data
fusion of the two features, based on the majority vote rule. For
three samples, as is the case, a minimum of three accept votes
is needed for acceptance. Also, for the final fusion, the AND
rule is used. Fig 5 shows the fusion applied in this study.
Fig. 5 Decision level fusion
IV. RESULTS
The results are given in terms of False Acceptance Rate
(FAR), False Rejection Rate (FRR) and the Receiver
Operating Characteristic (ROC) curve. FRR is the percentage
of users who claimed a specific identity, for which the system
has either falsely rejected them during decision process. FAR
is the percentage of users who claimed a specific identity, and
that the system has falsely accepted them for this claimed
identity during the decision process. Another way of analyzing
the performance of a biometric system is by making use of
Receiver Operating Characteristic (ROC) curves. An ROC
graph is a technique for visualizing, organizing and selecting
classifiers based on their performance [30]. The ROC curves
are an accepted method for summarizing the performance of
imperfect diagnostic, detection, and pattern matching systems
[31]. A ROC curve plots, parametrically as a function of the
decision threshold, the rate of “true positives” (i.e. genuine
attempts accepted) on the y-axis against the corresponding rate
of “false positives” (i.e. impostor attempts accepted) on the x-
axis, [32]. ROC curves are threshold independent, allowing
performance comparison of different systems under similar
conditions, or of a single system under differing conditions.
The results of face recognition on UOM database using
PCA technique provide the results as shown in Fig. 6.
Fig. 6 FAR and FRR against threshold for PCA algorithm
The Fig. 6 is used to choose the decision threshold. It helps
to choose the threshold that provide the lowest FAR and FRR
values. The experiment yields 90 % recognition rate, with
FAR 0.1 and FRR 0.1.
When the Local Binary Patterns (LBP) is applied to UOM
face database, the FAR and FRR values obtained at different
threshold are shown in Fig. 7.

Fig. 7 FAR and FRR values against threshold for LBP algorithm
The experimental results with LBP algorithm give a
recognition rate of 92 %, with FAR 0.12 and FRR 0.08. The
individual scores of both algorithms are fused and the
corresponding results are shown in a ROC curve in Fig. 8.
Fig. 8 ROC Curve for LBP, PCA and combined approach
The ROC curve shows an improvement in recognition rate
when the combined approach is used as compared to using
only PCA or only LBP. The recognition rate increases to 95 %
with lower error rates, FAR of 0.04 and FRR 0.05. This
approach is helpful when we have low resolution images and
only one algorithm cannot cater for such images.
V. CONCLUSION
Research on face recognition has produced quite
satisfactory result, however, the performance of the system
degrades when there is pose variance. The performance of
existing face recognition system can be further improved by
merging both local and global features. Two techniques have
been investigated, the Principal Component Analysis (PCA)
and Local Binary Patterns (LBP). The UOM database has
been created for the purpose of this study. Experimental
results have shown that the combination of PCA and LBP
yields a higher recognition rate than using a single algorithm
only.
References
[1] Bolle, R. M., Connell, J. H., Pankanti, S., Ratha, N. K. & Senior, A. W.,
2003. Guide to Biometric: Springer-Verlag, New-York.
[2] Jain, A. K., Ross, A. & Prabhakar, S., 2001, Fingerprint Matching Using
Minutiae and Texture Features. In: Proceedings of the International
Conference on Image Processing, Thessaloniki, Greece, pp 282-285.
[3] Choraś, M., 2005. Ear Biometrics Based on Geometrical Feature
Extraction, Electronic Letters on Computer Vision and Image Analysis,
Vol. 5, No. 3, pp. 84-95.
[4] Pankanti, S. & Jain, A. K. 2008. Beyond Fingerprinting, Scientific
American, pp. 78-81.
[5] Jain A. K., Ross A., & Pankanti S., 1999, A Prototype Hand Geometry-
based Verification System, In Proc. of 2nd Int'l Conference on Audio-
and Video-based Biometric Person Authentication (AVBPA),
Washington,D.C., pp.166-171, March 22-24, 1999.
[6] Prabhakar S., Pankanti S. & Jain A. K., 2003. Biometric Recognition:
Security and Privacy Concerns, IEEE Security and Privacy Magazine,
Vol. 1, No. 2, pp. 33-42
[7] Zhao, W., Chellappa, R., Phillips, J., &. Rosenfeld, A., 2003. Face
Recognition in Still and Video Images: A Literature Survey, ACM
Computing Surveys, Vol. 35, pp. 399-458
[8] Phillips, P.J., Moon, H., Rizvi, S.A. & Rauss, P.J., 2000. The FERET
Evaluation Methodology for Face Recognition Algorithms, In: IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 22, No.
10, pp. 1090-1104
[9] Bledsoe, W. W., 1964. The model method in facial recognition, Technical
Report, PRI 15, Panoramic Research, Inc., Palo Alto, California.
[10] Kirby, M. & Sirovich, L., 1990. Application of the Karhunen-Loeve
Procedure for the Characterization of Human Faces. In: IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 12, pp.
103-107.
[11] Turk, M., & Pentland, A., 1991. Eigenfaces for recognition, Journal of
Cognitive Neuroscience, Vol. 3, No. 1, pp. 71-86, 1991.
[12] Etemad K. & Chellappa R., 1997. Discriminant Analysis for Recognition
of Human Face Images, Journal of the Optical Society of America, Vol.
14, pp 1724-1733.
[13]Swets, D., & Weng, J., 1996. Using Discriminant Eigenfeatures for Image
Retrieval. IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 18, pp. 831-836.
[14] Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J., 1997, Eigenfaces vs.
Fisherfaces: Recognition Using Class Specific Linear Projection, IIEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 19, No.
7, pp 711-720
[15] Martinez A. & Kak A., 2001. PCA versus LDA, IEEE Transactions on
Pattern Analysis and Machine Intelligence, vol. 23, no. 2, pp. 228-233.
[16] Jaiswal S., Bhadauria S. S. & Jadon R. S., 2011, Comparison Between
Face Recognition Algorithm-Eigenfaces, Fisherfaces And Elastic Bunch
Graph Matching, Journal of Global Research in Computer Science, Vol.
2, No. 7
[17] Bartlett, M. S., 2001. Face Image Analysis by Unsupervised Learning.
Kluwer Academic Publishers.
[18]Liu C. & Wechsler H., 1999. Comparative Assessment of Independent
Component Analysis (ICA) for Face Recognition, In: Second
International Conference on Audio and Video-based Biometric Person
Authentication, AVBPA’99,Washington D. C. USA, pp. 22-24
[19] Rasied T.S.M., Khalifa O.O., & Kamarudin Y.B., 2005. Face
Recognition Based On Singular Valued Decomposition And
Backprogagation Neural Network, In: 2005 1st International Conference
on Computers, Communications and Signal Processing with Special
Track on Biomedical Engineering, CCSP 2005, pp. 304-309.
[20] Kaur M., Vashisht R., & Neeru N., 2010. Recognition of Facial
Expressions with Principal Component Analysis and Singular Value
Decomposition, In Proc. International Journal of Computer Applications,
Vol. 9, No.12, pp. 36-40.
[21] Scholkopf B., Smola A., & M¨uller K.-R.. 1999, Kernel Principal
Component Analysis. In Scholkopf B., Burges C. J. C., & Smola A. J.,
editors, Advances in Kernel Methods - Support Vector Learning, pages
327–352. MIT Press, Cambridge, MA, 1999.
[22] Ojala T., Pietikainen M., and Harwood D., 1996. A comparative study of
texture measures with classification based on featured distribution,
Pattern Recognition, vol. 29, no. 1, 51–59.
[23] Agarwal, M. Agrawal, H. Jain, & N. Kumar, M., 2010. Face Recognition
using Principle Component Analysis, Eigenface and Neural Network. In:
2010 International Conference on Signal Acquisition and Processing,
ICSAP '10. Bangalore, 9-10 Feb. 2010, pp. 310 – 314.
[24] Sellahewa, H. & Jassim, S. A., 2010. Image-Quality-Based Adaptive
Face Recognition, IEEE Transactions On Instrumentation And
Measurement, Vol. 59, No. 4, April 2010, pp. 805 – 813.
[25] Hsieh, C., Lai, S. & Chen, Y., 2010. An Optical Flow-Based Approach to
Robust Face Recognition Under Expression Variations, IEEE
Transactions On Image Processing, Vol. 19, No. 1, January 2010, pp. 233
– 240
[26] Chellappa, R. Sinha, P. & Phillips, P.J., 2010. Face Recognition by
Computers and Humans, Computer, Vol. 43 , No. 2, pp. 46 – 55.
[27] Park, U., Tong, Y., & Jain, A.K., 2010. IEEE Transactions On Pattern
Analysis And Machine Intelligence, Vol. 32, No. 5, pp. 947-954
[28] Cheng Z., Sun J. & Lei X., 2010, U-face Of Applied Research In The
Face Recognition, 3rd International Conference on Advanced Computer
Theory and Engineering (ICACTE), Vol. 5, pp. 15-18,20-22 Aug. 2010
,Chengdu .
[29] De Silva V. & Tenenbaum J.B., 2003, Global Versus Local Methods In
Nonlinear Dimensionality Reduction, pp. 721–728 in S. Becker, S.
Thrun,, K. Obermayer (Eds.), Advances in Neural Information Processing
Systems, 15, MIT Press, Cambridge, MA, 2003.
[30] Fawcett, T., 2006. An introduction to ROC analysis, Pattern Recognition
Letters, Vol. 27, pp. 861–874.
[31] Mansfield, A. J. & Wayman, J. L. 2002. Best practices in testing and
reporting performance of biometric devices, National Physical
Laboratory, Centre for Mathematics and Scientific Computing, Tech.
Rep. Version 2.01, 2002.
[32] Lachiche, N. & Flach, P. 2003, Improving accuracy and cost of two-class
and multi-class probabilistic classifiers using ROC curves, In: Twentieth
International Conference on Machine Learning (ICML'03), Washington
DC., pp. 416–423.
 Rapid Prototyping with a Local Geolocation API
Geerish Suddul, Kevin Nundran, Jonathan L.K.
Cheung
Dept. Business Informatics & Software Engineering
University of Technology, Mauritius (UTM)
Republic of Mauritius
g.suddul@umail.utm.ac.mu
Abstract—Geolocation technology provides the ability to
target content and services to users visiting specific locations.
There is an expanding growth of device features and Web
Application Programming Interfaces (APIs) supporting the
development of applications with geolocation services on mobile
platforms. However, to be effective, these applications rely on the
availability of broadband networks which are not readily
available in various developing countries, especially Africa. We
propose a geolocation API for the Orange Emerginov Platform
which keeps geolocation data in an offline environment and
periodically synchronises with its online database. The API also
has a set of new features like categorisation and shortest path. It
has been successfully implemented and tested with geolocation
data of Mauritius, consumed by mobile applications. Our result
demonstrates reduction of response time around 80% for some
features, when compared with other online Web APIs.
Keywords—micro-services; geolocation; Web API; OSM
geolocation API; mobile applications
I. INTRODUCTION
Geolocation technology makes it possible for a computing
device to acquire its physical location. The methods for
acquiring geolocation data can be categorised as device-based
or server-based [1]. The former implies that the device uses
Global Positioning System (GPS) and/or base station
triangulation techniques to generate its location, while server-
based techniques rely on third party services using IP lookup,
or WIFI service set identifiers (SSID).
Once the location information is available, applications
can connect to geolocation databases which provides geo-
positioning, geo-coding and geo-tagging functionalities in the
form of Application Programming Interfaces (APIs). There are
a number of commercial and open source available
geolocation databases, which stores huge amount of
unstructured and semi-structured data, providing varying
levels of accuracy, such as Google MAP APIs [2] and Open
Street Map APIs [3].
The emergence of mobile devices supporting both device-
based and server-based location techniques, and the
availability of several geolocation databases contribute to
make geolocation services significantly popular across several
types of application. It has become the foundation for location
aware mobile applications. Both the private and public sector
rely on geolocation data, for instance, businesses can track,
learn and target consumers based on their behaviours.
Therefore, the ability to identify a customer’s location at any
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Morgan Richomme
Research and Development
Orange Labs
France
morgan.richomme@orange-ftgroup.com
time opens up seemingly endless sales, marketing and
business opportunities.
It is mandatory for most mobile applications to have an
Internet connectivity in order to give meaning to the
geolocation data, such as a positioning user coordinates on a
map, finding nearby objects or services on a map or even to
tag geographic information around objects. In developing
countries, the ITU estimates that around 67% of the
population remains offline, and in Africa specifically only
20.7% of individuals are using the Internet, while mobile
broadband subscriptions is below 20% per 100 inhabitants [4].
Even though connectivity remains a hurdle, the adoption of
mobile phones in sub-Saharan Africa witnessed a growth of
around 62% since 2008. Countries like Mayotte, Gabon,
Mauritius, Botswana and Reunion have more than 75% unique
mobile subscribers [5]. The rate of mobile application usage
have also gone up, and we reasonably foresee the rise in
geolocation services in these applications.
In order to bridge the connectivity gap and promote
development of geolocation applications, we are proposing a
local geolocation API for rapid application development on
the France Telecom Orange Lab's Emerginov platform [6]. It
is an open platform which provides a rich set of APIs to
support the development of micro-services and aims at
empowering local communities to design and deliver
multimedia (vocal, SMS, USSD) and user-generated contents.
Being the usual hotspots of innovation, local universities in
the African region are especially targeted. They contribute to
enrich the platform with features like the USSD enabler [7]
and the Mobile Payment API [8]. The aim of this work is to
further contribute with an offline geolocation API.
II. BACKGROUND
A. Opportunities for Local Geo-coders
Over the years, consumers have witnessed a shift from
traditional applications of geolocation services to more
innovative, powerful and useful forms following different
business models. Geolocation-based applications provide
services that range from simple physical location relative to a
map, such as navigation, to searching for services like nearby
restaurants, to tagging geographic information onto objects,
such as photographs. This application landscape is still under
expansion and experimentation by various service providers
who see competitive advantages. New breeds of applications
provide services like location tagging shared socially among
friends or travellers. Geo-fencing applications, used by parents
to keep track of children or even neighbourhood connections
that help to bring together local communities. Mobile money,
which has been widely adopted by different African
communities can also largely benefit from geolocation
capabilities.
The need to provide geolocation services is well present,
and thus, the open offline geolocation API can be a real factor
of economic growth if it can be easily implemented by
developers.
The local geolocation service ecosystem consists of
developers in the African region, with a relatively low Internet
connection rate and high usage cost. Developers can be
students or start-up companies who do not have a fix revenue
or no revenue at all. They can rapidly prototype Web
applications for devices ranging from low-end to high-end
smartphones to desktop computers.
B. Existing Geolocation Services
There exist various geolocation services which can be
categorised into online and offline services as well as
commercial and non-commercial. Some existing geo-coding
solutions are:
 Google Map APIs
Google Map is the most popular geo-coder. Its Web
Services are a collection of HTTP interfaces to Google
services providing geographic data for map based
applications. Non-commercial applications use the
Google Map API free of charge. However, as soon as
it starts generating a high volume of traffic, usage
limits apply and a commercial licence is require.
Applications generating 25 000 map loads or more
each day, for more than 90 consecutive days fall in the
commercial category [2].
 Yahoo Boss PlaceFinder
Yahoo provides another popular geolocation API, the
PalceFinder[9] which supports 50,000 requests per
day and no restrictions are imposed onto how the data
can be used.
 Nominatim
Nominatim [10] is an API protocol that query request
via GET queries to return a set of information that are
the result of a semantic query on the database.
Nominatim performs a series of queries based on
certain tags, which contain information about the
names or addresses, and then returns the data in XML
or JSON. There are no restrictions on the number of
queries per day and no hidden clauses in “usage
terms”. However since the service run on donated
servers, there is no quality of service guarantee.
 Xapi
Xapi [11] is a read-only API protocol that provides
searching and querying facilities. As most other APIs
it also uses HTTP GET requests in which the
application keys are inserted as parameters of the
query. It returns a package containing an HTTP OSM
file in XML. The API has some limitations, as each
request can contain only one tag predicate and one
box predicate. Also, the request grammar is not
formally defined and is not processed using a proper
parser, so malformed requests may have unpredictable
results.
 Overpass API
Overpass API [12] is an extension of XAPI with a
query language known as Overpass QL. Requests are
always made in the HTTP GET format and are also
compatible with the XML format used in XAPI. With
the Overpass QL, the query must be sent into the field
data of the GET request. Overpass API has one
limitation, as it does not offer the possibility to query
the history data of an OSM object and also the change
set-based criteria. Moreover, since the size of a query
is unknown at the start of the download, the API
cannot provide an indication of the remaining time of
download during the process.
 Parsing libraries
Parsing libraries [13] are considered as local methods
that involve the use of an offline mapping tool to
display downloaded geo-coding data to the requesting
devices. Examples of some offline libraries are:
Libosmscout [14] which offers applications a simple,
high-level interface to offline rendering and routing
functionalities based on OpenStreetMap data.
Libosmscout is available for both desktop and mobile
platforms and has been developed under the LGPL
licence [15]. CartoType is another example of an
offline library which provides mapping and routing
services [16]. Compared to Libosmcout, CartoType
provides the same characteristics but it is developed
under a commercial license.
III. PROPOSED SOLUTION
The E-OSM API takes the form of a core component in the
Emerginov platform, and is therefore available like other
features, such as the USSD API or Money API, following
installation of the platform. During installation, the E-OSM
API can be configured by specifying data for regions to be
downloaded from the Open Street Map database[17] and
stored locally. The API can also be installed as a feature on an
existing Emerginov platform. Once installed and configured,
the API can be invoked using the REST style based Web
service approach which provides more than 50 powerful
methods that can be easily integrated into different
applications. It provides responses in XML [18] and JSON
[19] and uses a local database to store unstructured
geolocation data. As can be depicted in the architecture from
Fig. 1, there are three specific layers, which give access to
developers and the API administrator.

Fig. 1. E-OSM API Architecture
A. Developer Services
An interface is provided to allow developers to register for
an API key and manage their accounts. The API key needs to
be used in each of their micro-services/applications. Requests
are typically encoded in the HTTP GET method, within the
URL in the following format:
URL/ClassName/Method/Parameters
B. API Services
The API follows a modular design, which allows quick
and easy extension of new features. It consists of the following
modules:
 Points – These are nodes on a map, each points are
represented by a latitude and longitude value.
 Linestrings – These are defined as ways or lines on a
map. For example, it can take the form of a road on a
map.
 Polygons – These are a set of points that are linked to
form a shape (any number of sides). For example, a
polygon can be a building on a map.
 Review – Allows user to associate reviews to points,
linestrings and polygons.
 Categorization – Allows user to associate category
names to points, linestrings and polygons. For
example, regrouping all schools entity (kindergarten,
college, university, etc.) under the School category
name.
 Administration – Allows the registration of new
developer accounts and other administrative and
operational features of the API. For example, a
developer registers for an account, thereafter
activated by an administrator
C. Local Data Services
OpenStreetMap geolocation data is large set in semi-
structured format, and is available in XML. It is converted to
JSON format, which allows faster querying. Third party tools,
notably Osmosis [20] and 3angle[21] performs the
conversion, following which the data is stored in MongoDB
[22] which is a JSON-style data store. Unlike standard SQL
databases, it does not enforce document structures. This
flexibility facilitates the mapping of documents to an entity or
an object. Each document can match the data fields of the
represented entity, even if the data has substantial variation.
The API administrator can choose from different options to
synchronise with the offline database, usually performed off-
peak hours. A pre-configured automated script in the form of a
cron job can be scheduled to execute the update at specific
intervals.
IV. TESTING
An all-in-one geolocation client tool, called MapPie has
been developed to test all functionalities available for client
applications. We describe a few test scenarios below.
Scenario 1: Finding shortest path
The function takes two location parameters, either a
standard point or a point on a Polygon, each with a pair of
coordinates and it uses the Dijkstra’s[23] algorithm to find the
shortest path and distance between the two points. The
algorithm picks the unvisited vertex (node) with the lowest-
distance, calculates the distance through the edges (route)
connected to the vertex to each unvisited neighbour, and
updates the neighbour's distance if smaller.
Fig. 2. Shortest Path Result
Scenario 2:Searching a Polygon Category
The function searchPolyCategory takes a category name
defined by the application user, such as a school ,restaurant or
public building, and looks up all polygons in this category.
The category name groups multiple OSM keywords under a
single name which allows custom searches.
Scenario 3: Finding nearby objects
A user can find all nearby clothing shops within a distance
of 1km from the current location. The function is named
searchPointBound, in which the user provides a location, the
keyword (e.g. clothing) and the range/distance in meters (e.g.
1000). The API processes the operation according to the
parameters and displays the shops that are only within the
range of 1km.
Scenario 4: Reverse Geo-coding
The function searchLineCoord, takes as input a point on a
path to search the complete path (e.g. A coordinate on a road
can be used to find the road’s name). The function also
provides additional information such as the total distance of
the path, and the average time to travel the path via motor [3]
vehicle, cycling or by foot.
[4]
Following the implementation of MapPie, we conducted a
set of simple experiments to compare the response time of
accessing offline data from E-OSM with that of an existing [5]
online API. A subset of the functionalities from MapPie
dealing with Points, Linestings and Polygons have been
[6]
implemented to use the online Nominatim API. As can be
depicted in Fig.3, both for Point and Polygon functionalities,
E-OSM performs better with reduction of around 88% and
86% respectively in response time. As expected, clients using
E-OSM API only requires a local connection to the database [7]
as opposed to accessing the geographically far OSM server
over Nominatim. There are less intermediary network nodes
involved, and therefore less delay. As for the Linestring
[8]
functionality, the request consists of finding the shortest path
between two points. E-OSM shows negative results as it
returns the shortest path, where as Nominatim returns the first
set of points (route) which can connect the two points. [9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
Figure 3 Response Time Comparison for ESOM &
Namonatim [19]
[20]
V. CONCLUSION & FUTURE WORKS [21]
We successfully implemented an API for the Emerginov
platform which stores geolocation data in an offline mode, and [22]
provides a Restful interface to be consumed by micro-services. [23]
Test results demonstrate that offline mode clearly solves the
data rate and delay issues experienced over the Internet, and [24]
can thus be adopted in developing countries. Our future work,
on the one hand deals with improving the performance of the
[25]
shortest path algorithm which is based on categorisation tags
we introduced. On the other hand, in order to handle a bigger
[26]
data set like geolocation data for Africa, we are investigating
the opportunities provided by Big Data solutions, such as
Hadoop's Distributed File System (HSFS) [25] and
MapReduce [26] programming model.
OpenstreetMap (OSM) API, available at: http://wiki.openstreetmap.org
/wiki/API [Aug 2015]
ITU, "ICT Facts and Figures- The World in 2015", International
Telecommunications Union, Geneva, May 2015. Available at:
http://www.itu.int/en/ITU-D/Statistics/Pages/facts/default.aspx
GSM Association, "The Mobile Economy 2015", GSMA, London, UK,
2015. Available at: http://www.gsmamobileeconomy.com/
GSMA_Global_Mobile_Economy_Report_2015.pdf
M. Richomme, D. Blaisonneau, B. Herard, B. Ngom and G.Suddul,
“Emerginov: An Open PHP PaaS to Stimulate and Animate Local Co-
innovation”, In Proc. Of 4th International ICST Conference,
AFRICOMM 2012, Yaounde, Cameroon, Nov 2012. Springer-Verlag,
Berlin Heidelberg, pp. 126-132.
G. Suddul, A. Soobul, U. Bahadoor, A. Ramdoyal, N. Doolhur and M.
Richome, “An Open USSD-Enabler to simplify access to mobile
services in emerging countries”, 4th International Conference on
Emerging Trends in Engineering and Technology(ICETET), 2011
N. Doolhur, G. Suddul, R. Foogooa and M. Richomme, “An Open API
to Monetize Mobile Micro-Services for Emerging Countries”, In Proc.
of the IEEE 11th International Africon 2013 Conference, Mauritius, SEP
2013. IEEE USA, pp. 684-687.
Yahoo Boss PlaceFinder API. Available at: https://developer.yahoo.com
/boss/placefinder/. [Sep 2015]
Nominatim Search Tool. Available at: http://wiki.openstreetmap.org
/wiki/Nominatim [Jan 2015]
XAPI, OSM Extended API. Available at: http://wiki.openstreetmap.org
/wiki/Xapi [Jan 2015]
Overpass API. Available at: http://wiki.openstreetmap.org
/wiki/Overpass_API [May 2015]
Data Processing or Parsing Libraries, https://wiki.openstreetmap.org
/wiki/Frameworks [May 2015]
Libosmscout. Availabe at:https://wiki.openstreetmap.org/wiki/
Libosmscout [May 2015]
GNU Lesser General Public License, http://www.gnu.org/licenses/lgpl-
3.0.en.html [Jun 2015]
Cartotype. Available at:https://wiki.openstreetmap.org/wiki/CartoType
[APR 2015]
Open Street Map Database. Available at: http://wiki.openstreetmap.org/
wiki/Database [Aug 2015]
W3C XML Speficiation (version 1.0), 2008. Available at:
http://www.w3.org/TR/2008/REC-xml-20081126/ [Aug 2015]
JSON Specification. Available at: http://json.org/ [Sep 2015]
Osmosis Tool. Available at: www.wiki.openstreetmap.org/wiki/Osmosis
[Jun 2015]
Derick Rethans, 3angle, OSM to MongoDB, Available at:
http://derickrethans.nl/talks/osm-mongouk12.pdf [Mar 2015]
MongoDB. Available at: https://www.mongodb.org/ [Jun 2015]
Dijkstra, E.W. "A Note on Two Problems in Connection with Graphs.",
Numerische Math.1, 269-271, 1959.
A. Mannara,"Location-based games and the use of GIS information",
Master's Thesis, Norwegian University of Science and Technology,
2012.
Hadoop HDFS. Availble at: http://hadoop.apache.org/docs/
r1.2.1/hdfs_design.html [Sep 2015]
J. Dean and S. Ghemawat, " MapReduce: Simplified Data Processing on
Large Clusters ", In proceedings of the 6th International conference on
symposium on operating system design and implementation,
USENIX,2004.

REFERENCES
[1] ISACA, "Geolocation: Risk, Issues and Strategies",white paper, ISACA
USA, 2011
[2] Google Map APIs, available at: https://developers.google.com
/maps/?hl=en [Sep 2015]
 Role of Attributes Selection in Classification of
Chronic Kidney Disease Patients
Naganna Chetty Kunwar Singh Vaisla Sithu D Sudarsan
Research Scholar at UTU, Dehradun Dept. of Computer Science and Engg., ABB Corporate Research, India
and BTK Institute of Technology Dwarahat, sdsudarsan@gmail.com
Dept. of CSE, MITE, Mangalore, India Uttarakhand, India
nsc.chetty@gmail.com vaislaks@rediffmail.com

Abstract—In the present days the Chronic Kidney Disease
(CKD) is a common problem to the public. CKD is generally
considered as kidney damage and is usually measured with the
GFR (Glomerular Filtration Rate). Several researchers from
health care and academicians are working on the CKD problem
to have an efficient model to predict and classify the CKD patient
in the initial stage of CKD, so that the necessary treatment can be chronic diseases (such as congestive heart failure, and chronic
provided to prevent or cure CKD. In this work classification
models have been built with different classification algorithms,
Wrappersubset attribute evaluator and bestfirst search method to
predict and classify the CKD and non CKD patients. These
models have applied on recently collected CKD dataset
downloaded from the UCI repository. The models have shown
better performance in classifying CKD and non CKD cases.
Results of different models are compared. From the comparison
978-1-4673-9354-6/15/$31.00 focused
it has been observed that classifiers performed better on reduced
dataset than the original dataset.
Keywords—Data mining; classification; prediction; chronic
kidney disease; attributes reduction
I. INTRODUCTION
The real world data gets doubles every 20 months and
contain some amount of noise in it. Hence there is a need to
store, manage and process this data efficiently. With the rapid
advances in storage devices, it is easier to store and manage
the vast amount data. Even though continuous effort has been
made, the efficient processing of huge amount of data is still a
challenge to the researchers, academicians, etc. This challenge
can be handled with the data mining techniques. Data mining
is an essential activity in KDD (Knowledge Discovery in
Databases) process, which extracts patterns from observed
data.
Health Informatics is producing vast amount of data and
processing of this generated huge amount of data creates more
possibilities for knowledge to be gained. This gained
information can improve the service quality of healthcare to
patients. The number of issues arise when dealing with this
vast amount of data, one among them is how to analyze this
data in a reliable manner. The basic goal of Health Informatics
is to use real world medical data to improve our understanding
of medicine and medical practice [1].
The present work emphasizes on an application of data
mining, in particular, the classification techniques in health
informatics to detect Chronic Kidney Disease (CKD).
CKD is usually a silent condition. Signs and symptoms, if
present, are generally nonspecific and unlike several other
obstructive lung disease), they do not reveal a clue for
diagnosis or severity of the condition. Typical symptoms and
signs of uremia appear almost never in early stages (Stage 1 to
3, and even in Stage 4) and develop too late only in some
patients in the course of CKD. Still, all newly diagnosed CKD
patients, patients with an acute worsening in their kidney
function and CKD patients on regular follow-up should have a
©2015history
IEEEand physical examination. This will be the key
to perceive real implications of health associated with
decreased kidney function in CKD [2].
CKD is defined as damage to kidney or Glomerular
Filtration Rate (GFR) < 60 mL/min/1.73 m2 for 3 months or
more, irrespective of the cause. Kidney damage in kidney
related diseases can be caused by the presence of albuminuria,
defined as albumin to creatinine ratio >30 mg/g in two of three
spot urine specimens. GFR can be estimated from calibrated
serum creatinine and estimating equations, such as the
Modification of Diet in Renal Disease (MDRD) study
equation or the Cockcroft-Gault formula [3].
GFR is traditionally measured as renal clearance of an
ideal filtration marker, such as inulin from plasma. This
measured GFR is considered the gold standard but is not
practical for daily clinical use due to complexity of the
measurement procedure. Estimating GFR based on a filtration
marker (usually serum creatinine) is now widely accepted as
an initial test. Several GFR prediction equations that use
serum creatinine or some other filtration markers along with
certain patient characteristics (like age, gender, and race) are
giving precise estimates of GFR in various clinical settings
[4].
Different stages and action plan for CKD are shown in
TABLE 1. Here CKD stages from 1-5 along with the GFR
reading and actions required during each stage are described.

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

 TABLE 1 STAGES AND ACTION PLAN FOR CKD [5]
Stage Description GFR
(mL/min/1.73 m2)
- At increased risk for CKD >=90 with risk
factors
1 Diagnosis and treatment
Kidney damage with
normal or increased GFR >=90
2 Mild decrease in GFR 60-89
3 Moderate decrease in GFR 30-59
4 Severe decrease in GFR 15-29
5 Kidney failure < 15 or dialysis
II. LITERATURE SURVEY
Kusiak et al. [6] developed a method to predict survival
period for kidney patients with dialysis using data mining
approach. To obtain information on the interaction between
measured values and survival of patients, they used data
preprocessing, data transformations, and a data mining. Two
different data mining algorithms were employed for extracting
knowledge in the form of decision rules. These rules are
provided as input to a decision-making algorithm, which
predicts survival of new unseen patients with an accuracy of
75% to 85%. The approaches introduced in their research
have been applied and tested on data collected at four dialysis
sites.
Haubitz et al. [7] collected urine samples of 45 patients
and tested for protein/polypeptide patterns with a novel high
throughput method, capillary electrophoreses online coupled
to a mass spectrometer (CE-MS). CE-MS allows the fast and
accurate evaluation of up to 2000 polypeptides in one urine
sample. Then they compared igAN (igA Nephropathy) results
with findings in 13 patients with membranous nephropathy
(MN) and 57 healthy individuals.
Yeh et al. [8], integrated temporal abstraction with data
mining techniques to analyze dialysis patient’s biochemical
data to develop a decision support system with the purpose of
reducing hospitalization rate. These mined temporal patterns
are used by clinicians to predict hospitalization of
hemodialysis patients and to suggest immediate treatments to
avoid future hospitalization.
A first nationwide survey of predialysis CKD in Asian
children carried out by Ishikura and his team indicated that the
prevalence of CKD stage 3 and stage 5 in children in Japan is
2.98 for the age group 3 months to 15 years cases/100000
children. Most children with CKD presented themselves with
non-glomerular disease, most frequently CAKUT. They stated
improved management of CAKUT, including renoprotective
treatment and urological intervention, is needed [9].
Malluche et al. [10] demonstrated the value of assessing
bone from multiple views and hierarchical levels to
Action
Screening
CKD risk reduction
Slow progression of CKD
Treat comorbidities
Cardiovascular disease risk reduction
Estimate progression
Evaluate and treat complications
Prepare for renal replacement therapy
Replacement if uremic
understand CKD–MBD-related abnormalities in bone quality.
Knowing the relationships between variations in material,
structure, microdamage, and mechanical properties of bone in
patients with CKD–MBD should aid in the development of
new modalities to prevent, or treat, these abnormalities.
Chase et al. [11] developed prediction models with Naïve
Bayes and Logistics Regression classification algorithms to
predict the progression of stage 3 CKD. They used data
extracted from Electronic Health Record (EHR) to identify
two cohorts of patients in stage 3: progressors (eGFR declined
>3 ml/min/1.73m2/year; n = 117) and non-progressors (eGFR
declined <1 ml/min/ 1.73m2; n = 364). Initial laboratory
values recorded a year before to a year after the time of entry
to stage 3, reflecting metabolic complications (hemoglobin,
bicarbonate, calcium, phosphorous, and albumin) were
obtained. They made the comparison of average values in
progressors and non-progressors.
A hybrid decision support system is developed by Neves
and his team, allowed to consider incomplete, unknown, and
even contradictory information. This is complemented with an
approach to computing centered on Artificial Neural
Networks, in order to weigh the Degree-of-Confidence in
terms of reasoning procedures and knowledge representation
based on Logic Programming,. Their study involved 558
patients with an age average of 51.7 years and the chronic
kidney disease was observed in 175 cases. The dataset consist
twenty four variables, forming five main categories [12].
Vanlede et al. [13] examined the urinary excretion of
monosaccharides and polyols in children with different CKD
degrees, but without known metabolic or renal tubular
disorders. Urinary concentrations of 18 monosaccharides and
polyols were measured by gas chromatography–mass
spectrometry (GC-MS) in random urinary samples of 25
patients with CKD stage1– stage 5. A comparison of these
values are made with age-related reference values. Serum
creatinine was measured at the time of the urine sample, and
the height-independent estimated glomerular filtration rate
(eGFR-Pottel) was calculated. Urinary excretions of
monosaccharides and polyols were above the reference values
in 8–88 % of all patients. A significant difference between
CKD stage 1– stage 2 compared with CKD stage 3–stage 5
was found for allose, arabitol and sorbitol (p<0.05) and for
arabinose, fucose, myoinositol, ribitol, xylitol, and xylose
(p<0.01). They concluded that the excretion of polyols and
sugars depends on eGFR, which warrants a cautious
interpretation of the results in patients with CKD.
Dubey et al. [14] adopted the K-Means Clustering
Algorithm with a single mean vector of centroids, to classify
and make clusters of varying probability of likeliness of
suspect being prone to CKD. They observed and stated that
the suspects falling in clusters K1 or K3 are surely suffering
from CKD. The probability of a suspect lying in K2 cluster to
fall in the class of CKD is 0.50545, which implies that the
suspect cannot be classified by their L-factor classifier.
However, suspects from clusters K1 & K3 were found to be
falling in CKD class with full probability.
III. PROPOSED METHODOLOGY
In this section we propose a methodology for mining
patters from the CDK dataset. Fig. 1 shows the proposed
framework for mining patterns.
The proposed methodology consists of:
 A framework for pattern mining
 A methodology for attributes selection
 Applying classification method to original and
reduced datasets.
Attribute selection is a process of reducing the dimension
of a dataset by eliminating the attributes of less importance. A
framework for attribute selection is shown in Fig. 2. We have
used wrappersubseteval attribute evaluator with bestfirst
search method to select features for Naïve Bayes, SMO
(Sequential Minimal Optimization) and IBK classifiers using
WEKA data mining tool.
Naïve Bayes, SMO and IBk classifiers are used to classify
original and reduced CKD datasets. Naïve Bayes classifiers are
simple classifiers with probability based on Baye’s theorem.
This possess strong (naive) independence assumptions between
the features. SMO is an algorithm which solve problem
of quadratic programming that occurs at the time of training
support vector machines. IBk is one of the instance-based
learner using class of the nearest k training instances for the
class of the test instances. Parameter k is the number of nearest
neighbors to use for prediction.

Selected
Attributes Data Mining Patterns Knowledge
Attributes
Dataset Selection (Classification) Discovery

Fig. 1. Proposed Framework for Mining Patterns

Data with original attributes

Apply attribute selection methods

WrapperSusetEval attribute WrapperSusetEval attribute WrapperSusetEval attribute

evaluator with Naïve bayes evaluator with IBK classifier and evaluator with SMO classifier
classifier and bestfirst search bestfirst search and bestfirst search

Data with selected attributes Data with selected attributes for Data with selected attributes
for Naïve bayes classifier IBK classifier for SMO classifier

Fig. 2. Framework for Attributes Selection


IV. RESULTS
The dataset used in the present work is CKD,
downloaded from the UCI repository [15]. The CKD dataset
consists of 400 instances and 25 attributes viz. age, blood
pressure, sugar, etc., including class label. TABLE 2 shows
result of attributes reduction by the attribute evaluator with
different classifiers. WrapperSusetEval attribute evaluator
with Naïve bayes classifier and bestfirst search selects only
6 attributes from 25 total attributes with 76% of attributes
reduction. WrapperSusetEval attribute evaluator with SMO
classifier and bestfirst search results in 12 attributes from 25
total attributes with 52% of attributes reduction.
WrapperSusetEval attribute evaluator with IBK classifier
and bestfirst search selects only 7 attributes from 25 total
attributes with 72% of attributes reduction. The graphical
representation for the result of attributes selection is shown
in Fig. 3.
After reducing the attributes, different classifier models
have been applied on both original and reduced datasets.
The result of classification on the original dataset is shown
in the TABLE 3. The Naïve Bayes classifier classified 380
instances correctly and 20 instances incorrectly from total
400 instances of CKD dataset with the classification
accuracy of 95%. SMO classifier classified 391 instances
correctly and only 9 instances incorrectly from total 400
instances of CKD dataset with the classification accuracy of
97.75%. The IBK classifier classified 383 instances
correctly and 17 instances incorrectly from total 400
instances of CKD dataset with the classification accuracy of
95.75%. Graphical representation of classification
accuracies of different classifiers is shown in Fig. 4.
TABLE 4 shows the result of classification on reduced
dataset. Naïve Bayes classifier classified 396 instances
correctly and only 4 instances incorrectly from total 400
instances of CKD dataset with the classification accuracy of
99%. SMO classifier classified 393 instances correctly and
only 7 instances incorrectly from total 400 instances of
CKD dataset with the classification accuracy of 98.25%.
The IBK classifier classified all the 400 instances correctly,
this lead to the classification accuracy of 100%. Graphical
representation of classification accuracies of different
classifiers on reduced dataset is shown in Fig. 5.

TABLE 2 TABULAR REPRESENTATION OF ATTRIBUTES REDUCTION

Initial Selected Attributes
Attribute Evaluator
Attributes Attributes Reduction
WrapperSusetEval attribute evaluator with
25 06 76%
Naïve bayes classifier and bestfirst search
WrapperSusetEval attribute evaluator with
25 12 52%
SMO classifier and bestfirst search
WrapperSusetEval attribute evaluator with
25 07 72%
IBK classifier and bestfirst search

80
70
60 Attributes Selector with
50 Naïve Bayes
40 Attributes Selector with
30 SMO
20 Attributes Selector with IBK
10
0
Attributes Selected Attrbutes Reduction (%)

Fig. 3. Result of Attributes Reduction

TABLE 3 RESULT OF CLASSIFICATION ON ORIGINAL DATASET

Correctly Incorrectly
Classification
Classifier classified classified
accuracy
instances instances
Naïve Bayes 380 20 95%

SMO 391 9 97.75%

IBK 383 17 95.75%

 450
400
350
300
250
NB
200
SMO
150
100 IBK
50
0
Correctly classified Incorrectly classified Classification accuracy
instances instances

Fig. 4. Result of Classification on Original Dataset

TABLE 4 RESULT OF CLASSIFICATION ON REDUCED DATASET

Classifier Correctly Incorrectly Classification
classified classified accuracy
instances instances
Naïve Bayes 396 4 99%
SMO 393 7 98.25%
IBK 400 0 100%

450
400
350
300
250
NB
200
150 SMO
100 IBK
50
0
Correctly classified Incorrectly classified Classification accuracy
instances instances

Fig. 5. Result of Classification on Reduced Dataset

V. CONCLUSION better classification accuracy on the reduced dataset than the

Attributes evaluator and classification models have been
applied on CKD dataset, uploaded very recently to UCI
repository. From TABLE 2 it can be observed that attribute
evaluator model has performed well by reducing the
attributes from 25 to 6, 12 and 7 with NB, SMO and IBK
classifiers respectively. TABLE 3 shows the classification
accuracy for CKD original dataset as 95%, 97.75% and
95.75% for NB, SMO and IBK classifiers respectively. Here
SMO classifier performs better than other two. Table 4
shows the classification accuracy for CKD reduced dataset
as 99%, 98.25% and 100% for NB, SMO and IBK
classifiers respectively. Here IBK classifier performs better
than other two. Finally by referring to the results of Table 3
and Table 4, we can conclude that our models resulted in
original dataset.
REFERENCES
[1] M. Herland, T. M. Khoshgoftaar and R. Wald, “A review of data
mining using big data in health informatics,” Journal of Big Data, pp.
1-35, 2014.
[2] M. Arici, “Clinical Assessment of a Patient with Chronic Kidney
Disease,” Management of Chronic Kidney Disease, Springer-Verlag
Berlin Heidelberg, pp. 15-28, 2014.
[3] A. S. Levey, K. Eckardt, U. Tsukamoto, A. Levin, J. Koresh, J.
Rossert, D. D. Zeeuw, T. H. Hostetter, N. Lameire and G. Eknoyan,
“Definition and classification of chronic kidney disease: A position
statement from Kidney Disease: Improving Global Outcomes
(KDIGO),” Kidney International, Vol. 67, pp. 2089-2100, 2005.
[4] L. A. Stevens, J. Coresh, T. Greene, A. S. Levey, “Assessing kidney
function-measured and estimated glomerular filtration rate,” N Engl J
Med, 354(23), pp. 2473-83, 2006.
[5] M. A. Parazilla and R. F. Reilly, “Chronic Kidney Disease: A New
Classification and Staging System,” Hospital Physician, pp. 18-22, 45,
March 2003.
[6] A. Kusiak, B. Dixon, S. Shah, “Predicting survival time for kidney
dialysis patients: a data mining approach,” Computers in Biology and
Medicine 35, pp. 311–327, 2005.
[7] Haubitz et al., “Urine protein patterns can serve as diagnostic tools in
patients with igA nephropathy,” Kidney International, Vol.67, pp.
2313-2320, 2005.
[8] J. Y. Yeh, T. H. Wu, C. W. Tsao, “Using data mining techniques to
predict hospitalization of hemodialysis patients,” Decision Support
Systems, 50, pp. 439–448. 2011.
[9] Ishikura et al., “Pre-dialysis chronic kidney disease in children: results
of a nationwide survey in Japan,” Nephrol Dial Transplant, pp.1-11,
2013.
[10] H. H. Malluche, D. S. Porter, and D. Pienkowski, “Evaluating bone
quality in patients with chronic kidney disease,” Nat Rev Nephrol,
9(11), pp. 671-680, 2013.
[11] H. S. Chase, J. S. Hirsch, S. Mohan, M. K. Rao and J. Radhakrishnan,
“Presence of early CKD-related metabolic complications predict
progression of stage 3 CKD: a case-controlled study,” BMC
Nephrology, 15:187, 2014.
[12] J. Neves, M. R. Martins, J. Vilhena, J. Neves, S. Gomes, A. Abelha,
J. Machado, H. Vicente, “A Soft Computing Approach to Kidney
Diseases Evaluation,” J Med Syst, Springer, 39: 131, 2015.
[13] K. Vanlede, L. A. J. Kluijtmans, L. Monnens, E. Levtchenko, “Urinary
excretion of polyols and sugars in children with chronic kidney
disease,” Pediatr Nephrol, 30, pp. 1537–1540, 2015.
[14] A. Dubey, “A Classification of CKD Cases Using MultiVariate K-
Means Clustering,” International Journal of Scientific and Research
Publications, Vol. 5, Issue 8, 2015.
[15] P. Soundarapandian and L. J. Rubini,
http://archive.ics.uci.edu/ml/datasets/Chronic_Kidney_Disease, UCI
Machine Learning Repository, Irvine, 2015.
 Do you know Big Brother is watching you on
Facebook?
A study of the level of awareness of privacy and security issues among a selected sample of
Facebook users in Mauritius
Mrinal Sohoraye
Academic Affairs Division
Open University of Mauritius
Mauritius
Suniti Nundoo-Ghoorah
Academic Affairs Division
Open University of Mauritius
Mauritius
Abstract—Facebook has become the most powerful online
social network in recent years. The billions of users appreciate
the ease of use and the various applications of the network. It is
so pervasive in people’s personal life that it is difficult to imagine
life without Facebook. However, there have been a number of
concerns with regards to privacy and security. The present study
investigates the current use of Facebook in Mauritius and
assesses the level of awareness of privacy and security among a
selected sample of users. Data shows that literature is scarce in
Mauritius though the island aspires to be the Cyber Island of the
Indian Ocean. Thus, a large-scale study on the use of Facebook in
Mauritius is required. A comprehensive questionnaire was
designed to gauge the perception and to investigate privacy and
security concerns. Results demonstrate that users in Mauritius
blindly trust the network but are relatively concerned about the
usage of their personal information. It is also a matter of serious
concern that users have limited knowledge of the provisions of
the Computer Misuse and Cybercrime Act (2003) which is one of
the harshest in the world when it comes to misuse of computers.
The paper proposes a number of recommendations which need to
be applied urgently by online social networks and other
stakeholders and the need for an explicit informed consent from
users each time personal data is used.
Keywords—Facebook; privacy; law; security; extimacy;
awareness; Mauritius
I. INTRODUCTION
Facebook, the technological revolution that has invaded
the world and become an intimate part of the private lives of
millions of users, is increasingly „commodifying‟ the personal
into the public sphere [1]. Around the world, Facebook has
positioned itself as a trove of data from a wide spectrum of
users, at least 1.49 billion of whom are active each month [2].
Researchers state that „the revenue generated by Facebook
alone is approximately 3.91 million dollars‟ [3]. Facebook is
regarded as one of the wide-reaching and sharing information
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
1
Vandanah Gooria
Academic Affairs Division
Open University of Mauritius
Mauritius
Premanand Koonjal
Open School Division
Open University of Mauritius
Mauritius
platforms. At present, there are more than 370,000 Facebook
users in Mauritius and the 18-34 age group is the most active
making up 63% of total users [4]. Unfortunately, not much has
been published in the field of online social networks. The
literature is even scarcer for developing countries like
Mauritius which have embraced technological innovations
without any second thoughts. What is the current scenario in
Mauritius? Is Facebook private and secure? Very often, before
feeding some sensitive and personal information to the social
networks, users hardly give a deep thought about its
implications. This is precisely the reasons such a question
rather shakes an average trivial active user who pays no
attention to privacy and security issues on the online network.
In fact, there is no survey done in Mauritius where the
personal information dispensed through Facebook has been
assessed in terms of awareness of privacy and security.
AIMS AND OBJECTIVES OF PRESENT STUDY
The goal of this research is to contribute to the safety and
security of personal information on Facebook and to raise the
consciousness on distributing sensitive data which might
affect people‟s lives. In this work, we examine the currently
existing privacy and security policies of Facebook and look at
the users‟ awareness vis a vis privacy practices and their
knowledge of the existing laws of the country. The objectives
can be described as follows:
 To assess current awareness level of current terms of
use of Facebook
 To evaluate user privacy concerns with regards to
Facebook
 To gauge respondents‟ knowledge of the provisions
of the relevant legislations governing use of ICT in
Mauritius
 II. LITERATURE REVIEW
A. The ‘New World Order’ of Facebook
In his book „The Facebook Effect‟, Kirkpatrick [5] highlights
the founder of Facebook, Mark Zuckerberg‟s philosophy of
„publicness‟. Marichal [1] aptly describes the design of the
platform as „architecture of disclosure‟, whereby the human
instinct of „sharing‟ is networked as an act of „public
revelation‟. This „engineering of connectivity‟ [6] results into
generalised visibility of private life, often without the
conscious awareness of users. Facebookers, then, are very
vulnerable indeed in the hands of hackers, lurkers and other
unscrupulous forces. Tello [7] argues that certain Facebook
settings infringe on the right of people to private scrutiny and
the protection of personal data. She aptly refers to the erosion
of the boundaries of privacy as „extimacy‟ – the phenomenon
of exposed intimacy. Searches, likes, browses, links, and
other acts of virtual participation can be tracked by Big
Brother avatars such as (i) individuals obssessed with feeding
on the private lives of others (ii) profit-oriented companies,
and (iii) Governments collecting personal data in their own
servers. Ominously quoting Scott McNealy, the co-founder
of Sun Microsystems, „ Privacy has died. Get used to‟, Tello
[7] also calls for the review of certain features of Facebook‟s
design which limit users‟ ability to control the accessibility of
personal information.
B. Big Brother’s Eyes: Privacy and Security issues
According to Zhang & Sun [8], revealing personal information
in any online social networks (OSNs) is described as „double-
edged sword‟. Conflictingly, publicizing information, in
general, is a must as individuals desire to take part in social
communities and on the contrary, individual‟s identity may
muster malicious attacks like „stalking, personalized
spamming and phishing‟ from the virtual community. Shehab,
Squicciarini, Ahn & Kokkinou [9] are of opinion that the
growth of security and uncertainty owing to the mushrooming
of network sharing services via internet, booming amount of
messages, and the fast progress in ICT has induced the
protection of privacy to stand as one of the most arguable,
worrisome and daunting challenge. Leitch & Warren [10]
provide considerable empirical evidence to support their
argument on the reality of security and threats related to
Facebook and further stated that there are similarities of
security risks and threats between the common internet
community and Facebook. In various cases, it appeared that
these risks are significant and have commanding effect on
Facebook, for instance people who believe their Facebook
companion signifies that the possible effect for identify theft is
even more absolute. Shehab, Squicciarini, Ahn & Kokkinou
[9] pointed that the increase use of social networks like
Facebook, Twitter and others has caused many challenging
security and privacy difficulties. The increasing personal
information in Facebook marks up privacy matters and
necessitates deep understanding and awareness of security
problems. Kayes and Iamnitchi [11] propose an elaborate
2
taxonomy of privacy threats which they categorise into attacks
directed at the (i) user, and (ii) online social networking
platform. These range from identity thefts, fraud, virus spread,
„malware‟, phishing, to „de-anonymization‟ of user identity.
Albanesius [12] cites a Wall Street Journal report indicating
that many applications share personal data with advertising
agencies and other internet tracking companies. According to
Wall Street Journal this breach of privacy has impacted
millions of users. Companies are „mining trends‟ through
likes, and data surveillance has the power to swing election
results. In fact, many risks are involved that might rope in
users in complex situation about the protection of information
on Facebook and the freedom of people‟s privacy. Kulcu &
Henkoglu [13] explains that the cognizance and control of
digital information by the content owner has a decisive
responsibility in the assurance of personal data. It is
unbelievable and even worth considering how information
distributed around the world through Facebook, are used
momentarily and replicated in numbers. Furthermore,
Facebook and even other social networking sites keep on
polishing up their advertising strategies based on the personal
data obtained on a regular basis. Facebook follow the tracks
and sort ads accordingly with the personal interest of users. It
is crucial to secure private information through making legal
arrangement concerning the way Facebook might use personal
data and also without user‟s confirmation, the diffusion and
use of information be allowed. Nonetheless, the onus falls on
the users to protect their privacy since personal information is
imparted on Facebook by themselves. Private information is
being displayed and provided on Facebook by the user‟s own
will. The exposure of personal information was limited by
Facebook in 2005, yet everyone accessed these data in 2010.
In addition, users had the choice of keeping within bounds the
visibility of their intimate information [14]. From one point of
view, this adjustment in Facebook default settings empowered
everyone to watch the data in the user profile, on the contrary,
users had the chance to change the entire profile background
to defend their privacy. Anyhow, the changes related to the
services agreement have developed awkward conditions for
the users. Although users are being informed by mail about the
sporadic visit of their Facebook accounts, they pay no heed to
their account information, which may result an increase of
possible risks. Another concern is the complication to gain
access and apply the privacy settings which have to be
provided to the users owing to legal obligations, although the
backdrop differ from the basic sharing principle of Facebook.
Those who are not in the picture to protect their personal data
and those who have insufficient information about the settings
to protect their privacy on Facebook, is really impartial about
what confidential information is actually available. Based on
the information acquired through Facebook, anyone can get in
touch with an individual‟s private data [15]. Facebook may act
correctly and control our social lives but it can also damage
our romantic relationships. In fact, Facebook does gather data
on a large-scale about its users in spite of copyright issues.
Moreover, Facebook is aware when one look at another
person‟s timeline, communicate a message and even know the
time, date and place where your picture is taken. For example
if you log on or post anything from a mobile phone, it can spot
where you are or else if you access the site from the computer
it immediately tracks record through the IP address you are
using [16].
C. Legal framework for using Facebook in Mauritius
As an online social networking service Facebook has been
the subject of much debate especially with regard to the right
to privacy and freedom of expressions. Article 22 of the Civil
Code of Mauritius lays the foundation of the right to privacy
[17]. It reads; ‘Chacun a droit au respect de la vie privée.’ In
other words, the constitution of Mauritius jealously protects
and guarantees the right to privacy. The parent Act which is
the Information and Communication Technologies Act 2001
in Mauritius has been enacted with the objective of making the
dissemination of information easy and at the same time
guaranteeing the citizens the necessary protection of the
inalienable right to privacy. The question that has perplexed
many people here and elsewhere is how far we have
compromised our right to privacy under the garb of freedom
of expression and right to information through the medium of
Facebook. It is pertinent to note that by just clicking to the
terms and conditions appended to the use of Facebook we are
voluntarily or inadvertently relinquishing our right to privacy.
In this context we must remind ourselves that recourse to the
authorities for redress would become difficult if not
impossible legal battle. Thus, „any person who knowingly
sends, transmits or causes to be transmitted a false or
fraudulent message; uses an information and communication
service, including telecommunication service, for the
transmission or reception of a message which is grossly
offensive, or of an indecent, obscene or menacing character; or
for the purpose of causing annoyance, inconvenience or
needless anxiety to any person; or for the transmission of a
message which is of a nature likely to endanger or
compromise State defence, public safety or public order; will,
if found guilty, will be required to pay a fine not exceeding
1,000,000 rupees and to imprisonment for a term not
exceeding 5 years‟ in Section 46 of the Computer Misuse and
Cybercrime Act. Citizens need to be educated that Facebook,
being an international social network does not or is not in a
position to check the veracity of the information shared among
the users. The responsibility to do so resides on the user of the
network and the consequences of publicizing the information
through the medium remains exclusively on the user.
Moreover the jurisdiction where a case may be lodged against
Facebook is California which may prove difficult to some to
achieve justice. However, in March 2015 the French court
ruled that a litigant can take recourse to French jurisdiction in
the event of a dispute arising between Facebook and its user in
France. This is a matter of jurisprudential importance but as
regard Mauritius it is still to be tested how far a Mauritian
citizen can take a matter to Mauritian court against Facebook.
D. Existing Research and limitations
The usual objectives of most research previously was to gauge
the existing level of privacy and user awareness and to reveal
the dangers and risks that social network users encounter. In
3
spite of the extensive and alarming privacy and security clause
on the online social networks user agreement, users are
continually sharing, liking, entertaining, connecting blindly all
sorts of personal information with enthusiasm and curiosity
and as a result blunting the criticism made in past research
papers against online social networks with a smile. Past
research showed that social network users are often not aware
of the bulk data that is actually being divulged as privacy
threat is often hidden or unclear to them. Complaints from
Facebook users were related to failing with the compliance of
privacy matters and practices. The improper use of both
private information on social networking sites and the
viewable content by other users have become the key reason
for the need to be responsive about the use of social
networking sites and attributing importance to privacy.
Various studies emphasized on the important position of users
on social networking sites. In Mauritius, there is no study
analysing user profiles which puts forward the sensitivity and
awareness level of the users on the protection of personal data
and privacy. One study in Turkey showed the shortfall in the
legal arrangements but the online users receive an advantage
from this study apparently to act more cautiously against the
risks in the general use of social networks [13]. Since the
existence of social platform, the protection of user‟s privacy
on Facebook has been an element of discussion, but privacy
policies change each year on these platforms. [13]. However,
in the past studies, a decrease was noticed in the number of
personal data available to everyone but an increase also was
noted in the amount of sharing with the users in the friend list.
It becomes easy to access the personal data of prime
importance by Facebook users unless the privacy settings are
changed. Anyone must contemplate the restrictions on the
default settings which normally allow other users to show the
Facebook profile [13]. To what extent are privacy settings
used, and how aware are Facebook users of the various
privacy and security threats? It is commonly assumed that
users could not be bothered about using privacy controls.
Albanesius [18] reports a projection that 13 million users do
not apply or are not aware of privacy controls. She cites a
survey showing that 28% of respondents open their profiles to
others beyond friends. 4.8% reveal their whereabouts on
Facebook. Do they any knowledge of the dangers lurking on
the net? Interestingly, Grimmelman [19] argues that people are
generally not cognisant of the consequences of revealing
personal information on Facebook. He posits that users
„massively misunderstand Facebook‟s privacy architecture.‟
Grimmelman [19] debunks „myths‟ about users‟ attitudes
towards the sharing of personal information, asserting that
‟social networking sites activate the subconscious cues that
make users think that they are interacting within bounded,
closed, „private spaces‟.
III. METHODOLOGICAL APPROACH
The current study adopts a quantitative approach. This
research is based on a two-pronged outlook on:
1. Terms of use of Facebook
2. The Computer Crime and Cybercrime Act (2003) of
Mauritius
The survey method has been chosen to gather data from a
wide audience. This method is recommended in several The majority of users were employees of either the private or
studies on online social networks as it allows collection of a public sector making up 66% of the total.
large amount of data [20]. The sampling technique used is a In general, people make use of the network on a quasi-daily
probability method namely random sampling. A questionnaire basis as shown below:
was designed and disseminated online and especially on the
Facebook walls of the authors of the study and disseminated to
random users in Mauritius. The first part of the questionnaire
assessed the level of awareness of the contents of the „terms of
use‟ of Facebook. In the second part, the knowledge of the
laws of Mauritius pertaining to use of ICT was assessed. The
„friends‟ of the respective authors were encouraged to
participate in the study and to disseminate to a maximum
number of people online. A number of printed questionnaires
were distributed randomly to Facebook users of Mauritius
from a wide-ranging background. According to statistical
tables, a population of 370, 000 yields 384 responses at a 95%
confidence interval. About 620 responses were obtained, out
of which 600 were analysed after cleaning and coding of data.
Data was analysed using SPSS for descriptive and inferential
outputs. An open-ended question was also included and this
was analysed qualitatively in the form of a word cloud.
IV. FINDINGS AND INTERPRETATION

A. Demographic Information
Fig 3. Usage of Facebook
The following shows the demographic distribution of the
sample.
B. General interpretation
Results show that users had serious concerns with regards to
the use of private data but were ignorant of the type of
information being collected. Thus, there is no correlation (r:
0.046) about knowledge of terms of use of Facebook and
amount of information being collected by the network. This
clearly shows that though the terms of use are public
knowledge, no one reads about them and thus accounts for
their ignorance about the fact that Facebook collects all data
pertaining to their contents, chats and other activities.

Correlation Which amount

of information
Fig 1. Gender Distribution does Facebook
collect?
It can be observed that out of 600 participants, females Pearson .046
Have you read
accounted for more two-thirds of the sample. Correlation
Facebook terms
Sig. (2- .269
of use?
The occupations were as follows: tailed)
Occupation Percent
Prefer not to say 4.5 Fig 4. Correlation between terms of use and information
Student 19.3 collected
Unemployed 7.8
Public officer 23.5 Another important research finding concerns the number of
Employed in private sector 42.5 Facebook friends and meeting up with them. There is a weak
Self-employed 2.3 positive correlation (r= 0.400) between the number of friends
Total 100.0 and the fact of meeting up. This can be explained by the fact
that people normally add individuals who know them already
Fig 2. Occupation types

4
but in general, some are called just „Facebook friends‟, which  Inadequate knowledge of the provisions of the law
is equivalent to the online aspect only.  Superficial insight into the fines associated with
A more in-depth analysis was carried out on the available data transgressing the law
to test if there was any difference between male and female  Erroneous ideas about the term of imprisonment
respondents concerning knowledge of the terms of use. It associated with specific offences
shows that there is a difference (significant at 0.000, less than
0.05).
ANOVA
Terms of use and Gender

Sum of df Mean F Sig.

Squares Square

Between 31.922 2 15.961 16.812 .000

Groups
Within 558.254 588 .949
Groups
Total 590.176 590

Fig 6. Awareness of legal aspects

Fig 5. Analysis of Variance
The above chart shows clearly the relationship between use of
C. Awareness of Facebook privacy settings the network and ignorance of the way forward when a legal
Research findings converge when it comes to privacy settings. issue arises. In order to prosecute Facebook, the only legal
Users are unaware of most features like automatic collection recourse is to lodge a complaint at a US District Court. The
of data as carried out by Facebook. Users were unaware of the survey demonstrates that only 3% of the respondents were
terms of use. Though not exhaustive, the terms of use does aware about this. This can pose some serious problems
provide ample information. Main areas of lack of knowledge considering that Facebook is an international network with
were as follows according to the data collected: users all over the world. While the terms of use makes this
 Possession of more than one account clear, it obviously restricts considerably the legal recourse of
 Usage of real name for any Facebook account the users. The process is not only costly but obviously time-
consuming.
 All information pertaining to a chat with a friend, in
The data was analysed mostly in a quantitative manner. Only
terms of duration, content and shared items
one open-ended was included. This was analysed using the
 Tagging of photos if no privacy settings have been
word cloud to get an idea about the emerging themes related to
activated
Facebook. This is depicted below:
 Likes are likely to be compiled and shared with third
parties. Facebook makes this clear by arguing that
this is important for an enhanced user experience and
in driving the right content into newsfeed.
 Advertisement is embedded in the business model of
Facebook though in general. Users seem quite
divided on the issue of seeing advertisement of their
walls (Yes: 39.5% and No: 48%).

D. Knowledge of current laws governing ICT in Mauritius

Knowledge of laws and regulations is not the forte of any
population. The study further confirms this. Indeed, the users
seem unaware of how their misuse of online systems including
Facebook can impact negatively on their rights. A typical Fig 7. Word Cloud of user concerns
example is the maximum fine for accessing someone‟s
account illegally. It is a notable fact that only 10% got the The word cloud shows that concerns revolved around personal
right answer. Thus, 9 out of 10 persons do not know the legal information, use of data and privacy.
framework. The findings are summarised below:

5
 V. RECOMMENDATIONS
Risks are never ending when it comes to technology. What
people have to understand is that technology is not perfect and
certain aspects of it are not fool proof. Technology is
manmade and glitches may happen at any time. Incidents may
happen with any website and not just Facebook. Hence it is up
to the public to be reasonable on what information they share
and keep on social networking sites. Furthermore, we should
be cautious and remember to delete personal conversation if it
is no longer needed. We choose to share information through
Facebook at our own peril - it is a willing risk we are taking.
In the light of the findings (Section V, parts B, C & D of the
paper) of this study, the following measures are proposed to
protect user privacy:
•awareness raising campaigns in various media to
sensitise the population about basic security measures
regarding Facebook usage
•Enforcement of our legal systems vis a vis security and
privacy issues.
•integrating safe Facebooking principles into secondary
and tertiary curricula
•changing certain privacy policies of Facebook and re-
designing elements of its architecture which make users
vulnerable to privacy attacks (for instance user consent
should be sought prior to using personal information.
•dissemination of information security news to
organisations and public
•awareness sessions for organisations on information
security issues
•conduct IT security audits for organisations based for
social networks
•dedicated cyber security portal
•sensitisation campaigns in schools and cyber-caravans
•dedicated portal for online reporting of privacy breaches
VI. CONCLUSION
The human race is continually going through a lot of changes
nowadays, and it could be said that there is something elusive
happening at every second. New inventions, new gadgets,
new mobiles, new fashion, new social platforms and much
more are the rage of the day. Hacking and interception of
users‟ profile by other unknown users are not uncommon. The
technological processes starting from the „click or touch
screen to like or share” operations are indeed very complex to
users and this complexity is hidden from them via a „window‟
like interface.
Facebook research has started only very recently. Considering
the importance of this social network on the lives of
individuals, it is worth noting that people have been overtaken
by the novelty of this phenomenon. The majority of users are
willingly joining networks without understanding the risks of
doing so. The terms of use, though publicly available, are not
being scrutinised and applied by the users. However,
paradoxically aspect, users feel strongly about privacy issues.
6
People on the social networks want to have private
conversation without anyone, even the site owner being aware
of the contents. Unfortunately, this remains a utopia with the
new technology which seems to capture everything, even
before we even join the network.
Facebook‟s architecture promotes extimacy and a global
culture of publicness. Within this superstructure of shared
disclosure Big Brother avatars pose serious threats to the
personal information willingly shared every day by millions of
users unaware of the dangers lurking in the maze of networks.
Whether innocently or naively or foolishly or negligently
shared – the fact remains that intimate data, and indeed every
move on Facebook can be spied on and fraudulently misused.
Although users don‟t read the terms and conditions of usage
and profess ignorance of privacy threats as well as existing
laws, paradoxically they are concerned about the various
security issues surrounding private data in social networks.
This calls for the implementation of widespread awareness
raising endeavours to deflate the myth that privacy can be
safely shared on Facebook. Whilst policy changes aimed at
giving users some degree of control on their data will require
global lobbies, in the short term enlightened and responsible
„facebooking‟ can go a long way towards correcting a utopic
vision of Facebook.
Acknowledgments
The authors would like to put on record the invaluable
contribution of Miss Lorna Ramsamy and Miss Nivita
Mannick, Research Assistants, for their significant input.
VII. REFERENCES
[1] Marichal, J. Facebook Democracy: The Architecture of Disclosure and
the Threat to Public Life. Ashgate. 2012
[2] International Business Times, Beyond Lol –Facebook collection analysis
user data- no laughing matter.http://www.ibtimes.com/beyond-lol-
facebooks-collection-analysis-user-data-no-laughing-matter-204671
2015.
[3] Kumar, A.M, Sharma, B.N. & Shrivastava, S.K. Online Social
Networks: Privacy Challenges and Proposed Security Framework for
Facebook. International Journal of Soft Computing and Engineering
(IJSCE). 2014,Vol. 4 Issue 1.
[4] Le Defi Quotidien, Le nombre des utilisateurs de Facebook, 2014.
[5] Kirkpatrick, D. The Facebook Effect: The Real Inside Story of Mark
Zuckerberg and the World's Fastest-growing Company. 2011. Virgin
books.
[6] Van Dijk, J. The Culture of Connectivity: A Critical History of Social
Media. Oxford. 2013. Oxford University Press.
[7] Tello, L. Intimacy and “Extimacy” in Social Networks: The Ethical
Boundaries of Facebook. Scientific Journal of Media Studies
Communicar. 2013. No. 41, v XXI
[8] Zhang, C. & Sun, J. Privacy and Security for Online Social Networks:
Challenges and Opportunities. 2010. IEEE Network
[9] Shehab, M., Squicciarini, A. Ahn, G-J. & Kokkinou, I. Access control
for online social networks third party applications. Computer &
Security. ScienceDirect Journal, 2012. pg 897-911 Elsevier Ltd.
[10] Leitch, S. & Warren, M. Security Issues Challenging Facebook.
Facebook Proceedings of the 7th Australian Information Society
Australian Information Security Management Conference. 2009.
[11] Kayes, I. & Iamnitchi, A. A Survey on Privacy and Security on Online
Social Networks.http://arxiv.org/pdf/1504.03342.pdf . 2015.
[12] Albanesius, C. A Privacy Breach for Facebook? PC Magazine. 2010.
Vol. 29 Issue 12, p1-1. 1p.
[13] Kulcu, O. & Henkoglu, T. Privacy in social networks: An analysis of
Facebook. Internationl Journal of Information Management. 2014. pg
761-769 Elsevier Lts
[14] Boyd, D. Hargittai, E. Facebook privacy settings: Who cares? 2010.
Vol. 15 No. 8 http://firstmonday.org/article/view/3086/2589#p2
7
[15] Kosinski. M, Stillwell, D. & T. Graepel, Private traits and attributes are
predictable from digital records of human behaviour. Proceedings of the
National Academy of Sciences (PNAS), 2013.
[16] Le Defi Quotitidien. Facebook viral copyright warning is useless.
http://www.defimedia.info/dimanche-hebdo/dh-magazine/item/22845-
facebook-%E2%80%93-viral-copyright-warning-is-useless.html. 2012
[17] No name. Code Civil Mauricien – Article 22
http://www.ilo.org/dyn/natlex/docs/ELECTRONIC/88152/114145/F-
172904586/MUS88152%20Fre.pdf. 2012.
[18] Albanesius, C.13 Million U.S. Facebook Users Not Using Privacy
Controls. PC Magazine. 08888507, 2012, p1-1. 1p.
[19] Grimmelmann, J. Privacy as Product Safety . Widener Law Journal.
2010. Vol. 19 Issue 3, p793-827.
[20] Fasola O. S. Perceptions and acceptance of librarians towards using
Facebook and Twitter to promote library services in Oyo State, Nigeria",
The Electronic Library , 2015. Vol. 33 Issue: 5.
 UOM Multimodal Face and Ear Database
Nazmeen B. Boodoo-Jahangeer* & Sunilduth Baichoo
Department of Computer Science, University of Mauritius,
Reduit, Mauritius
nazmeen182@yahoo.com*
Abstract—Research in biometrics have received increased
attention in the past decades. Existing face and ear dataases have
been described and compared. Some databases may not have
enough number of subjects or images per subject to allow proper
evaluation of algorithms. In order to support the development of
algorithms for biometrics, especially for face and ear, a new
database, the UOM Multimodal database has been created. This
database involves images taken by more than 100 volunteers at
the University of Mauritius. The images are taken under
different lighting and various pose. At least 80 images per subject
are available in the database.
Keywords—Face Recognition; Biometrics; Ear; Database
I. INTRODUCTION
Face recognition plays an important role in our society
since it is currently being used successfully in several
applications including Identity Cards, Passports, credit card,
Driver’s licenses and access control, crowd surveillance,
among others. Researchers have studied face recognition
extensively in the past few decades and have developed
several algorithms that allow identification of an individual
based on their facial image. Nevertheless, face recognition still
remains a challenge since researchers are trying to get
improved performance rate. One way of improving the
performance of an authentication system is by fusing face
features with ear features, thus creating a multi-modal
biometric system.
Ear biometric, on the other hand is a newer technology
compared to face biometric. The potential of the human ear for
personal identification was recognized and advocated as long
ago as 1890 by the French criminologist Alphonse Bertillon
[1]. In machine vision, ear biometrics has received little
attention compared to the more popular techniques of
automatic face, eye, or fingerprint recognition. However, ears
have played a significant role in forensic science for many
years, especially in the United States, where an ear
classification system based on manual measurements has been
developed by Iannarelli, and has been in use for more than 40
years [2].
Ears have certain advantages over the more established
biometrics; as Bertillon pointed out, they have a rich and
stable structure that is preserved from birth well into old age.
The ear does not suffer from changes in facial expression, and
is firmly fixed in the middle of the side of the head so that the
immediate background is predictable, whereas face
recognition usually requires the face to be captured against a
controlled background. Ear image acquisition does not have an
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
1
associated hygiene issue, as may be the case with direct
contact fingerprint scanning, and is not likely to cause anxiety,
as may happen with iris and retina measurements. The ear is
large compared with the iris, retina, and fingerprint and
therefore is more easily captured.
The main drawback of ear biometrics is that they are not
usable when the ear of the subject is covered [2]. In the case of
active identification systems, this is not a drawback as the
subject can pull his hair back and proceed with the
authentication process. The problem arises during passive
identification as in this case no assistance on the part of the
subject can be assumed. In the case of the ear being only
partially occluded by hair, it is possible to recognize the hair
and segment it out of the image.
A common problem that researchers face is the non-
availability of a large dataset to test the reliability of face and
ear recognition algorithms. In this work, datasets of face and
ear have been created to cater for the testing of the different
algorithms. The next sections outlines the different databases
available, evaluates them and explain the framework of the
UOM Multimodal face and ear database.
II. BACKGROUND STUDY
In this section, the existing face databases used in literature
by other researchers are outlined, together with features of
each database
A. Face Databases
Database consists of a number of images used for testing
of an algorithm. There are several criteria that affect the
choice of database, depending upon the tasks required. Some
features of database are:
 Number of Subjects: Each database has different number
of subjects.
 Number of images per subject: Also, the number of images
per subject will help in determining the number of training
and test sets.
 Gender: Certain databases cater for both male and female
subjects, and help for research in gender classification.
 Image Resolution: The resolution of an image determines
the storage space required and also has an impact on the
processing time.
 Pose Variation: Images of different poses help in testing
for algorithms that cater for variance in pose.
  Light Variation: In real-life scenario, light cannot be
controlled, images with lighting variance help to test
algorithms that cater for light variance.
 Variation in Facial Expression: Face images can be subject
to different expressions, which can be used for research,
such as determining drowsy drivers.
 Occlusion: In real-life system, at times, the image of a
subject is partly occluded by scarf, sunglasses or jewelry.
A database providing occluded face images help to test
robustness of algorithms with respect to occlusion.
The following paragraphs give details about the existing
facial databases commonly used in literature.
1) AR Face database
The AR face database [3] contains over 4,000 color images
corresponding to 126 people's faces (70 men and 56 women).
Images feature frontal view faces with different facial
expressions, illumination conditions, and occlusions (sun
glasses and scarf). The pictures were taken under strictly
controlled conditions. No restrictions on wear (clothes,
glasses, etc.), make-up, hair style, etc. were imposed to
participants. Each person participated in two sessions,
separated by two weeks (14 days) time.
2) MIT database
The MIT-CBCL face recognition database contains face
images of 10 subjects [4]. It provides two training sets. The
first one consists of high resolution pictures, including frontal,
half-profile and profile view while the second set consists of
synthetic images (324/subject) rendered from 3D head models
of the 10 subjects. The test set consists of 200 images per
subject. We varied the illumination, pose (up to about 30
degrees of rotation in depth) and the background.
3) BANCA face database
The BANCA database is a multi-modal database intended
for training and testing multi-modal verification systems [5].
The BANCA database was captured in four European
languages in two modalities (face and voice). The subjects
were recorded in three different scenarios, controlled,
degraded and adverse over 12 different sessions spanning
three months. In total 208 people were captured, 104 men and
104 women.
4) CASPEAL
The CAS-PEAL face database provides a large-scale
Chinese face images [6]. Currently, the CAS-PEAL face
database contains 99,594 images of 1040 individuals (595
males and 445 females) with varying Pose, Expression,
Accessory, and Lighting (PEAL). Each subject is also asked to
look up and down to capture 18 images in another two shots.
The authors also considered 5 kinds of expressions, 6 kinds
accessories (3 glasses, and 3 caps), and 15 lighting directions.
It is partly made available to the public.
5) CMU Hyperspectral Face Database
The CMU database contains visible and near infrared
(NIR) images from 450nm to 1100nm containing 65 spectral
bands and a spatial resolution of 640x480 pixels [7].
Acquisition of the 65 images took an average of 8 seconds.
2
Each of the 54 subjects was then imaged under four
illumination conditions (three lamps individually and then
combined). Subjects were recorded between one and five
times over a 6–week period.
6) CMU Pose, Illumination, and Expression (PIE)
Database
The CMU PIE is a database of over 40,000 facial images
of 68 people [8]. Using the CMU 3D Room, the authors
imaged each person across 13 different poses, under 43
different illumination conditions, and with 4 different
expressions.
7) Equinox IR
The Equinox IR database consists of images from 340
individuals [9]. For each individual, Images are
simultaneously acquired in visible, Short Wave InfraRed
(SWIR), Mid Wave InfraRed (MWIR) and Long Wave
InfraRed (LWIR). For each of three illumination conditions,
40 frame contiguous sequences are taken while subject recites
a standard phrase and 3 different expressions.
8) FERET
The FERET database was collected in 15 sessions between
August 1993 and July 1996 [10]. The database contains 1564
sets of images for a total of 14,126 images that includes 1199
individuals and 365 duplicate sets of images. A duplicate set is
a second set of images of a person already in the database and
was usually taken on a different day. For some individuals,
over two years had elapsed between their first and last sittings,
with some subjects being photographed multiple times. This
time lapse was important because it enabled researchers to
study, for the first time, changes in a subject's appearance that
occur over a year.
9) FRGC database
The database contains 3D face scans [11] of 4950 face
images of 466 different subjects. Each subject session consists
of images taken under well-controlled conditions (i.e., uniform
illumination, high resolution) and images taken under fairly
uncontrolled ones (i.e., non-uniform illumination, poor
quality). The database also contains various facial expressions
(happiness, surprise). The subjects are 57% male and 43%
female.
10) GTAV database
This database has been created at the Universitat
Politècnica de Catalunya (UPC) [12]. This database
emphasizes on facial expression variations, pose variations,
illumination variations and in partially occluded faces. It
includes a total of 44 persons with 27 pictures per person
which correspond to different pose views (0º, ±30º, ±45º, ±60º
and v90º) under three different illuminations (environment or
natural light, strong light source from an angle of 45º, and
finally an almost frontal mid-strong light source. Furthermore,
at least 10 more additional frontal view pictures are included
with different occlusions and facial expression variations.
11) Harvard RL
The Harvard Robotics Lab database consists of 10 subjects
taken in a wide range of illumination [13]. In each image, the
subject held his head steady while being illuminated by a
dominant light source. The space of light source directions,
which can be parameterized by spherical angles, was then
sampled in 15 increments. 75 images for each subject were
recorded, the images were of 193 × 254 pixels each.
12) KFDB database
The Korean Face Database (KFDB) contains images of
1000 subjects. Images are taken with varying pose,
illumination and facial expressions [14]. The subjects were
imaged in the middle of an octagonal frame carrying seven
cameras and eight lights (in two colors: fluorescent and
incandescent) against a blue screen background. Also, pose
images were collected in three styles: natural (no glasses, no
hair band to hold back hair from the forehead), hair band, and
glasses. Separate frontal pose images were recorded with each
light turned on individually for both the fluorescent and
incandescent lights. The subjects were also asked to display
five facial expressions: neutral, happy, surprise, anger, and
blink, which were recorded with two different colored lights,
resulting in 10 images per subject.
13) MPI database
The MPI (Max Planck Institute) face database [15]
contains images of 200 subjects (100 men and 100 women).
Subjects were wearing bathing caps at the time of recording
that are later automatically removed. The faces are free of
makeup, accessories, or facial hair. Part of the database is
available for researchers.
14) ND HID
The Notre Dame HumanID database contains images of
more than 300 subjects [16]. The images were recorded
multiple times over a period of 13 weeks. . A minimum of
four high-resolution color images were obtained during each
session under controlled conditions. In addition to that, images
were recorded under two lighting conditions and two facial
expressions, resulting in more than 15 000 images.
15) NIST MID
The NIST Mugshot Identification Database [17] contains
images of 1573 subjects. The images include both frontal and
profile view and have different resolutions. The subjects were
mostly male in this database.
16) ORL
The ORL face database is a set of face images taken
between April 1992 and April 1994 at the Olivetti Research
Laboratory in Cambridge, UK [18]. There are 10 different
images of 40 distinct subjects. For some of the subjects, the
images were taken at different times, varying lighting slightly,
facial expressions (open/closed eyes, smiling/non-smiling) and
facial details (glasses/no-glasses).
17) UMIST
The UMIST face database [19] contains images of 20
subjects taken in different poses. In fact, the images sequences
were taken as the subjects were slowly rotating their head
from profile to frontal view.
18) University of Essex Face database
The University of Essex Face Database [20] contains 400
images for 20 persons (20 images per person). Face images
were of 180 × 200 pixels. There were small variations in
3
facial expression, luminance, scale and viewing angle and
were shot at different time. Limited side movement and tilt of
the head were tolerated
19) University of Texas database
The University of Texas contains images [21] contains
images of 284 subjects, including 76 males and 208 females.
The database contains both still images and video clips of
faces and people. The still images and videos were taken at
close range, under controlled lighting conditions in an indoor
laboratory environment. Another duplicate session includes a
full set of these still and video images, taken at an average
interval of 24 days were taken for testing purposes.
20) University of Oulu
The University of Oulu Face database contains images of
125 subjects in 16 different camera calibration and
illumination condition and an additional image if the person
has glasses [22]. The faces were in frontal position captured
under Horizon, Incandescent, Fluorescent and Daylight
illuminant. The images also included 3 spectral reflectance of
skin per person measured from both cheeks and forehead
21) XM2VTS databasse
The Extended M2VTS (XM2VTS) multi-modal face
database [23] includes still colour images, audio data, video
sequences and 3D Model. It contains four recordings of 295
subjects taken over a period of four months. Each recording
contains a speaking head shot and a rotating head shot.
22) Yale database A
The first Yale face database [24] contains 165 grayscale
images in GIF format of 15 individuals (14 male, 1 female).
There are 11 images per subject, one per different facial
expression or configuration: center-light, w/glasses, happy,
left-light, w/no glasses, normal, right-light, sad, sleepy,
surprised, and wink.
23) Yale database B
The second version, Yale Face Database B [25], contains
5760 single light source images of 10 subjects each seen under
576 viewing conditions (9 poses x 64 illumination conditions).
For every subject in a particular pose, an image with ambient
(background) illumination was also captured.
24) Wild LFW
The main aim of this database was to allow testing of
unconstrained face recognition system. The database contains
13233 face images [26], out of which 1680 people have more
than one image in the database. The images were basically
taken from the web and detected by the Viola Jones Detector,
and have been re-scaled and cropped to a fixed size. The
images have variation in pose, lighting, expression,
background, race, ethnicity, age, gender, clothing, hairstyles,
camera quality, color saturation, focus, and other parameters
A summary of face databases used by researchers for
testing of algorithms have been listed in Table I. The
available databases are listed in the table, each one with
different specifications, in terms of number of subjects,
number of images per subject, gender, image resolution and
variances such as Pose, Illumination, Time and occlusion.
B. Evaluation of Face Database
As detailed in above section, several databases have been
created in the past for testing of face recognition algorithms.
The advantages of these databases are:
 Most of them are publicly available for use for research
purposes.
 The databases have varying number of subjects and
researcher can choose depending on applications
 Some have images of both males and females
 Databases, such as CAS-PEAL, CMU PIE, FERET,
GTAV, KFDB and Yale DB provides images taken at
different angles.
 Some databases, including AR, CAS-PEAL, CMU PIE,
Harvard RL, KFDB, Oulu and Yale offer images taken
under different lighting conditions.
 A few databases also offer variations in facial
expressions, such as neutral, smiling, sleepy, and sad,
among others.
 The databases, such as AR, CAS-PEAL, CMU PIE,
GTAV, ORL, UMIST and Yale provide images with
occlusion due to caps, scarves, glasses and hands.
 Database such as Wild LFW allows testing for
algorithms in unconstrained environment.
The above databases are definitely an asset to researchers
since they support research in face recognition and helps to
test robustness of such system with variance in lighting
conditions, pose, expressions, as well as occlusions. An
algorithm may work well when tested with images under
controlled environment. However good the algorithm may be,
there may be variations in performance if images are in
unconstrained environment.
However, in this research, the database created will have:
 At least 10 images per person, as per tests carried out
(refer to section 4.2.3)
 A resolution of at least 200 by 200 pixels in order to
extract the features precisely.
 Frontal images of the subjects without variation in pose
and lighting
 Profile images of subjects to be able to extract the features
of ear
C. Ear Database
Several databases have been used in literature for testing of
ear recognition system. Examples include the University of
Science and Technology Beijing (USTB) database and the
University of Notre Dame (UND) database, which are publicly
available for research purposes. Due to lack of a standard set,
several researchers have built their proprietary ear datasets
[27]. Thus, the results cannot be compared if experiments are
done on different datasets. Many researchers have used face
database, such as FERET, containing face profile images in
order to extract the ear image. Also, researchers have created
their own database for testing purposes.
4
An ear database allow ear detection and ear recognition
algorithms to be evaluated. Criteria for choosing ear database
can include: Number of subjects, No. of images per subject,
Image resolution, Pose variation, Light Variation and
Occlusion. The following paragraphs give details about
commonly-used ear databases.
1) IIT Delhi ear database
The Indian Institute of Technology (IIT) Delhi ear image
database [28] was used for the purpose of this research. The
IIT Delhi ear image database was acquired from 125 different
subjects and each subject has at least 3 ear images. All the
subjects in the database are in the age group 14-58 years. The
resolution of these ear images is 272 x 204 pixels.
2) IIT Kanpur ear database
The IIT Kanpur database is available in two subsets. The
first subset contains 801 side view images of 190 people, 2 to
10 images per person. The second subset contains images of
89 subjects. It contains 9 images for every person were taken
at 3 dissimilar postures. Each pose was captured at three
different scales.
3) UND databases
The University of Notre Dame provides several databases
that can be use for ear recognition [29]:
CollectionE: 464 visible-light profile ear images from 114
human subjects;
CollectionF: 942 3D ,plus corresponding 2D profile ear
images from 302 human subjects;
CollectionG: 738 3D, plus corresponding 2D profile ear
images from 235 human subjects;
CollectionJ2: 1800 3D, plus corresponding 2D profile ear
images from 415 human subjects.
4) USTB databases
USTB DB1
The USTB database involves 60 subjects, whereby the right
ear of each subject is photographed with a digital camera [30].
Three different images of each subject are taken. They are
normal frontal image, frontal image with trivial angle rotation
and image under different lighting conditions. Each of them
has 256 gray scales
USTB DB2
The USTB DB2 contains images of 77 volunteers [30]. The
purpose of the database was to support the research about ear
recognition under illumination variations and angle variations.
The images are taken under different lighting and angle
variations
USTB DB3
The USTB DB3 contains images from 79 [30]. The main
purpose of this database was to support research concerning
about steps of ear recognition system including ear detection,
the robustness of recognition methods under depth variation,
ear recognition under partial occlusion and multi-modal
biometric feature recognition based on the fusion of
information from ear and face. The database contains both
face and ear images with varying pose and occlusion.
USTB DB4
This database contains images from 500 volunteers [30]. The
images were captured with different pose. Each pose is
photographed with 17 CCD cameras simultaneously. The
interval between the cameras is 15 degrees. The integral
images of face and ear are captured. Both grayscale and color
images are available.
5)
6) UBEAR
The UBEAR ear database has over 4430 images [31], having.
The images were captured under varying lighting conditions.
Also, the subjects were taken under different pose, such as
looking towards the camera, upwards or downwards. The
subjects were not asked to remove hair, jewelery and
headdresses during image capture. the images had varying
quality to enable testing for robustness of algorithms.
The Table II below summarises details about the commonly-
used ear image database. The table outlines the different ear
databases in terms of the number of subjects, no. of images,
image resolution, pose variation, light variation, occlusion
details as well as no. of face images corresponding to the ear
image.
D. Evaluation of Ear databases
Some of the commonly-used databases are detailed above.
These databases provides
 Images of different resolutions
 Images are captured under various pose and illumination
conditions.
 UBEAR USTB DB4 and IIT Delhi also provides images
with partial occlusions with hair and jewelry.
 IIT Kanpur also provides the corresponding face image,
two of them, but it is not enough for testing.
E. Rationale for New Face and Ear database
A new database has been created for this study, because it
offers
 Images of more than 100 subjects
 Subjects include both males and females
 Images are taken at different poses and lighting
conditions to support further research.
 Several cameras are required to take photographs at the
same time
 A controlled environment is required
 Images of both frontal and profile are required for the
same person
 A good resolution of about at least 200 by 200 pixels are
required.
5
III. FRAMEWORK OF THE UOM MULTIMODAL DATABASE
A multi-modal database consists of different traits of a single
person. Existing Multi-modal databases are either not free or
does not have enough number of images to get valid results.
Some multi-modal database involving face and ear are not
made available to the public. Most researchers make use of
profile face image to get the ear, while others use chimeric
databases [32].
According to the investigations done by Poh and Bengio [33],
generating multiple chimeric databases does not degrade nor
improve the performance of a fusion operator when tested on a
real-user database with respect to using only a real-user
database. However, fusion models can only be evaluated on
real multimodal biometric data. For the purpose of this
research we have created a multi-modal face and ear database.
In this study the database is named UOM Multimodal
Database. Database availability helps to validate a given
algorithm, compare between different algorithms and develop
new algorithm [34].
Considerations have been given to the following factors:-
(a) Number of subjects: At least 100 subjects will be taken at
the first instance
(b) Number of images per subject. The system will take 10
images per subject.
(c) Gender: Both Males and Females will be considered.
(d) Image Resolution: Face image: 252 × 288
Ear image: 80 × 150
(e) Pose variations: 5 poses will be taken: -900, -450, 00, 450,
900 on the x-axis.
(f) Light variations: 3 lighting conditions will be considered.
(g) Occlusion: Minor occlusion due to earrings and glasses
have been tolerated. However, subjects were asked to make
sure the ear is not occluded.
A. Experiment to determine Number of Training and Testing
images
An investigation is done to determine the number of training
and test images required for getting maximum recognition
rate. Details of the experiment are given below:
Aim of experiment: To determine how many test images and
train images are required to obtain a stable recognition rate.
Dataset: Frontal face images from the UOM database
Algorithm: The Principal Components Analysis (PCA) is
applied to recognize faces.
Methodology: The testing is done in two parts:
Experiment 1: Firstly the training set is kept constant and the
test set is changed incrementally from 1 to 10. Here the train
images consist of 3 images per subject.
Experiment 2: Secondly, the training set is changed
incrementally from 1 to 10 while the test set is kept
unchanged. Here the test set consists of 3 images per subject
Results: The results of the experiments are as follows:
1) Experiment 1 result:
The recognition rate increases initially with more test images,
as shown in Table III. A graph is drawn to show the trend as in
Fig. 1
training images, here the performance is stable with 5 training
images per subject.
As per the above experiment, a stable performance is obtained
with 5 training images and 5 testing images per subject. Thus,
for the UOM Multimodal database, a minimum of 10 images
per subject is required for testing of face or ear recognition
system.
TABLE IV RESULTS OF EXPERIMENT 2: TEST IMAGES ARE KEPT
UNCHANGED

As shown in Fig. 1, the recognition rate increases with an No of train No. of test images Recognition Rate
increase in the number of images per subject in testing set, till images (per (per subject)
a stable value is reached. Here the value 5 is considered to be subject)
the appropriate value 1 3 images 67
2 71
TABLE III: RESULTS OF EXPERIMENT 1: TRAIN IMAGES ARE KEPT
UNCHANGED TABLE STYLES
3 75
4 79
5 85
No. of train No of test images Recognition Rate 6 87
images (per (per subject) 7 88
subject) 8 89
3 images 1 65 9 90
2 70 10 91
3 75
4 80 A graph is drawn to study the trend in recognition rate, as
5 87 shown in Fig 2
6 90
7 90.2
8 90.3
9 90.4
10 91

Fig. 2. Trend showing recognition rate verses train images

Fig. 1 Trend showing recognition rate versus test images
2) Experiment 2 Result
As far as the training images are concerned, a similar trend is
noticed as shown in Table IV. There is an increase in
performance with an increase in the number of training images
per subject.
As shown in Fig. 2, the performance of a system increases up
to a stable value when there is an increase in the number of
B. Imaging framework
The database has been created at the University of Mauritius
to support further research in involving pose and light
variance. However, the scope of this research is limited to
frontal face and ear image. The equipment used includes:
(a) Five CCTV cameras to take videos of a subject
simultaneously from 5 angles: -900, -450, 00, 450, 900 on the x-
axis.
(b) Three light sources
(c) A computer connected to the DVR system to view the
videos.

6
The Fig. 3 shows the set-up of the camera. The subject was
required to look at the middle camera, and the video is taken
for a few minutes. To get the images, corresponding frames at
specific time intervals were extracted from the different videos
taken, resulting in 5 images at any particular time. 10 images
per camera were extracted within one second to avoid
variations in pose, giving a total of 50 images per subject in
normal light. Also, images were captured under three lighting
conditions, resulting in 30 additional images per subject.
Above 100 subjects volunteered to provide their image. Since
for some of the subjects, there were problems of pose and
occlusions, 100 subjects have been selected, resulting in a total
of 5000 images in the UOM .
-900 -900
-900
-900 -900
Subject
Fig. 3 Setup of cameras
C. Sample Images
Sample face images and ear images from the UOM
Multimodal database are shown in Fig.4 and Fig. 5
respectively.
Fig. 4. Sample face images from UOM Multimodal database
Fig. 5. Sample ear images from UOM Multimodal database
7
Five images are included in the test set and five in training set
for each subject. Images of 80 subjects were used for training
and images for 20 subjects were used for testing of imposters.
IV. CONCLUSION
This work provides details about the different factors that
needs to be considered when choosing a database for research.
The database commonly-used for face recognition and ear
recognition are outlined and evaluated. The reasons for new
database have been outlined and framework for the new
database has been explained. The UOM Multimodal database
contains images light variance and pose variance to encourage,
research in pose and lighting for ear and face recognition.
Sample images of all databases used in this research have been
given.
References
[1] Bertillon, A., 1890. La Photographie Judiciaire, Avec Un
Appendice Sur La Classification Et L’identification
Anthropometriques, Gauthier-Villars, Paris.
[2] Iannarelli, A., 1989. Ear Identification. Forensic Identification
Series. Paramont Publishing Company, Fremont, California.
[3] Martinez A.M. & Benavente R.. 1998, The AR Face Database.
CVC Technical Report #24, June 1998
[4] Weyrauch B., Huang J., Heisele B., & Blanz V., 2004,
Component-based Face Recognition with 3D Morphable Models,
First IEEE Workshop on Face Processing in Video, Washington,
D.C., 2004.
[5] Bailly-baillire E., Bengio S., Bimbot F., Hamouz M., Kittler J.,
Mariéthoz J., Matas J., Messer K., Porée F., Ruiz B., 2003, The
BANCA Database and Evaluation Protocol, In Proc. International
Conference on Audio- and Video-Based Biometric Person
Authentication (AVBPA03), pp. 625 - 638.
[6] Gao W., Cao B., Shan S., Zhou D., Zhang X., & Zhao D., The
CAS-PEAL Large-Scale Chinese Face Database and Evaluation
Protocols, Technical Report No. JDL_TR_04_FR_001, Joint
Research & Development Laboratory, CAS, 2004
[7] Denes L. J., Metes P., & Liu Y., Hyperspectral Face Database,
Technical Report CMU-RI-TR-02-25, Robotics Institute,
Carnegie Mellon University, October, 2002
[8] Sim T., Baker S., & Bsat M., 2002, The CMU Pose, Illumination,
and Expression Database. In: IEEE International Conference on
Automatic Face and Gesture Recognition, pp. 46-51
[9] Socolinsky D., Wolff L., Neuheisel J., & Eveland C., 2001.
Illumination Invariant Face Recognition Using Thermal Infrared
Imagery. In Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, 2001
[10] Phillips, P.J., Moon, H., Rizvi, S.A. & Rauss, P.J., 2000. The
FERET Evaluation Methodology for Face Recognition
Algorithms, In: IEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 22, No. 10, pp. 1090-1104.
[11] Phillips P. J., Flynn P. J., Scruggs T., Bowyer K. W., Chang J.,
Hoffman K., Marques J., Min J., & Worek W., 2005, Overview
of the Face Recognition Grand Challenge, Proc. 2005 IEEE
Computer Society Conference on Computer Vision and Pattern
Recognition, No. 1, pp. 947 – 954.
[12] Tarrés F. & Rama A., 2005, A Novel Method For Face
Recognition Under Partial Occlusion Or Facial Expression
 Variations, 47th International Symposium of ELMAR, 2005, pp.
163 - 166.
[13] Hallinan P., 1995, A Deformable Model for Face Recognition
under Arbitrary Lighting Conditions. PhD thesis, Harvard
University, 1995
[14] Hwang B. W., Byun H., Roh M.C., &. Lee S.W., 2003,
Performance Evaluation Of Face Recognition Algorithms On The
Asian Face Database, KFDB. In Audio- and Video-Based
Biometric Person Authentication (AVBPA), pp. 557–565.
[15] Troje, N. and H. H. Bülthoff, 1996, Face recognition under
varying poses: The role of texture and shape. Vision Research 36,
1761-1771.
[16] Flynn P., Bowyer K., & Phillips P. J., 2003. Assesment of Time
Dependency in Face Recognition: An Initial Study. In Audio and
Video-Based Biometric Person Authentication (AVBPA), pp. 44
–51.
[17] Watson C. I. 1994, NIST mugshot identification database.
Available at http://www.nist.gov/srd/nistsd18.htm, Retrieved on
15th May 2015.
[18](Samaria F., & Harter A., 1994. Parameterisation of a Stochastic
Model for Human Face Identification, Proceedings of 2nd IEEE
Workshop on Applications of Computer Vision, Sarasota FL, 5-7
December 1994, pp. 138 – 142
[19] Graham, Daniel B. and Allinson, Nigel (1998) Characterizing
virtual Eigensignatures for general purpose face recognition. In:
Face recognition: from theory to applications. NATO ASI Series
F, Computer and Systems Sciences, Vol. 163. Springer, pp. 446-
456
[20] Spacek L., 2007, Computer Vision Science Research Projects,
Accessed on July 20 2013, from Essex University website:
http://cswww.essex.ac.uk/mv/allfaces/faces94.html.
[21] O'Toole, A.J., Harms, J., Snow, S. L., Hurst, D. R., Pappas, M.
R. & Abdi, H., 2005. A Video Database of Moving Faces and
People. IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 27, No. 5, pp. 812-816.
[22] Martinkauppi B., Soriano M., Huovinen S. & Laaksonen M.,
2002, Face video database. Proc. First European Conference on
Color in Graphics, Imaging and Vision (CGIV'2002), April 2-5,
Poitiers, France, pp. 380-383.
[23] Messer K., Matas J., Kittler J., Luettin J., & Maitre G., 1999.
XM2VTSDB: The Extended M2VTS Database. In Second
International Conference on Audio and Video-based Biometric
Person Authentication, 1999.
TABLE I. DATABASES COMMONLY-USED BY RESEARCHERS
No. of No. of Gender Resolutio
Subjects images n on
per
subject
1 AR Face 126 26 70 Male, 576 × 768
database 56 Female
8
[24] Belhumeur, P.N., Hespanha, J.P., Kriegman, D.J., 1997,
Eigenfaces vs. Fisherfaces: Recognition Using Class Specific
Linear Projection, IIEEE Transactions on Pattern Analysis and
Machine Intelligence, Vol. 19, No. 7, pp 711-720
[25] Georghiades, A.S., Belhumeur, P.N. & Kriegman, D.J., 2001.
From Few to Many: Illumination Cone Models for Face
Recognition under Variable Lighting and Pose IEEE Transactions
on Pattern Analysis and Machine Intelligence, Vol. 23, No. 6, pp.
643-660.
[26] Huang G. B., Ramesh M., Berg T., & Learned-Miller E., 2007
Labeled Faces in the Wild: A Database for Studying Face
Recognition in Unconstrained Environments, University of
Massachusetts, Amherst, Technical Report 07-49, October, 2007.
[27] Abate, A.F., Nappi, M., Riccio, D., & Ricciardi, S., 2006. Ear
Recognition By Means Of A Rotation Invariant Descriptor. In:
18th International Conference on Pattern Recognition, 2006.
ICPR 2006, Hong Kong, pp. 437 – 440.
[28] Kumar A, Wu C., 2012, Automated human identifcation using
ear imaging, Pattern Recognition, Vol. 45, No. 3, pp. 956 – 968.
[29] UND. (2014). CVRL DATA SETS ( University of Notre Dame
UND database),. Retrieved from
http://www3.nd.edu/~cvrl/CVRL/Data_Sets.html
[30] USTB 2005 The ear recognition laboratory at the University of
Science and Technology Beijing, China.
[31] Raposo R, Hoyle E, Peixinho A, Proenca H. ’UBEAR: A dataset
of ear images captured on-the-move in uncontrolled conditions’.
In: Computational Intelligence in Biometrics and Identity
Management (CIBIM), 2011 IEEE Workshop on; 2011. p. 84 –
90.
[32] Ross A., & Jain A., 2003. Information Fusion in Biometrics,
Pattern Recognition Letters, Vol. 24, No. 13, pp. 2115–2125.
[33] Poh N. & Bengio S., 2005, Can chimeric persons be used in
multimodal biometric authentication experiments?, In: MLMI'05,
Second international conference on Machine Learning for
Multimodal Interaction, pp. 87-100
[34] Fabregas J., & Faundez-Zanuy M., 2008. Biometric Face
Recognition with Different Training and Testing Databases,
Lecture Notes in Computer Science, Vol.5042, pp 44-55.
Pose Light Expressi Time Occlusion
1 4 4 2 Scarves
and
sunglasses
2 MIT 114 7 108 Male, 480 × 640 5 3 2 1 ++
database 6 Female

3 BANCA 52 5 28 male, ++ 1 ++ 1 12 ++
28 Female

4 CAS- 1040 Varied 595 male, 360 × 480 21 9 – 15 6 2 3 caps and

PEAL for 445 3 glasses
different females
subject

5 CMU 54 ++ 1 4 1 1–5
Hyper
6 CMU PIE 68 640 × 486 13 43 3 1 Glasses

7 Equinox 91 240 × 320 1 3 3 1

IR
8 FERET 1199 256 × 384 9 – 20 2 2 2

9 FRGC 688 1200 × 1 2 2

1600

10 GTAV 44 27 5 3 Varying Hand,

Glasses

11 Harvard 10 193 × 254 1 77 – 1 1

RL 84
12 KFDB 1000 640 × 480 7 16 5 1

13 MIT 15 120 × 128 3 3 1 1

14 MPI 200 3 3 1 1

15 ND HID 300+ 1600 × 1 3 2 10/13

1200

16 NIST 1573 Varying 2 1 ++ 1

MID
17 ORL 40 10 92 × 112 1 ++ 3 ++ Glasses

19 UMIST 20 220 × 220 ++ 1 ++ 1 Glasses

20 U. of 20 180 × 200 1
Essex

9
 21 U. Texas 284 76 male, 720 × 480 ++ 1 ++ 1
208
female

22 U. Oulu 125 428 × 569 1 16 1 1

23 XM2VTS 295 720 × 576 ++ 1 ++ 4

24 Yale 15 11 320 × 243 1 3 6 1 Glasses

25 Yale B 10 756 640 × 480 9 64 1 1

26 Wild 5749 250 × 250 Varyin Varyin Varying Varyin

LFW g g g

TABLE II: EAR IMAGE DATABASES

Database No. of No. of Images Image Pose Illumination Occlusion No of face

Subjects per Subject Resolution images

IIT Delhi 121 varying 272 x ++ ++ Earing

204

IIT Kanpur 400 2 640 x 480 ++ ++ ++ 2

UND > 400 varying varying varying varying ++

USTB DB1 60 3 256 gray scale 2 2 ++ ++

USTB DB2 77 4 300 x 400 3 2 ++ ++

USTB DB3 79 varying 768 x 576 18 ++ By hair

USTB DB4 500 500 x 400 17 ++ ++ ++

UBEAR 126 Varying Occlusion

by hair and
earrings

10
Review Current CRM Architectures and Introducing
New Adapted Architecture to Big Data
A. Daif1, F. Eljamiy2, M. Azzouazi3 and A. Marzak3
Laboratoire Mathématiques Informatique et Traitement de l’Information MITI, Casablanca, Maroc
1
daif.abdou@gmail.com, 2fatima.eljamiy-etu@univh2m.ma, 3azouazii@gmail.com
Abstract—The business model as known by the majority
of specialists has moved from product concentration to the
customer concentration. And as we all know, electronic
commerce and generally the world of technology has
exponentially believed that principal. In this era, many
companies have begun to permeate the Electronic Customer
Relationship Management (CRAG / E-CRM) more than
conventional CRM to better understand their customers. As
early as the first definitions of the CRM, we found among
the main three axes - technology - Currently the world data
is also outstanding to have a change with the advent of Big
Data paradigm. There is a need to rethink and reconsider
the validity of existing architectures management of
electronic customer relationship.
In this paper, we will see the existing architectures, and we
subsequently present an architecture with features capable
to respect and exploit the new data based on research works.
Keywords—CRM; Big Data; CRM; e-CRM; Architecture
I. INTRODUCTION
Customer Relationship Management (CRM) is a business
strategy that integrates organizational culture, human
resources, processes and technology, to acquire and retain
high value customers.
With the most demanding and critical consumers vis-à-vis
trademarks by their access to information, organizations
can no longer do without a new reflection on the customer
relationship. If the digital revolution and the social web
renew the opportunities to enrich the customer
relationship strategies, pitfalls are to be avoided to not
lose the customer who now does not hesitate to turn to the
competition if he finds more advantages.
The customer relationship management handles all
aspects of customer interactions. It offers a view on the
company's performance and employees and bring
productivity. After a good consolidation of all possible
data sources within the enterprise, commercial and
marketing department can manage profitably the service
and targeting of customers and prospects.
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
In fact, it is necessary to consider the challenges of the
implementation of the CRM, as justified by the
commercial market studies, which have shown that
approximately 70% of CRM projects result either losses
or no improvement in performance of the company [1].
These negative results are well described in academic and
professional level. [2] What we can add is that there are
many actors and startups that are placed on the new
operating niche of big-data and CRM, but there was
therefore a gap between their positioning and state of
evolution of their systems due to the rapid development,
diversification and heterogeneity of the data available
today. This requires a review of the start of the
development cycle of these solutions, and review the
structure and architecture that give a clear view of the
modules and the possible interfaces between them.
It is in this sense that we started our research on managing
the customer relationship, precisely the part of the
recommendation, and in this paper we have made a study
of some existing architectures of the main market
industries and we proposed our own architecture which
will include a module with features that could add value
to the final results of an electronic customer relationship
management system. This paper is organized as follow: in
the second section, we will present Big-Data
fondamentals and concepts, then, we will present some
architectures, in section 4 we will present our proposed
architecture, and our document will be concluded in
Section 5.
II. BIG DATA FONDAMENTALS AND CONCEPTS
Big Data is more or less structured data sets that become
so large that they are difficult to work with conventional
management database tools.
The associated definition is: "structured and unstructured
data requires the very large volume of adapted analytical
tools ". some states that are also found the term "massive
data" also often preferred in recent years.
The excitement around this phenomenon appeared a few
years ago, generates definition’s muddle.
While it is true that the Big Data word seems sometimes
used to excess, we have to believe that the phenomenon is
real. Let us adduce some notions, for insiders and those
who are not in context.
So quite generally accepted, Big Data is defined by 3V or
4 or 5V.
The concept of big data has been defined through the 3V
model, which was defined in 2001 by Laney [14]
This is the schematic and can synthetically to explain
what is in this notion.
II.1. BIG DATA V’S
Volume: The volume describes the amount of data
generated by businesses or individuals.
With the coming on of the Internet and social networks,
quantity of electronic data generated has explored and
continues to rise.
The companies, all confused sectors and activities, must
now find ways to manage a data volume gradually
increasing. It is now in 1000 terabytes that we name
Petabytes.
Variety: These data further being massive, are
characterized as unstructured or very variable formats,
unlike the dominant data which were formatted in
conventional databases.
Thus, social media, video, semantic content, web logs,
location data ... are weighty data sources that do not
present a "standard format".
Velocity: describes the speed with which the data is
generated and processed. We are focused on getting
knowledge from the data arriving as streams in real time.
More we focus on real time; more we are in big data
problem. Gradually, the immediate treatment of data
would be the key element of a model big data. [15]
Other organisations, and big data practitioners
(e.g.,researchers, engineers, andsoon), have extended this
3V model to a 4V model by including a new “V”:Value
[16]. Thismodel can be even extended to 5Vs if the
concepts of Veracity is incorporated into the big data
definition.
Value: exploring and scrutinizing huge amount of non
structured data in real/short time is big chalenge, but
without creating new value, it wont be necessary at all.
There is a vast opportunity offered by Big Data
technologies to discover new insights that drive
significant business value. Industries are seeing data as a
market differentiator and have started reinventing
themselves as “data companies”, as they realise that
information has become their biggest asset
Veracity: the data is it right or true. Can I believe in this
information? As social, or internet give that autonomy of
expression, and precisely because it is issued directly by
clients, blogers..., they require work on the quality and
interpretation. These external data have a behavioral
dimension. The process of the new "black gold" that is the
Big data is not without difficulty and pretenses new
challenges. Indeed, all the information will not have the
same value and some may be little or irrelevant (or
wrong)
II.2. BIG DATA CHALLENGES:
In this part, we will present some of big data challendes
since it big topic currently against many challenges.
Storage:
The future is to store more amounts of data than today
and it is not trivial because such data is valuable. The
sectors that succeed are those that use the data in the most
intelligent way and today this principle seems to apply to
all business sectors. [17]
During the last decade emerged a new kind of software,
Open Source, which has made Linux the most popular
operating system in the world. There is no doubt that open
source is becoming the norm and that major developments
are expected to use this type of platform.
Data Privacy:
Big Data platforms have been specifically designed to be
used on advanced form of analytics and respinding to
strict performance and scalability requirements. However,
no proper consideration has been dedicated so far to data
protection. Indeed, although the analyzed data often
include personal and sensitive information, with relevant
threats to privacy indirect by the analysis, so far Big Data
platforms integrate quite basic form of access control, and
no support for privacy policies [18].
Analyzing Big Data:
Nowadays generated data through heterogenous systems,
is very different from the old one, further more it
complicates analyts work, almost all used algorithms and
methods should be adapted to this new era of big data, for
example, researchs touch now the way of exploring social
medias,
the exponential growth of social media has created serious
problems for traditional data analysis algorithms and
techniques (such as data mining, statistics, machine
learning, and so on) due to their high computational
complexity for large datasets. This type of methods does
not properly scale as the data size increases. For this
reason, the methodologies and frameworks behind the big
data concept are becoming very popular in a wide number
of research and industrial areas Currently, the exponential
growth of social media has created serious problems for
traditional data analysis algorithms and techniques (such
as data mining, statistics, machine learning, and so on)
due to their high computational complexity for large
datasets[19].
Big Data is gradually transforming organizations around
the information’s valorization. With the information’s
finesse about past transactions and more and more
forward-looking information, big data will also allow the
producing most relevant predictive models. and it is in
this sense that our research is directed to review the Social CRM:
customer relationship management systems, and more
precisely digging in the recommendation in future The Social Customer Relationship Management is a
research business strategy considered as a natural evolution
imposed by the new social customers in the sense of
creating a bilateral relationship between the customer and
the company [10]. It can be conveyed through different
III. PRINCIPLES AND ARCHITECTURE technological platforms and customer relationship tools
such as a CRM system.
III.1. DEFINITIONS Consumers took power on social networks and brands
within thousands of simultaneous conversations of these
CRM : "consum-actors" who seek to intrude.
Customer Relationship Management is a set of tools and III.2. ARCHITECTURES
techniques that :
Studies of architectures and CRM conceptual models [5]
- Conduct an historical points of contact of a customer or [6] [8] show that there is a resemblance and similarities
a prospect with a company between the CRM global architectures. The Figure 1
- provide support to customers interactions management shows the layers CRM:
through appropriate tools
- store structured information on these customers and
prospects
- Re-use this information to optimize the next customer
contact
The CRM based mostly on technology tools to process
large amounts of data on customers disseminate this
information and automate some of the exchanges.
However, it is not just tools, it is a project management,
substantive policy which often requires changes in the
structures and business processes [3] [4]
Figure 1 Conceptual layers of CRM

The technology architecture and internal and external

business processes are the critical part of the system. And
following an official publication of HP [7] and that of
Gartner, there are three classes of technologies:

Figure 2 Classes of technology

 - Collaborative technology
Thanks to new communication technologies,
collaborative technology allows customers to interact
with the organization (Meta Group, 1999). This
provides central communication and coordination
network or the main links between the customer and
the entire supply chain as suppliers and / or partners
(Xu and Walton, 2005: 961). It also represents a point
of contact between the customer and one of the other
communication channels. [11] It may show as portals,
or center of interaction with the customer.
This class is the backbone of the development of e-
CRM [12], which is centered on the Web. It allows
companies to have more contact with their customers
across intranet, extranet and mobile applications.
Today, the new CRM technologies must be adapted
and re-think to have the ability to interact with new
channels and collect data of different types and
different sources by following this opaque wave of Big
Data
- Operational Technology
Represents all the internal order of business
applications and customer-driven, and enable to
automate and manage:
- The marketing plan (contact management, track leads,
opportunities and companions)
- Sales source of different channels as well as the rest
of the process.
Sales forces (registration of visits and surveys, marking
potential customers)
Figure 3 CRM Oracle Architecture
The customer call centers are also an operational CRM
component and have been identified as the dominant
aspect in CRM systems (Xu and Walton, 2005: 960;
Anon, 2000). All customer interactions are recorded
allowing the agency to collect data on the customer and
thus follow the customer (Xu and Walton, 2005: 961)
- Analytical Technology
We can say that without this class the data will be
useless since the data collected are unclean, and that
policy makers can not make an interpretation to an
amount of data that does not cease to increase at a
significant rate in recent years, hence the need to have
the tools to process and analyze that data. Analytical
CRM involves the capture, storage, retrieval,
processing, interpretation and presentation of customer
data stored in data warehouses (Xu and Walton, 2005:
961)
The analytical part of CRM includes data marts or data
warehouses such as customer databases that are used
by applications that implement algorithms for cutting
the data and display them in a form that is useful to
policy makers [6].
As mentioned above for collaborative CRM, analytical
CRM can be enhanced through use of new types and
amount of data, while putting the focus on the creation
of the competitive value for the company [9] and
improved sales and customer loyalty.
III.3. INDUSTRIAL ARCHITECTURE
In this part, we will present some architectures of
specialists industries in this segment [7] [11] [12] and
we will finish with a proposal in section 4
ORACLE
Oracle launched on the market the Siebel product as a
more feature-rich solution that is designed for several
industries. The figure below shows the application
component elements, but basically, if we analyze the
grouped elements (rectangles), we see very well the
presence of three technologies classes. Siebel 8.1 has
the Siebel user interface, which can be analytical,
highly customized embedded uses, and focused on the
task.
SAGE

Figure 4 Sage CRM Architecture

Sage CRM, however, is powered by a pure Internet database access), but in fact is the analytical part
architecture that surpasses client / server technology injected into the operational part that allows to export
which is an advantage today. As shown in the figure data to other advanced analysis tools. [13]
below, the CRM Internet architecture includes three
distinct layers that communicate using Internet
standards (Interface services, business logic services,
HP CRM

Figure 5 HP CRM Architecture

The same as other industries, the model of HP is
represented in the form of six layers, starting from the
highest layer, we'll have a match still in three classes, the
first and second layer repent collaboration technology ;
the third layer for the operational technology and the last
three layers are the analytical
technology. If we scrutinize these architectures we have
the same remarks made on [6] which concluded that for a
global visibility we end up with a named representation:
three envelopes
III.4. MODEL THREE ENVELOPES
Figure 6 Model three envelopes
This architecture is simple and clear without doubt vis-à-
vis the connections and interactions between the three
envelopes, but how these architectures can operate with
the Big Data paradigm? Are they able to interact with new
solutions on the market?
IV. PROPOSED ARCHITECTURE
Figure 7 Proposed architecture
In our architecture we have added a whole module as
illustrated in Figure 7, which represents a big-data
ecosystem that will have as main task the preparation of
data to be processed converged with other existing and
those baptized unknown, which are actually the results of
previous interpretations; while respecting the maximum
possible five V (Velocity, Variety, Accuracy and Value,
Speed).
We have made a design with three main sub-modules:
real-time analysis, storage tools and processing tools.
These three modules have a complementary relationship,
it means that for the processing of these data (excluding
the techniques and algorithms used) and the additional in
real time we need special and adapted mechanisms in
terms of storage . The synchronization between these
three sub modules will be the key to perfect the overall
interpretation process of the 5V data. If we break these
three modules, we will see that the complexity can
decrease by collaborating, technically speaking, between
the types of tools available (analysis, storage, processing)
such as hadoop, pentaho, Hbase, MongoDB .... This
module will be as a complex platform that malaxera
between the solutions that have weighed in the fields of
processing and storage of data as well as new models of
CRM, namely: Social CRM and Open CRM. The
interaction and data analysis of social networks, blogs,
customer reviews will create value to enrich and
eventually improve the product and recommend (draw
satisfaction) at best, thanks to Client-a-customer
experience based on analysis of web data by any of the
existing methods, citing as examples to be sentimental and
social listening to new data.
CONCLUSION
In this work we have made a comparative and synthetic
study of architectures offered by industries of the field
and we have proposed an architecture that will incorporate
a layer or ecosystem for the big-data and consequently
cover the need to create more competitive value and
follow current trends.
The research may be continued on the technical
integration of available tools in the market, taking
advantage of the strengths of each and assess the added
value of the knowledge extracted.
Our architecture is based on heneritated architectures of
large market industries and it will be more enhanced
while taking into account the evolution of the
aforementioned tools at the mercy of events and technical
developments.
REFERENCES
[1] Gartner Group, "CRM Success Is in Strategy and Implementation,
Not Software," available at http://www.gartner.com. 2003.
[2] J. Reynolds. A Practical Guide to CRM. New York: CMP Book,
2006
[3] Intro au Social CRM : le grand-frère du Community Management,
pix-geeks.com, Fabien Elharrar 2012
[4] Mercator 8ème édition Chapitre 17 le marketing relationnel,
Landrevie, Lévy, Lindon pp. : 25
[5] Radcliffe , John, Eight Building Blocks of CRM: A Framework for
Success, Gartner research, 2001
[6] S. Rezaiian Fardoie, M.A Saniee Monfared, A new design
architecture for e-CRM systems (Case study: tour package choice
in tourism industry), IEEE ICMIT, Proceedings of the 2008,
pp.463-468
[7] Janjicek, R., CRM architecture hp, technical white paper, 2005.
[8] M Viljoen, JA Bennett, AD Berndt & CR van Zyl, the use of
technology in customer relationship management (crm), Acta
Commercii 2005, pp.106-116
[9] Zhang Guozheng*, Chen Yun, Fu Chuan, A Study On The
Relation Between Enterprise Competitive Advantage And Crm
Based On Data Mining, International Technology And Innovation
Conference, 2006, pp.1710-1714
[10] Christos Giannakis-Bompolis, Christina Boutsouki, Customer
Relationship Management in the Era of Social Web and Social
Customer: An Investigation of Customer Engagement in the Greek
Retail Banking Sector, Procedia - Social and Behavioral Sciences,
2014, pp.67-78
[11] M.I. Kramer, ,"Comparing CRM Architectures, How the
Architectures of Epiphany’s, Oracle’s, PeopleSoft’s, SAP’s and
Siebel’s CRM Product Suites Stack Up against Each Other,
Seybold_Comparing CRMArchitetures_Report.pdf, Patricia,
Seybold Group, 2003.
[12] David Puglia, Building E-CRM Architecture: Customer,
Relationship Portals, 2004.
[13] Sage, the benefits of Sage CRM internet architecture, 2009
[14] Doug Laney, 3D Data Management: Controlling Data Volume,
Velocity, and Variety, Technical Report, 2001.
[15] A. DAIF, The potential and challenges of Big data:
Recommendation systems next level application, IJCSI
International Journal of Computer Science Issues, Vol 11, pp: 21,
2014
[16] I.A.T.Hashema, I. Yaqooba, N.B.Anuara, S. Mokhtara, A. Gania,
S.U.Khanb, The rise of big data on cloud computing: review and
open research issues , Inf.Syst.47(2015)98–115.
[17] M. Fazio, A. Celesti, A. Puliafito, M. Villari, Big Data Storage in
the Cloud for Smart Environment Monitoring, Elsevier, pp:500-
506, 2015
[18] Pietro Colombo, Elena Ferrari, Privacy aware access control for
Big Data: A research roadmap, Big Data Research Accepted
manuscrit, 2015.
[19] Gema Bello-Orgaz , Jason J. Jung ,∗ , David Camacho, Social big
data: Recent achievements and new challenges, Elsevier, pp:20,37,
2015
 An enhanced software tool to aid novices in learning
Object Oriented Programming (OOP)
Yeeshtdevisingh Hosanee
University of Technology,
Mauritius
nousha_study@yahoo.com
Abstract—Because of the complexity of Object Oriented
Programming (OOP), various applications have been
implemented as a learning support. However, results show that
the novice learners still face problems in learning OOP. This
paper has undergone an empirical research, aiming at
understanding the problems faced by the novice learners in
grasping the OOP concepts. This research has therefore the aim
of determining requirements for a software tool which caters for
the needs of the students. As methodology, a survey was carried
out to identify the learners’ difficulties and the types of supports
they require while interacting with the OOP supporting tool. The
latter also takes into consideration valid features of prior OOP
teaching software tools. A particular feature, in-built in the tool,
also caters for e-assessment which will allow the user to know his
proficiency level. Novice learners are thus able to monitor their
progress and take relevant actions accordingly. Thus, this will
encourage novice learners to engage and be motivated to learn
OOP at their own pace.
Keywords—Graphical User Interface (GUI); Object-Oriented
Programming (OOP); teaching; learning, software tools; engage;
proficiency level; Integrated Development Environment (IDE)
I. INTRODUCTION
Today, there are many undergoing researches regarding
teaching and learning [1] in order to research on the most
efficient ways to teach and learn. However, learning goes
beyond classrooms. Students should be able to study at their
own pace and place [2]. Classroom technology has evolved
from the handbook (1650) to iPad (2011) [3]. Despite the need
to have more programmers in the future, students are not
willing to overcome the difficulties of programming. Object-
oriented programming (OOP) remains difficult to teach [4, 5]
and to learn [5]. Throughout years, researchers have been
trying to find the best strategy to teach OOP so that novices
are motivated to learn it [6].One of the strategies being used is
the usage of software tool to facilitate teaching and learning.
Recent studies both in 2013 and 2014 confirmed that
existing tools do not meet the expectation of novice learners
[7, 8]. The technical University Kosice in Slovakia has been
using existing tools for many years. They concluded that
existing software tools still do not prepare novice learners for
the job market. Novice learners still find OOP difficult as they
have to switch from a novice tool to a real IDE [7]. In 2014
[8], an experiment was carried out with students at the
University of Surrey in United Kingdom. Although the
students like the simplicity of the tool being used, they still
Shireen Panchoo
University of Technology,
Mauritius
s.panchoo@umail.utm.ac.mu
prefer a tool which is similar to real-industry IDE where they
can work with the real codes [8]. Existing OOP tools do not
allow a progressive learning style. Existing tools are either too
simple without displaying real codes or they are found to be
too complex as they display too much of information. A
compromise between the different features to be displayed to a
user should be considered. Existing tools should take into
account the learners’ growth. This means they should allow a
novice learner to progressively outreach to an advanced level.
This will ultimately reduce the problem of switching to a real-
IDE [7].
It is thus imperative to implement relevant OOP tools
which will support novice learners. The aim of this paper is to
understand the problem faced by the novice learners in
grasping OOP concepts and proposing an efficient software
tool. The later will contribute in both teaching and learning of
OOP.
II. LITERATURE REVIEW
A. Evolution of classroom tools[3]
Different classroom tools and technology have been used
over more than five decades to help teachers to teach
efficiently in their classes. In 1650, the introduction of the
horn-book allowed students to follow lessons with wooden
paddles printed. In 1850, the ferrule, some kind acting as a
pointer, was being used. The “magic lantern” in 1870 allowed
teachers to project lectures onto glass slides so that students
were able to follow the class in large group. The chalkboard in
1890 had been very beneficial for many educational
institutions. It had been used for more than hundred years. In
the same year, the invention of a small chalkboard commonly
known as “the school slate” allowed students to learn concepts
more effectively. The pencil in 1900 was more widely used
when introduced compared to the school slate. The
stereoscope in 1905 allowed classroom lectures to be more
visually animated. The filmstrip projector in 1925 allowed
students to view motion-pictures. In that same year itself, the
transmission of class lessons through the radio made a
tremendous impact for 2 decades. The mimeograph (1940),
the language lab headset (1950), the reading accelerator
(1957) , the skinner teaching machine (1957), the educational
television (1958), the liquid paper (1960), the filmstrip viewer
(1965), the hand-held calculator (1972), the scantron (1972),
the PLATO computer (1980), the cd-rom drive (1985), the
hand-held graphing calculator (1985), the interactive
whiteboard (1999), the iclicker (2005) ,the XO laptop (2006)
and ipad (2011) are tools which had contributed a lot in the
progress of teaching and learning. Software programming
tools are now being run on ipad or laptops to teach
programming. Today, learners can eventually learn OOP
through these devices.
B. Evolution of software OOP teaching tool
With the aim of supporting students in understanding the
concept of programming, LOGO concept [11] was conceived
in 1967.The latter supported learning and formal thinking to
solve problems by using micro-chips to build small robots.
Instructions were given to these micro-tools to motivate
learners to grasp the concept of programming [11]. After logo,
the most used software tool which contribute to the teaching
and learning of computer programming was “Karel the robot”,
developed by Richard E.Pattis [12]. Xinogalos [13] made
another version of karel (objectKarel) which was used to
introduce OOP to first year university students. ObjectKarel
made a positive impact on novice learners. The latters were
able to overcome their difficulties in understanding OOP
concepts. They were more motivated with objectKarel.
However, objectKarel was not being used on the job market. It
was not a professional programming language. [13]
Further researches done on programming educational tools
were developed to help teachers deliver quality standards
programming classes. These tools have been built to motivate
students so that they do not drop programming classes [14].
These software tools have been categorized into 5
categories: narrative tools, visual programming tool(static and
dynamic visualization), flow-model tools, specialized output
realizations and tiered language tools [15] [16]. Narrative tools
allow the user to follow some story telling tutorials to learn
programming. In the end, the user shall be able to build the
story in the software. Visual programming tools can be either
static or dynamic. Static visual programming tools do not
allow us to see the state and behavior of classes [17]. Whereas,
dynamic visual tools allow us to see the state and the behavior
of the objects [18]. However, both static and visual tools allow
the user to drag and drop objects to learn OOP programming.
Flow-model tools allow the user to connect program elements
together in such a way that the user can represent the order of
computation. Specialized output realizations provide
feedbacks in non-textual way which guides the user more
appropriately. Finally, tiered tool language allows the user to
work with only features he is allowed to, depending on his
proficiency skills. For instance, a novice user will have less
features appearing on his screen compared to an advanced
user.
Alice, an OOP educational software tool, was introduced
in 1995 [19]. Alice is a 3D tool. The user is not required to
understand 3D to use the software. The user will drag and
drop objects to code visually and graphically. He does not see
the real java codes. [20, 21]. Alice aim was to bring virtual
reality for students to work in group [19]. It has tremendously
ease learning process of novices [8]. Teachers have also found
the tool to be effective for them. However, learners had
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
complained that they would prefer to see the real java codes so
that they would be more accustomed to industry coding [8].
BlueJ [22, 17] in 1996 allowed students to learn OOP in a
more structured way. One of the strength of blueJ is the clear
separation of the concepts of OOP classes and objects and the
possibility to interact with and analyse these objects and
classes. One other of its strength is its simple user interface
which makes it very suitable for novice learners. A research
[22] conducted in 2007 at the University of Piraeus in Greece
favored blueJ for its simplicity. The simplicity of the software
allowed students to be more engaged to learn OOP. However,
the research highlighted some disadvantages with its usage.
One of them has been the fact that BlueJ is a static visualizing
tool. This means blueJ does not allow us to monitor the state
and behavior of objects. This has limited the teaching and
learning process of OOP.
Dynamic visualization software tools allow us to see the
state and behavior of objects. Both JPIE [23] and Greenfoot
[18] [24] are dynamic visualizing tools. Throughout years,
various experiments [25] have been carried out to test the
effectiveness of Greenfoot in teaching and learning of OOP.
Greenfoot still has many limitations which prevent novices to
self-learn. They desperately need assistance for complicated
classes [25]. Other educational software tools also have
brought out new features which have been contributing to the
ease of the learning process of OOP. Jeeroo is a multi
language tool and narrative tool [26]. However, scratch is not
multi language but is a narrative tool as well [27]. Jeeroo had
reduced withdrawal rate of programming class in the USA.
Students maintain a high interest in learning OOP with the
introduction of Jeeroo to them [28,32]. Baltie displays
components on the screens as the user develops his
proficiency skills (tiered language). For instance, for a novice
user, lesser components will be displayed compared to an
expert user. JHAVE displays pop-up questions and
instructions to guide the user throughout his learning process.
Therefore, this feature makes it a specialized output
realizations software tool type [26]. Raptor is a flow-model
software tool. [26]
III. ANALYSIS
A. Analysis of existing software tools
Alice [20, 21] being a 3D tool is good for visual
programming. However, 3D is seen to be strainful for the eyes
of the users. For a novice user, the first-time exposure should
be a 2d-image [29]. Upon the level of expertise of the user, the
option 3-D can be made available.
BlueJ being a simple software object-oriented teaching
tool still lacks both animated feedback (specialized output
realization features) and tiered-language features (displaying
components according to user expertise level). The same
improvements for both JPIE and Greenfoot would be
recommended as suggested for BlueJ. Baltie’s main strength is
that it allows users to learn concepts only at their proficiency
level. Unfortunately, baltie is not free. We need to purchase
the software to acquire its full version [30] [28]. JHave [26]
allows users to have feedback as a guide to be able to build the
software correctly. However, it lacks tiered language features.
Raptor has flowchart diagrams as a way to teach object-
oriented [26]. Class-diagrams were more preferred according
to our survey.
As it can be seen, each of the existing software tools lack
one or some particular features which are essential for a
novice to learn object-oriented programming at his own pace.
The following table shows us a comparison of the nine
existing software tools having previously discussed in section
2. All the tools in the table below allow users to visualize their
features either statically or dynamically. Out of the nine tools,
three of them (Alice, Jeeroo and Scratch) are of narrative
type. Out of the six remaining tools, two of them use class-
diagram to teach object-oriented , one uses uml diagrams and
one uses flowchart diagram. Jhave is a specialized output
realization tool where users are able to received feedback and
TABLE I Comparison analysis of each software tool existing features
As mentioned in the literature review section, software
tools have been categorized into 5 categories: narrative tools,
visual programming tool (static and dynamic visualization),
flow-model tools, specialized output realizations and tiered
language tools [15]. According to figure 4, Alice, Scratch and
Jeeroo are categorized as a narrative and visualize tool.
Moreover, Jeeroo is multi-language as well. BlueJ, Greenfoot
and Raptor are visual and flow-model tools. JPIE is a visual
tool. Baltie is a visual and tiered-language tool. Lastly, Jhave
the specialized output realization tools guiding the users
through their learning process. As it can be seen, different
functionalities are being researched, implemented, assessed in
order to give as much supports to the learners as possible. No
software seems to have given the effective support to the
novice learners yet.
B. Methodology
A qualitative analysis was done in order to identify the
problems that novice learners faced in understanding OOP. An
online survey was carried out. Thirty two surveyees of age
between 18 to 28 participated in this survey. Some of them
already learnt OOP some years ago and some of them are
beginners to OOP. The survey consisted of four sections: the
profile of the surveyee , his motivation to learn OOP, his and
others’ difficulties faced in learning OOP and the tools they
guidelines while using the tool to learn OOP. Baltie is a tiered
language tool where the user is able to work with different
graphical user interface according to his proficiency level.
Jeeroo allows users to learn OOP in different programming
languages(Java, c ++ and so on).
Tools consisting of a blank cell do not contain that particular
feature(characteristic) which is in row 1 of the blank cell
column. For example, Alice does not contain flow-model,
specialized output realization, tiered language, multi-language
features and it does not have a simple user interface. The
designing aim of each of these software are different from
each other. Hence, each of them has different focussing areas.
For example, Alice main focus was to allow users to learn
games. Whereas, blueJ main focus was to allow users to use a
simple user interface.
have been using or using while learning OOP for the first
time.
In the profile section (first), data collected allowed us to
identify whether he is a male or female respondent. Morever,
it allows us to know the age and the highest qualitication of
the person. The variables in the profile section allowed us to
be able to know whether the person filling the form is an
expert or a beginner to OOP.
In the motivation section (second), the data collected
allowed us to know whether the surveyee still does want to
learn OOP or he is discouraged by the complexity of OOP
concepts.
The difficulties faced with OOP section (third) allowed us
to know how the surveyee finds OOP for novice learners. He
was asked to rate the complexity of the following topics for a
novice learner: inheritance, association and polymorphism.
The fourth section, allows us to identify those software
which are currently being used by most people. It was seen
that most of people who do not like OOP have been using
advanced IDE when they first started learning OOP. These
tools were complicated for them to work with and to learn
OOP.
The approach taken was to enhanced blueJ because it has a
simple user interface compared to others. Furthermore,
findings from both the survey and background study
concluded that class-diagram is a better choice for learning
OOP. BlueJ consists of class-diagrams which look simple to
users. As blueJ is the simplest tool among all the existing
ones, it was easier to progressively add features to it. This will
ultimately allow us to achieve our goal of displaying features
according to user’s proficiency level. Newer is the learner
with OOP concepts, the lesser features will be displayed to
him. Easier is the learner with OOP concepts, more features
will be displayed to him.
Fig. 3. rating for introducing Polymorphism to novices

C. Results of the findings

Programming is not an easy task. We should not only do Furthermore, OOP learners were asked to choose among 4
the efforts of understanding the problem to solve, but we flow-model diagrams (flowchart, use-case diagrams, class
should also structure it and code it in defined syntax. This diagrams and state diagrams) the most effective for a novice
complexity of programming is found to be difficult by learner to grasp OOP concepts. 81% of the responses opted for
students. Students like to code. 27 (84%) out of 32 people class-diagrams as a flow-model for a beginner level course.
surveyed, like object-oriented programming, 26 (81%)
respondents find object-oriented programming difficult for
novice learners. 16% of the students claimed that they do not
like OOP and they find it difficult to learn.
Concepts such as polymorphism, inheritance and
association are found to be complex for novice learners. Our
survey required respondents to rate the level of complexity of
each of these OOP concepts. The rating set (1-3) was of
increasing difficulty. Rating 1 meant the concept was easy.
Rating 2 meant the concept was fairly difficult. Rating 3
meant the concept was very difficult to be taught to a novice.
15 people found inheritance relatively simple compare to
association and polymorphism. Association is seen to be more
difficult compared to inheritance. It is perceived to be less Fig. 4. flow-model chosen: class-diagrams
difficult compared to polymorphism. Finally, Polymorphism is
seen to be a very difficult topic to be introduced to a novice.
The following 3 figures show the result of the survey carried The learners were asked to select their preferences from a
out: set of 8 characteristics. These characteristics can be used to
improve existing software tools features. They were found to
be important to novice learners in the literature review section.
The following table illustrates the 8
characteristics(narrative, visualization, flow-model,
specialized-output realization, tiered language, the
programming language taught should be use in the industry,
has a debugger, has simple user interface) which can be used
to enhanced existing software tools.
TABLE II. Showing the 8 Characteristics (features) which can be used to
enhanced existing tools.
Fig. 1. Rating for introducing inheritance to novices
Characteristics Acronym Chosen
(features) percentage
(%)
1. Narrative F1 3
2. Visualisation F2 5
3. Flow-model F3 16
4. Specialised- F4 17
output
Fig. 2. Rating for introducing association to novices
realisation
5. Tiered language F5 17 Static visualization of inherited classes, dynamic visualization
of inherited classes, static visualization of associated classes,
6. The F6 12 dynamic visualization of associated classes, static
programming visualization of polymorphism concepts and lastly, dynamic
language taught visualization of polymorphism concepts. These 6 learning
should be used topics will be the 6 proficiency user levels of the software to
in the industry be built. Level 1(static visualization of inherited classes) will
7. Has a debugger F7 9 be for a novice. Level 6 (dynamic visualization of
polymorphism concepts) will be for an expert.
8 Has simple user F8 21
. interface for For instance, at level 1, novice learners will be able to
ease of view only inheritance components displayed on their graphical
interaction user interface of the software to be built. The following figure
illustrates the options to be displayed at level 1:

Fig. 5. Piechart representing the percentage of preferences regarding the 8
features which are seen to be vital according to the literature review
(distribution of data from table II).
It was concluded that most of the surveyees wanted a
simple user interface tool, with tiered-features possibility,
flow-model and specialized output realization features. 21%
preferred a simple user interface for ease of interaction. 17 %
would like to be guided during their learning process by
having feedbacks, assessment to test their level of proficiency
and to be able to move to more advanced OOP concepts
progressively. The programming language taught should be
used in the industry.
IV. SOFTWARE TOOL PROPOSED
As it can be seen, each of the existing software tools lack
one or some particular features which are essential for a
novice to learn object-oriented programming at his own pace.
The solution proposed is to build a software OOP educational
tool combining all the seven functionalities which have been
discussed earlier in this paper. Therefore, the software which
is to be built should be narrative, visual, flow-model,
specialized output realization, tiered language and multi-
languages tool with a simple user interface. The programming
language taught should be used in the industry.
Fig. 6. Level 1 functionalities(features)
Novice learners will be able to create inherited (parent-
child) class-relationship between class-diagrams, define
methods and variables of each class. Tutorials can be in the
form of text and images. These tutorials should be part of the
software to be built. Furthermore, the user should be given the
choice to choose between narrative style tutorials or some
other types of tutorials. The users will be allowed to play
around graphically with the class-diagrams while following
these tutorials. Upon creation/modification of each class-
diagram and their respective relationship with other classes,
real programming code will be generated in the back-end of
the software tool. The user will ultimately have an option
where he can view and modify these codes in real-time. If
name or method or variable of a class is to be changed in the
code, this change will be reflected on the class-diagram. This
is often known as static visualization. If a user wishes to move
to another level of proficiency, he will have to click on the
“change level button” as shown in the following figure 7.
Upon clicking on the “change level button”, the learner will be
asked to answer a set of questions. The latter will allow the
software application to assess the user. In this way, if the user
fails the test, the application will not allow the user to proceed
to further topics.

A. Description of the software proposed
Important OOP topics are inheritance, association and
polymorphism. The application to be built will sub-divide
these topics into 6 topics for the learning process of a user.
The following figure illustrates one of the e-assessment
questions to test the user knowledge before allowing him to go
to the next level:
 Fig. 7. E-Assessment of OOP concepts by typing java codes in the textboxes and validate it by using the “Next” button.
The application interacts with the user by providing him
feedbacks. If the user successfully passes the tests, he will be
allowed to proceed to level 2. Level 2 will be where a user
will be able to view the state and behaviors of each class
dynamically. The following figure shows the option at level 3.
The user is allowed to learn both inheritance (2nd button) and
association (4th button)”:
Fig. 8. Level 3 functionalities(features)
The tiered language feature will allow the user to switch
between the 6 proficiency levels of the application. It is
recommended to avoid cognitive overload as the brain
memory is limited. We need to keep cognitive load at a
manageable level as far as possible [31]. Feedbacks, tutorials
and the assessments carried out form part of the specialized
realization output features.
V. CONCLUSION
According to the survey carried out among students,
96.9% agrees that everybody can learn OOP concepts with an
appropriate software tool. Therefore, this paper has proposed
the most promising tool to contribute in the teaching and
learning of OOP. The tool is a suitable solution aimed at
beginners. It will combine all existing features found in
existing software tools in an effective way. No existing tools
have yet considered combining existing functionalities to
promote learning and teaching. Therefore, the software being
proposed in this paper will attempt to teach and learn OOP
progressively. It will be equipped with simple features such as
a simple graphical user interface (GUI) according to user-level
expertise, simple class-diagrams and pop-up messages
providing feedback on the user’s performance.
An OOP course should be set to evaluate the software with
novice learners. A survey should be carried out at the end of
the course to receive feedbacks from these students.
The tool having proposed in this paper caters only for OOP
topics such as inheritance, association and polymorphism.
Other OOP topics such as encapsulation can be included in the
application. 3d modeling was not considered in this paper.
Future work may consider switching between 2d and 3d user-
level to maximise learning growth.
Acknowledgment
We thank all the surveyees who have spare some of their
free-time to answer questions in the survey forms. Without
them, it would have been very difficult to validate the gap of
knowledge found in the literature review section.
REFERENCES
[1] J. Pritchard, "Introduction to learning and teaching," 2008.
[2] J J. Deneen, 2013. [Online].
Available: http://www.opencolleges.edu.au/informed/news/teachers-or-
facilitators-10-reasons-why-educators-should-step-out-of-the-way-and-
encourage-independent-learning/. [Accessed 01 07 2015].
[3] S. Gervais, "The Evolution of Educational Tools and Technology,"
2012. [Online]. Available:
https://marcplamondon.nipissingu.ca/wiki/2011_Education.The-
Evolution-of-Educational-Tools-and-Technology.ashx. [Accessed 20 06
2015].
[4] M. Kölling, "The problem of teaching object-oriented programming,"
1999
[5] S. M. Biju, "Difficulties in understanding object oriented programming
concepts," University of wollongong research online, 2013.
[6] E. Janke and s. Wagner, "Does Outside-In Teaching Improve the
Learning of Object-Oriented Programming?," 02 2015.
[7] j. livovsky, m. binas and j. poruban, "Teaching Object-oriented
programming with object benches: practical experience," 2013.
[8] A. Allinjawi, H. Nuaim and P. Krause, "Evaluating the effectiveness of a
3D visualisation environment while learning object oriented
programming," Journal of Information Technology and Application in
Education, vol. 3, 2014.
[9] W. H.Throw and E. A.Smith, "Filmstrip techniques for individualised
instruction".
[10] W. clark and R. Luckin, "ipads in the classroom," 2013.
[11] C. Solomon and S. Papert, "A case study of a young child doing turtle
graphics in logo," in national computer conference and exposition, 1976.
[12] R. E.Pattis, Karel the Robot: A Gentle Introduction to the Art of
Programming, vol. 1st edition, New York: John Wiley & Sons, 1981.
[13] S. Xinogalos, M. Satratzemi and V. Dagdilelis, "An objects- first
approach to teaching object orientation based on object Oriented
Programming," in In Proceedings of the 5th WSEAS Interna- tional
Conference on Education and Educational Technology (EDU’06), 2007.
[14] F. E. Castillo-Barrera, P. D. Arjona-Villicana, C. A. Ramirez-Gamez, F.
E. Hernandez-Castro and S. M. Sadjadi, "Turtles, Robots, Sheep, Cats,
Languages what is next to teach programming? A future developer's
crisis?," in International Conference on Frontiers in Education:
Computer Science and Computer Engineering, 2013.
[15] K. Powers, S. Cooper, K. J.Goldman and M. Cartlisle, "Tools for
teaching Introductory Programming," Proceedings of the 37th SIGCSE
Technical Symposium on Computer Science Education, pp. 560-561,
2006.
[16] S. SILESSI, H. VAROL, O. KOLAWOLE, D. KEY and B. HOUCK,
"Dragon Drop Pictorial Programming," 2013.
[17] M. Kolling and J. Rosenberg, "BLUE-A language for teaching object-
oriented programming," n Proceedings of 27th SIGCSE Technical
Symposium on Computer Science Education, ACM, , pp. 83-87, March
1996.
[18] P. Henriksen and M. Kolling, "Greenfoot: Combining Object
visualisation with interaction," Conference on Object Oriented
Programming Systems Languages and Applications, 2004.
[19] R. Pausch, T. Burnette, A. Capeheart, M. Conway, D. Cosgrove, R.
DeLine, J. Durbin, R. Gossweiler, S. Koga and J. White, "Alice: Rapid
Prototyping System for Virtual Reality," IEEE Computer Graphics and
Applications, vol. Volume 15 , no. 3, pp. 8-11, 1995.
[20] W. Dann, D. Cosgrove, D. slater, D. Culbya, S. Cooper and L. Paoletti,
"Alice 3-How to guide," 2012.
[21] W. Dann and S. Cooper, ""Education Alice 3: Concrete to Abstract","
Communications of the ACM 52.8, 2009
[22] S. Georgantaki and S. Retalis, "Using educational tools for teaching
object oriented design and programming," Journal of information
technology impact, vol. 7, pp. pp 111-130, 2007.
[23] K. J.Goldman, "An interactive Environment for Beginning Java
Programmers," Washington University Department of Computer Science
and Engineering, 2003.
[24] M. Haungs, Creative Greenfoot, Packt Publishing, 2014, p. pg107.
[25] M. Al-Bow, D. Austin, J. Edgington, R. Fajardo, J. Fishburn, c. lara, s.
leutenegger and s. meyer, "Using Greenfoot and Games to Teach Rising
9th and 10th Grade Novice Programmers," 2004.
[26] D. Tebring, "Using Introductory Programming Tools to teach
programming concepts: a literature review," The Journal for Computing
Teachers, 2009.
[27] MIT, "About scratch," 2003. [Online]. Available:
https://scratch.mit.edu/about/. [Accessed 29 August 2015].
[28] D. Sanders and B. Dorn, "Object-Oriented Programming with Jeroo in
the Information Technology classroom," 2004.
[29] D. Sanders and B. Dorn, "Object-Oriented Programming with Jeroo in
the Information Technology classroom," 2004.
[30] SGP, "Educational programming tools," 1978. [Online]. Available:
https://www.sgpsys.com/en/download.asp. [Accessed 29 08 2015].
[31] M. Kolling, "The greenfoot programming environment," ACM
Transactions on Computing Education, vol. 10, no. NO 4, 2010.
[32] D. Sanders and B. Dorn, "Jeroo: A tool for teaching object-oriented
programming," In SIGCSE '03: Proceedings of the 34th SIGCSE
Technical Symposium on Computer Science Education, pp. pages 201-
204, 2003.
 Design and Implementation of an Automatic Staff
Availability Tracking System
Kyle Stone, Jan Spies
Computer Systems Engineering
Tshwane University of Technology
Pretoria, South Africa
StoneK@tut.ac.za, SpiesJJ@tut.ac.za
Abstract—This paper presents the design and implementation
of an automatic lecturer availability tracking system
incorporating various software and hardware technologies. The
system is aimed at reducing the amount of time that students,
visitors and faculty members spend looking for specific
employees without being hindered by the ethical issues
surrounding RFID based staff tracking. It focuses on the
availability of the staff members within their allocated office
space. The system is currently being tested and is performing
well enough to provide clients and other staff members with real-
time information regarding staff member availability and
essential contact information.
Keywords—Staff tracking; Staff availability; Privacy;
I. INTRODUCTION
There are many systems available for tracking staff
members. A good percentage of them involve the personnel
wearing RFID (Radio-frequency identification) tags that are
scanned as they move from one part of a building to another.
This has brought about ethical issues regarding the privacy of
the staff members being tracked [1]–[4]. This proposed system
aims to bypass these privacy issues by focusing on the
availability of the staff member in their physical office space or
lab space. It does not factor in the movement of staff members
once they leave their offices.
II. OBJECTIVES
The system needs to track staff members unobtrusively in
an attempt to provide other staff members and visitors with an
easy means of obtaining information regarding staff members’
whereabouts and availability. The system needs to be
accessible via a web interface for easy, ubiquitous access. It
also needs to be operated in an automated fashion with as little
input from the staff members themselves as possible. This will
be achieved by installing sensors that monitor the doors of staff
members. The sensors should be able to give data about the
state of the door – either open or closed. This data should then
enable the system to give information about the availability of
the staff member concerned. A means of visually representing
the availability of staff members needs to be designed and the
representation should conform to both mobile and desktop
platforms. This visual representation should then be displayed
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
in the lobby or waiting area of the office so that clients can
easily identify the availability of the staff members that they
are looking for. Changes in staff member availability need to
be reflected client-side in real-time.
III. TECHNOLOGY DESCRIPTION
A. Software
 Apache: This serves as the web host for all the web
pages used by the system.
 PHP (Hypertext Preprocessor): For the web component
of the project, PHP will be used. PHP was chosen due
to it being open source and having very good
performance statistics when compared to other web
service engines [5].
 MySQL (My Structured Query Language): MySQL
was chosen as the database system to be used as it is
one of the most successful database systems used
worldwide [6]. It also performs well when compared to
other database systems even when hosted on a limited
resource server [6] and can integrate easily with a PHP
based project
 JavaScript: JavaScript is used for client-side scripting
and performs the visual changes with the help of PHP
and AJAX.
 AJAX (Asynchronous JavaScript and XML): AJAX
was chosen so that the web interface could be updated
dynamically without having to refresh the entire page
every time a change in staff availability was detected.
 HTML (Hyper Text Markup Language): Chosen due to
its simplicity and the fact that it is supported on such a
wide range of devices.
 C#: This was chosen as the platform for a desktop
application that allows staff members to make manual
changes to their availability status when needed.
B. Hardware:
 Magnetic Switches: These switches are inexpensive
and provide clear binary data on the status of the door.
  Arduino Ethernet Board: The Arduino Ethernet board
is what gives our system the capability of transferring
the data from the staff members’ doors into an online
database for further use.
IV. SYSTEM ARCHITECTURE
The system , as shown in Fig. 1, makes use of client-server
architecture. The server hosts a database and the multiple
clients access a web server which reads from the database. The
web server waits for requests for data and when a request is
received it reads all the data from the table containing the staff
members’ availability status and basic information. This data is
then compiled into an appropriate HTML and JavaScript
display which is served to the client. Due to the display page
running simple HTML and JavaScript, it can be viewed from
almost any networked device that is capable of displaying
HTML and running JavaScript.
After the initial connection, the client only requests data
from the web server regarding a single column in the database
that lets it know if a change has been made to the data. If a
change is detected, it updates its visual data along with the
local variable that keeps track of when last a change was made.
Changes are tracked using a timer in JavaScript and can be
set to run at a custom value effective up to at least 150
milliseconds.
A number of methods exist whereby a staff member can
change his/her availability status. By opening or closing their
office door, a magnetic switch, attached to the door is
triggered. This change feeds into an Arduino Ethernet board
Fig. 2. Web Interface
which submits the state change, along with a timestamp,
through a page on the web server into the database. Another
method of making changes is via the web application where The last method for staff members to make changes to the
staff members can set or update all stored information. This system is provided in the form of a desktop application which
interface is shown in Fig. 2. allows them to change their availability status manually. The
desktop application also communicates with the web server
which then interfaces with the database server. The desktop
application interface is shown in Fig. 3.
V. VISUAL REPRESENTATION
For the visual representation of the staff member
availability and basic information, a basic HTML interface
needed to be designed. Various iterations were tried and
ultimately the interface shown in Fig. 4 was decided.
Each staff member is assigned a card that contains the
information required to get into contact with them as well as a
customizable message. Each card is color-coded using one of
three colors: Red indicates unavailable, green indicates
available and orange indicates busy.
All the cards are then displayed in an HTML table on a
single kiosk that will be available in the lobby (Fig. 4). This
display is dynamically resized according to the resolution of
the screen it is viewed on, allowing it to be viewed across
multiple devices.

Fig. 1. System Architecture

 Responsiveness of system
25
20
15
10
5
0
200ms 150ms 100ms 50ms
Intervals Intervals Intervals Intervals

Number of changes detected per 20 changes

made

Fig. 5 System Responsiveness

Fig. 3.Desktop Interface
VI. SYSTEM RESPONSIVENESS
A. What was assessed
Response time was measured in terms of how frequently
the system can check for a change in the database and then
successfully update the visual representation of that data on the
client’s side.
1) Variables:
The JavaScript timer was set to run at different intervals for
each test case.
2) Method used:
A web page was set up that automatically made 20 changes
to a staff member’s status in five seconds. The view on the
client’s side was then observed and the number of changes that
occurred were tabulated.
3) Results
From looking at Fig. 5. it can be seen that changes will be
successfully reflected on the client’s side within at least 150
milliseconds. This is fast enough to meet the systems objective
of being real-time.
VII. CONCLUSION
The system successfully manages to provide availability of
staff members in an unobtrusive way. The staff members also
have ultimate control over what availability status is selected,
thereby eliminating the ethical problems associated with
monitoring a user by means of RFID. The system is available
via the web and is therefore easily accessible to anyone with a
device capable of loading an HTML web page. For real-time
feedback they need a JavaScript enabled browser. A visual
representation was designed in HTML that showed the
availability of staff members along with their contact
information. The HTML was designed to conform to the screen
size of the device that it is viewed on, and is therefore viewable
on both mobile and desktop platforms.
VIII. REFERENCES
[1] S. Kurkovsky, E. Syta, and B. Casano, “Continuous RFID-enabled
authentication: Privacy implications,” Technology and Society
Magazine, IEEE, vol. 30, no. 3, pp. 34–41, 2011.
[2] G. Kaupins and R. Minch, “Legal and ethical implications of employee
location monitoring,” in System Sciences, 2005. HICSS’05. Proceedings
of the 38th Annual Hawaii International Conference on, 2005, p. 133a–
133a.
[3] V. Samaranayake and C. Gamage, “Employee perception towards
electronic monitoring at work place and its impact on job satisfaction of
software professionals in Sri Lanka,” Telematics and Informatics, vol.
29, no. 2, pp. 233–244, 2012.
[4] M. Workman, “A field study of corporate employee monitoring:
Attitudes, absenteeism, and the moderating influences of procedural
justice perceptions,” Information and Organization, vol. 19, no. 4, pp.
218–232, 2009.
[5] T. Suzumura, S. Trent, M. Tatsubori, A. Tozawa, and T. Onodera,
“Performance comparison of web service engines in php, java and c,” in
Web Services, 2008. ICWS’08. IEEE International Conference on, 2008,
pp. 385–392.
[6] M. Ahmed, M. M. Uddin, M. S. Azad, and S. Haseeb, “MySQL
performance analysis on a limited resource server: Fedora vs. Ubuntu
Linux,” in Proceedings of the 2010 Spring Simulation Multiconference,
2010, p. 99.

Fig. 4. Lecturer Information

 Proposed Model for Evaluation of mHealth Systems
Chel-Mari Spies
Department: Computer Systems Engineering
Tshwane University of Technology
Pretoria, South Africa
spiesc@tut.ac.za
Abstract—Advances in mobile health (mHealth) has been
notable in the last few years, but so has the problems associated
with the development, implementation and sustained use of
mHealth systems. After implementation, various difficulties arise
contributing to being considered a failure. The purpose of the
study is to create a model by which e-Health systems can be
evaluated in order to ensure development for sustained use, with
attention focused on aspects identified through literature. The
author examines problems identified in previous research to
establish difficulties and shortfalls regarding the perceived
success of said systems. Patient and healthcare practitioner
points-of-view, along with software and hardware considerations
are taken into account. The investigation determines that the use
of IT in mHealth is still dependent on serious factors influencing
the realization of success of well-established as well as newly
developed systems. These concerns undermine the effectiveness
and usefulness of the improvement and efficiency of healthcare
facilities and knowledge as a whole. The author proposes a model
against which developed and developing mHealth systems can be
measured in order to promote continued usage of implemented
systems. This study sheds new light on little recognized issues for
bringing concept and practice together in a useful, uninterrupted,
and unending manner.
Keywords—mHealth; system design; system evaluation
I. INTRODUCTION
There has been a steady increase in the use of IT in the
health sector. Hospitals and healthcare practitioners have
come a long way from simply entering patient data onto a
computer. The implementation of information technology
takes on many forms – from taking simple measurements as
implemented in electronic healthcare systems (eHealth), to
more complex systems and incorporating wireless technology
in mobile health (mHealth, to wearable and implantable
devices as described in personalised electronic healthcare and
medicine (pHealth).
Through the involvement of healthcare practitioners and
engineers alike, many systems have been designed and
implemented with various degrees of success. In the instances
where the continuous use of such systems failed, various
reasons have been cited as to why this was the case.
It becomes clear from the literature referred to throughout the
paper that no single person, institution, or aspect can be
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
blamed, but instead a collaboration of culprit issues have been
identified.
In this study, the author aims to identify the problematic areas
and propose a solution in the form of a model that can be used
throughout the design and implementation phases to try to
improve sustainability in terms of implementability and
continued use.
A. Background
Thorough research for development and implementation
has been done by various companies, institutions and/or
individuals in order to develop useful systems which aims to
improve service delivery, ensure ease-of-use for patients, and
lighten the load of healthcare practitioners, but despite best
efforts, most systems are still open to improvements. Studies
show that problems regarding various human and
technological aspects are still experienced and the evidence
shows that many flaws can be avoided by taking a few things
under advisement during the design phase.
In this study, the difficulties are studied from various
points-of-view: patient, healthcare practitioner, and technical
(which has been divided into software and hardware).
Problems experienced by the patient are consumer
oriented: users are concerned with comfort and ease-of-use.
Healthcare practitioners have different needs as they are
focused on information that can be collected from the system.
When considering the technical (or design) side, it
becomes clear that a distinction must be made between the
hardware and software aspects. The software must be accuracy
and security driven, and the hardware scheme must consider
user needs.
II. PROBLEMS FACING IMPLEMENTATION OF MHEALTH
SYSTEMS
Bearing in mind previous research and case studies, it
becomes evident that the following issues poses major
concerns not only in the design of new tools, but also when
sustained use is intended:
A. Patient point-of-view
From a user outlook, the tool must be unobtrusive to
ensure comfort. It must also be easy to learn to use and the
procedural process should be uncomplicated. The system will
most probably contain a wearable component which must be
considered correspondingly.
The following aspects have been identified from literature:
 Acceptance – Patient perception is of utmost importance
when it comes to predicting success of a tool and the
patient point-of-view must be considered throughout the
development process [1].
 Convenience – The tool must be convenient for the user
to wear [2], easy to learn and uncomplicated to use [3].
 Discretion – When the user is in the process of using the
device, it must be unobtrusive [4], i.e. hard to detect by
other people.
 Usefulness – The issue of usefulness can be very broad
(from a healthcare perspective), or it can be very specific
(from a patient perspective). Examples of tools that were
specifically developed include sending reminders [5],
Parkinson‟s Disease induced disability monitoring [6],
and alcohol usage treatment [7].
 Wearability – If the device is of such a nature that it must
be worn on the user‟s body, it must be comfortable[8],
lightweight [9], and allow for maximum patient mobility
[10].
B. Practitioner point-of-view
Healthcare practitioners, whether physicians, nurses or
other, need fast, accurate and reliable information from the
system in order to make quick, effective decisions regarding
patient care. They are mainly focussed on:
 Acceptance – The perception of the healthcare
practitioner is influenced by several factors, including
perceived ease of use, perceived usefulness, perception
other people has, and personal initiative according to a
study on a Mobile Clinical Decision Support (mCDS)
system conducted at a UK Teaching hospital [1].
 Activity monitoring – Healthcare practitioners need to be
able to monitor both patients who are stationary and
patients who are not [11].
 Quality of service – From a practitioner‟s perspective, the
intended use should be easily achievable, and the system
should provide reliable data [12].
C. Hardware design point-of-view
Hardware must undeniably be designed with the user in
mind, since the system can consist a multitude of sensors,
nodes and communicating devices. Desired characteristics
focus on:
 Compatible architecture – The specifications of the
different hardware components should be examined
throughout the development of a new system. The
components, even from different suppliers, should be able
to integrate seamlessly, or employ an additional layer to
allow hindrance-free operation [13].
 Electromagnetic compatibility – Devices should be able
to function correctly and efficiently without being
susceptible to electromagnetic signals from other devices.
If such interference is experienced, crucial and non-
crucial medical equipment may not yield the desired
outputs and this will affect measurements, diagnosis and
treatments [14].
 Mobility – mHealth systems are made more favorable and
less obtrusive when patients are have a certain amount of
mobility. It is therefore a motivating point to include this
feature in a system [15].
 Power – Power constraints are clear when the sizes of
components in a wireless environment is considered,
therefor it is imperative that attention is paid to power
consumption of the components that are used [16].
 Physical parameters – Weight and size are important
considerations. Keeping these measurements as low as
possible will be more power efficient, more cost effective
on a large scale and will benefit the patient in terms of
comfort [9].
D. Software design point-of-view
Requirement of the design of software will have to
stipulate the needed outcomes concerning:
 Communication – Communication protocols, such as Wi-
Fi and Bluetooth [14], and hand-over protocols [17] are
key considerations in a IT system.
 Data management – Good data quality [18], efficient
exchange [12] and effective data management has been
proven to have multiple advantages.
 Functionality – The desired functions, such as movement
monitoring, will dictate the design and ultimately the
information protocols [19] that will be implemented to
reach definitive functionality.
 Interoperability – The system must allow the use of
various types of sensors [10] and employ appropriate
techniques to allow the system to cope with the
integration of different standards [20].
 Network management – Network management, and
especially node authentication facilities, should be
provided for [21].
 Patient identification – The system needs to be able to
identify a patient through certain inputs, whether
manually fed in, sensed automatically (e.g. by means of
RFID tags [14]), or via the decryption of physiological
parameters [22].
 Processing capabilities – Software efficiency should be a
major concern in the design of the system. Computing
power [23], sampling rate [9], and delay time [24] are
focal issues.
 Table 1. Aspects identified from patient point-of-view
Aspect Features Standardising body, number & description
“ISO 10004:2012 Quality management – Customer satisfaction – Guidelines for
Acceptance Perception
monitoring and measuring” [25]
“DIN ISO 20282-1:2008 Ease of operation of everyday products - Part 1: Design
Learning curve
requirements for context of use and user characteristics” [26]
Convenience
“ISO/TR 16982:2002 Ergonomics of human-system interaction - Usability methods
Usability
supporting human-centered design” [27]
“ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411:
Discretion Unobtrusive
Evaluation methods for the design of physical input devices” [27]
Useful i.t.o. medical “IEC 62366 Medical devices – Application of usability engineering to medical
Usefulness
condition devices” [28]
“ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411:
Comfortable
Evaluation methods for the design of physical input devices” [27]
“ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411:
Wearability Lightweight
Evaluation methods for the design of physical input devices” [27]
“ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411:
Allows mobility
Evaluation methods for the design of physical input devices” [27]

 Security – Data in a health system must be protected from  “Institute of Electrical and Electronics Engineers
malicious attacks and disclosures and patient information Standards Association (IEEE)” [31]
must be protected at all cost [29].  “International Electrotechnical Commission (IEC)
 Sensing methods – Various sensing methods have been – International Electrotechnical Commission –
implemented in developed systems in the past. Examples Technical Report (IEC/TR)” [32]
include electrocardiogram (ECG) electrodes using  “International Organization for Standardization (ISO)
textiles, global positioning (GPS) system modules, and – International Organization for Standardization -
temperature sensors that are optical fiber based [30]. Technical Report (ISO/TR)” [33]
 “German Institute for Standardisation (DIN – “Deutsches
Institut für Normung”)” [34]
After consideration of the identified concerns, it becomes
clear that there is a need for a single model that can be used to In the tables, the standardising body and standard number
scrutinize the projected efficiency of a developed system, is given as identification method of each standard. The
together with its accompanying tools. description (printed in Italic script) of each is the indorsed
This paper proposes that the evaluation problem be solved name of each standard.
by means of identifying endorsed methods of standardization.
A. Patient point-of-view
III. UNIVERSAL STANDARDS AS BASIS FOR THE PROPOSED Aspects in this section cover acceptance by the patient,
MODEL convenience for the patient, discretion when the patient is
wearing or using the device, the degree of perceived
After considering the aspects described in the previous usefulness, and wearability in terms of comfort for the patient.
paragraphs, judgement methods, and official engineering (and A summary is given in Table 1.
other) standards were identified as basis for the proposed
model.
Tables 1 to 4 show existing standards that serve as B. Practitioner point-of-view
guidelines in development of tools, systems and other related Healthcare practitioners consider ease-of-use, usefulness,
aspect in the development and implementation processes. the perception of others, activity monitoring and the quality of
the service that the system yields to be among the most
Standards from the following bodies have been included: important aspects, as shown in Table 2.
 Clinical and Laboratory Standards Institute (CLSI)
Table 2. Aspects identified from practitioner point-of-view
Aspect Features Standardising body, number & description
Ease-of-use “DIN ISO 20282-1:2008 Ease of operation of everyday products - Part 1: Design
requirements for context of use and user characteristics” [26]
Usefulness “DIN ISO 20282-1:2008 Ease of operation of everyday products - Part 1: Design
Acceptance
requirements for context of use and user characteristics” [26]
Others‟ perception Survey
Personal initiative Electronic monitoring
“ISO/IEC 18092:2013 Information technology - Telecommunications and information
Stationary patients exchange between systems - Near Field Communication - Interface and Protocol (NFCIP-
Activity 1)” [35]
monitoring “ISO/IEC TR 8802-1:2001 Information technology – Telecommunications and information
Roaming patients exchange between systems – Local and metropolitan area networks specification
requirements – Part 1: Overview of Local Area Network Standards” [36]
Intended use “ISO 10004:2012 Quality management – Customer satisfaction – Guidelines for monitoring
Quality of and measuring” [25]
service Data reliability “ISO 8000-130:2009 Data quality – Part 130: Master data: Exchange of characteristic data:
Accuracy” [37]
 Table 3. Aspects identified from hardware design point-of-view
Aspect Features Standardising body, number & description
Compatible Multi-vendor “ISO/IEC 7498-1 Information technology - Open Systems Interconnection - Basic
architecture compatibility Reference Model” [38]
“ISO/TR 21720:2007 Health informatics - Use of mobile wireless communication and
Electromagnetic Electromagnetic computing technology in healthcare facilities - Recommendations for electromagnetic
compatibility susceptibility compatibility (management of unintentional electromagnetic interference) with medical
devices” [39]
“ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411: Evaluation
Favourable
methods for the design of physical input devices” [27]
Mobility
“ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411: Evaluation
Unobtrusive
methods for the design of physical input devices” [27]
Low power “IEC 62018 ED. 1.0 B:2003 Power consumption of information technology equipment.
Power
consumption Measurement methods” [40]
“ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411: Evaluation
Weight
Physical methods for the design of physical input devices” [27]
parameters “ISO 9241-411:2012 Ergonomics of human-system interaction – Part 411: Evaluation
Size
methods for the design of physical input devices” [27]

By summarising and ordering the standards (by

C. Hardware design point-of-view
In order to set up an ideal model for scrutinizing hardware
in mHealth, compatibility of architectures from different
vendors, electromagnetic compatibility, degree of allowed
mobility, power consumption and physical parameters such as
weight and size are describing factors. See Table 3 for details.
D. Software design point-of-view
The software scheme would be dictated by characteristics
(Table 4) such as communication protocols, data quality,
exchange and management, the ability to function as intended,
components and standards interoperability, network
management, chosen patient identification methods, sensing
methods, processing capabilities, and robust security.
All standards have been chosen due to their applicability to
each concern, thus the standards summarized above forms the
basis of the evaluation method proposed in this paper. By
considering each standard individually, and in terms of an
over-all influence on the development and implementation
phases of new technologies, a well-rounded model for use in
practice can be established.
The aim is to have as few as possible standards that must
be adhered to in order to create the most desirable system
under acknowledged standards.
IV. PROPOSED MODEL FOR EVALUATION
In light of concerns emphasised by previous research, and
after making a study of recognised standards that dictate
design, operation and use of technical matters, a model can be
proposed to aid in the accurate and effective planning and
design of healthcare systems, and as a result can be used to
predict acceptance and continued use of newly implemented
systems.
standardising body and number), duplicated standards were
eliminated. The final model was compiled in the form of a list
as shown in Table 5.
17 main fields of study were identified from among the
adherence standards. Contributions made by each standard
towards the final model are indicated by means of calculated
percentages. The underwriting standards are Information
technology (contributing 14.3%), Health informatics (10.7%),
Security (10.7%), Ergonomics of human-system interaction
(7.1%), Risk management for IT networks making use of
medical devices (7.1%), Systems and software engineering
(7.1%), Data quality (3.6%), Delay and power calculation
(3.6%), Ease of operation of everyday products (3.6%),
Industrial automation systems and integration (3.6%),
Information and documentation (3.6%), Local and
metropolitan area networks (3.6%), Medical devices (3.6%),
Patient and sample identification (3.6%), Power consumption
(3.6%), Quality management (3.6%), and other areas (7.1%).
In the final model, the 17 fields of application covers the
41 initially identified problematic features in the following
manner (also see fig 1):
 Ergonomics of human-system interaction (contributing
22%)
 Information technology (9.8%)
 Ease of operation of everyday products (7.3%)
 Health informatics (7.3%)
 Security (7.3%)
 Data quality (4.9%)
 Delay and power calculation (4.9%)
 Industrial automation systems and integration (4.9%)
 Quality management (4.9%)
 Risk management for IT networks making use of medical
devices (4.9%)
 Systems and software engineering (4.9%)
 Information and documentation (2.4%)
 Table 4. Aspects identified from software design point-of-view
Aspect Features Standardising body, number & description
Communication “ISO/IEC 19794-7:2007 Information Technology – Biometric Data Interchange
protocols Formats – Part 7: Signature/Sign time series Data” [41]
Communication
Hand-over “IEEE 802.21-2008 IEEE standard for local and metropolitan area network – Part 21:
protocols Media independent handover services” [42]
“ISO 8000-130:2009 Data quality – Part 130: Master data: Exchange of
Data quality
characteristic data: Accuracy” [37]
Data “ISO 13606-1:2012 Health Informatics – Electronic health record communication –
Data exchange
management Part 1: Reference model” [43]
“ISO 15489-1:2001 Information and documentation – Records management – Part 1:
Data management
General” [44]
Ability to function “IEEE STD 1175.3 – 2004 IEEE Standard for CASE Tool Interconnections –
Functionality
well Reference model for specifying software behaviour” [45]
Component “IEC 80001-1 ED. 1.0 B:2010 Application of risk management for IT-networks
interoperability incorporating medical devices – Part 1: Roles, responsibilities and activities” [46]
Interoperability
Standard “ISO/IEEE 11073-20602:2010 Health informatics - Personal health device
interoperability communication - Part 20601: Application profile - Optimized exchange protocol” [47]
Network “IEC/TR 80001-2-3 ED. 1.0 EN:2012 Application of risk management for IT-networks
Network management incorporating medical devices – Part 2-3: Guidance for wireless networks” [48]
management Node “ISO/IEC 9798-1:2010 Information tecnology – Security techniques – Entity
authentication authentication – Part 1: General” [49]
Patient
Patient “CLSI GP33-A Vol. 30 No. 7 Accuracy in Patient and Sample Identification;
identification
identification Approved Guideline” [50]
methods
“ISO/IEC 25010:2011 Systems and software engineering – Systems and software
Software
Quality Requirements and Evaluation (SQuaRE) – System and software quality
efficiency
models” [51]
Computing power “IEC 61523-1:2012 Delay and Power Calculation Standards – Part 1: Integrated
Processing
efficiency circuit delay and power calculation systems” [52]
capabilities
Sampling rate “ISO 10303-1:1994 Industrial automation systems and integration – Product data
efficiency representation and exchange” [53]
Minimized delay “IEC 61523-1:2012 Delay and Power Calculation Standards – Part 1: Integrated
time circuit delay and power calculation systems” [52]
Secure against “ISO/IEC 15408-1:2009 Information technology – Security techniques – Evaluation
attacks criteria for IT security” [54]
Security
Secure data “ISO/IEC 27040:2015 Information technology – Security techniques – Storage
storage security” [55]
Sensing Application “ISO 10303-1:1994 Industrial automation systems and integration – Product data
methods specific methods representation and exchange” [53]

 Local and metropolitan area networks (2.4%) As a point of interest it might be highlighted that this study
 Medical devices (2.4%) does not cover telemedicine and teleconsulting, or implantable
 Patient and sample identification (2.4%) medical devices as described by pHealth. These aspects might
 Power consumption (2.4%) be included additionally in further development of the
 Other areas (4.9%) proposed model.

A. Future work V. CONCLUSION

In future work, the model might be refined further to
include alternative standards that would cover more of the
identified aspects, the result of which would be to include less
standards and to construct a more compact model.
The model could be tested against successful as well as
unsuccessful mHealth systems in an attempt to confirm its
accuracy in predicting sustainable use.
The purpose of the study is to create a model by which e-
Health systems can be evaluated in order to ensure
development for sustained use. The areas of patient frame of
reference, healthcare practitioner needs, as well as software
and hardware requirements have been the focus of
determining the optimal requirements for an ideal system.
After taking into account patient and healthcare
practitioner points-of-view, along with software and hardware
 considerations, a model is proposed which can be used to
guide developers in their focus when developing new systems
and to help executive decision makers to scrutinize
prospective systems when considering implementation in their
environments. Fields of study that are identified as most
important factors in the continued (or interruption) of use, and
should be considered are information technology, health
informatics, security and ergonomics of human-system
interaction.

VI. ACKNOWLEDGMENT
Chel-Mari Spies is the principal researcher and author of
this study. This work is partially supported in terms of
resources by Tshwane University of Technology.
The author declares no conflict of interest.

VII. REFERENCES
[1] R. Patel, W. Green, M. W. Shahzad, and C. Larkin, “Use of Mobile
Clinical Decision Support Software by Junior Doctors at a UK
Teaching Hospital: Identification and Evaluation of Barriers to
Engagement,” JMIR mHealth and uHealth, vol. 3, no. 3, p. e80, 2015.
[2] I. Kim, P.-H. Lai, R. Lobo, and B. J. Gluckman, “Challenges in
wearable personal health monitoring systems,” in Engineering in
Medicine and Biology Society (EMBC), 2014 36th Annual
Figure 1. Fields of application in final model International Conference of the IEEE, 2014, pp. 5264–5267.
[3] M. Ji, Y. Wu, P. Chang, X. Yang, F. Yang, and S. Xu, “Development
and Usability Evaluation of the Mobile Delirium Assessment App
Table 5. Adherence standards for universal design – Final proposed model
Standardising body Standard number Aspects covered Point-of-View
CLSI GP33-A Vol. 30 No. 7 Patient identification Software
DIN ISO 20282-1:2008 Convenience Patient
IEC 61523 Processing capabilities Software
IEC 62018 ED. 1.0 B:2003 Power consumption Hardware
IEC 62366 Usefulness Patient
80001-1 ED. 1.0
IEC Interoperability Software
B:2010
IEC/TR 80001-2-3 ED. 1.0 EN:2012 Network management Software
IEEE 802.21-2008 Communication Software
IEEE STD 1175.3 – 2004 Functionality Software
Data management Practitioner
ISO 8000-130:2009
Quality of service Software
Discretion
Mobility Hardware
ISO 9241
Physical parameters Patient
Wearability
Acceptance Patient
ISO 10004:2008
Quality of service Practitioner
Processing capabilities
ISO 10303-1:1994 Software
Security
ISO 13606-1:2012 Data management Software
ISO 15489-1:2001 Data management Software
ISO/IEC 7498 Compatible architecture Hardware
ISO/IEC 9798-1:2010 Network management Software
ISO/IEC 15408-1:2009 Security Software
ISO/IEC 18092:2013 Activity monitoring Practitioner
ISO/IEC 19794-7:2007 Communication Software
ISO/IEC 25010:2011 Processing capabilities Software
ISO/IEC 27040:2015 Security Software
ISO/IEC TR 8802-1:2001 Activity monitoring Practitioner
ISO/IEEE 11073-20602:2010 Interoperability Software
ISO/TR 16982:2002 Convenience Patient
ISO/TR 21720:2007 Electromagnetic compatibility Hardware
Electronic monitoring --- Acceptance Practitioner
Survey --- Acceptance Practitioner
 Based on Confusion Assessment Method for Intensive Care Unit
(CAM-ICU).,” Studies in health technology and informatics, vol. 216,
pp. 899–899, 2014.
[4] A. Pantelopoulos and N. G. Bourbakis, “A survey on wearable sensor-
based systems for health monitoring and prognosis,” Systems, Man,
and Cybernetics, Part C: Applications and Reviews, IEEE
Transactions on, vol. 40, no. 1, pp. 1–12, 2010.
[5] N. Leon, R. Surender, K. Bobrow, J. Muller, and A. Farmer,
“Improving treatment adherence for blood pressure lowering via
mobile phone SMS-messages in South Africa: a qualitative evaluation
of the SMS-text Adherence SuppoRt (StAR) trial,” BMC family
practice, vol. 16, no. 1, p. 80, 2015.
[6] W. Maetzler, J. Domingos, K. Srulijes, J. J. Ferreira, and B. R. Bloem,
“Quantitative wearable sensors for objective assessment of
Parkinson‟s disease.,” Mov. Disord., vol. 28, no. 12, pp. 1628–37,
2013.
[7] F. Muench, “The promises and pitfalls of digital technology in its
application to alcohol treatment,” Alcohol Research: Current Reviews,
vol. 36, no. 1, pp. 131–142, 2015.
[8] C. C. Poon, Y. M. Wong, and Y.-T. Zhang, “M-health: the
development of cuff-less and wearable blood pressure meters for use
in body sensor networks,” in Life Science Systems and Applications
Workshop, 2006. IEEE/NLM, 2006, pp. 1–2.
[9] J. Lӧnnblad, J. Castano, M. Ekstrӧm, M. Lindén, and Y. Bäcklund,
“Optimization of wireless BluetoothTM sensor systems,” in
Engineering in Medicine and Biology Society, 2004. IEMBS’04. 26th
Annual International Conference of the IEEE, 2004, vol. 1, pp. 2133–
2136.
[10] K. Wac, R. Bults, B. van Beijnum, I. Widya, V. Jones, D. Konstantas,
M. Vollenbroek-Hutten, and H. Hermens, “Mobile patient monitoring:
the MobiHealth system.,” Conf Proc IEEE Eng Med Biol Soc, vol.
2009, pp. 1238–41, 2009.
[11] F. Felisberto, N. Costa, F. Fdez-Riverola, and A. Pereira,
“Unobstructive Body Area Networks (BAN) for efficient movement
monitoring.,” Sensors (Basel), vol. 12, no. 9, pp. 12473–88, 2012.
[12] E. Jovanov, “Wireless Technology and System Integration in Body
Area Networks for m-Health Applications.,” Conf Proc IEEE Eng
Med Biol Soc, vol. 7, pp. 7158–60, 2005.
[13] A. B. Waluyo, I. Pek, X. Chen, and W.-S. Yeoh, “SLIM: a Secured
Lightweight Interactive Middleware for wireless body area network.,”
Conf Proc IEEE Eng Med Biol Soc, vol. 2008, pp. 1821–4, 2008.
[14] P. Yu, M. X. Wu, H. Yu, and G. Q. Xiao, “The challenges for the
adoption of M-health,” in Service Operations and Logistics, and
Informatics, 2006. SOLI’06. IEEE International Conference on, 2006,
pp. 181–186.
[15] A. Darwish and A. E. Hassanien, “Wearable and implantable wireless
sensor network solutions for healthcare monitoring,” Sensors, vol. 11,
no. 6, pp. 5561–5595, 2011.
[16] M. Kuroda, S. Qiu, and O. Tochikubo, “Low-power secure body area
network for vital sensors toward IEEE802. 15.6,” in Engineering in
Medicine and Biology Society, 2009. EMBC 2009. Annual
International Conference of the IEEE, 2009, pp. 2442–2445.
[17] M. D. Jara‟\iz-Simón, J. A. Gómez-Pulido, M. A. Vega-Rodr‟\iguez,
and J. M. Sánchez-Pérez, “Fast decision algorithms in low-power
embedded processors for quality-of-service based connectivity of
mobile sensors in heterogeneous wireless sensor networks,” Sensors,
vol. 12, no. 2, pp. 1612–1624, 2012.
[18] E. Borycki, H. Monkman, J. Griffith, and A. Kushniruk, “Mobile
Usability Testing in Healthcare: Methodological Approaches.,”
Studies in health technology and informatics, vol. 216, pp. 338–342,
2014.
[19] E. Kyriacou, P. Constantinides, C. S. Pattichis, M. S. Pattichis, and A.
Panayides, “eEmergency health care information systems.,” in
Conference proceedings : ... Annual International Conference of the
IEEE Engineering in Medicine and Biology Society. IEEE
Engineering in Medicine and Biology Society. Annual Conference,
2011, vol. 2011, pp. 2501–4.
[20] M. Lee and T. M. Gatton, “Wireless health data exchange for home
healthcare monitoring systems.,” Sensors (Basel), vol. 10, no. 4, pp.
3243–60, 2010.
[21] K. Han, K. Kim, and T. Shon, “Untraceable mobile node
authentication in WSN.,” Sensors (Basel), vol. 10, no. 5, pp. 4410–29,
2010.
[22] M.-P. Gagnon, P. Ngangue, J. Payne-Gagnon, and M. Desmartis, “m-
Health Adoption by Healthcare Professionals: A Systematic Review,”
Journal of the American Medical Informatics Association, p. ocv052,
2015.
[23] M. K. Khan and K. Alghathbar, “Cryptanalysis and security
improvements of „two-factor user authentication in wireless sensor
networks‟.,” Sensors (Basel), vol. 10, no. 3, pp. 2450–9, 2010.
[24] Y.-W. Ma, J.-L. Chen, Y.-M. Huang, and M.-Y. Lee, “An efficient
management system for wireless sensor networks.,” Sensors (Basel),
vol. 10, no. 12, pp. 11400–13, 2010.
[25] Quality management - Customer satisfaction - Guidelines for
monitoring and measuring, ISO Standard 10004, 2012.
[26] Ease of operation of everyday products - Part 1: Design requirements
for context of use and user characteristics, DIN ISO standard 20282-1,
2008
[27] Ergonomics of human-system interaction - Usability methods
supporting human-centered design, ISO/TR standard 16982, 2002
[28] Medical devices - Application of usability engineering to medical
devices, IEC standard 62366,
[29] P. K. Sahoo, “Efficient security mechanisms for mHealth applications
using wireless body sensor networks.,” Sensors (Basel), vol. 12, no. 9,
pp. 12606–33, 2012.
[30] K. Hung, C. C. Lee, W. M. Chan, S.-O. Choy, and P. Kwok,
“Development of a wearable system integrated with novel biomedical
sensors for ubiquitous healthcare.,” Conf Proc IEEE Eng Med Biol
Soc, vol. 2012, pp. 5802–5, 2012.
[31] Institute of Electrical and Electronic Engineers. (2015). About IEEE
[Online]. Available: http://www.ieee.org/index.html
[32] International Electrotechnical Commission. (2015). Who we are
[Online]. Available: http://www.iec.ch/about/profile/
[33] International Organization for Standardization. About ISO [Online].
Available: http://www.iso.org/iso/home/about.htm
[34] German Institute for Standardization. (2015). DIN Standards
[Online]. Available: http://www.din.de/en/about-standards/din-
standards
[35] Information technology - Telecommunications and information
exchange between systems - Near Field Communication - Interface
and Protocol (NFCIP-1), ISO/IEC Standard 18092, 2013.
[36] Information technology – Telecommunications and information
exchange between systems – Local and metropolitan area networks
specification requirements – Part 1: Overview of Local Area Network
Standards, ISO/IEC Technical Report 8802-1, 2001.
[37] Data quality – Part 130: Master data: Exchange of characteristic data:
Accuracy, ISO Standard 8000-130, 2009.
[38] Information technology - Open Systems Interconnection - Basic
Reference Model” ISO/IES Standard 7498-1, 1994.
[39] Health informatics - Use of mobile wireless communication and
computing technology in healthcare facilities - Recommendations for
electromagnetic compatibility (management of unintentional
electromagnetic interference) with medical devices, ISO Technical
Report 21720, 2007.
[40] Power consumption of information technology equipment.
Measurement methods, IEC Standard 62018 ED. 1.0 B, 2003
[41] Information Technology – Biometric Data Interchange Formats – Part
7: Signature/Sign time series Data, ISO/IEC Standard 19794-7, 2007.
[42] IEEE standard for local and metropolitan area network – Part 21:
Media independent handover services, IEEE Standard 802.21, 2008.
[43] Health Informatics – Electronic health record communication – Part 1:
Reference model, ISO Standard 13606-1, 2012.
[44] Information and documentation – Records management – Part 1:
General, ISO Standard 15489-1, 2001.
[45] IEEE Standard for CASE Tool Interconnections – Reference model
for specifying software behaviour, IEEE Standard 1175.3, 2004.
[46] Application of risk management for IT-networks incorporating
medical devices – Part 1: Roles, responsibilities and activities, IEC
Standard 80001-1 ED.1.0 B, 2010.
[47] Health informatics - Personal health device communication - Part
20601: Application profile - Optimized exchange protocol, ISO/IEEE
Standard 11073-20602, 2010.
[48] Application of risk management for IT-networks incorporating
medical devices – Part 2-3: Guidance for wireless networks, IEC
Technical Report 80001-2-3 ED. 1.0 EN, 2012.
[49] Information tecnology – Security techniques – Entity authentication –
Part 1: General, ISO/IEC Standard 9798-1, 2010.
[50] Accuracy in Patient and Sample Identification; Approved Guideline,
Clinical and Laboratories Institute Standard GP33-A Vol. 30 No. 7,
2010.
[51] Systems and software engineering – Systems and software Quality
Requirements and Evaluation (SQuaRE) – System and software
quality models, ISO.IES Standard 25010, 2011.
[52] Delay and Power Calculation Standards - Part 1: Integrated circuit
delay and power calculations, IEC Standard 61523-1, 2012.
[53] Industrial automation systems and integration – Product data
representation and exchange, ISO Standard 10303-1, 1994.
[54] Information technology – Security techniques – Evaluation criteria for
IT security, ISO/IEC Standard 15408-1, 2009.
[55] Information technology – Security techniques – Storage security,
ISO/IEC Standard 27040, 2015.
 An Improved Image Steganography Scheme with
High Visual Image Quality
Sumit Laha, Rinita Roy
Department of Computer Science and Engineering
Future Institute of Engineering and Management
Kolkata, India
sumitlaha@icloud.com, rinita5493@gmail.com
Abstract—With the advent of internet, secured transmission
of data over networks offers a great challenge. This problem of
secrecy of information has drawn a lot of attention due to its
immense demand in real life applications such as information
security systems. The objective of this study consist of three-fold.
In the first phase, we propose an embedding algorithm to hide a
secret image within the cover image. In the second phase, the
stego image quality is optimized using GA. Then, in the third
phase, we have presented the procedure to extract the secret
image from the optimized stego image with 100% data lossless.
The experimental results demonstrate that the proposed method
significantly outperforms the five state-of-the-art steganography
techniques known to date in terms of enhanced peak signal to
noise ratio (PSNR) of the stego image.
Keywords—Image Steganography; Embedding algorithm;
XOR; Genectic Algorithm; 8-connected neighbor; Peak Signal to
Noise Ratio
I. INTRODUCTION
With the development of internet, secured transmission of
data over networks provides a great challenge to the
researchers and practitioners. This problem of secrecy of
information has long drawn a lot of attention due to its
immense demand in real life applications such as information
security systems and rapid development of communication and
network. There are two methods, namely, cryptography and
steganography in order to keep the data secret. Cryptography
comprises two phases - encryption and decryption of data.
However the existence of the data cannot be hidden from the
third party. On the other hand, steganography aims to keep the
existence of the data (in the format of image, video, audio or
text) secret in the cover image [1]. The steganography
procedure [2] consists of three important elements: the data to
be hidden (here, image), the cover image, and the resulting
stego image (after embedding the secret image into the cover
file). Two important steganography approaches frequently in
the literature, namely spatial domain-based [3-6] and frequency
domain-based [7, 8]. In case of spatial domain approach,
embedding of data is carried out using Least Significant Bit
(LSB) substitution method, whereas in the frequency domain,
the data is embedded in the frequency coefficient of images.
This paper deals with the problem of embedding of the
secret image, optimizing this stego image as close as to the
cover image without altering the embedded bits and extracting
the original secret image from the stego image. Although there
have been several studies on the embedding of secret images,
most of them deals with normal LSB substitution. But the LSB
substitution approach suffers from certain drawbacks. For
example, if a third party tries to read the LSB positions from
the stego image, a secret image hidden using this technique
can easily be intercepted. Moreover, this method completely
replaces the original values in the LSB positions, while
substituting the secret image. This results in low visual quality
of the stego image as compared to the original cover image.
Inspired by this motivation, in this paper, we propose an
embedding algorithm to hide a secret image within the cover
image. Then, the stego image quality is optimized using GA
and finally, we have presented the procedure to extract the
secret image from the optimized stego image. As a result, the
secret image is obtained with 100% data lossless and also,
provides a high visual image quality with respect to the cover
image.
The remainder of the paper is organized as follows:
Section II gives a brief description of a literature review of
previous solution methods; Section III presents the proposed
embedding technique, GA and the extraction procedure;
Section IV provides the experimental results and finally, the
conclusions are drawn in Section V.
Nomenclature
pc cover image pixel
ps stego image pixel
ip initial population array
C cover image represented as 2D matrix
S stego image represented as 2D matrix
MAXI maximum intensity of a pixel, i.e., 255
k population size of GA
II. LITERATURE REVIEW
Recently, there have been growing number of research on
steganography methods especially, based on soft computing
methods like GA and particle swarm optimization in both
spatial and frequency domains. In the spatial domain,
embedding of message is done by LSB substitution method,
whereas in the frequency domain, the message is embedded in
the frequency coefficient of images. In this section, we focus
on a review of the previous studies on the steganography
methods.
Kanan and Nazeri [9] presented a novel steganography
technique based on GA to obtain high embedding capacity and
enhanced PSNR of the stego image. The high embedding
capacity is achieved by modifying secret bits and then finding
the best place for embedding it into the cover image based on
GA.
Roy and Laha [10] proposed a GA to optimize the stego
image based on a fitness function using 8-connected neighbor
of each pixel. They have shown that the proposed scheme is
found to be relatively better than some existing steganography
techniques.
Li and Wang [11] presented a stenographic scheme to hide
data JPEG images securely along with high message capacity
and good image quality in two phases. In the first phase,
particle swarm optimization algorithm is applied to improve
the stego image quality by deriving the optimal substitution
matrix for transforming the secret images. In the next phase,
the hiding the secret image into the cover image using a
modified JPEG quantization table.
Tseng, Chan, Ho and Chu [12] proposed an effective
steganography scheme to enhance the quality of stego image
by using an improved GA and optimal pixel adjustment
process. The experimental results show that the mean square
error of the stego image of the proposed scheme is much lower
than those produced by the existing methods.
Wang, Lin and Lin [13] developed a steganography method
based on GA to embed the secret message in the moderately
significant bit of the cover image. They also applied a global
substitution step and a local pixel adjustment process for
improvement of the quality of the stego image. Later, in
another research work, Wang, Lin and Lin [14] proposed
optimal LSB substitution to embed the data secret in the host
image. They also proposed GA to the data hiding problem in
the k LSBs of the host image.
Recently, there have been some noteworthy studies in
connection with the development of image sharing scheme
with steganography and authentication ability to protect the
integrity stego images from dishonest participants [15-18].
Lin and Tsai [15] proposed a technique for the secret image
sharing by applying Shamir method. Their proposed scheme
consists of three levels of security protection, namely, sharing
the secret among the participants, embedding the data hiding,
and authentication capability, resulting in high security and
efficiency of the system.
Chang, Hsieh and Lin [16] developed a novel image
sharing scheme considering both steganography and
authentication based on Chinese remainder theorem. The
performance of this scheme is shown to be superior to the
existing methods in respect of both high authentication ability
and good stego image quality.
Wu and Kao [17] presented a secret image sharing scheme
by employing optimal pixel adjustment process to obtain
enhanced stego image quality under different payload capacity
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
and authentication bit conditions and showed its improved
performance compared with the recent steganography schemes.
Yang, Chen, Yu and Wang [18] proposed an improved
image secret sharing scheme by hashing the four-pixel block,
block ID and image ID to increase the authentication ability
and by arranging the bits in the for-pixel square block to
improve the stego image and finally, by using the GF (28) to
obtain a lossless version without requiring additional pixels.
III. PROPOSED WORK
In this section the proposed work comprises three phases.
The first phase deals with a new embedding algorithm to hide
the secret image, the second phase addresses the proposed GA
that uses a heuristic approach to find superior solution and the
third phase provides the extraction algorithm and these are
presented in subsequent sections.
A. Embedding of data
In this section, we have proposed a new embedding
technique to hide the secret image. In the conventional LSB
substitution method, the LSBs of the cover image are just
replaced with the bits of the secret image. But there is a flaw in
terms of security. If an intruder scans the LSBs, he can easily
intercept the secret image. Furthermore, this approach yields
low PSNR values.
Fig. 1. Embedding of secret image into cover image.
So a new approach has been proposed, where instead of
simply substituting the secret image bits, XOR operation is
performed between the cover image bits and the secret image
bits. This is illustrated in Fig. 1. The proposed algorithm
(considering 2 bits per pixel for embedding secret bits) is given
as follows:
Input: Cover and secret grey scale images of sizes m×m and
n×n, respectively, where n ≤ m/2
Output: Stego image of size m×m
Step 1: Read the cover and secret image files.
Step 2: 2 bits from the MSB side of first pixel of the secret
image are taken and XOR-ed respectively with the 6th and 7th
bit of the each pixel of the cover image and the results are
copied into the 6th and 7th bit positions of the corresponding
pixel of the stego image.
Step 3: Step 2 is repeated for all the pixels of the secret image,
taking 2 bits at a time of the pixel of the secret image.
Step 4: Write the output stego into the output file.
 The proposed approach has proved to be better than normal
LSB substitution. Since the embedded bit positions of the stego
image does not contain real values of secret image, it is
difficult to intercept the secret message. In order to extract the
secret image, the original cover image is needed. This is
explained further in subsection C. Furthermore, the PSNR
value of the stego image obtained is also higher.
B. Optimization using GA
Genetic algorithms (GAs) originally developed by John
Holland [19] are robust structured stochastic global search
algorithm based on biologically inspired computational
intelligence that have been employed to a wide range of
problems in the areas of science, engineering, business and
other fields [20-23]. These algorithms can be successfully
applied for generating efficient solutions from large, complex
and multi-modal search space. While implementing GA for
solving a particular problem, one has to amend the coding of
the solution in the search space along with the appropriate
tuning of parameters pertaining to GA to make use of the
problem specific available information effectively. The
important components of GA consists of an initial population
of candidate solutions (chromosomes), selection of new
population based on a suitable fitness function, recombination
based on the genetic operator as crossover and randomly
perturbation of candidate solution using mutation operator. The
GA is repeated in subsequent generations until it reaches the
maximum number of generations as the stopping criterion.
While applying the selection operator, it attempts to evolve
good quality solutions based on the fitness function and also
picking some inferior quality solutions so that the diversity of
the solution quality in the population is retained.
In this paper, we present a modification of the GA [10]. In
the proposed GA, initial population, fitness function, selection,
crossover, mutation, evaluation and termination are described
as follows.
1) Initial Population
We select each pixel of the stego image (say, ps)
individually and ps is optimized using GA. Let pc represent the
corresponding pixel (having the same coordinates as ps) of the
cover image. The population size is taken as 30. It is generated
based on 20% of the individuals using heuristic approach and
80% of the individuals randomly. In the proposed GA, for the
heuristic initialization part, we take the embedded pixel, and
flip the i) 2nd -7th bits, ii) 2nd - 6th bits, iii) 2nd – 5th bits, iv)
2nd – 4th bits, v) 2nd and 3rd bits, and vi) 2nd bit respectively,
so as to obtain the six individuals of the initial population. In
order to illustrate the heuristic consider a pixel (say, 186). The
binary equivalent of this pixel is 10111010. The six most
significant bits are flipped to obtain 01000110 (decimal
equivalent being 70) as the first candidate solution of the initial
population. Similarly, 198, 134, 166, 182 and 190 become rest
of the population generated from the heuristic initialization.
This is illustrated in Fig. 2.
Fig. 2. Heuristic Initialization for a single pixel.
Similarly, in the random initialization phase, the remaining
part of the initial population is generated randomly using
random permutation from the entire range of possible solutions
(0 – 255), while retaining the embedding in the two LSBs
unaltered. The solutions generated are unique from the
solutions obtained heuristically in the initial population.
2) Fitness Function and Selection
In this GA, the fitness function based on 8-connected
PSNR [10] is obtained from the following equations:
 x 1 y 1 2 (1)
mse '(k )     C (i, j )  S (i, j )  C (x, y)  S (x, y)  C (x, y)  i p (k)  / 9
2 2
i  x 1 j  y 1 
 MAX I 2 
psnr '(k )  10log10   (2)
 mse '(k ) 
Using Equation (1), we compute the MSE of the stego-
image for each of the solutions in the population with respect to
the corresponding pixel of the cover image. Next ewe compute
the fitness values in terms of PSNR from Equation (2).
Based on the above fitness function, roulette wheel
selection is used to choose two parents from the initial
population.
3) Crossover and Mutation
In this study, the one-point crossover operator is employed
to combine good portions of different individuals to generate
diverse solutions. Depending on the crossover probability, it is
decided whether new offspring will be generated or not.
Selection of the random crossover point and swapping of bits
on either side is done without altering the embedded bits.
In mutation, mutation probability is used to have diverse
solution so that local optimum does not get stuck. Unlike
crossover probability, the mutation probability is set very low
value. The mutation operator used in the proposed GA
randomly swaps 2 bits excluding the embedded bits.
4) Evaluation and Termination
In each generation evaluation is performed by updating the
best-so-far solution amongst the two offspring and the best-so-
far solution obtained from the previous generation.
The termination criterion is reached after executing a
predefined number of generations. At the end of the GA, we
obtain the optimized value for each pixel from the best-so-far
array.
C. Extraction of secret image
As explained in subsection A, the stego image contains the
secret image in encrypted form. So in order to extract the secret
image, the original cover image is required. The extraction
procedure is same as the embedding one. This is illustrated in
Fig. 3.

Fig. 5. Histogram of Stego Lena image (simple LSB embedding).

Fig. 6. Histogram of Stego Lena image (XOR embedding).

Fig. 3. Extraction of secret image from stego image.
Once again XOR operation is performed individually
between the stego and cover image bits respectively. The
extraction algorithm is given as follows:
Input: Optimized stego and cover gray scale images of sizes
m×m each.
Output: Secret grey scale image of size n×n.
Step 1: Read the optimized stego image and the cover image
files.
Step 2: The 6th and 7th bits of each pixel of the stego image is
XOR-ed with that of the corresponding bits of the cover and
the two resultant bits are kept in the output array.
Step 3: Step 2 is repeated for all the pixels of the stego image
so as to get the all the pixels of the secret image.
Step 4: Write the output array into the output file.
In simple LSB embedding, PSNR is 43.797 and that in case
of XOR embedding is 45.076. It is evident from the
experimental results that the new embedding approach not only
ensures security of secret data, but also guarantees better visual
quality (high PSNR value) with respect to the cover image.
B. Proposed Genetic Algorithm
In this study, we have taken three 512×512 standard gray
scale images (8-bits per pixel), “Lena”, “Jet”, “Pepper”,
“Sailboat” and “Baboon” as cover images as shown in Fig. 7.
The secret message is a 256×256 gray scale “General test
pattern” image in each case. The program was implemented
using C running on 2.3 GHz i3 core processor under Windows
7.

IV. EXPERIMENTAL RESULTS

A. Embedding Algorithm
To explain the effectiveness of the new embedding
technique, an experiment has been carried out. A 512×512 a b c
grayscale “Lena” image and a 256×256 “General test pattern”
is taken as cover and secret images respectively. Fig. 4
represents the histogram of the original “Lena” image, while
Fig. 5 and Fig. 6 represent that of the stego images.

d e f
Fig. 7. Secret image – (a) General test pattern; Cover images – (b) Lena, (c)
Jet, (d) Peppers, (e) Sailboat, (f) Baboon.

1 m 1 n 1
MSE   C (i, j )  S (i, j )
2
(3)
mn i  0 j 0
Fig. 4. Histogram of original Lena image.
  MAX I 2  There are some issues important for future research
PSNR  10log10   (4) consideration. First, an attempt can be made to develop better
 MSE  embedding algorithm that can produce high visual quality
while keeping the secret message intact. Second, as an
The comparisons of the above methods are done on the
alternative to the existing soft computing algorithm, other
basis of the PSNR values (from equations (3) and (4)) of the
algorithms such as cuckoo search algorithm and bacteria
optimized images obtained by applying the proposed GA with
foraging optimization algorithm can be implemented to obtain
respect to the corresponding cover images. The GA parameters
accurate modeling of steganography. Finally, the proposed
has been summarized in table 1.
scheme can also be tried for color images.
TABLE I. GA PARAMETERS
ACKNOWLEDGMENT
Parameter Value
Number of generations 100
We are grateful to Professor Amrita Khamrui for her
helpful comments and suggestions in writing this paper.
Population size 30
Crossover rate 0.9
Mutation rate 0.05
REFERENCES
[1] A. Cheddad, J. Condell, K. Curran, and P. McKevitt, “Digital image
Table 2 compares the PSNR values of different grey scale steganography: Survey and analysis of current methods,” Signal
images for the proposed method as against the Lin and Tsai’s Processing, vol. 90, pp. 727-752, 2010.
method [15], Yang, Chen, Yu and Wang’s method [18], [2] L.M. Marvel, C.G. Boncelet, and C.T. Retter, “Spread spectrum image
Chang, Hsieh and Lin’s method [16], Wu and Kao’s method steganography,” IEEE Trans. Image Process, vol. 8, pp. 1075-1083,
[17] and Kanan and Nazeri’s method [9]. 1999.
[3] H. Sajedi, and M. Jamzad, “BSS: Boosted steganography scheme with
TABLE II. COMPARISON OF PROPOSED METHOD WITH DIFFERENT cover image processing,” Expert Systems with Applications, vol. 37, pp.
STEGANOGRAPHY TECHNIQUES 7703-7710, 2010.
[4] W.-J. Chen, C.-C. Chang, and T. Le, “High payload steganography
Images mechanism using hybrid edge detector,” Expert Systems with
Methods Avg.
Lena Jet Peppers Sailboat Baboon Applications, vol. 37, pp. 3292-3301, 2010.
Lin and Tsai [5] A. Ioannidou, S.T. Halkidis, and G. Stephanides, “A novel technique for
39.2 39.25 39.17 39.16 39.18 39.19 image steganography based on a high payload method and edge
[15]
Chang, Hsieh detection,” Expert Systems with Applications, vol. 39, pp. 11517-11524,
40.37 40.73 39.3 38.86 39.94 39.84 2012.
and Lin [16]
Yang, Chen, [6] B.E. Carvajal-Gamez, F.J. Gallegos-Funes, and A.J. Rosales-Silva,
Yu and 41.6 41.66 41.56 41.51 41.55 41.58 “Color local complexity estimation based steganographic (CLCES)
Wang [18] method,” Expert Systems with Applications, vol. 40, pp. 1132-1142,
Wu and Kao 2013.
43.54 43.53 43.56 43.55 43.54 43.54
[17] [7] W.-Y. Chen, “Color image steganography scheme using DFT, SPIHT
Kanan and codec, and modified differential phase – shift keying techniques,”
45.12 45.18 45.13 45.1 45.12 45.13
Nazeri [9] Applied Mathematics and Computation, vol. 196, pp. 40-54, 2008.
Proposed [8] R. Jafari, D. Ziou, and M.M. Rashidi, “Increasing image compression
47.03 47.037 47.033 47.039 47.029 47.034
method rate using steganography,” Expert Systems with Applications, vol. 40,
pp. 6918-6927, 2013.
It is revealed from the results of Table 2 that the proposed [9] H.R. Kanan, and B. Nazeri, “A novel image steganography scheme with
method yields average PSNR value of 47.034 which is high embedding capacity and tunable visual image quality based on a
genetic algorithm,” Expert Systems with Applications, vol. 41, pp.
considerably higher in comparison with those produced by all 6123–6130, 2014.
other algorithms. [10] R. Roy, and S. Laha, “Optimization of Stego Image Retaining Secret
Information Using Genetic Algorithm with 8-connected PSNR,”
V. CONCLUSIONS Procedia Computer Science, vol. 60, pp. 468-477, 2015.
[11] X. Li, and J. Wang, “A steganographic method based upon JPEG and
In this paper, the problem of steganography is considered to particle swarm optimization algorithm,” Information Sciences, vol. 177,
optimize the stego image close to the cover image without pp. 3099–3109, 2007.
affecting the secret image embedded within it. We proposed an [12] L.-Y. Tseng, Y.-K. Chan, Y.-A. Ho, and Y.-P. Chu, “Image hiding with
embedding algorithm to hide a secret image within the cover an improved genetic algorithm and an optimal pixel adjustment
image and then, the stego image quality is optimized using GA. process,” In Intelligent Systems Design and Applications, ISDA'08.
Eighth International Conference on, IEEE, vol. 3, pp. 320-325, 2008.
Finally, we have presented the procedure to extract the secret
image from the optimized stego image with 100% data lossless. [13] R.Z. Wang, C.F. Lin, and J.C. Lin, “Hiding data in images by optimal
moderately significant-bit replacement,” IEEE Electronics Letters, vol.
The proposed scheme optimizes some benchmark images 36, pp. 2069-2070, 2000.
taken from USC-SIPI image database to get images with higher [14] R.Z. Wang, C.F. Lin, and J.C. Lin, “Image hiding by optimal LSB
substitution and genetic algorithm,” Pattern Recognition, vol. 34, pp.
PSNRs and these PSNRs are greater than those produced by 671-683, 2001.
the existing algorithms.
[15] C.-C. Lin, and W.-H. Tsai, “Secret image sharing with steganography
and authentication,” The Journal of Systems and Software, vol. 73, pp.
405-414, 2004.
[16] C.-C. Chang, Y.-P. Hsieh, and C.-H. Lin, “Sharing secrets in stego
images with authentication,” Pattern Recognition, vol. 41, pp. 3130-
3137, 2008.
[17] C.-C. Wu, S.-J. Kao, and M.-S. Hwang, “A high quality sharing with
steganography and adaptive authentication scheme,” The Journal of
Systems and Software, vol. 84, pp. 2196-2207, 2011.
[18] C.-N. Yang, T.-S. Chen, K.H. Yu, and C.-C. Wang, “Improvements of
image sharing with steganography and authentication,” The Journal of
Systems and Software, vol. 80, pp. 1070-1076, 2007.
[19] J. Holland, “Adaptation in natural and artificial systems: an introductory
analysis with applications in biology, control, and artificial intelligence,”
Ann Arbor, MI: University of Michigan Press, 1975.
[20] D.E. Goldberg, “Genetic algorithms in search, optimization and machine
learning,” Addison-Wesley Longman Publishing Co., Inc., Boston, MA,
1989.
[21] M. Gen, and R. Cheng, “Genetic algorithms and engineering design,”
NY: John Wiley & Sons, 1997.
[22] T.D. Gwiazda, “Genetic algorithms reference. Crossover for single-
objective numerical optimization problems (Vol I),” Berlin: Springer;
2006.
[23] T.D. Gwiazda, “Genetic algorithms reference. Mutation operator for
single-objective numerical optimization problems (Vol II),” Lomianki,
Poland: Tomasz Gwiazda, 2007.
 A Quantum-inspired Cuckoo Search Algorithm for
the Travelling Salesman Problem
Sumit Laha
Department of Computer Science and Engineering
Future Institute of Engineering and Management
Kolkata, India
sumitlaha@icloud.com
Abstract—The quantum algorithm and the cuckoos search
algorithm as emerging novel evolutionary techniques has recently
drawn a lot of research interest due to their capability to search
globally as well as locally by exploring the search space more
efficiently in various applications of engineering and
management. To the best of our knowledge, this paper first
considers the application of quantum-inspired cuckoo search
algorithm to solve the classic travelling salesman problems. In
this paper, we present a quantum embedded cuckoo search
algorithm for the travelling salesman problem. To accelerate the
search process for better solution quality, some neighborhood
search based construction and stochastic heuristic approaches is
utilized in the simulated annealing algorithm. The proposed
method is tested with several benchmark test problem instances
taken from the TSP library in the literature. The computational
results demonstrate that the proposed hybrid method is very
competitive with the state-of-the-art procedures in the literature.
Keywords—Travelling salesman problem; quantum algorithm;
cuckoo search algorithm; simulated annealing; quantum-inspired
cuckoo search algorithm; combinatorial optimization
I. INTRODUCTION
The travelling salesman problem (TSP) is a classic
example of NP-hard combinatorial optimization problem. The
objective of this problem is to obtain the shortest possible path
so that the salesman visits each city only once and comes back
to the destination. These TSP problems are frequently
encountered in various applications in industry and real-life
situations such as transportation problems, assignment
problems, scheduling, production systems, vehicle routing,
computer systems, and image processing and pattern
recognition. Since the TSP belongs to the class of NP-hard,
there is great interest among researchers to develop efficient
algorithms to find the optimal or near-optimal solutions in
reasonable computational time.
During last three decades, there exists numerous studies on
the development of different optimization techniques such as
exact, heuristics, and metaheuristics to solve the TSP. The
exact algorithm includes branch and bound method and linear
programming, however, these algorithms are feasible only for
solving small-sized TSP, whereas for large-sized problems,
the computational times required by them is excessively high.
Heuristics are based on problem specific information and
produce near-optimal solutions, especially for large-sized
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
problems in reasonable computational times due to their
polynomial time complexities. On the other hand,
metaheuristics are robust and stochastic based adaptive search
optimization techniques and perform better than the heuristics
to generate optimal or near-optimal solutions for large-sized
TSP problems in reasonable times, however, at the cost of
additional computational times.
Noteworthy metaheuristics to solve the TSP include
simulated annealing [1], neural network [2], particle swarm
optimization [3], ant colony optimization [4], genetic
algorithm [5], self-organizing maps [6], Lagrangian relaxation
[7], and elastic net [8].
Apart from the above metaheuristics, recently, quantum
computing as new research filed has drawn much attention and
wide applications in combinatorial optimization. Later, it is
found that that there is a growing interest in emerging
quantum computing with evolutionary algorithms and
quantum-inspired evolutionary algorithms have developed as
promising alternative global optimization techniques due to
their capability to search globally as well as locally by
exploring the search space more efficiently in various
applications of engineering and management. Since late
1990‟s a variety of quantum-inspired evolutionary algorithms
such as genetic-quantum algorithm [9], quantum-inspired
gravitational search algorithm [10], quantum differential
evolution algorithm [11] and quantum-inspired cuckoo search
algorithm (QCSA) [12] have been successfully utilized to
solve different engineering and combinatorial optimization
problems.
Although there exists few research on the application of
the QCSA to knapsack problems [12], application of QCSA to
TSP was not reported in the literature. However, there some
literature on other quantum-inspired genetic algorithm to solve
TSP problems [13].
This paper offers a novel evolutionary algorithm called
QCSA for the travelling salesman problem. The important
feature of the cuckoo search algorithm [14, 15] is its global
random walk governed by Levy flights, rather than standard
isotropic random walk to converge global optimal solution. To
improve the solution further, some neighborhood search based
construction heuristic approaches is utilized in the simulated
annealing (SA) algorithm.
 II. QUANTUM REPRESENTATION OF TSP 1‟s in positions 5 and 6, we replace them by 3 and 5
The smallest unit of information in the quantum computing respectively in order to get a feasible path) or [3-1-2-5-4-6].
is known as quantum bit or qubit. A qubit may be in the „1‟
state or „0‟ state or superposition of these two states. The state III. THE PROPOSED ALGORITHM
of a qubit is expressed as In this section, we propose a QCSA for the travelling
salesman problem. To accelerate the search process for better
   1   0  (1) solution quality, some neighborhood search based on
construction heuristics as well as random neighborhood is
where α and β are denoted as complex numbers that represent utilized in the simulated annealing algorithm. The details of
the probability amplitudes of the respective states. The these algorithms are mentioned below.
probability that the qubit is in the „0‟ state and „1‟ state are
measured by |α|2 and |β|2 respectively. Normalization of these
A. Quantum-inspired cuckoo search algorithm
two states guarantees the condition as given below.
We present the QCSA for the solving the TSP. The
   1
2 2
(2) principles of the quantum algorithm such as qubit and
superposition of states, measurement, mutation and
For n-city TSP, the path can be represented as follows. interference are embedded in the core of the cuckoo search
algorithm. The steps of the QCSA are given below.
Quantum representation:
Step 1: A set of N number of n-city paths are randomly
1  2         n 1k 1   n 1k  2   (3) constructed based on quantum representation of Equation (3).
  k   k 1 k  2  2 k     n
        n  Step 2: The binary representation of the N paths are generated
 1 2 k   k 1 k 2 2k 
  n 1k 1  n 1k  2 
from step 1 and from these paths, a set of N feasible paths with
Binary representation: the corresponding path distance are obtained (given in Section
2).
b1 b2  bk  bk 1 bk  2  b2 k    (4)
 b n1k 1 b n1k  2  bn  Step 3: The best path of the N paths obtained from Step 2 is
computed and if it is equal to the optimal path distance, then
0 1  0 11  1
 0 0  1 (5) stop.
Step 4: Set k = 1.
For example, consider 6-city TSP and use 3-bit for each Step 5: Get a random path S (k ) from N paths from step 1.
city of a path (we use 3-bit for each city for number cities up to Step 6: The new path S (k  1) using Levy flights is obtained
7. Similarly, we consider 4-bit and 5-bit for each city with from Equation (8) and update the path based on quantum
cities up to 15 and 31 respectively.). Suppose, the quantum representation. However, each qubit of the S (k  1) violates
representation of a path is given as follows.
the condition of normalization given is Equation (2). Therefore,
a | b | c | d | e | f  (6) to ensure the guarantee of normalization of the states of each
 
qubit, it is updated by selecting the larger one between  1 
where,  1 
0.88 0.53 0.67   0.83 0.55 0.89  1  1 1  1
2 2
and replacing by either or corresponding
a  ; b  to the value of 1 or 1 respectively.
 0.47 0.85 0.74   0.56 0.84 0.46 
S (k  1)  S (k )   Levy   
0.37 0.96 0.54 0.09 0.68 0.80 
c  ; d   or, S (k  1)  S (k )  10 * cities * k (1    3) (8)
5 

 0.93 0.28 0.84  1.00 0.73 0.61 Step 7: Apply quantum intra-bit mutation operation by
selecting randomly a qubit with mutation probability (pm) from
0.88 0.53 0.67   0.73 0.61 0.31 the path obtained in step 6 and interchange the qubit
e  ; f   probability amplitudes α and β in the qubit.
 0.47 0.85 0.74   0.69 0.80 0.95 Step 8: Apply quantum measurement operation to obtain the
The binary representation of the path from (6) can be feasible path and the corresponding path distance.
obtained following Equation (4) as: Step 9: Apply quantum interference operator by using
 
appropriate quantum gates U ( ) at each position of   in
0 1 0 11 1 1 0 0 0 0 0 11 0 0 0 1 (7)  
the path. Among different types of gates, rotation gates, U ( )
The decimal equivalent for the path given in Equation (7)
can be written as: [2-7-4-0-6-1] or [2-1-4-0-1-1] (we change 7 is frequently used by the researchers [12, 16]. In this paper, we
and 6 values which are greater than 5 by considering 7 (mod 6) consider the rotation gates for solving the TSP. The rotation
=1; 6 (mod 6) =0) or [2-1-4-0-3-5] (since there are two more gate is given as:
 cos   sin   Step 4: If C Y   C  X  , set X  Y .
U      (9) Step 5: If C (Y )  C ( X ) , compute   C (Y )  C ( X ) and
 sin   cos  
where θ is the rotation angle. The value of rotation angle   
Tk  Tk 1 and set X  Y with probability  e Tk  .
should be well taken to avoid getting stuck of the local  
optimum solution. A big value of θ can converge to premature Step 6: If k < max_iteration, return to step 2. Otherwise go to
local optima, whereas, a small value of θ increases the step 7.
convergence time [12]. Also, the lookup table of θ to obtain the Step 7: Output the current best solution as the final solution.
direction of θ is given in table 1. In this study, θ is taken as
 same as used by Layeb [12], which is used to solve C. Initial solution
20
similar problem like bin packing problem. The initial solution for the SA is considered as the output
Step 10: Replace some paths (50%) among N paths randomly solution of the QCSA (Section III A).
by the generated path in Step 9 if it is better. D. Neighborhood solutions
Step 11: Delete a fraction (pa) of the inferior paths and replace
them following Step 6. In order to improve the neighborhood solutions of the
Step 12: Keep the best path and rank the paths. proposed SA, two construction heuristics are proposed.
Step 13: Set k  k  1 . If k  maxcount , return to step 5. Applying these two heuristics on a current n-city path, a total
Otherwise, stop. of 2(n-1) paths are generated and the best among them is taken
as the updated current path. Next, in order to randomly explore
the solution space more efficiently to converge to the global
TABLE I. LOOKUP TABLE OF THE ROTATION ANGLE Θ solution instead of getting stuck to local optimum, we use four
Reference
random neighborhood search techniques on the updated current
αij βij Direction of θ αij path obtained from the construction heuristics. The details of
binary value
>0 >0 1 +θ >0 these techniques are presented below.
>0 >0 0 -θ >0
>0 <0 1 -θ >0 1) Heuristic 1
>0 <0 0 +θ >0 Consider an example of a 5-city TSP. Assume the current
<0 >0 1 -θ <0 path as 3-5-1-4-2. Pick the first city from the path and append
<0 >0 0 +θ <0 to each of the remaining cities in the path two generate four
<0 <0 1 +θ <0 partial paths, 5-3, 1-3, 4-3 and 2-3. Next select the partial path
<0 <0 0 -θ <0 with least distance (say, 4-3). Next, to generate the new path,
B. Proposed SA heuristic insert this partial path 4-3 at the beginning and append the
remaining cities from the previous path (here, 3-5-1-4-2)
SA [17] is a stochastic neighborhood search technique sequentially. Thus the new path becomes 4-3-5-1-2.
based on the ideas of physical annealing process drawn from Proceeding in the similar manner, repeat this steps till a total of
the principle of statistical mechanics. It has been found four new paths are generated and the best path among them is
effective to solve many combinatorial optimization problems. taken as the current path. The detailed illustration of the
It begins with an initial solution, an initial high temperature, Heuristic 1 is given in Fig. 1.
and number of iterations. A neighborhood solution is
generated from the initial solution, which is considered as the
current solution. If the neighborhood solution is better than the
current solution, it is accepted and updated as the current
solution. However, if the neighborhood solution is inferior to
the current solution, it is also accepted with a probability based

on e T and updated as the current solution, thereby enhancing
the possibility of global search out of a local optimum. At high
temperature in the beginning, there is a high probability to
move to a worst solution. However, as the iteration increases,
the probability is reduced. As the iteration increases, the
temperature is gradually reduced following a particular
annealing schedule and continues till the stopping criterion is
reached. The proposed SA-based heuristic is presented as
follows.
Fig. 1. Illustratio of Heuristic 1
Step 1: Initialization: Obtain the initial solution (X) with path
distance (C(X)) and consider it as current solution. Set the 2) Heuristic 2
initial temperature (T1); Set iteration number k =1. In order to explain heuristic 2, let us take an example of a
Step 2: Set k  k  1 . 5-city TSP. Assume the current path as 3-5-1-4-2. To generate
Step 3: Generate new neighborhood solution Y from X (Section the first path same procedure is applied as heuristic 1 and the
III D). new path becomes 4-3-5-1-2. Now the partial path 4-3 in the
path is locked and its terminal cities (4 and 3) are considered
for further processing. Applying similar procedure for both the
cities individually, new sets of partial paths (5-4, 1-4, 2-4; 5-3,
1-3, 2-3) are generated. Next the partial path with least distance
is selected and it is appended before or after the previously
locked partial path with the remaining cities added sequentially
from the previous path. Proceeding in the similar manner,
repeat this steps till a total of four new paths are generated and
the best path among them is taken as the current path. The
detailed illustration of the Heuristic 2 is given in Fig. 2.
A. Experimental Framework
In order to measure the performance of the proposed
method over other existing algorithms, two metrics, namely,
percentage error of the average solution obtained by an
algorithm with respect to the optimal solution, PEavg (%) and
percentage error of the best solution obtained by an algorithm
with respect to the optimal solution, PEbest (%) are usually used
for each problem instance in the TSP literature. These
performance measures for an algorithm for a particular
problem instance are defined as follows:

PEavg 
average solution by an alorithm - optimal solution
100
(10)
optimal solution

PEbest 
best solution by an algorithm - optimal solution
100
(11)
optimal solution

The parameters of the QCSA selected are given as:

Population size = 30; quantum mutation probability (pm) =
0.40; fraction of worst nests (pa) = 0.40; number of generations,
maxcount = 2000 for first 15 instances, whereas maxcount =
1000 for the rest in table 2. The parameters of the proposed SA
are takes as: initial temperature (T1) = 5000; temperature
reduction factor (α) = 0.9 and max_iteration = 30,000.

B. Comparison with Existing Algorithms

The proposed algorithm is compared with the best-known
evolutionary algorithms in the literature [2-8, 21]. Table 2
presents different TSP instances.
Table 4 displays the comparative evaluation of the Dis-PSO
[3] and the proposed method based on PEavg. The result show
that the PEavg of the proposed method is significantly less than
Fig. 2. Illustration of Heuristic 2 the existing method. Similarly, as depicted in tables 5-10, an
identical level of performance of the proposed method
3) Random neighborhood techniques continues as well compared to the respective algorithms [2, 4-
In this study, three commonly used random neighborhood 6, 8, 21]. While comparing with the ELC-LR [7] based on the
techniques applied to the path obtained from the two heuristics results of table 11, the proposed method shows similar
are pairwise interchange [18], inverse [19], and shift (forward performance with the ELC-LR. However, if we consider the
or backward) [18]. In addition to these methods, we present PEbest, the proposed method shows improved results over the
another interchange neighborhood method. To explain this ELC-LR.
method, consider an n-city path and two distinct random
integers are generated between 1 and n (inclusive). For each of Table 3 shows the CPU times (s) taken by the proposed
the random integers, the parts of the path before and after the method for different TSP instances. Comparing the
random integer are interchanged, resulting in a new path. Thus, computational times required by the existing algorithms, it is
using the four random neighborhood techniques, all the seen that the proposed method takes slightly more time except
complete paths generated are compared and the best among that of RABNET-TSP [21].
them is selected.
TABLE II. DIFFERENT TSP INSTANCES
IV. COMPUTATIONAL EXPERIMENTATION Sl. No. Instances Sl. No. Instance Sl. No. Instance
We now present the experimental setup and the 1 eil51 10 kroC100 19 pr144
2 berlin52 11 kroD100 20 ch150
computational results to evaluate the performance of the 3 st70 12 kroE100 21 kroA150
proposed hybrid QCSA. For this purpose, the proposed method 4 eil76 13 eil101 22 kroB150
has been coded in C programming and executed on a PC with 5 pr76 14 lin105 23 pr152
Intel Core 2 Duo CPU with 2 GB RAM running at 2.93 GHz 6 rat99 15 pr107 24 u159
processor. In order to evaluate the solution quality and 7 rd100 16 pr124 25 rat195
efficiency of the proposed method, the experiments are 8 kroA100 17 bier127 26 d198
9 kroB100 18 ch130
performed based on 26 benchmark problems from TSPLIB
with five trials [20].
 TABLE III. RESULTS OF THE PROPOSED ALGORITHM FOR THE TABLE VII. COMPARISON OF THE IMPROVED-EN [] WITH THE PROPOSED
BENCHMARK PROBLEM INSTANCES FROM TSPLIB METHOD

Sl. PEavg PEbest CPU Improved-EN Proposed method

Opt Best Avg Worst Problem
No. (%) (%) time (s) PEavg (%) PEavg (%) PEbest (%)
1 426 426 427 428 0.32 0.00 4.96 8 2.09 0.32 0.01
2 7542 7544 7544 7544 0.02 0.02 5.01 9 3.02 0.60 0.31
3 675 677 677 677 0.29 0.29 7.95 10 2.33 0.47 0.34
4 638 544 547 548 1.63 1.11 8.92 12 4.22 0.44 0.35
5 108159 108159 108202 108266 0.03 0.00 8.89 22 5.56 3.52 2.99
6 1211 1223 1226 1228 1.25 0.99 12.50 23 3.80 2.83 2.21
7 7910 7918 7975 7998 0.82 0.10 12.76
Average 3.24 1.06 0.80
8 21282 21285 21352 21377 0.32 0.01 12.73
9 22141 22210 22274 22297 0.60 0.31 12.76
10 20749 20820 20848 20883 0.47 0.34 12.70 TABLE VIII. COMPARISON OF THE RABNET-TSP [21] WITH THE
PROPOSED METHOD
11 21294 21526 21560 21584 1.24 1.08 12.70
12 22068 22146 22166 22181 0.44 0.35 12.67 RABNET-TSP Proposed method
13 629 646 646 647 2.73 2.70 12.96 Problem
PEavg (%) PEavg (%) PEbest (%)
14 14379 14382 14384 14394 0.03 0.02 13.68
1 2.69 0.32 0.00
15 44303 44480 44523 44562 0.49 0.39 14.37
2 5.18 0.02 0.02
16 59030 59030 59150 59261 0.20 0.00 0.00
4 3.41 1.63 1.11
17 118282 118882 119263 119547 0.82 0.50 11.56
18 6110 6159 6181 6211 1.16 0.80 12.59 7 3.66 0.82 0.10
19 58537 58615 58625 58638 0.15 0.13 14.85 8 1.13 0.32 0.01
20 6528 6624 6635 6640 1.64 1.47 16.07 9 2.35 0.60 0.31
21 26524 27318 27458 27540 3.52 2.99 15.45 10 1.07 0.47 0.34
22 26130 26709 26871 26936 2.83 2.21 15.40 11 1.89 1.24 1.08
23 73682 74472 74556 74691 1.18 1.07 15.98 12 2.93 0.44 0.35
24 42080 42394 42775 43396 1.65 0.74 17.45 13 3.12 2.73 2.70
25 2323 2365 2382 2393 2.54 1.80 24.04 14 0.15 0.03 0.02
26 15780 16042 16114 16182 2.11 1.66 24.60 17 2.20 0.82 0.50
18 2.82 1.16 0.80
TABLE IV. COMPARISON OF THE DIS-PSO [3] WITH THE PROPOSED 20 3.22 1.64 1.47
METHOD 21 3.14 3.52 2.99
22 1.92 2.83 2.21
Problem Dis-PSO Proposed method
Average 2.55 1.16 0.87
PEavg (%) PEavg (%) PEbest (%)
eil51 2.57 0.32 0.00
berlin52 3.84 0.02 0.02 TABLE IX. COMPARISON OF THE GCGA [5] WITH THE PROPOSED
METHOD
st70 3.34 0.29 0.29
eil76 4.16 1.63 1.11 GCGA Proposed method
pr76 3.81 0.03 0.00 Problem
PEavg (%) PEavg (%) PEbest (%)
Average 3.54 0.45 0.28 1 0.94 0.32 0.00
3 0.44 0.29 0.29
TABLE V. COMPARISON OF THE ACOMAC-DNN [4] WITH THE 4 2.42 1.63 1.11
PROPOSED METHOD
5 0.72 0.03 0.00
Problem ACOMAC-DNN Proposed method 6 2.23 1.25 0.99
PEavg (%) PEavg (%) PEbest (%) 7 1.53 0.82 0.10
eil51 0.94 0.32 0.00 8 1.23 0.32 0.01
eil76 2.71 1.63 1.11 0.60 0.31
9 1.81
kroA100 0.59 0.32 0.01
10 1.33 0.47 0.34
eil76 1.11 2.11 1.66
11 2.42 1.24 1.08
Average 1.34 1.09 0.69
12 1.41 0.44 0.35
TABLE VI. COMPARISON OF THE MGSOM [6] WITH THE PROPOSED 13 2.70 2.73 2.70
METHOD 14 1.15 0.03 0.02
15 1.37 0.49 0.39
MGSOM Proposed method
Problem 16 0.19 0.20 0.00
PEavg (%) PEavg (%) PEbest (%)
1 1.39 0.32 0.00 17 1.80 0.82 0.50
3 1.18 0.29 0.29 19 0.04 0.15 0.13
4 3.38 1.63 1.11 21 2.92 3.52 2.99
7 1.17 0.82 0.10 22 2.11 2.83 2.21
14 0.02 0.03 0.02 23 1.22 1.18 1.07
15 0.17 0.49 0.39 24 2.28 1.65 0.74
17 1.09 0.82 0.50
25 4.18 2.54 1.80
23 0.74 1.18 1.07
26 1.93 2.11 1.66
25 5.98 2.54 1.80
Average 1.70 1.16 0.81 Average 1.66 1.11 0.81
 TABLE X. COMPARISON OF THE MEMETIC-SOM [2] WITH THE
PROPOSED METHOD References
Memetic-SOM Proposed method [1] K. Meer, “Simulated annealing versus metropolis for a TSP instance,”
Problem Information Processing Letters, vol. 104, pp. 216-219, 2007.
PEavg (%) PEavg (%) PEbest (%)
1 2.14 0.32 0.00 [2] J. C. Creput, and A. Koukam, “A memetic neural network for the
0.29 0.29 Euclidean travelling salesman problem,” Neurocomputing, vol. 72, pp.
3 0.99 1250-1264, 2009.
4 2.88 1.63 1.11
[3] X. H. Shi, Y. C. Liang, H. P. Lee, C. Lu, and Q. X. Wang, “Particle
7 2.65 0.82 0.10
swarm optimization-based algorithms for TSP and generalized TSP,”
8 1.14 0.32 0.01 Information Processing Letters, vol. 103, pp. 169-176, 2007.
13 3.15 2.73 2.70 [4] C. F. Tsai, C. W. Tsai, and C. C. Tseng, “A new hybrid heuristic
14 0.34 0.03 0.02 approach for solving large travelling salesman problem,” Information
15 0.67 0.49 0.39 Sciences, vol. 166, pp. 67-81, 2004.
16 1.52 0.20 0.00 [5] J. Yang, C. Wu, H. P. Lee, and Y. Liang, “Solving travelling salesman
17 2.78 0.82 0.50 problem using generalized chromosome genetic algorithms,” Progress in
3.52 2.99 Natural Science, vol. 18, pp. 887-892, 2008.
21 2.73
23 2.60 1.18 1.07 [6] Y. Bai, W. Zhang, and Z. Jin, “A new self-organizing maps strategy for
solving the travelling salesman problem,” Chaos, Solitons and Fractals,
25 6.59 2.54 1.80
vol. 28, pp. 1082-1089, 2006.
Average 2.32 1.14 0.84 [7] R. Zamani, and S. K. Lau, “Embedding learning capability in
Lagrangian relaxation: an application to the travelling salesman
TABLE XI. COMPARISON OF THE ELC-LR [7] WITH THE PROPOSED problem,” European Journal of Operational Research, vol. 201, pp. 82-
METHOD 88, 2010.
ELC-LR Proposed method [8] J. Yi, Z. Yang, and Z. Tang, “An improved elastic net method for
Problem travelling salesman problem,” Neurocomputing, vol. 72, pp. 1329-1335,
PEavg (%) PEavg (%) PEbest (%)
2009.
1 0.94 0.32 0.00
3 0.74 0.29 0.29 [9] K. –H. Han, and J. -H. Kim, “Quantum-inspired evolutionary algorithm
4 0.37 1.63 1.11 for a class of combinatorial optimization,” IEEE Transactions on
5 2.82 0.03 0.00 evolutionary Computation, vol. 6, pp. 580-593, 2002.
6 0.50 1.25 0.99 [10] H. Nezamabadi-pour, “A quantum-inspired gravitational search
7 0.16 0.82 0.10 algorithm for binary encoded optimization problems,” Engineering
8 1.63 0.32 0.01 Applications of Artificial Intelligence, vol. 40, pp. 62-75, 2015.
9 1.39 0.60 0.31 [11] A. Draa, M. Souham, T. Hichem, and B. Mohamed, “A quantum-
10 1.34 0.47 0.34 inspired differential evolution algorithm for solving the N-queens
11 0.72 1.24 1.08 problem," Neural networks, vol. 1, pp. 12, 2011.
12 1.22 0.44 0.35 [12] A. Layeb, “A novel quantum inspired cuckoo search for knapsack
13 0.32 2.73 2.70 problems,” International Journal of Bio-Inspired Computation, vol. 3,
14 0.06 0.03 0.02 pp. 297-305, 2011.
15 4.09 0.49 0.39 [13] J. Xiao, Y. Yu, J. Ping, J. Zhang, and Y. Tang, "A quantum-inspired
16 1.30 0.20 0.00 genetic algorithm for k-means clustering," Expert Systems with
17 0.66 0.82 0.50 Applications, vol. 37, pp. 4966-4973, 2010.
18 0.59 1.16 0.80 [14] X. S. Yang, and S. Deb, “Cuckoo search via Levy flights. In: Nature &
20 0.61 1.64 1.47 biologically Inspired Computing,” NaBIC 2009, World Congress on,
21 0.85 3.52 2.99 IEEE, pp. 210-214, 2009.
22 1.53 2.83 2.21 [15] X. S. Yang, and S. Deb, “Cuckoo search: recent advances and
23 2.58 1.18 1.07 applications,” Neural Computation and Applications, vol. 24, pp. 169-
24 0.37 1.65 0.74 174, 2014.
25 1.29 2.54 1.80 [16] J. Xiao, Y. Yan, Zhang, and Y. Tang, “A quantum-inspired genetic
Average 1.13 1.13 0.83 algorithm for k-means clustering,” Expert systems with Applications,
vol. 37, pp. 4966-4973, 2010.
V. CONCLUSION [17] S. Kirkpatrick, C. D. Gelatt, and M. P. Vecchi, "Optimization by
simulated annealing," Science vol. 220, pp. 671-680, 1983.
Recently, a new hybrid algorithm combining the concept of
[18] D. Laha, and U. K. Chakraborty, "An efficient stochastic hybrid
quantum computing and metaheuristic search algorithms has heuristic for flowshop scheduling," Engineering Applications of
attracted research interests. Quantum-inspired metaheuristic Artificial Intelligence vol. 20, pp. 851-856, 2007.
algorithm employs computational methods based on principles [19] O. Engin, and A. Döyen, "A new approach to solve hybrid flow shop
of quantum computing such as quantum bits, superposition, scheduling problems by artificial immune system," Future Generation
mutation and interference operators, and quantum Computer Systems, vol. 20, pp. 1083-1095, 2004.
measurement. In this paper, a hybrid of QCSA and SA is [20] X. Geng, Z. Chen, W. Yang, D. Shi, and K. Zhao, “Solving the
considered as a novel evolutionary computing for the travelling travelling salesman problem based on an adaptive simulated annealing
algorithm with greedy search,” Applied Soft Computing, vol. 11, pp.
salesman problems. This approach has shown to produce 3680-3689, 2011.
solutions that are either better than or competitive with the
[21] T. A. S. Masutti, and L. N. de Castro, “A self-organizing neural network
other notable evolutionary algorithms. For future using ideas from the immune system to solve the travelling salesman
investigations, it would be useful to assess the ability of the problem,” Information Sciences, vol. 179, pp. 1454-1468, 2009.
proposed algorithm in practical applications such as bin
packing problems and other image processing problems.
 EWSA: An Enriched Workflow Scheduling
Algorithm in Cloud Computing
Shilpi Saxena
Department of Computer Science and Information
Technology
Graphic Era Hill University
Dehradun, India
25shilpisaxena@gmail.com
Abstract—Cloud computing environment is one of the latest
research area where research is taking place at a very fast pace.
This concept of computing which is based on cloud is expanding
its umbrella day by day, taking almost every computing activity
under its shed either it is related to resources, tasks or data
storage. This hiking interest of users has made it quite
challenging for the providers to meet up the consumer desires.
Scheduling of tasks plays an important role in the
performance of any cloudlet scheduling algorithm. Tasks
scheduling if observed closely in realistic world is based on
dependent tasks i.e., they represents a workflow. A number of
scheduling algorithms have been proposed for independent tasks
but very few are there which operates on dependent tasks. Here,
in this paper we propose a new scheduling algorithm for
workflow, which are having dependencies among tasks, taking
into consideration important parameters of transfer time and
bandwidth along with basic requirements of optimizing the
execution time and cost. The simulation is experimented using
cloudSim toolkit. Our algorithm provides better results over
other existing algorithm like PSO and CSO and is more closely
related to real world scenario.
Keywords—Cloud computing; Workflow; Tasks scheduling;
CloudSim; Virtual Machine
I. INTRODUCTION
Cloud computing is very popular these days. The reason
behind this popularity is the type of services offered by it. It
has made it possible for everyone to access and use its services
as per user’s choice and requirements. The provision of
services on rental basis is one of the major reasons for its
popularity. Besides this, the services it offers is also an
important reason as it not only offers ready to use applications
as a service but it also does provide the platform for
application development and its own infrastructure to build
and run those applications on rental basis. These three
services, Software-As-A-Service (SAAS), Platform-As-A-
Service (PAAS) and Infrastructure-As-A-Service (IAAS) are
unique in the world of computing. Cloud computing is
basically a computing which is performed over a number of
resources that are widely distributed at various datacenters
throughout the world. Some of the common challenges of any
computing are storage management, execution time, cost,
latency, energy, data security, virtualization and scheduling.
Scheduling of tasks is always been the primary concern for
Deepika Saxena
Department of Computer Science and Information
Technology
Dayanand Mahila Mahavidhyalaya
Kurukshetra, India
13deepikasaxena@gmail.com
better performance of any computing scheme. Scheduling of
tasks means defining the order in which tasks are going to be
executed on virtual machine. Cloud computing do provide
several special characteristics to workflows like:
1. Resources are available in the form of standardized
services and can be availed as per the user’s the
choice.
2. The number and type of resources allocated to a
workflow is decided as per the customer’s
requirement.
3. The number of resources allocated to a workflow can
be changed dynamically at runtime and hence rightly
said elastically scalable.
Broadly, the workflows can be categorized under two
categories, a) Business workflow and b) Scientific workflow.
A scientific workflow is basically represented by a Direct
Acyclic Graph (DAG). Thus, the tasks are dependent in nature
and while scheduling tasks this dependency needs to be taken
special care. The allocation of tasks and resources at runtime
and dynamically mapping resources to meet the performance
expectations is to be handled with deep care.
Scientific workflows usually represent a complex and huge
form of data. Handling, managing and performing calculations
over such type of data is a big challenge. Computation of such
type of data in an efficient way is surely going to contribute a
lot in handling the data of latest emerging technologies.
There are various scheduling algorithms which takes into
account Qos factors, but most of them are proposed for
independent tasks which is quite impractical in consideration
with the real time scenario. In this paper, we have proposed an
optimized workflow scheduling algorithm which takes into
account almost every kind of important QoS factors like
energy, cost of execution, meeting deadline, etc. that’s why
the algorithm is named as “EWSA: An Enriched Workflow
Scheduling Algorithm”. Our algorithm shows improved
performance results over existing algorithms.
The rest of the paper comprises of following sections,
Second section enlists the previous work related to this. Third
section describes the proposed methodology, its algorithm and
flowchart. Fourth section shows the results in tabular form as
well as in graphical format. Finally, the fifth section concludes
our work and shows future perspectives.
II. RELATED WORK
Mapping of tasks to machines for their execution is termed
as scheduling. This scheduling can be done using various
scheduling strategies. The performance of any machine
depends truly on the scheduling strategy followed for directing
the tasks on the machine. Again it should also be considered
that which type of task is to be scheduled, as the tasks can be
dependent one or independent one. It is quite easy to schedule
independent tasks but scheduling of dependent tasks i.e.,
workflows is bit challenging, as you have to keep track of the
dependency of tasks. It is to be closely observed and handled
that which task is dependent on which one, and which task
needs to be performed before its dependent tasks. This thing is
bit complex about the workflows. As the workflows are
closely related to real time tasks and are dynamic in nature.
Thus more pragmatic algorithms should be designed to deal
with this in an efficient way.
[7] shows scheduling of workflows depending upon
different QoS requirements. It takes into consideration the
type of service a ready task is demanding and based upon its
requirement the scheduler checks if that service can be
provided and assigns the task to that resource. Otherwise, the
task have to wait for its turn. After regular intervals this
availability of resources is checked, as a resource becomes
available it is assigned to the waiting task. The problem with
this approach is that, a lot of time is wasted in calculating the
status of resources after every round. Apart from this, there is
no strategy defined that in which order the waiting tasks will
be assigned to the machine.[8] provides a compromised
approach for time and cost based scheduling of workflows.
This technique computes cost of execution and time of
execution for each task on every resource, and then assigns the
task to the resource which executes the task within deadline
and incurs minimum cost. The major flaw of this approach is
that, after every round it finds out the tasks that have failed to
execute successfully and gives priority to them over new tasks
to be assigned to a resource for execution that is, the new tasks
in the workflow suffers starvation. Due to this, the workflows
may fail to achieve their deadlines. [9] proposed a Particle
Swarm Optimization (PSO) algorithm for workflows. It works
on QoS features time and cost, and improves the distribution
of workload and achieves cost optimization over existing
algorithm Best Resource Selection (BRS). The technique
works on heuristic based PSO, it should be modified for more
realistic approaches. [10] Implements a workflow scheduling
algorithm which surpasses many other existing algorithms in
terms of execution time and scaling elastically.
[11] a three bi-criteria complementary approach has been
devised which focuses on optimization of cost of execution
and time of execution while scheduling of workflows. Cat
Swarm Optimization (CSO) workflow scheduling technique is
more close to realistic scheduling as it considers both cost of
data transmission between two dependent resources as well as
cost of execution of tasks on different resources [12]. [12] the
technique shows improved results in terms of task to resource
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
scheduling cost, number of iterations and load distribution
among available resources. [13] a cloud based workflow
architecture for Smart city is been introduced here and some
of the workflow scheduling algorithms are also been reviewed
to define a taxonomy of workflow scheduling and
management in cloud environment. [14] an optimized
scheduling scheme is proposed for highly dependent tasks to
minimize the energy consumption and enhance the lifetime of
the network.
Thus, various schemes have been proposed with time to
schedule workflows in the cloud environment. Every scheme
considers one or the other QoS factor and tries to provide
better results over the existing ones. But in the proposed
algorithm we have undertaken all the good points of existing
algorithm with the added improvement which is provided by
assigning level to each task in the workflow, while
considering their parent child relationship.
III. WORKFLOW SCHEDULING
The workflow is considered in form of Directed Acyclic
Graph (DAG) G=(Vr,E), where Vr is defined as the set of
tasks T={T1,T2,......,Tn} and E is set of edges that represents
dependency between resources. It is assumed that a set of
processing resources is represented by R= {R1, R2,......, Rn}.
The processing resources are distributed throughout the
network.
The experimental workflow assumed here consists of 11
tasks which are divided into 5 different levels to be scheduled
on six different virtual machines. Fig.1 represents the
workflow considered in our proposed work.
Fig.1 Workflow for Experiment
Task T1 is the initial task which is at first level, task T2
and T3 are dependent on task T1 and are placed at level 2.
Task T4, T5 and T6 are dependent on task T2 and task T7 and
T8 are dependent on task T3. All these are placed at level 3.
Further task T9 is dependent on task T4 and T5 and task T10
is dependent on T7 and T8, it is at level 4. Finally, the
workflow completes at task T11which is dependent on task
T9, T6 and T10 and is at level 5.
A. Methodology of Proposed Algorithm
1. User submits workflows at datacenter. Sorting of all the
available workflow is done in ascending order of deadline
is done.
2. Take smallest deadline workflow from top of list. For
scheduling each workflow follow the below steps.
 a) Prepare transfer time matrix for the workflow and The proposed algorithm is clearly and concisely
also assign levels to each task in the workflow. explained through the flowchart given in figure 2.
b) Child task belonging to same parent task are at Simultaneously multiple workflows are submitted at cloud
same level and total number of task at each level datacenter, but which workflow is to be executed first is
are counted. decided on the basis of deadline for its completion. The
3. For assigning task belonging to a workflow to resource workflow with shortest deadline is selected and its transfer
VM for execution, do follow below steps: time matrix is calculated. Also, the tasks are divided into
a) At each level one by one each task is checked that levels based upon their dependency on another task. The
if it is having parent task, then first execute its independent tasks are placed at level 1, the task which takes
parent task, and calculate its file output size. input from level 1 tasks only are placed at level 2 and the
Otherwise, simply assign the task to execute on tasks dependent on either level 2 tasks or both level 1 and
minimum turnaround machine. level 2 tasks are placed at level 3 and so on. Flowchart in
b) Now, if parent task execution is done, then all child Fig. 2 is showing steps for scheduling each workflow.
task belonging to same parent task and at same
level are executed in parallel on same or different IV. EXPERIMENTAL SETUP AND RESULTS
VM. A. Experimental Data
c) Data transfer time from parent task to its respective
The experiment is conducted using the CloudSim simulator.
child task is calculated as follows:
The experiment is performed for the workflows which have
Transfer time = Cloudlet.outputSize /
been scheduled on six different virtual machines having
Vmlist.get(id).getBw()
Here parent task’s output size and VM’s bandwidth configuration as given below in TABLE I.
is use for calculation of transfer time. Total
TABLE I. Configuration of VMs
execution time for each cloudlet is calculated as Vms Vm1 Vm2 Vm3 Vm4 Vm5 Vm6
follows: RAM 5024 5024 5024 5024 5024 5024
Totalexecution time = (cloudlet.getCloudletLength Processing 22000 19800 15200 14400 11000 10098
( ) / processing power) + transfer time Power
4. Cost of transfer time will be zero if child task is executed (MIPS)
on same VM as parent task. Otherwise, cost per Processing 1 2 2 4 2 4
Element
bandwidth is calculated. (CPU)
Transfer cost =costPerBw * getCloudletOutputSize( ) Bandwidth 1000 900 850 725 500 500
5. Finally, checks if all task levels are executed, then stop,
otherwise go to step 3. The transfer time between tasks calculated using above
formula for the given workflow is given below TABLE II.
Prepare transfer time matrix
and calculate levels and TABLE II. Transfer time between tasks
number of task at each level T1 T2 T3 T4 T5 T6 T7 T8 T9 T10 T11
T1 0 0.2 0.5 0 0 0 0 0 0 0 0
T2 0 0 0 0.8 0.7 0.1 0 0 0 0 0
One by one task are selected T3 0 0 0 0 0 0 0.9 0.4 0 0 0
for execution at each level T4 0 0 0 0 0 0 0 0 0.6 0 0
T5 0 0 0 0 0 0 0 0 0.4 0 0
Assign selected task to be T6 0 0 0 0 0 0 0 0 0 0 0.8
executed on minimum T7 0 0 0 0 0 0 0 0 0 0.2 0
Turnaround Vm T8 0 0 0 0 0 0 0 0 0 0.4 0
T9 0 0 0 0 0 0 0 0 0 0 0.6
T10 0 0 0 0 0 0 0 0 0 0 0.8
No
T11 0 0 0 0 0 0 0 0 0 0 0
If task, has
Yes parent task?
B. Results Obtained

Total time of execution= Total execution
processing time + transfer time processing time
Cost of execution = task length x Cost of execution = task
cost of Vm + transfer cost length x cost of Vm
Combine result
Fig. 2 Flowchart of EWSA
time = The EWSA algorithm is compared with Cat Swarm
Optimization (CSO) technique [12]. The results obtained by
running the simulation clearly shows that EWSA algorithm
provides a more optimized solution in terms of scheduling
workflow tasks in a more realistic way. The experiment is
performed for different set of workflows as mentioned in table
3 and table 4. The tabular as well as graphical comparisons are
shown among both the algorithms in terms of Execution time
and Cost of execution.
 TABLE III. Execution time Comparison workflow. For each number of task, we have executed the
Number of task in CSO algorithm EWSA algorithm algorithm 20 times and mean value of that 20 execution
workflow was taken as final value for each respective number of task.
20 98.7135 76.5634 Fig. 3 and Fig. 4 shows distribution of workload for
30 283.8921 214.5672
different number of tasks with respect to execution time and
execution cost respectively, which is different each time
40 729.4284 576.3049
depending upon the file size, output size of file, bandwidth
50 799.6712 652.8310 of Vm on which it is executed. But the results are improved
60 819.3914 734.2735 each time as compared to CSO algorithm.

900 V. CONCLUSION AND FUTURE WORK

800 The paper discusses the workflow scheduling technique
700 which takes into consideration important QoS factors time
of execution, cost of execution and the transfer time
600
between the dependent tasks. The algorithm provides more
500 optimized results as compared to other techniques like
CSO
400 CSO. Thus, the paper concludes with the fact that our
EWSA proposed algorithm performs better over CSO and other
300
existing techniques. This is been clearly proven by the
200 simulation results obtained, as each time the results are
100 improved. In future, the work can be extended taking into
0 consideration more factors like energy consumption which
gives it a more realistic approach.
20 30 40 50 60

Fig. 3 Execution time Comparison REFERENCES

[1] J. Geelan, "Twenty-one experts define cloud computing," Cloud
TABLE IV. Cost of Execution Comparison Computing Journal, vol. 2009, pp. 1-5, 2009.
Number of task [2] R. Mikkilineni and V. Sarathy. (2009). Cloud Computing and the
Lessons from the Past in 18th IEEE International Workshop on Enabling
in workflow CSO Algorithm EWSA Algorithm Technologies: Infrastructures for Collaborative Enterprises,
WETICE'09,2009.
20 77.842 75.325
[3] S. Ma, "A Review on Cloud Computing Development," Journal of
30 129.316 110.566 Networks, vol. 7, no.2, pp. 305-310, 2012.
[4] P. Mell and T. Grance, "The NIST definition of cloud computing,"
40 145.001 139.458 National Institute of Standards and Technology, vol. 53, no.6, 2009.
[5] I. Foster, Z. Yong, I. Raicu, and S. Lu, "Cloud computing and grid
50 211.395 188.739
computing 360-degree compared," 2008, pp. 1-10.
[6] CloudComputing[Online]:
60 385.568 298.526
http://en.wikipedia.org/wiki/Cloud_computing
[7] Meng Xu, Lizhen Cui, Haiyang Wang, Yanbing Bi, “ A Multiple QoS
Constrained Scheduling Strategy of Multiple Workflows for Cloud
450 Computing”. 2009 IEEE International Symposium on Parallel and
Distributed Processing with Applications.
400 [8] Ke Liu1,2, Hai Jin1, Jinjun Chen2, Xiao Liu2, Dong Yuan2 and Yun
350 Yang 2, “A Compromised Time-Cost Scheduling Algorithm in SwineDe
W-C for Instance-Intensive Cost-Constrained Workflows on a Cloud
300 Computing Platform”. International Journal of High Performance
250 Computing Applications OnlineFirst, published on May 24, 2010 as
CSO doi:10.1177/1094342010369114
200 [9] Suraj Pandey1, LinlinWu1, Siddeswara Mayura Guru2, Rajkumar
EWSA Buyya1, “A Particle Swarm Optimization based Heuristicfor Scheduling
150
Workflow Applications in Cloud Computing Environment”. 2010 24th
100 IEEE International Conference on Advanced Information Networking
and Applications.
50
[10] Cui Lin and Shiyong Lu, “Scheduling Scientific Workflows Elastically
0 for Cloud Computing”. 2011 IEEE 4th International Conference on
20 30 40 50 60 Cloud Computing.
[11] Kahina∗∗Bessai ∗, Samir Youcef∗∗, Ammar Oulamara∗∗, Claude Godart
and Selmin Nurcan, “Bi-crieteria Workflow task allocation and
Fig. 4 Cost of Execution Comparison scheduling in Cloud Computing Environment”. 2012 IEEE Fifth
International Conference on Cloud Computing.
We experimented by taking different number of [12] Saurabh Bilgaiyan, Santwana Sagnika, Madhabananda Da, “ Workflow
cloudlets (different amount of data) processed by the Scheduling in Cloud Computing Environment Using Cat Swarm
 Optimization” 2014 IEEE International Advance Computing Conference
(IACC).
[13] Li Liu, Miao Zhang, Yuqing Lin and Liangjuan Qin, “A Survey on
Workflow Management and Scheduling in Cloud Computing”. 2014
14th IEEE/ACM International Symposium on Cluster, Cloud and Grid
Computing.
[14] Elaine N. Watanabe and Pedro P. V. Campos and Kelly R. Braghetto
and Daniel M. Batista, “Energy saving algorithms for workflow
scheduling in cloud computing”. 2014 Brazilian Symposium on
Computer Networks and Distributed Systems.
[15] R. Mikkilineni and V. Sarathy. (2009). Cloud Computing and the
Lessons from the Past in 18th IEEE International Workshop on Enabling
Technologies: Infrastructures for Collaborative Enterprises,
WETICE'09,2009.
[16] S. Ma, "A Review on Cloud Computing Development," Journal of
Networks, vol. 7, no.2, pp. 305-310, 2012
[17] A.Fox and R. Griffith, "Above the clouds: A Berkeley view of cloud
computing," Dept. Electrical Engineering and Computer. Sciences,
University of California, Berkeley.
[18] S. Singh and K. Kant, "Greedy grid scheduling algorithm in dynamic job
submission environment," in Emerging Trends in Electrical and
Computer Technology (ICETECT), 2011 International Conference on,
2011, pp. 933-936.
 An Analysis of the Suitability of Cloud Computing
Services in the Nigerian Education Landscape
Sanjay Misra and Adewole Adewumi
Department of Computer and Information Sciences
Covenant University
Ota, Nigeria
{sanjay.misra,wole.adewumi}@covenantuniversity.edu.ng
Abstract—Cloud computing is fast gaining popularity in
educational institutions of developing countries like Nigeria.
Software as a Service, Platform as a Service and Infrastructure
as a Service are the three key models through which cloud
computing services are delivered to end-users. A number of
studies have been conducted to identify the enabling factors as
well as the issues being faced as regards the adoption of cloud
computing in the Nigerian context. In this study however,
Strength, Weakness, Opportunity and Threat analysis of the
service delivery models in the Nigerian Education landscape has
been presented. In addition, the issues that an educational
institution needs to consider when adopting cloud computing is
discussed.
Keywords—cloud computing; SWOT analysis; suitability;
Nigerian education; service delivery models
I. INTRODUCTION
Cloud computing has some characteristics that distinguish
it from other technologies [1] [2]. They include the following:
Users do not necessarily have to own the information
technology (IT) resources they utilize. For instance, the servers
they interact with might be hosted in data centers at remote
locations from them. Also, services are provided on-demand to
the end-users and the end-users only pay for what is used.
Cloud services can be delivered as software (SaaS), platform,
(PaaS), or infrastructure (IaaS) [1]. SaaS is a model in which
application software is delivered via the Internet [3]. PaaS is a
model where, the service providers supply services to the users,
such as development environment, and server platforms
through which the users can develop custom applications. In
IaaS, computer infrastructure such as servers, and storage
devices are remotely delivered through the Internet.
Cloud computing has moved from just being a topic of
interest and debate to one that is being adopted and applied to
various aspects of the economy in sub-Saharan Africa -
prominently the enterprise [4] and in recent times the education
landscape. A recent study has identified the need for Nigerian
educational institutions to have a plan of action for the
adoption of cloud services [5]. The need arises thus for an
analysis of the applicability of the various cloud service
offerings within the Nigerian educational institution context.
This is the main motivation behind this study. Also, the various
challenges that may be faced are outlined as well as possible
solutions. In addition, the issues the institutions need to
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
consider when making the move to the cloud are also
discussed. Hence, the rest of this paper is structured as follows:
Section 2 reviews related works. In Section 3, Strengths,
Weakness, Opportunities and Threats (SWOT) analysis is
performed to determine the applicability of the applicability of
the various cloud service offerings in the Nigerian educational
institutions. In Section 4 discusses the issues to be considered
by institutions intending to adopt any of the service delivery
models. Section 5 concludes the paper.
II. RELATED WORKS
In [6] an empirical study was conducted to determine the
potential for the adoption of grid computing in tertiary
institutions of Nigeria. Although, grid computing is not exactly
cloud computing, they share a lot in common especially in
terms of vision, architecture and technology [7]. The study
revealed a significant lack of awareness about the benefits of
grid computing particularly in tertiary institutions, which
generally prevented its low adoption and suggested investing in
awareness initiatives, workshops as well as the acquisition of
grid resources to facilitate adoption.
In [8] an empirical study was also conducted to assess
affordances of selected cloud computing tools for language
teacher education in Nigeria. The study revealed that
participants were able to perceive the opportunities inherent in
the use of cloud computing for classroom learning as well as
the unintended affordances. The study however was focused on
language Teacher Educators in Colleges of Education and the
scope of tools studied comprised mainly of SaaS tools such as
DropBox and Google Drive.
In [9] a critical analysis of the benefits and challenges of
the adoption and usage of cloud computing in Nigeria was
performed. In the study, the relationship between key
stakeholders in the Nigerian cloud ecosystem and proposed
methods for optimizing the benefits of cloud computing while
reducing the inherent adoption challenges was presented. This
study however, focused on businesses and corporate
organizations as being the consumers of cloud technology and
not educational institutions. The study also deemphasizes the
cloud service delivery channels (IaaS, PaaS, and SaaS) arguing
that the rise of cloud ecosystem would render them irrelevant.
Therefore, the challenges identified in this study were for the
cloud ecosystem and not for the service delivery models.
 In [10] a framework was proposed that would help
organization of public and private sectors in Bangladesh to
adopt cloud computing technology opportunities and prevent
its obstacles. Although SWOT Analysis was performed in this
study, it was not directed at the service delivery channels of
cloud computing.
In the study [11], the author investigated cloud service
adoption extent in Nigeria so as to identify the motivating
factors and current issues affecting the adoption of cloud
computing in Nigeria. The participants comprised of
employees in Information Technology as well as
Telecommunication companies in Nigeria. The outcome of the
study showed that the level of the adoption was still low. Also
of all the available cloud computing services, SaaS was the
most used in Nigeria. Increased focus on primary services,
collaboration, easy access to data and provision of basic
infrastructure were identified as the motivating factors for
cloud computing adoption in Nigeria. Poor awareness of cloud
computing, unstable power supply and high cost of Internet
bandwidth and unreliability of Internet services were the
negative factors affecting adoption. This study is limited in the
sense that the participants are mainly from the Industry and not
academia and so the results may not necessarily be true
representation of the Nigerian scenario.
In [12] the opportunities and challenges affecting the
adoption of cloud computing in Nigeria was also discussed.
The author made recommendations and stressed the need for
the nation to determine whether it was going to function as
services provider, service implementer or content creator.
The study by [13] investigated the impact and challenges of
the adoption of cloud computing by ten public universities in
the Southwestern part of Nigeria. The study revealed 90%
adoption rate of cloud services. The most used cloud service
was SaaS (seven out of the ten universities) followed by PaaS
(two of the universities) and IaaS (by only one of the
universities). However, the authors call for further studies
focusing on readiness assessment of institutions' adoption of
the various cloud service delivery models.
III. SWOT ANALYSIS
A. IaaS in the Nigerian Education Context
This service delivery model although popular in some
African countries such as South Africa, it is yet to gain traction
in the Nigerian Education context due to the state of existing
infrastructure [13].
1) Strength: With IaaS, educational institutions will have
full control and are able to manage their computing resources.
2) Weakness
 Skilled personnel may be required: Since this service
delivery model has not been widely adopted in Nigeria
as a whole and particularly in the educational context,
the need for skilled personnel becomes imperative.
 Increased cost: IaaS providers to Nigeria can charge
premium prices especially where there is limited
competition and increasing demand. Institutions that
adopt IaaS would need to own data centers as part of
best practices for sustainable development. These data
centers require the provision of constant electric power
with the requisite output and stability for the IT
equipment as well as provision of adequate cooling
systems for the equipment. As such institutions would
require alternative means of generating power to
complement what is currently obtainable.
3) Opportunity: If the problem of power is addressed in
Nigeria, then this has great prospects for the Nigerian
education sector as well as other sectors. Also service
providers such as IBM [14], Amazon [15] and Google have
developed a number of solutions for the educational context at
subsidized rates, which might encourage more educational
institutions to buy into IaaS.
4) Threat: One major threat to the wide spread acceptance
of IaaS in the Nigerian education context will be power
supply. Until this is addressed, the question of setting up a
data center for sustainable development within Nigerian
educational institutions will not arise.
B. PaaS in the Nigerian Education Context
With PaaS, programmers that work in educational
institutions can develop cloud services and applications.
Google being an education-focused organization have
introduced Google App Engine as their PaaS, which supports a
number of programming languages. Although, it is difficult to
tell how many institutions actually use Google App Engine, a
recent literature shows that there is possibility of it being used
for research purposes in some Nigerian institutions [16].
1) Strength: PaaS provide development environments and
familiar programming languages for developers that can be
leveraged to develop custom applications either for
commercial – thereby helping the educational institution to
generate revenue; or research purposes as seen in [16].
2) Weakness: Vendor lock-in - when developing on a
particular platform, developers are usually restricted to what
obtains in that platform
3) Opportunity: Being an emerging area in the Nigerian
software development ecosystem, it could be of value
especially at the undergraduate level to incorporate PaaS
development into the curriculum.
4) Threat: Once IT staff (who are developers) adopt a
platform, it might be difficult to migrate to another platform.
C. SaaS in the Nigerian Education Context
In 2012, Microsoft launched Liv@edu (now integrated into
Office 365) for a number of secondary schools in Nigeria [17].
The service allows the use of Microsoft Office apps on
Windows and OS X, provides storage space on Microsoft’s
cloud storage service OneDrive. Also, a number of tertiary
institutions in Nigeria have subscribed to Google Apps for
Education [13] - a suite of free productivity tools for classroom
collaboration. A recent introduction into the suite is Classroom
[18], which was designed to save time, and keep classes
organized. It allows for the creation, sharing and grading of IaaS PaaS SaaS
assignments with ease. with development bandwidth by
technology technology giants
1) Strength giants such as
 Requires little physical infrastructure: An educational IBM,
institution adopting this service delivery model will not Amazon and
Google
require more than an Internet-enabled device to access Lack of control or
the services. access to
Migration of
underlying
 No need to purchase licenses or install any software: Threat
In consistent data in the event
technology
Educational institutions that adopt this platform will power supply of platform
Probability of
change
require no software licenses or updates as this will be service continuity
taken care of by the SaaS provider. Some providers by provider
may not even charge individual users within an
educational institution (e.g. Google).
IV. DISCUSSION
2) Weakness: Internet connectivity required to operate: The goal and mission of an institution will determine the
Majority of SaaS products do not operate offline. In fact, SaaS service delivery model that best suits it. An institution wanting
was never designed for offline use and so requires stable and to be in full control of and manage its computing resources can
constant Internet connectivity. Also functionalities are limited opt for the IaaS delivery model. Such an institution should
to what the provider implements. explore partnerships and collaborations with technology
3) Opportunity: The increasing adoption of Internet- companies such as IBM and Google (who currently have
enabled mobile devices in the Nigerian educational sector is a branches in Nigeria and offer IaaS services) to help in setting
great opportunity. For instance, the University of Ilorin has up the institution's data center and training the institution's
personnel. However, such an institution should ensure that it is
launched a scheme where Internet-enabled tablets are provided
able to provide constant power supply or has alternative power
to all students. It serves as a medium for receiving and storing supply sources such as generator sets.
lecture materials. Google has also embarked on initiatives to
make more bandwidth available to institutions subscribed to For an institution not necessarily wanting to own its data
the Google Apps for Education. center but requiring a platform that allows for some form of
4) Threat: If a SaaS provider ceases to operate then the customization or development either for the sake of research or
teaching can opt for the PaaS service delivery model. However,
software may become unavailable for use. Also, end-user
such an organization should do the needed research on the
institutions have no control or access to the underlying platform to be adopted. For instance, the institution should seek
technology. to know how long the provider has been in existence as well as
its customers' satisfaction. This is important so as to avoid
Table 1 provides a summary of the SWOT analysis of IaaS, vendor lock-in.
PaaS and SaaS in the Nigerian Education Landscape.
An institution that is seeking primarily to cut cost by
eliminating the need to constantly renew or purchase licenses
can opt for the SaaS model. Such an institution may not be
TABLE I. SWOT ANALYSIS interested in owning or running its own data center due to the
IaaS PaaS SaaS cost involved. The institution may also not be interested in
Reduced cost as a developing custom solutions to meet its need and is content
Ability to
Full control
develop custom
result of not with using ready-made solutions that are available on demand.
and access to needing to
applications for
Strength manage
commercial or
purchase licenses V. CONCLUSION
computing or possess too
research- This paper set out to analyze the suitability of the cloud
resources many physical
purposes computing service delivery models (IaaS, PaaS and IaaS) in the
infrastructure
Functionalities Nigerian education landscape. The motivation stemmed from
available in the the gap identified in literature, as previous studies did not
software is
address this issue. In order to carry out this task, SWOT
Skilled limited to what a
personnel provider analysis was conducted across the three platforms and the
Weakness Vendor lock-in results presented in Table 1. From the results, three kinds of
required implements
Increased cost Internet institution could be identified that would be seeking to adopt
connectivity often cloud computing. Issues that need to be attended to by the
required to various institutions were also discussed.
operate
Possible Reviewing the Ubiquity of
partnerships present Internet-enabled
Opportunity and undergraduate mobile devices as
collaboration curriculum to well as the
of institutions include PaaS provision of
 REFERENCES
[1] P. Mell, and T. Grance, The NIST definition of cloud computing
[Online]. Retrieved from http://csrc.nist.gov/publications/nistpubs/800-
145/SP800-145.pdf. Date accessed: 19th August 2015
[2] S. Jain, R. Kumar, and S. K. Jangir, "A Comparative Study for Cloud
Computing Platform on Open Source Software", International Journal of
Engineering & Technology (AIJET), vol. 1, 2014, pp. 28-35
[3] A. F. Mohammad and H. Mcheick, "Cloud services testing: An
understanding". Procedia Computer Science, vol. 5, 2011, pp. 513-520
[4] C. P. Uchenna, E. N. Godwin, E. O. Oliver, O. O. Moses, and J. Ahaiwe,
"Overcoming the Barriers to Enterprise Cloud Adoption within Nigerian
Consumer Constituency", British Journal of Mathematics & Computer
Science, vol. 8, 2015, pp. 39-56.
[5] E. Aginam, NIALS to NUC: Upgrade ICT curriculum with cloud
computing services [Online], Vanguard Newspaper Available from:
http://www.vanguardngr.com/2015/01/nials-nuc-upgrade-ict-
curriculum-cloud-computing-services/ Date accessed: 17th August 2015.
[6] F. M. E. Uzoka, B. A. Akinnuwesi, S. O. Olabiyisi, and A. Demilade,
"An empirical study of potentials of adoption of grid computing as a
vehicle for tertiary institutions collaboration," International Journal of
Business Information Systems, vol. 10, 2012, pp. 245-263
[7] I. Foster, Y. Zhao, I. Raicu, and S. Lu, "Cloud computing and grid
computing 360-degree compared". In Grid Computing Environments
Workshop, GCE'08, 2008, pp. 1-10
[8] A. Y. Ofemile, "Assessing Affordances of Selected Cloud Computing
Tools for Language Teacher Education in Nigeria", Journal of Education
and Practice, vol. 6, 2015, pp. 1-10
[9] F. M. Dahunsi, and T. M. Owoseni, "Cloud Computing in Nigeria: The
Cloud Ecosystem Perspective", Nigerian Journal of Technology, vol. 34,
2015, pp. 209-216.
[10] S. Islam, M. H. Kabir, M. J. Hossain, A. Chakraborty, and N. Majadi,
"Cloud Computing Technology in Bangladesh: A Framework of Social
and Economic Development", European Scientific Journal, vol. 11,
2015, pp. 393-410.
[11] R. K. Awosan, "Factor Analysis of the Adoption of Cloud Computing In
Nigeria", African Journal of Computing & ICT, vol. 7, 2014, pp. 33-42.
[12] E. M. Dogo, A. Salami, and S. Salman, "Feasibility Analysis of Critical
Factors Affecting Cloud Computing in Nigeria". International Journal of
Cloud Computing and Services Science (IJ-CLOSER), vol. 2, 2013, pp.
276-287.
[13] C. A. Oyeleye, T. M. Fagbola, and C. Y. Daramola, "The Impact and
Challenges of Cloud Computing Adoption on Public Universities in
Southwestern Nigeria", International Journal of Advanced Computer
Science and Applications, vol. 5, 2014, pp. 13-19.
[14] IBM, “Introducing the IBM Cloud Academy,” [Online] Available:
http://www.ibm.com/solutions/education/cloudacademy/us/en. Accessed
September 10, 2015.
[15] Amazon, “AWS in Education,” [Online] Available:
http://aws.amazon.com/education Accessed September 10, 2015
[16] J. O. Daramola, F. Olajide, A. O. Adewumi, and C. K. Ayo, "An
Experimental Validation of Public Cloud Mobile Banking" Research
Journal of Applied Sciences, Engineering and Technology, vol. 7,
(2014) pp. 5304-5314.
[17] E. Aginam, Microsoft takes Live@edu cloud service to LBIHS [Online],
Vanguard http://www.vanguardngr.com/2012/03/microsoft-takes-
liveedu-cloud-service-to-lbihs/#sthash.AhhAJ3Ye.dpuf Date accessed:
29th April, 2015.
[18] Google for Education [Online] Available at:
https://www.google.com/edu/products/productivity-
tools/classroom/index.html Date accessed: 7th September, 2015
 Developing a Mobile Application for Taxi Service
Company in Nigeria
Adewole Adewumi, Victor Odunjo and Sanjay Misra
Department of Computer and Information Sciences
Covenant University
Ota, Nigeria
{wole.adewumi, victor.odunjo, sanjay.misra}@covenantuniversity.edu.ng
Abstract—Transportation is an issue of concern in big cities
of many developing countries today. Due to the large population
in these cities, there is constant traffic congestion and pollution.
As a result taxi services are common. In Nigeria, companies
offering these services have discovered that they can better serve
the large population by providing their services through the
mobile platform. Given the wide spread adoption of smart
phones in these regions, we designed, developed and deployed an
Android-based application for one of the taxi service company
called Red Cab. The application makes it easier for Red Cab to
cater for its current customers and also reach out to newer ones.
Keywords—Android, Mobile application, Taxi booking,
transportation
I. INTRODUCTION
Transportation plays a vital role in the day-to-day activities
of the society. In most communities, a large fraction of the
working population commutes to work daily [1]. Commuting
may not only be for business purposes but also for relaxation,
shopping and other social activities. Of all the means of
transportation, land transport, comprising of the use of vehicles
– both private and commercial is common especially in
developing countries of Africa [2]. However, a key advantage
of commercial transportation over owning a personal vehicle is
that it is less expensive and economical considering the high
poverty rate in this region [3].
Taxi services are becoming prominent especially in big
cities of sub-Saharan Africa [4] [5]. Prior to this time, the
companies offering these services have not been able to reach
out to as many people as they would have desired especially
remote locations [6], [7], [8]. However, given the wide spread
adoption of mobile devices (particularly smart phones) in this
region, there is a pressing need to reach out to more customers
and also better cater for existing ones [9], [10]. Some of the
companies based in Nigeria – the nation with the largest black
population in sub-Saharan Africa – have already deployed
mobile applications to better serve their customers but Red Cab
Taxi is yet to successfully deploy a mobile app to support their
services. Instead the company relies on its Web-based taxi
booking system. In order to maintain a competitive edge with
its competitors, Red Cab Taxi has recognized the need for a
mobile version of its taxi booking service. The aim of this
paper therefore is to discuss the development of an Android-
based mobile app for Red Cab Taxi Company. The rest of this
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
paper is thus structured as follows: In Section 2 a review of
existing systems is conducted. In Section 3, we focus on the
design of the proposed system using the Unified Modeling
Language (UML) diagrams. In Section 4, we discuss the
implementation of the mobile application and we compare it
with the existing systems in Section 5. Section 6 concludes the
paper.
II. REVIEW OF EXISTING SYSTEMS
Easy Taxi was founded in Brazil and has since been
expanding. Bankole Cardoso introduced the company to
Nigeria in July 2013 [11]. Its main goal is to create efficiencies
in the Nigerian transportation network by changing the
perception of Nigerians about taxis. With Easy Taxi, users can
call a taxi anytime and anywhere and the app would
automatically search for the taxi closest to the vicinity of the
customer. As soon as a taxi is selected, a live map of the taxi’s
movement is shown on the screen of the customer’s mobile
device. The drawback of Easy Taxi however is that because it
is new, Nigerians are still rather skeptical about its ability to
always deliver hence it is not yet widely used.
Afro Cab is another Nigerian based taxi hailing app, which
emerged in 2012 [12]. The company is currently servicing
Lagos and Abuja with plans to expand their reach to more
major cities in the country. The company has over 600
registered drivers in Abuja and Lagos. The application works
by allowing users to input their location and then a menu of
registered locations/streets are shown. If the location keyed-in
is not known, an alternative pickup location is shown. The type
of car the user wants to ride is listed and the user is also able to
specify the range of amount that s/he is willing to pay before
tapping the Get Cab button. The app sends the request to a
number of the drivers within the vicinity. A driver can either
accept or reject the request and the feedback is fed to the
customer in real time. However, once a driver accepts the offer,
a live map view of the taxi’s movement is shown on the screen
of the user’s mobile device. A drawback of this application is
that it fails to recognize some locations and so customers
requiring the services in those areas may be denied the service.
Uber Taxi on the other hand can be described as a taxi
service for the elite. It portrays its users as successful
individuals who have style, class and elegance [13]. The Uber
smartphone application that is available on all mobile platforms
connects passengers to drivers in an area. All passengers have
to do is to download the Uber application on their smartphone.
When a request is made, a GPS based dispatcher feature simply
sends the location of the requesting passenger as pinpointed on
a Google map, in specifics to the country. The user then selects
the type of car s/he wants to use. A picture of the driver and the
vehicle's registration detail is sent so the passenger and a
suitable price is then agreed upon. When a passenger is in the
car, s/he can send a link of an online map to a friend or loved
one, who can track their entire journey via GPS. Driver rating
is also a feature of the app, in which a rider can rate the driver
and also give a testimonial on the service provided by Uber.
However, this service is not yet functional in Nigeria.
Tranzit, unlike other mobile enabled taxi service is not just
a transportation service, but also a delivery service that helps a
person locate required and interesting places, events, by
suggestions based on the person’s current location. It can also
be described as an intermediary between passengers and Figure 1. Use case diagram of the Red Cab Mobile Application
drivers [14]. Tranzit claims that with a few taps, goods can be
delivered to a customer’s doorstep, and can with those same TABLE I. USE CASE NARRATIVE
taps find a license-vetted taxi. Tranzit, which was first called
Red Cab
Taxi Park, was launched in August 2013, bridging the gap
between passengers and drivers. Tranzit was borne of the This Use Case describes how a user places
desire to better serve our steadily growing database of client. Brief Description an order to a Taxi driver and how the user
The service is currently running in Lagos, though there are monitors the taxi until its arrival
plans to expand the coverage. Lagos traffic is legendary and Actor The owner of the mobile phone (the user)
booking a taxi with ID and details of drivers unknown is very
Basic Flow
risky. Tranzit’s plan is to change all this, is by developing a
The use case starts when the actor inserts
system of service that moves people and objects from one place his details on a registration page which is
to another in the city of Lagos. displayed on installation
1. The actor can update the initial
Red Cab Taxi, which is one of the leading taxi service registered details
companies in Nigeria currently, has no mobile application that 2. The actor places an order,
allows for booking as seen in the other existing systems. The tapping the icon on the application menu.
company relies on its website for providing customers with a 3. A page with different text fields
means of booking taxis. In order to maintain its competitive is displayed requiring the user to insert his
location and its description.
advantage, the need for such a mobile application that would 4. Details of placed orders are
complement their website was identified. The process of displayed in the order history on the
developing the initial version of the application is the subject of application menu
discussion in this paper. Flow of Events 5. The location of the Taxi man
can be viewed on a Google map by tapping
III. DESIGN OF THE RED CAB MOBILE APPLICATION the ‘Where is taxi’ button on the
application menu
Based on interaction with relevant stakeholders (customers, 6. A call can be made to a taxi
personnel in the company) and a study of similar existing man
applications, we identified the following requirements as core 7. A user double taps the exit
button to exit
for the application:
 A module for registering on the platform Alternative Flow
If in the Basic Flow when Registering and
 A module that allows a user to update his/her profile Placing an order, the actor clicks the Save
or Submit button without having filled out
 A module for placing order for a Taxi the compulsory fields, an error message
will be flagged.
 A module for calling the Taxi man Special
None
Requirements
 A module for locating the Taxi
The actors must have filled the Place Order
Pre-Conditions
A use case model was used to capture these basic functions requirements completely.
and is shown in Fig. 1. Table 1 is use case narrative describing Post-Conditions
A message would be received by the user
the flow of activity in the mobile application. from the assigned driver

Extension Points None

 IV. IMPLEMENTATION OF THE MOBILE APPLICATION B. Place Order Module
The mobile application was implemented using Eclipse This module is the core activity of the mobile application. It
IDE for the Android platform and was deployed to the Google allows a user to input his/her current location with specific
Play Store. It can be retrieved at details. This information is then sent to a potential driver’s
(https://play.google.com/store/apps/details?id=com.red.redcab phone as text message. This is depicted in Fig. 4.
&hl=en). Android is officially the most widely used mobile
operating system in Nigeria [15]. The application is able to
interface with Google maps to provide detailed information
about a Taxi’s location. The modules of the mobile application
include: a menu page - that is displayed right after the
application loads. If a user is downloading the application for
the first time then s/he would have to register but subsequently
the menu screen is displayed as shown in Fig. 2.

Figure 4. Red Cab Place Order Module

C. Order History
This module shows in ascending order, the various orders
previously placed. The user can also delete an order history by
pressing down an order for three seconds that is to be deleted.
The module is depicted in Fig. 5.

Figure 2. Red Cab Menu Screen

A. Registration Module
This is the first interactive screen that is displayed after the
splash screen if the user is downloading the application for the
first time. It provides fields that capture the name (surname) the
mobile phone number and password for login to the
application. This is depicted in Fig. 3.

Figure 5. Red Cab Order History Module

D. Taxi Locator
This module is integrated with Google Maps showing the
location of the assigned driver on the map. It is used to monitor
the position of the driver as shown in Fig. 6.

Figure 3. Red Cab Registration Page

 V. COMPARATIVE ANALYSIS AND DISCUSSION
To check that the mobile application developed in this
study provides features that can compete favorably with the
existing mobile applications, we compare the application with
the existing systems discussed in Section 2. The comparison is
done based on the following criteria namely: whether or not
registration is required in order to use; the size of the app when
installed on a device; whether or not it works with Google
Maps; the support for call services as well as the support for
short messaging services (SMS). Table II gives the
comparative analysis between the Red Cab app and similar
existing apps.

TABLE II. A COMPARATIVE ANALYSIS BETWEEN THE RED CAB APP AND
OTHER EXISTING APPS

EasyT Afro Uber Red

Tranzi
Criteria axi Cab Taxi Cab
t
Taxi
Requires registration
Yes Yes Yes Yes Yes
to use
App size when 10.68 11.52 40.11 17.52 2.60
installed MB MB MB MB MB

Payment Integration No Yes Yes Yes No

Works with Google

Yes Yes Yes Yes Yes
Maps
Call services
supported within the
Yes No Yes No Yes
Figure 6. Taxi Locator Module app (to driver and
customer care)
E. Call Module SMS messaging
No No No No Yes
support
The application can allow a user to call the Taxi man that
receives his order or the Customer Care of the company. It is
depicted in Fig. 7.
From Table II, we observe that all the mobile applications
considered require their users to register initially after the
application is downloaded from the store. The various mobile
applications also integrate Google Maps. When installed on a
device, Red Cab App happens to occupy less space compared
to the other applications and so requires far less time to
download. For payment integration, EasyTaxi app de-
emphasizes payment focusing rather on hailing a taxi. Also,
given that the application developed in this paper is an initial
version, payment module is yet to be incorporated compared to
the other existing mobile applications (Afro Cab, Uber Taxi
and Tranzit) that incorporate payment. For call services within
the app, Red Cab app makes it possible to call the driver
assigned as well as customer care. Call services are also
obtainable in Tranzit and Easy Taxi. SMS messaging is
supported in the Red Cab application in instances where it
might be impossible to get across via mobile data due to low
signal strength.
VI. CONCLUSION AND FURTHER WORKS
This study has shown the process of designing and
developing a mobile application for Red Cab Taxi Company
based in Nigeria. Prior to this time, the company has had to
rely solely on its website to provide user's with the booking
Figure 7. Red Cab Call Module
functionality. To ensure that the possible deployment of this
system is progressive and that it reaches its maximum [7] J. C. Aker, and I. M. Mbiti, Mobile phones and economic development
capabilities in the nearest future, the following will be in Africa, Center for Global Development Working Paper, (211), 2010.
considered in future work: Usability evaluation of the mobile [8] G. Porter, Mobile phones, livelihoods and the poor in Sub‐Saharan
Africa: Review and prospect. Geography Compass, vol. 6, pp. 241-259,
application so as to elicit feedback from users that will help to 2012.
further improve the quality of the mobile application. Also, the [9] R. Sietchiping, M. J. Permezel, and C. Ngomsi, Transport and mobility
capabilities of the system will be extended by incorporating a in sub-Saharan African cities: An overview of practices, lessons and
payment module into the app to cater for users who prefer to options for improvements. Cities, vol. 29, pp. 183-189, 2012.
use cards for transactions. All of these efforts will be geared [10] Techloy, "AfroCab Is The Newest Way To Hail A Taxi In Lagos"
towards giving Red Cab Taxi the leading edge in Nigeria. [Online], Telecomist, Available from:
http://techloy.com/2014/04/22/afrocab/ Date accessed: 25th August,
REFERENCES 2015.
[1] N. McGuckin, and N. Srinivasan, The Journey-to-work in the Context of [11] B. Oluwafemi, "Uber Lagos Is Slashing Uber X Pricing By Up To 25
Daily Travel, In Census Data for Transportation Planning Conference, Percent" [Online], Available from:
2005 http://techcabal.com/2015/05/01/uber-lagos-is-slashing-uber-x-pricing-
by-up-to-25-percent/ Date accessed: 25th August, 2015.
[2] E. A. Vasconcellos, Urban Transport Environment and Equity: The case
for developing countries, Routledge, 2014 [12] S. Wakoba, "Nigeria’s TaxiPark Founders To Launch Tranzit.ng To
Take On Rocket Internet’s EasyTaxi" [Online], TechMoran, Available
[3] R. Carruthers, M. Dick, and A. Saurkar, Affordability of public transport at: http://techmoran.com/nigerias-taxipark-founders-to-launch-tranzit-
in developing countries, 2005 ng-to-take-on-rocket-internets-easytaxi/#sthash.rUGTySGA.dpuf Date
[4] D. F. Bryceson, T. C. Mbara, and D. Maunder, "Livelihoods, daily accessed: 25th August, 2015
mobility and poverty in sub-Saharan Africa," Transport Reviews, vol. [13] Techloy, "Android Is Officially The Most Widely Used Mobile OS In
23, pp. 177-196, 2003 Nigeria [STATS]" [Online], Available from:
[5] G. Porter, Mobility constraints and their implications for rural women http://techloy.com/2014/09/27/android-is-officially-the-most-widely-
and girls in sub-Saharan Africa, Gender, place and culture, vol. 18, pp. used-mobile-os-in-nigeria-stats/ Date accessed: 12th October, 2015
65-81, 2011. [14] G. Porter, Transport planning in sub-Saharan Africa, Progress in
[6] G. Porter, K. Blaufuss, and F. Owusu Acheampong, Youth, mobility and development studies., vol. 7, pp. 251-257, 2007.
rural livelihoods in sub-Saharan Africa: perspectives from Ghana and [15] J. Kujenya, "How technology boosts enterprise" [Online], The Nation,
Nigeria. Africa insight., vol. 37, pp. 420-431, 2007. Available from: http://thenationonlineng.net/how-technology-boosts-
enterprise/ Date accessed: 25th August, 2015.
 Enhancing Data Security in Open Data Kit as an
mHealth Application
Paul MACHARIA Betsy SAMBAI David BUKUSI
National AIDS & STIs Control Kenyatta National Hospital Kenyatta National Hospital
Programme Nairobi, Kenya Nairobi, Kenya
Ministry of Health betsysambai07@gmail.com davidbukusi@gmail.com
Nairobi, Kenya
paulmachariah@gmail.com Bourke BETZ Peter CHERUTICH
University of Washington Ministry of Health
Matthew D. DUNBAR Seattle, WA USA Nairobi, Kenya
Center for Studies in Demography bbetz@uw.edu pcheru@uw.edu
and Ecology
University of Washington Anne NJOROGE Carey FARQUHAR
Seattle, WA USA Kenyatta National Hospital Departments of Medicine,
mddunbar@uw.edu Nairobi, Kenya Epidemiology and Global Health
anjoroge@uw.edu University of Washington
Seattle, WA USA
cfarq@uw.edu

Abstract—mHealth as an emerging technology in healthcare is

making it easier to streamline processes, reduce cost, accelerate
completion of research or health outcomes. However, privacy and
security concerns are hampering integration of mobile
technologies in healthcare due to insufficient or lack of guidelines
in securing data during transfer and storage. Securing data
includes ensuring the confidentiality, integrity, and availability of
the electronic data. Security solutions should attempt to be a
balance between individual’s privacy and making the information
accessible when necessary. This study describes an operations
research piloting the enrollment processes of patients into a case-
based registry. We assess the feasibility and time constraints
associated with administration of this secure mHealth system.
Keywords—Privacy; Security; Confidentiality; mHealth
I. INTRODUCTION
Health information is considered as one of the most
sensitive about an individual, confidence has to be built on
mHealth solutions for clients to be sure their information is
both secure and confidential. Security of data includes its
storage and transmission [1]. The World Health Organization
(WHO) indicates that mHealth initiatives are being used for
appointment reminders, accessing patient records, measuring
treatment compliance, patient monitoring and treatment
decision support [2]. There are concerns on the use of
handhelds or mobile devices compromising confidentiality of
medical information, potential threats on privacy and security
of confidential data has become a barrier to effective
implementation [3].
II. LITERATURE REVIEW
Data security is defined as “the extent to which personal
information can be stored and transmitted in such a manner that
access to the information is limited to authorized parties”. The
Canadian Office of Health and the Information Highway
defines security as “The procedures and systems used to restrict
access and maintain the integrity of information.” Data
breaches could have serious consequences to reputation,
financial and personal harm to patients. Poor privacy and
security practices increase vulnerability and risk health
information system [4].
mHealth, the use of mobile technologies in healthcare and
public health for data collection and clinical decision support
systems [5] is one of the options to streamline data capture
processes and reduce cost. Mobile technologies could
accelerate completion of research or healthcare outcomes,
make it easier to store and backup electronic files in secure
locations reducing loss or damage [6, 7]. However, privacy and
security concerns are hampering integration of mobile
technologies in healthcare due to insufficient or lack of
guidelines in securing data in transit and at rest [8]. Challenges
of mitigating breach and compliance is affecting the potential
use of mobile health applications, this is caused by the need to
use well tested security protocols. There is need for further
research on the privacy and security risks of mHealth
application to establish policy guidelines [9].
This is especially the case owing to the fact that most
mHealth solutions transmit Personal Health Information (PHI).
The American Health Insurance Portability and Accountability

978-1-4673-9354-6/15/$31.00 ©2015 IEEE

Act (HIPAA) define PHI as “individually identifiable health
information that is transmitted or maintained in electronic
media or in any other form”. Privacy and confidentiality is of
great concern for systems involving patient records, the
electronic systems could be vulnerable to hacking and theft,
this is due to the remote access ability of mHealth applications
and amount of data that could be potentially accessed during a
breach [10].
The HIPAA Privacy Rule defines a data breach as “the
acquisition, access, use, or disclosure of protected health
information in a manner not permitted compromising health
information”. Breaches could lead to loss or compromise of
identifiable healthcare records. Recent evidence shows
insufficient security process have led to data breaches resulting
in mental anguish to patients [11]. Security of confidential data
on the portable devices applies also to information contained
on lost or stolen devices. Documented electronic data breaches
include a USB drive stolen from Alaskan Health and Human
Services employee’s vehicle [12] and unencrypted hard drives
stolen from a storage facility [13]. Both led to large financial
losses.
An mHealth application should have features that mitigate
security risks to healthcare providers and patients [14]. Security
features should be enforceable on both the hardware and
software providing device and information security with
specific features such as remote data wipe and up-to-date virus
protection on both bring-your-own-device (BYOD) programs
and company-issued equipment. Secure Communication
channels should be encrypted and require user authentication
and enforces password policies [15].
Securing PHI includes ensuring the confidentiality,
integrity, and availability the electronic data. This would
include restricting access to data centre, encryption of both
stored and data on transit and the applicable data transmission
channels [16]. Security solutions should strike a balance
between individual’s privacy and making the information
accessible when necessary. The mHealth solutions should also
be interoperable, easy-to-implement, and scalable. An ongoing
process of monitoring an mHealth application is key to
maintaining the integrity, confidentiality, availability,
reliability and security of data [17]. Research for health
purposes using personal data must only take place within a
robust ethical framework with clear governance rules to
mitigate breach of privacy on participant’s data [18].
Privacy and security policies to protect data need to be
formulated as well as a risk management process that
prioritizes securing information systems to ensure enforcement
of policies. Lack of skills to personnel also poses a risk to
privacy and security of data. Data security is based on a
systematic assessment of threats and risks of breach or loss
which should be considered when developing a system’s
design requirements. Risk assessment results should inform
security controls, risk analysis as an ongoing process can help
an organization detect threats and also evaluate effectiveness of
their current security in securing PHI [19].
A. Open Data Kit
Open Data Kit (ODK) is an extensible, open-source suite of
tools designed to build data collection tools for resource limited
settings. The ODK platform includes three pieces of software:
Collect, Aggregate, and Build. ODK Collect runs on android
devices rendering application logic and supports the
manipulation of data. ODK Aggregate provides a data
aggregation server for storage in a centralized, remote location.
ODK Build is used to develop data collection forms [20, 21].
ODK is simple to use and no extensive training is required
to operate. ODK has been deployed by health and research
organizations to author, field, and manage mobile data
collection at health facilities. ODK enhances the quality of
data, and reduces turn-around time in data collection and
analysis, all of which facilitates immediate feedback [22].
ODK has also been used for home-based care services to take
healthcare to people in cost-effective way. Community health
workers were able to administer ODK surveys in screening
efforts in western Kenya [23]. In addition, this mHealth
application can acquire data in the form of text, images, and
real time geo-coded coordinates, working together as evidence
to support faster decision-making [24].
1) Open Data Kit Data Encryption
Leading organizations implement Public Key Infrastructure
(PKI) technology for electronic health applications [25, 26].
ODK supports form encryption applying asymmetric public
key encryption, which makes it possible for the application to
comply with data privacy laws. ODK Collect generates a
random 256-bit symmetric key when a form is finalized. The
encrypted form’s contents are attached to the random key, a
submission manifest is constructed describing the encrypted
submission, and an asymmetric-key encryption of the
symmetric key used for the encryption. The forms remain
encrypted when submitted to the ODK Aggregate server. The
encrypted forms cannot be decrypted on the server. Aggregated
forms that are still encrypted must be downloaded using ODK
Briefcase, and then decrypted by supplying the asymmetric
private key. ODK Briefcase then generates an unencrypted
tabular data set, in a Comma Separated Values (CSV) file [27].
B. Hyper Text Transfer Protocol Secure (HTTPS)
HTTPS is a protocol that was developed by Netscape that
uses a Secure Socket Layer (SSL) or Transport Layer Security
(TLS) when accessing a website server. HTTPS encrypts and
decrypts web end user communications using 128, 192 or 256-
bit long key size for encryption and decryption [28]. Data
packets captured during transmission are not recognizable [29].
 The ODK data encryption capabilities and the HTTPS
ability to securely transmit data form the study facilities to the
aggregate server informed the design and setup of our mHealth
system of data collection.
III. METHODOLOGY
Case-based surveillance systems are not standard in sub-
Saharan African, and most surveillance data come from facility
records or periodic surveys, such as Demographic & Health
Surveys (DHS) and AIDS Indicator Surveys. However, each of
these has significant limitations. Facility records are often
unreliable or poorly kept and are not generalizable beyond the
facility-level. Surveys provide higher quality, national-level
data using sound methodology, but they only occur once every
5 years. Thus, it is difficult for a Ministry of Health (MoH) to
use these data to assess the effectiveness of changes in health
system policies and practices.
The main purpose of this implementation science effort is
to design a case-based registry which regularly incorporates
patient records stored in facility-based electronic medical
record (EMR) systems. This will be used to improve disease
surveillance and promote health policies and practices to
improve patient outcomes. Aim 1 of the study entails piloting
the enrollment of phase of our electronic case-based
surveillance system by enrolling subjects into a case-based
registry. This process begins with administration of the MoH
HIV Testing and Counseling (HTC) intake form to individuals
with HIV and assignment of a 12-digit unique identification
(UID) number to patients using iris recognition biometrics.
This study component is operations research which
attempts to pilot the enrollment processes assess the feasibility
and time constraints associated with administration of the
system. The aim of this study component is as follows:
Establishing and piloting an electronic case-based reporting
system at the Kenya National AIDS and STI Control Program
(NASCOP) using Open Data Kit (ODK) forms and biometric
identification numbers.
The NASCOP ODK Aggregate server runs on a virtualized
CentOS 7 a Red Hat Enterprise Linux (RHEL) Linux
environment. The ODK Aggregate server connects to the
facilities phones on a 256-bit Secure Socket Layer (https)
Encryption providing a secure transmission of data. The
aggregate server also has restricted system users and all users
have passwords that met minimum stands of security. The
HTC forms are encrypted on finalization during data collection
on the android devices before submission to the aggregate
server. ODK Briefcase is used to decrypt this data. Further data
exchanges of de-identified among the research team uses
zipped encrypted format. Figure 1 below shows the Case-based
surveillance system.
Fig.1 HIV Case-Based Surveillance System
IV. RESULTS
Between April and Sept 2015, over 8000 unique records
have been securely collected, transferred, and stored using
ODK tools by a team of 8 Health Advisors at 4 Health
Facilities in Kenya. On a weekly basis our study team use
scripts to unencrypt and perform Quality Assurance (QA)/
Quality Control (QC) checks on the data. The QA/QC
benchmark is based on the completeness, accuracy, privacy and
confidentiality of collected data. The security measures have
made no noticeable impact on the ease of data collection or
analysis, but have significantly increased the privacy and all-
but eliminated the potential for data loss or misuse. The
enhanced security features of the study mHealth setup meet
international standards of collection, transmission and storage
of PHI using mobile technologies.
V. CONCLUSION
Although our study is still ongoing, to-date we have
demonstrated the feasibility of implement a secure case-based
registry data collection system using the ODK platform. The
only cost of adding the increased security measures of data
encryption on both storage and transfer has been the minimal
up-front time required to implement these systems. From an
operational perspective, these additional security measures
have born no additional cost on any other aspect of the research
process or the functionality of a centralized digital case-based
surveillance system.
Acknowledgment
We thank the Kenyan Ministry of Health for their support
in the implementation of the study. We are also grateful to the
study participants, as well as the team of health advisors and
program coordinators for their work on this project. This
research is supported by a grant from the US National Institutes
of Health (NIH), PA-14-024 (P.I. Farquhar). The funding
agencies played no role in the study conception or design,
writing of the manuscript, or decision to submit the manuscript Participants Potential Effects on the Legal and Ethical Conduct of
Clinical Trials. Therapeutic Innovation & Regulatory Science, 47(6),
for publication. 656-664.
[16] The mHealth Revolution: A Clash of Innovation & the Law (2015)
https://www.healthlawyers.org/events/programs/materials/documents/hit
References 14/f_gatling.pdf [accessed on 27th August 2015]
[17] Schwab, K., Marcus, A., Oyola, J. O., Hoffman, W., & Luzi, M. (2011).
Personal data: The emergence of a new asset class. In An Initiative of
[1] The Roadmap mHealth Interventions For At Risk Women (2015) the World Economic Forum.
http://www.fzr.ru/materials/250.pdf?view [accessed on 27th August
2015] [18] Healthcare Coalition on Data Protection (2015)
http://www.datasaveslives.eu/media/1122/healthcarecoalitionondataprot
[2] West, D. (2012). How mobile devices are transforming ection_2014_jointstatementpublished.pdf [accessed on 27th August
healthcare. Issues in technology innovation, 18(1), 1-11. 2015]
[3] ames Katz and Ronald Rice, “Public Views of Mobile Medical Devices [19] Health Care Data Breaches and Information Security Addressing Threats
and Services”, International Journal of Medical Informatics, Volume 78, and Risks to Patient Data (2015)
2009, pp. 104-114. http://www.americanbar.org/content/dam/aba/publications/books/healthc
th
[4] Guide to Privacy and Security of Electronic Health Information (2015) are_data_breaches.authcheckdam.pdf [accessed on 27 August 2015]
http://www.healthit.gov/sites/default/files/pdf/privacy/privacy-and- [20] Hartung, C., Lerer, A., Anokwa, Y., Tseng, C., Brunette, W., &
security-guide.pdf [accessed on 27th August 2015]
Borriello, G. (2010, December). Open data kit: tools to build
[5] Phillips, G., Felix, L., Galli, L., Patel, V., & Edwards, P. (2010). The information services for developing regions. In Proceedings of the 4th
effectiveness of M-health technologies for improving health and health ACM/IEEE International Conference on Information and
services: a systematic review protocol. BMC research notes, 3(1), 250. Communication Technologies and Development(p. 18). ACM.
[6] Can mHealth be Thought Of as the Default Solution for Accelerating [21] Brunette, W., Sundt, M., Dell, N., Chaudhri, R., Breit, N., & Borriello,
Completion and Increasing Success Outcomes in Clinical Trials? (2015) G. (2013, February). Open data kit 2.0: expanding and refining
http://www.jforcs.com/wp-content/uploads/2015/02/12.-Can-mHelath- information services for developing regions. In Proceedings of the 14th
be-Thought....pdf [accessed on 27th August 2015] Workshop on Mobile Computing Systems and Applications (p. 10).
[7] World Bank. Information, Communication Technologies, & infoDev ACM.
(Program). (2012). Information and Communications for Development [22] Macharia, P., Muluve, E., Lizcano, J., Cleland, C., Cherutich, P., &
2012: Maximizing Mobile. World Bank Publications. Kurth, A. (2013, May). Open Data Kit, A solution implementing a
[8] Melnik, T. A. T. I. A. N. A. (2011). Mobile tech: is it right for your mobile health information system to enhance data management in public
organization.J Health Care Compliance, 13(6), 49-52. health. In IST-Africa Conference and Exhibition (IST-Africa), 2013 (pp.
[9] mHealth Security (2015) https://columbitech.com/wp- 1-6).IEEE.
content/uploads/2013/11/mobile-security-in-healthcare.pdf [accessed on [23] Rajput, Z. A., Mbugua, S., Amadi, D., Chepnǵeno, V., Saleem, J. J.,
27th August 2015] Anokwa, Y., ... & Were, M. C. (2012). Evaluation of an Android-based
[10] Wittet, S. (2012). eHealth mHealth reproductive health. Outlook, 29(2), mHealth system for population surveillance in developing
8. countries. Journal of the American Medical Informatics
Association, 19(4),655-659.
[11] Appari, A., & Johnson, M. E. (2010). Information security and privacy
in healthcare: current state of research. International journal of Internet [24] Singh, H. (2013). Mobile data collection using an android
and enterprise management, 6(4), 279-314. device. IJCST, 4(1).
[12] US Department of Health and Human Services. (2013). Alaska settles [25] Clayton, P. D., Boebert, W. E., Defriese, G. H., Dowell, S. P., Fennell,
HIPAA security case for $1,700,000 [news release]. June 26, 2012 M. L., Frawley, K. A., ... & Rindfleisch, T. C. (1997). For the record:
protecting electronic health information. National Research
[13] Mobile Health And Security (2015)
Council.(Washington, DC: National Academy Press, 1997).
http://www.diversinet.com/downloads/whitepaper/Mobile_Health_Feb2
5.pdf [accessed on 27th August 2015] [26] Brands, S. (2003). Privacy and Security in Electronic
Health.PricewaterhouseCoopers CCE Journal.
[14] Malvey, D. M., & Slovensky, D. J. (2014). MHealth: Transforming
Healthcare. Springer [27] Encrypted Forms (2015) https://opendatakit.org/help/encrypted-forms/
[accessed on 27th August 2015]
[15] Arquiza, M. (2013). An Exploratory Study on the Policies of Health
Care Institutions in the New England Region of the United States
Regarding Third-Party Access to Medical Records of Research
[29] HTTPS definition (2015) http://techterms.com/definition/https [accessed
[28] HTTPS (HTTP over SSL or HTTP Secure) definition (2015) on 27th August 2015]
http://searchsoftwarequality.techtarget.com/definition/HTTPS [accessed
on 27th August 2015]
 Performance Evaluation of Classification
Techniques for Devanagari Script
M.S.Sonawane, C.A.Dhawale, Sanjay Misra
SGBAU, Amravati (M.S.), P.R. Pote College of Engineering and Management, Amravati (M.S.)
Covenant University, Nigeria
manojkumar.sonawane@rediffmail.com, cadhawale@rediffmail.com, sanjay.misra@covenantuniversity.edu.ng
Abstract—Enormous applications and the future necessities
of image processing area open new paths for researchers. The
analysis and recognition of numerous documents in the form of
images are the challenging task. The classification is one of the
vital phases in the image processing. The methods of
classification must possess the consistency and accuracy. This
research analyzes the different classification techniques and the
performance analysis is carried out with the help of testing the
image classification principles. In this research paper few
classification techniques for Devanagari script are considered
and evaluated in MATLAB R2014a.
Keywords—OCR; Classification; Nearest neighbor; Neural
network; SFSA, MLP
I. INTRODUCTION
In current scenario interdisciplinary exploration grooming
very rapidly, this includes two or more areas to do the
research. One kind of interdisciplinary study area is machine
learning which gives a communal stand to the scholar from the
field of computer science, artificial intelligence, biology,
medicine, economics, statistics, mathematics etc. to effort
collectively. Optical Character Recognition (OCR) is a
mission of programmed computer identification of electronic
character or probed images to spawn an automated edition.
The accuracy rate of an OCR system varies and it’s reliant on
input document. For a decent superiority image gets from a
sanitary, the result is typically an accurate dictation of the text
blocks of an input. In case a noisy image taken from a
document which is damaged, unclear, or else in meager
condition, the resultant eminence is likely to be deprived with
errors. In such scenarios, an image itself is more realistic
representation of the original as compared to OCR outcome.
There are numerous OCR systems existing commercially in
the sell [1]. These OCRs are extensively used to renovate
documents into an electronic archive [2] with an aim to
automate record saving in office [3] or to circulate the words
on a Web page [4]. Because of OCR it is promising to revise
the text, seek out for a text/phrase [5] [6], save text more
efficiently and apply methods like text to speech translation
[7], text mining [8], machine translation [9] etc. Consequently
OCR is arduous, contemporary pattern recognition field,
predominantly in document image analysis expanse. Various
applications of OCR are in natural language processing, data
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
compression, health care, banks, post offices, etc. Number of
researchers enthusiastically participated in the emergence of
OCR systems [10] -[13] and investigation regarding
document images pattern exploration has been vigorous since
30 years [14]. OCR work is continuously going on for several
Indian scripts [15] -[17]. Devanagari screenplay is practical
for Marathi , Hindi, Nepali, Sanskrit, and sundry linguistics
which are used by more than 450 million persons around the
world. Other Indian languages like Punjabi, Bengali, and
Gujarati follows scripts analogous to Devanagari. Indian
scripts, including Devanagari invented from the primordial
Brahmi script after diverse modifications. A Devanagari script
has 13 vowels as well as 34 consonants with 14 vowel
modifiers. Compound characters are made by joining two or
more core characters, it not distinguish upper, lowercase
characters. Researchers have contributed their efforts for
Devanagari script. Bhattacharya and Chaudhuri [35] have used
different multilayer perceptron classifier at every stage of their
scheme of recognition. Govindaraju [46] has thought about
gradient features to do the characters feature extraction.
Method for recognition, dissection of join Devanagari
machine made texts are explained by Grain and Chaudhuri
[47], it relies on fuzzy multifactorial analysis. Recognition of
words is done by devising the stochastic finite state
automaton, which considers classifier scores and frequencies
of character [26]. Sinha [48] explained how the spatial
association among ingredient characters of the Devanagari
script do crucial function for word recognition. In recognition
of Devanagari characters, classifier like support vector
machine have been reported by Jayadevan et al. [15].
Currently significance of bilingual [18] -[21] and multi-lingual
OCR is also raising [22]. Numbers of editorial published
before 2000 are stored in the study carried out by Pal and
Chaudhari [17]. Because of exploration, solicitation and
expansion of machine learning algorithms, newer and newer
classification techniques are impending from research.
Numerous classification methods have been developed which
might belong to categories like neural network based, rule
based, fuzzy based concept, perceptron based method,
statistical method, symbolic method, logical method, etc. This
research paper is an attempt to present a review of the OCR
work on Devanagari script classification available from year
2000 and onwards. The review focuses efforts associated with
identification of printed characters, numerals and it provides
an evaluation of few reported techniques in table form. The
appraisal is investigated on feature set, recognition methods
and get a precision ratio.
The remaining part of the paper is arranged as follows.
Discussion about classification phase is in section II;
classifiers considered are described in section III. In section
IV experimental results are conveyed, and our efforts are
summarized in Section V.
II. CLASSIFICATION
OCR systems comprehensively use the techniques of
pattern recognition, which allocate an unidentified sample of
an identified class. The features extracted are provided as an
input to the classification method. The production of feature
extraction phase in image processing is a feature vector which
is given as an input to classification phase and classified by
means of supervised, unsupervised technique. Here the data
set is alienated into training set, test set for each character.
Bag-of-key points take out from the feature extraction are
employed for recognition aim. Numerous tactics [15] are
employed to catalogue the character structures like k-Nearest
Neighbor approach, neural network, and SVM classifier etc.
Phases of document image processing including classification
are shown in Fig. 1.
III. CLASSIFIERS CONSIDERED
The automation of Devanagari script recognition begins
in the early 1970s. Numerous classifiers have been used to do
the Devanagari text identification work. In the training period
of such approaches, patterns are assumed as input and the
scheme legalizes themselves to weaken the misclassification
gaffes of these patterns. Such trained schemes are employed to
do the taxonomy of enigmatic test patterns. One prominent
example of such a technique is the Neural network, that revise
the weight associations from the training configurations with
the purpose of enhanced classification. Classifiers considered
here are as follows.
3.1 Schema dependant minimum distance classifier
Here nearness of an incoming pattern to the patterns
of the probable pattern classes gives a measure in deciding the
pattern class of the pattern under deliberation. As the
classification is depends on the minimum distance
computation, this technique is known as minimum distance
classification technique. For classification varied information
supply is used. The schema states classifier hierarchical
structure and afford the relative significance, role of varied
information supply in the hierarchy. The benefit of this
methodology is, classifier structure can be effortlessly
customized by changing the schema [23] [24].
3.2 Nearest neighbor
Preprocessing
Skew Detection/Correction, Binarization, Filtering etc.
Image Segmentation
Paragraphs/ Lines/Words/Characters
Feature Extraction
Statistical/Structural etc.
Classification
Nearest neighbor, Neural network, SFSA, MLP etc.
Fig. 1. Phases of Document Image Process.
Nearest neighbor is an instance dependent learning
method to classify objects find on the basis of closest training
examples in the feature space. K-NN is incredibly easy in the
calculation. For every data point from k neighboring training
data points having known class values and nearest distance are
chosen. The class which comes most frequently in the
neighborhood is allocated to the new data point depends on an
assured distance metric. K-NN classifiers are extremely
effortless to train, to catalogue the input mockup feature by
means of the mutual assortment of its k adjacent neighbors
[25] -[27]. K-NN is utilized with concavity, structural,
gradient features [27]. One type of nearest neighbor classifier
is Minimum Distance Classifier [28] which is used for
evaluation of priority matching of characters portions, contour
sets.
3.3 Neural network
The neural network is a computing design that
encompasses analogous interconnection of adaptive neural
processors. Because of comparable flora, it can accomplish
calculations at a greater rate in comparison to the traditional
procedures. Due to its adaptability spirit, it discerns how to
acclimate data amendments and discover the input signal
appearances. Node productivity assumed to another node in
the linkage and the decisive consequence trusts on the
multifarious associations of all nodes. Neural networks cover
excellent learning, generalization skills. These skills are
crucial for trade with input arrays fuzziness and triumph
tolerably in the noise manifestation, fractional data and they
could invent from the illustrations [26], [29] -[31]. For
character recognition, one of the admired classifiers is the
multilayer perceptron [32] -[41].
3.4 Stochastic finite state automaton
These models are competently learnable from data,
efficient for decoding and are related with a calculus for
composing models which allows for rigid integration of
constraints from diverse levels of language processing.
Stochastic finite state automaton recommends the gain of a
unified framework to consider classifier grooves, character
regularities for word identification. It is employed to merge
probabilistic n-gram language model, rule-based script
conformation validity scrutiny in a particular structure [26].
3.5 String matching
String matching methods signify the easy primitive
patterns in the character resembling string constituents that, if
gathered rendering to precise rules construct the multifaceted
pattern symbolized by the unabridged string. This method is
helpful, since it is forthright to diagnose the humble
primitives. Dynamic programming is typically employed for
string toning [43] [44].
3.6 Multilayer perceptron
Multilayer perceptron is also known as Feed forward
network and it’s a kind of artificial neural network that
comprises of a nonlinear activation function in the hidden
layer. Multilayer perceptron network offer nonlinear mapping
concerning input and output vectors. Two vital functions with
neural networks are nonlinear adaptive filters and pattern
classifiers. A common structure of the neural network
comprises of t h r e e layer design; input layer which state the
input value, secondly one or more hidden layers which state
the mathematical function and finally an output layer which
state final outcome. Every layer involves of a huge number of
neurons which are interrelated through weights. Every neuron
has mathematical function also called as activation function
which receives input from preceding layer and creates output
for subsequent layer. Therefore, in neural networks the
forecast is stated by the activation function [45].
IV. RESULT AND DISCUSSION
From this review paper, it is noted out that every
character recognition system is good whenever precision of
recognition rate is high. For character recognition system, lots
of classifiers are available. Every one of them has their own
advantages and drawbacks. Table 1 shows results analysis in a
tabular form [42]. As shown in the table, Bansal and Sinha
[23] have used statistical features, preferred schema dependent
minimum distance classifier and achieved 90% recognition
accuracy. Dhurandhar et al. [28] extracted curve dependent
features, interestingly, they used a combination of neural
network and nearest neighbor approaches. The accuracy rate
was 94%. Kompalli et al. [30] have selected gradient features
and nearest neighbor, neural network classifiers. They got
91.11% and 94.17% accuracy rate respectively. Kompalli et
al. [26] have focused on extraction of gradient features, GSC
and combination of k-NN, SFSA classifiers. In this case they
got very effective results. Bag et al. [44] used structural,
topological features and string matching classifier in their
work. They got a 98.64% accuracy rate. In our experiments we
found that pair of pixel-shape based features and MLP
classifier gives excellent results with accuracy 99.20%.
V. CONCLUSION
With an extensive use of digitized documents, there is an
exceptional mandate for competent, effective techniques for
retrieving the information confined in these documents.
Classification is one crucial step in optical character
recognition and numerous classification techniques are
offered. In this review paper, we have studied few of them and
their combinations. For work text images of Devanagari script
are considered and classifiers are executed in MATLAB
R2014a. Depends on circumstances, it is needed to extract
appropriate features and use of classifiers with an aim to get
better recognition accuracy. For future work one may find a
new classification technique, modifications in existing
techniques with better accuracy or appropriate combination of
classification techniques.

TABLE I. PERFORMANCE EVALUATION OF CLASSIFIERS REFERENCES

Authors Features Classifiers Result in % [1] Fujisawa H. ―Forty years of research in character and document
Bansal, Sinha [23] Statistical Schema 90% recognition—an industrial perspective,‖ Pattern Recogn.41: 2435–2446,
dependant 2008.
minimum [2] Sarkar P. ―Document image analysis for digital libraries,‖ Proceedings
distance of the International Workshop on Research Issues in Digital Libraries,
Dhurandhar et.al. Curve Neural 94% Article 12, 2006.
[28] dependent network & [3] Doucet A, Kazai G, Dresevic B, Uzelac A, Radakovic B and Todic N.
nearest ―Setting up a competition framework for the evaluation of structure
neighbor extraction from OCR.‖ Ed books. Int. J. Doc. Anal. Recognit.14: 45–52,
2011.
Kompalli et.al. [30] Gradient Nearest 91.11% [4] Zagoris K, Papamarkos N and Chamzas C. ―Web document image
neighbor retrieval system based on word spotting,‖ Proceedings of the
Kompalli, Setlur Gradient Neural 94.17% International Conference on Image Processing, 477–480, 2006.
[25] network [5] Kumar A, Jawahar C V and Manmatha R. ―Efficient search in document
Kompalli et.al. [26] GSC & SFSA & k- 94%…97% image collections,‖ Asian Conference on Computer Vision, 586–595,
Gradient NN 2007.
[6] Rodriguez-Serrano J A and Perronnin F. ―Handwritten word-spotting
Bag et.al. [44] Structural- String 98.64%
using hidden Markov models and universal vocabularies,‖ Pattern
Topological Matching
Recogn.42: 2106–2116, 2009.
Majumdar, Shape, Pixel MLP 99.20%
Chaudhari [45] dependent
[7] Bahrampour A, Barkhoda W and Azami B Z. ―Implementation of three
texts to speech systems for Kurdish language,‖ Iberoamerican Congress
on Pattern Recognition, 321–328, 2009.
[8] Fuentes F A, Garcia R G and Contelles J M B. ―A high-dimensional
access method for approximated similarity search in text mining,‖
International Conference on Pattern Recognition, 3155–3158, 2010.
[9] Genzel D, Popat A C, Spasojevic N, Jahr M, Senior A, IE E and Tang F
Y. ―Translation-inspired OCR,‖ International Conference on Document
Analysis and Recognition, 1339–1343, 2011.
[10] Mantas J. ―An overview of character recognition methodologies,‖
Pattern Recogn.19: 425–430, 1986.
[11] Govindan V K and Shivaprasad A P. ―Character recognition – A
review,‖ Pattern Recogn.23: 671–683, 1990.
[12] Mori S, Suen C Y and Yamamoto K. ―Historical review of OCR
research and development,‖ Proceedings of IEEE 80: 1029–1058, 1992.
[13] Plamondon R and Srihari S N. ―On-line and off-line handwritten
character recognition: A comprehensive survey,‖ IEEE Trans. Pattern
Anal. Mach. Intell.22: 62–84, 2000.
[14] Nagy G. ―Twenty years of document image analysis in pattern analysis
and machine intelligence,‖ IEEE Trans. Pattern Anal. Mach. Intell.22:
38–62, 2000.
[15] Jayadevan R, Kolhe S R, Patil P M and Pal U. ―Off-line recognition of
Devanagari script: A survey,‖ IEEE Trans. Syst. Man Cybern., Part C:
Appl. Rev.41: 782–796, 2011.
[16] Kannan R J. ―A comparative study of optical character recognition for
Tamil script,‖ Eur. J. Scientific Res.35: 570–582, 2009.
[17] Pal U and Chaudhuri B B. ―Indian script character recognition: A
survey,‖ Pattern Recogn. 37: 1887–1899, 2004.
[18] Pal U and Chaudhuri B B. ―Printed Devnagari script OCR system,‖
Vivek 10: 12–2, 1997.
[19] Jawahar C V, Pavan Kumar M N S S K and Ravi Kiran S S. ―A
bilingual OCR for Hindi-Telugu documents and its applications,‖
Proceedings of the International Conference on Document Analysis and
Recognition, 408–413, 2003.
[20] Lehal G S and Bhatt N. ―A recognition system for Devnagri and English
handwritten numerals,‖ Proceedings of the International Conference on
Advances in Multimodal Interfaces, 442–449, 2000.
[21] Mohanty S, Dasbebartta H N and Behera T K. ―An efficient bilingual
optical character recognition (English-Oriya) system for printed
documents,‖ Proceedings of the International Conference on Advances
in Pattern Recognition, 398–401, 2009.
[22] Aradhya V N M, Kumar G H and Noushath S. ―Multilingual OCR
system for South Indian scripts and English documents: An approach
based on Fourier transforms and principal component analysis,‖
Eng.Appl. Artif. Intell.21: 658–668, 2008.
[23] Bansal V and Sinha R M K. ―Integrating knowledge sources in
Devanagari text recognition system,‖ IEEE Trans. Syst. Man Cybern.,
Part A, Syst. Humans 30: 500–505, 2000.
[24] Bansal V and Sinha R M K. ―A complete OCR for printed Hindi text in
Devanagari script,‖ Proceedings of the International Conference on
Document Analysis and Recognition, 800–804, 2001.
[25] Kompalli S and Setlur S. ―Design and comparison of segmentation
driven and recognition driven Devanagari OCR,‖ Proceedings of the
International Conference on Document Image Analysis for Libraries,
96–102, 2006.
[26] Kompalli S, Setlur S and Govindaraju V. ―Devanagari OCR using a
recognition driven segmentation framework and stochastic language
models,‖ Int. J. Doc. Anal. Recognit.12: 123–1308, 2009.
[27] Majumdar A. ―Bangla basic character recognition using digital curvelet
transform,‖ Journal of Pattern Recognition Research 2: 17–26, 2007.
[28] Dhurandhar A, Shankarnarayanan K and Jawale R. ―Robust pattern
recognition scheme for Devanagari script,‖ Proceedings of the
International Conference on Computational Intelligence and Security,
1021–1026, 2005.
[29] Dutta A and Chaudhury S. ―Bengali alphanumeric character recognition
using curvature features,‖ Pattern Recogn.26: 1757–1770, 1993.
[30] Kompalli S, Nayak S and Setlur S. ―Challenges in OCR of Devanagari
documents,‖ Proceedings of the International Conference on Document
Analysis and Recognition, 408–413, 2005.
[31] Singh D, Dutta M and Singh S H. ―Neural network based handwritten
Hindi character recognition system,‖ Proceedings of the Bangalore
Annual Compute Conference, article no. 15, 2009.
[32] Das N, Das B, Sarkar R, Basu S, Kundu M and Nasipuri M.
―Handwritten Bangla basic and compound character recognition using
MLP and SVM classifier,‖ J. Comput.2: 109–115, 2010.
[33] Sural S and Das P K. ―An MLP use the Hough transform based fuzzy
feature extraction for Bengali script recognition,‖ Pattern Recogn.
Lett.20: 771–782, 1999.
[34] Bhattacharya U, Das T K, Datta A, Parui S K and Chaudhuri B B.
―Recognition of hand printed Bangla numerals using neural network
models,‖ Proceedings of the AFSS International Conference on Fuzzy
Systems, 228–235, 2002a.
[35] Bhattacharya U, Das T K, Datta A, Parui S K and Chaudhuri B B. ―A
hybrid scheme for hand printed numeral recognition based on a self-
organizing network and MLP classifiers,‖ Int. J. Pattern Recogn.Artif.
Intell.16: 845–864, 2002b.
[36] Bhattacharya U, Shridhar M and Parui S K. ―On recognition of
handwritten Bangla characters,‖ Proceedings of the Indian Conference
on Computer Vision, Graphics and Image Processing, 817–828, 2006.
[37] Rahman M A and Saddik A E. ―Modified syntactic method to recognize
Bengali handwritten characters,‖ IEEE Trans. Instrum. Meas.56: 2623–
2632, 2007.
[38] Bhowmik T K, Bhattacharya U and Parui S K. ―Recognition of Bangla
handwritten characters using an MLP classifier based on stroke
features,‖ Proceedings of the International Conference on Neural
Information Processing, 814–819, 2004.
[39] Basu S, Das N, Sarkar R, Kundu M, Nasipuri M and Basu D K. ―A
hierarchical approach to recognition of handwritten Bangla characters,‖
Pattern Recogn.42: 1467–1484, 2009.
[40] Bajaj R, Dey L and Chaudhury S. ―Devnagari numeral recognition by
combining decision of multiple connections classifiers,‖ Sadhana 27:
59–72, 2002.
[41] Verma B K. ―Handwritten Hindi character recognition using multilayer
perceptron and radial basis function neural networks,‖ Proceedings of
the IEEE International Conference on Neural Network, 2111–2115,
1995.
[42] Bag S. and Harit G. ―A survey on optical character recognition of
Bangla and Devanagari scripts,‖ Sadhana Vol. 38, Part 1, pp. 133–168,
2013.
[43] Rahman M A and Saddik A E. ―Modified syntactic method to recognize
Bengali handwritten characters,‖ IEEE Trans. Instrum. Meas.56: 2623–
2632, 2007.
[44] Bag S, Harit G and Bhowmick P. ―Topological features for recognizing
printed and handwritten Bangla characters,‖ Joint Workshop on
Multilingual OCR and Analytics for Noisy Unstructured Text Data,
article no. 10, 2011b.
[45] Majumdar A and Chaudhuri B B. ―A MLP classifier for both printed and
handwritten Bangla numeral recognition,‖ Proceedings of the Indian
Conference on Computer Vision, Graphics and Image Processing, 796–
804, 2006.
[46] Govindaraju V, Khedekar S, Kompalli S, Farooq F, Setlur S and
Vemulapati R. ―Tools for enabling digital access to multilingual indic
documents,‖ Proc. 1st Int.Workshop Document Image Anal. Libraries,
pp. 122–133, 2004.
[47] Garain U and Chaudhuri B B. ―Segmentation of touching characters in
printed Devnagari and Bangla scripts using fuzzy multifactorial
analysis,‖ IEEE Trans. Syst. Man, Cybern. Part C: Appl. Rev., vol. 32,
no. 4, pp. 449–459, Nov. 2002.
[48] Sinha R M K. ―Rule based contextual post-processing for Devanagari
text recognition,‖ Pattern Recognit., vol. 20, pp. 475–485, 1987.
 Enhanced Rule Extraction by Augmenting Internal
Representation of Feed Forward Neural Network
Vinita Srivastava Chitra Dhavale
Dept. of I.I.C.C., Department of Computer Science,
R.T.M. Nagpur University, Amravati
Nagpur, India Email: cadhavale@rediffmail.com
Email: vinni.sara12@gmail.com
Abstract—Human readable symbols are extracted from a
trained neural network using Rule Extraction Techniques. In
this paper internal representation of feed forward neural
network is augmented by a distance term to produce fewer rules.
This paper presents an efficient method to extract fewer rules
from multilayer feed forward neural network. The proposed
method calculates distance between activation values of hidden
units for a given input values and moves them depending on the
calculated distance value. The method shows fewer rules on
three publicly available data sets without compromising
classification accuracy.
Keywords—Rule extraction, Feed Forward Neural Network,
Hidden units, Activation value.
I. INTRODUCTION
The application of Artificial Neural Network (ANN)
technology is now extended in fields as diverse as commerce,
science, industry and medicine, offers a clear testament to the
capability of the ANN paradigm. The following three salient
characteristics of ANNs makes it capable for this success. The
first is the ability to learn dynamically from the network by
adjusting weights in training of the network. i.e. dynamically
acquiring information/knowledge about a given problem
domain through a training phase. The second characteristic is
the ability to store knowledge in numerical form which makes
it compact. The third characteristic is its robustness, the
ability to provide solution of noisy input data.
In addition to these characteristics, one of the most important
advantages of trained artificial neural networks is the high
degree of accuracy on generalization of solution [2], extracted
on a sample data of a problem domain, over a set of
previously unseen examples from the problem domain. The
effectiveness of artificial neural network as a tool to aid
human decision making variety of business application
problems and other industries are increasing widely.
Classification and pattern recognition is one of the important
application of ANN [8][9] . Text Classification uses several
tools from Information Retrieval (IR) and Machine Learning
[13]. Similarly Gait for human activity recognition also can be
done using ANN [14]. To detect and diagnosis of mental
disorders, researcher used neural network pattern recognition
tool for classification task and designed a network [15].
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Sanjay Misra
Covenant University
Ota, Nigeria
Email:ssopam@gmail.com
Though efficient rule extraction algorithms makes it possible
for decision process of a trained network to be expressed as
classification rules, concepts learned by neural networks are
difficult to understand because they are represented using
large assemblages of real valued parameters. These rules if
more comprehensive to a human user leads better, quicker and
accurate classification of decision. Greedy Rule Generation
algorithm produces concise and accurate rules [6].
Fig. 1. Fully connected feed forward neural network.
Figure 1 shows a network consisting of simple
interconnected units, or neurons, where each unit/node
receives input from other units and sends its output to others.
Neurons of a layer receives the inputs, which is weighted and
biased through an transfer/activation function, usually a step
threshold or a sigmoid function, and then sends the output
value to the neuron of the next layer which will be an input
for that layer. Neuron of a layer is connected to the other
neuron via a weighted connection. The simplified process for
training a Feed Forward Neural Network (FFNN): (1) present
the input to the network (2) propagate through the network
until it reaches the output layer. This forward process
produces a predicted output. (3) Calculate the error value for
the network i.e. the difference between the predicted output
and the actual output. (4) Adjust the weights of connections to
minimize the error value. Since the networks internal layers
and weights at internal nodes are a set of floating points, their
behavior are not human readable. Hence ANN are also known
as Black box [3]. In past research work, many researchers and units are made to reduce distance between each other, the
scholars have worked on this feature of ANN [1][3][4][7][16]. hypothesis of the proposed method is that it will result in
Rules are extracted in human readable format. The approach fewer rules.
of extracting rules with input-output relation without
considering units and weights is known as Pedagogical IV. THE DISTANCE TERM
approach.RE-RX algorithm is used [5]. In Decompositional
approach, units and weights at internal representation is being Figure 2 shows the proposed method uses the
taken into consideration. The rules are extracted initially dissimilarities or distances between objects when forming the
between the hidden units and output units and then rules are clusters. These distances can be based on a single dimension
extracted between input and hidden units to get the rules in or multiple dimensions. The most straightforward way of
terms of input and output after combining them. The hybrid computing distances between objects in a multi-dimensional
approach of the above two approaches is known as Eclectic space is to compute Euclidean distances. Which is the most
method. commonly chosen type of distance. Euclidean distance is the
geometric distance in the multidimensional space. The
squared Euclidean distance squares the standard Euclidean
II. SIGNIFICANCE OF HIDDEN LAYER REPRESENTATION
distance in order to place progressively greater weight on
A network is trained by a training algorithm which works objects that are further apart. This distance is computed as
on the principle of minimizing the error term. The error term
is calculated at output layer by squaring the difference
between the target output and the actual output. The goal of
the training algorithm is to minimize this error term by
changing representation of hidden units at each iteration,
which makes it complex and needs more rules to explain.
Researchers and scholars have worked on improvisation on
training algorithms but not much on the internal
representation of a network, the hidden layers and weights
calculated at hidden neurons in comparison to training
algorithms. Since in decompositional approach of rule
extraction, rules are extracted from hidden units, therefore the
number of rules mainly depends on hidden units and internal
layer representation. During training hidden unit activation
values can take their values anywhere in the space to achieve
minimum squared error term calculated by the learning
algorithm. Hence to get fewer rules, it is important to Fig. 2. Internal Representation of Feed forward Neural Network.
minimize the scattered activation values at hidden units in the
network.

At hidden layer, activation value a of ith hidden unit for jth

III. MOTIVATION input sample is calculated as
The researcher [10] has proposed an Error term in hidden N aji = f ∑ ∑ (wipj + bi) (1)
layer representation that makes the hidden unit activation i=1 j=1
representation sparser. The novelty of his method is in the use
of the new error term that pushes the hidden activation vectors Where f is a transfer function. The transfer function can
away from each other, so that their component values tend to be sigmoid or logistic function.
cluster toward the two ends of the interval [0, 1]. The method
only performed well on four out of the five large data sets. Error term E at output unit is calculated as:
The approach was not successful on a fifth data set of Image N Output
segmentation which was taken from UCI Machine Learning
E = XX (t Pk − aPOK)2 (2)
Repository [16].
P=1 K=1
The proposed method attempts to reduce hidden unit
activation values distance from each and every other
Where is the target output at kth output unit for Pth input
activation values for an input pattern at hidden layer of three
sample and is the activation value at kth output unit for Pth
layered feed forward network without sacrificing the accuracy
of classification. Since activation values of hidden input value.
The input-hidden layer connection weights changes in the
space during training and results a complex and scattered
values of hidden units. By reducing the space between this
values, the attempt has been made to bring hidden unit values
closer to each other. (10)

Since Distance term D incorporates the sum of distances

between the hidden unit activation values of all hidden units (11)
for an input sample, its effect is to push activation value of all
hidden units for an input sample closer to each other or in
other words minimize the distance between them. With this Similarly for Q ≠ j
effect they converge to an intended class label and give fewer
rules by overcoming from intermediate rules. For example, In
the illustrative example, three hidden units are considered at
internal layer. For a given input sample P, apHk represents the
activation value at hidden unit Hk. The distance between Adding both Eqn
activation values of hidden units for an input sample P is
calculated as:

(3)

The gradient of D with respect to weight wji : for K=Q

(4)

Since for k ≠ j, ≠

(5) The activation value which is farther from activation

values of other hidden unit for a given input sample will have
large value of the distance term D. To bring activation values
(6) of all hidden units for an input sample, closer to each other,
the farthest activation value is brought closer to other
activation values by travelling the calculated distance for that
hidden unit if the distance travelled is less than 1. Since
(7) activation values falls between 0 and 1 after sigmoid transfer
function, the Distance term for a hidden unit, having value
more than 1 signifies that the activation value of the hidden
unit is towards the other end of the space , hence could intend
to different class label. To preserve the accuracy of the rules,
such hidden unit activation values are not moved closer to
other hidden units activation values. To illustrate the
With K ≠j and Q ≠ j significance of value of distance term for taking decision of
moving a hidden units activation value, closer to other hidden
units activation value or not, we consider that there are k
number of hidden units at internal layer of three layered feed
forward neural network. Out of k hidden units, a disjoint
subset of hidden unit’s activation values where number of
With K ≠j and Q ≠ j hidden unit’s activation values in the subset is greater than or
equal to 1 and less than or equal to k-1, for an input sample,
falls in the space closer to each other. Similarly other disjoint
subsets of hidden units activation values also falls on the
space taking values between 0 and 1. The activation value of a
(9) hidden unit of a sparser subset will have larger value of
distance term in comparison to denser subset of hidden unit’s
activation values. If all subsets of activation values are not
for k ≠ j
near to two opposite ends of the space (value of distance term
is more than 1), i.e few fall at intermediate place in the space,
the sparser hidden unit activation values will be moved to
denser group of hidden units activation values. Hence the
scattered activation values are converged towards the class
label for which more number of hidden units intends to. This
in effect gives less number of rules without compromising the
accuracy of rules.

V. RULE EXTRACTION
The three layered feed forward network is trained using
Levenberg Marquardt training algorithm. After training the
hidden unit activation values are clustered using Chi2
algorithm [12]. It clusters these hidden unit activation values
in the intervals [0,b1),(b1,b2),(b2,b3),(bk,1] such that each
activation values falls in interval li where li represents
(bi−1,bi)[12]. Explaining the output in term of clustered hidden
unit activation values using CART algorithm. Rules are
extracted in the form If (Hj1 = l1 and Hj2 = l2) then class =
class1
Which means if j1th hidden unit’s activation value is in
Fig. 3. Pattern of hidden unit activation throughout the space for the ILPD
l1interval and if j2th hidden unit’s activation value is in l 2 problem after training without Euclidean distance term.
interval. Then the rule is classified as class1. Since after
introducing the distance term, hidden unit activation values space more rules are required to explain the hidden activation
are moved closer, the values will fall in less number of - output relationship.
different intervals li and hence it will result fewer rules. In the
next step of rule extraction from input values and hidden unit Figure 4, shows the pattern of hidden unit activation
values, rules can be generated directly for continuous input as values after training the network with Euclidean Distance
follows: If jth hidden unit activation is in an interval li i.e. term. It can be seen that the hidden unit activation values are
not scattered in the space. Values are brought closer to each
If aHj falls to li then wj1x1 + wj2x2 + ... + wjn xn falls to li other to make cluster. This will result in more number of
Since extracted hidden output rules are in terms of the hidden unit activation values to fall in an interval and hence
intervals a hidden unit value falls, the input output rules can the extraction of rules of hidden unit values, belonging to an
be extracted by combining it with above input and hidden interval, to be classified to a class will be fewer.
value rule.

VI. ILLUSTRATIVE EXAMPLE

The ILPD (Indian Liver patient Data set) is used to
illustrate the effect of proposed distance term on the hidden
unit encoding learned by the neural network. The ILPD data
set consists of 583 instances of patients. Each patient is
characterized by 10 continuous inputs. The problem is to
classify these patients into one of the two classes. First a three
layer feed forward neural network with 3 hidden units is
constructed. The tangent sigmoid and logarithmic functions
are used on input and hidden layers as activation function.
The sixty percent of the total data instances are taken for
training and twenty percent each for testing and validation of
the network. After training, the computed hidden unit
activations of three hidden units for each ith instance are
represented as aiH1 ,aiH2, aiH3 . This vector is an encoding of
the 10dimension vector input. After Euclidean distance
between each vector, it is plotted in three dimensional space.
Fig. 4. Pattern of hidden unit activation throughout the space for the ILPD
Figure 3 is the plot of clustered hidden unit vectors problem after training with Euclidean distance term.
without the Euclidean distance term, trained by Levenberg
Marquadt training algorithm. The figure depicts the scattered
cluster values in the space. Since the values are spread in the
 VII. EXPERIMENTAL RESULTS
The proposed rule extraction method is experimented on
three datasets taken randomly from the UCI Machine
Learning Repository [16]: the ILPD, waveform and Image-
segmentation problems. The dataset is large having multiple
attributes and classes. These three datasets has continuous
inputs. ILPD dataset is a problem of patients having test
results of lipid profile to determine diabetes in the patient. The
waveform problem involves classifying waves into one of
three classes based on 21 noisy features and the image-
segmentation data described by high-level numeric-valued Fig. 5. Experimentation Results Average over 150 Runs.
attributes, 7 classes. The instances were drawn randomly from
a database of 7 outdoor images. The images were hand
segmented to create a classification for every pixel. Each VIII. DISCUSSION AND FUTURE SCOPE
instance is a 3x3 region. From each dataset, 80 percent of the
data is used for the training the training, 10 percent is used for Following observations has been identified on comparing
testing and 10 percent for validation. The initial weights and experiment results of all datasets.
weight decay term are kept same during entire ILPD and Waveform has highest percentage of reduction in
experimentation of all datasets. . Experiments are performed number of rules whereas Image Segmentation has lowest. The
on these datasets to evaluate and then compare the number of difference between them is mainly the number of classes.
rules extracted from the dataset from a trained network Number of parameters of Image Segmentation dataset is
without the distance term and after the inclusion of distance highest.
term. ILPD and Waveform dataset has almost same number of
classes for classification. Waveform dataset has more number
Table I shows the description of the data sets used in the of attributes and instances in comparison to ILPD dataset and
experimentation and Table II shows the results of experiments it resulted in slight fall in percentage of reduction in number
on these data sets with and without distance term. Results are of rules.
compared between the column actual, Figure 6 and Figure 7 With respect to result of previous work done[10], the
shows the results of experiment on data sets without distance proposed method has shown positive results even for Image
term and new, the results of experiment on data sets with Segmentation dataset which was a demerit of the work in
distance term. [10]. The proposed method is experimented on datasets with
continuous inputs only. Hence future scope of the work done
The training algorithm was set to 100 epochs for ILPD and its neighborhood values. Experimental results on three large
400 for waveform and Image-segmentation depending on the public data sets show that the method has helped to reduce the
number of instances in each dataset. The minimum gradient number of rules significantly without sacrificing the
was set to 1e-10. classification accuracy. The Movement of an activation value
of is to experiment the proposed method for discrete attributes
TABLE I. DATA SETs also.
Data set No of attribute No of Classes No of instance
ILPD 10 2 583
Waveform 21 3 5000 IX. CONCLUSION
Image Segmentation 18 7 2310
In this paper, a simple, improved and efficient method is
TABLE II. EXPERIMENTAL RESULTS proposed for extracting rules from Multi Layered Feed
forward Neural Network. The method mainly works on
Data set No of Rules Rule Accuracy
internal representation of the network where activation values
New Rule Regular Reduced Actual New
are moved to hidden unit, having distance term value less than
ILPD 22.54 15 33.45% 76.14% 76.5
% 1, for stronger inclination towards a class label compromises
Waveform 69.04 39.8 42.35% 80.06% 80 on the accuracy due to approximation. Applications where
%
Image Segmentation 83 79 5% 55.5% 56.9
decisions are many in a small space or probability to
% proneness for a class label is also to be extracted, there such
scattered activation values will help in driving probabilities of
class labels of that activation value at two ends. The
application considered in this research work does not require
such probability estimations. Hence approximation does not
compromise the decision. The moved activation value
pronounces the decision more boldly.

REFERENCES
[1] C.W.Omlin, C. L. Giles. Extraction of Rules from Discrete-Time
Recurrent Neural Networks, Neural Networks,Vol 9,No 1,pp 441-52,
1996.
[2] A.Gupta. Generalized Analytic Rule Extraction for Feed forward Neural
Networks, IEEE Transactions on Knowledge and Data Engineering, Vol.
11, No. 6, pp 985-991, 1999.
[3] S.M. Kamruzzaman and A. R. Hasan, Rule Extraction using Artificial
Neural Networks, ICTM, 2005.
[4] D. Dancey, Z.A.Bandar and D. McLean.Logistic Model Tree extraction
from Artificial Neural Network. IEEE Transactions on Systems, Man,
and Cybernetics, Vol.37, No.4, pp 794-803, 2007.
[5] R. Setiono, B. Baesens, and C. Mues. Recursive Neural Network Rule
Extraction for data with mixed attributes, IEEE Transactions on Neural
Networks, Vol.19, No.2, 2008.
[6] K. Odajimaa, Y. Hayashi, G. Tianxia, R. Setiono. Greedy rule generation
from discrete data and its use in neural network rule extraction, Neural
Network, Vol 21, pp 1020-1028, 2008.
[7] M.G. Augasta, T. Kathirvalavakumar. Reverse Engineering the Neural
Networks for Rule Extraction in Classification Problems, Neural Process
Lett, Springer, Vol 354, pp 131-150, 2011.
[8] S.M. Kamruzzaman and A.M. Jehad Sarkar. A new data mining scheme
using artificial neural networks, Sensors, Vol 11, pp 4622-4647, 2011.
[9] L. Ai-sheng, Z.Qi. Automatic modulation classification based on the
combination of clustering and neural network, Science Direct, 2011.
[10] T. Q. Huynh and J. A. Reggia. Guiding Hidden Layer Representations
for Improved Rule Extraction from Neural Networks, IEEE, Neural
networks, Vol 22. No 2, Pg No 264-275, 2011.
[11] S. Kulluk, L. Ozbakr, A. Baykasoglu. ”Fuzzy DIFACONN-miner: A
Novel approach for fuzzy rule extraction from neural networks ”,Expert
System with Applications, Vol 40, No 3, pp 938946,2013.
[12] H. Liu and R. Setiono, Chi2: Feature selection and discretization of
numeric attributes, Proceedings of the Seventh International Conference
on Tools with Artificial Intelligence, IEEE Computer Society
Washington, DC, USA,1995.
[13] V. Bijalwan , V. kumar, P. Kumari KNN based Machine Learning
Approach for Text and Document Mining , International Journal of
Database Theory and Application, Vol 7, pp 61-70, 2014.
[14] A. Gupta, J. Prakash. Human activity recognition using gait pattern,
International Journal of Computer Vision and Image Processing, Vol 3 ,
pp 31-53,2013.
[15] P. Kumari, A. Vaish Individual Identification based on Neurosignal
using Motor Movement and Imaginary Cognitive Process, International
Journal for Light and Electron Optics. DOI:
http://dx.doi.org/doi:10.1016/j.ijleo.2015.09.020.2015.
[16] A. Asuncion and D.J. Newman, UCI Machine Learning Repository,
http://www.ics.uci.edu/ mlearn/MLRepository.html.
 An Investigation on Residential Exposure to
Electromagnetic Field from Cellular Mobile Base
Station Antennas
Amar Renke
Shivaji University: Department of Electronics Engineering
KIT’s College of Engineering
Kolhapur, India
amarrenke@hotmail.com
Abstract—The increasing number of antennas on the cellular
base station towers led to growth of public concern related to
human exposure to electromagnetic fields (EMF) and various
health effects. The study investigated the levels of
electromagnetic radiation emitted from cellular base stations in
Kolhapur, a district located in west Maharashtra. The paper
focused on human exposure to electromagnetic radiation from
base station antennas in residential areas. Various exposure
situations were considered such as hall, kitchen, bedroom, and
terrace, where electromagnetic field exposure was measured in
terms of power density and electric field. The average height of
the antenna tower was around 150 feet and average numbers of
antennas on the base station towers were around 16. Results were
tabulated in terms of power density and electric field. The
measured values of power density were well below the maximum
permissible exposure levels. The electromagnetic field exposure
was maximum at terrace, medium at hall and bedrooms and low
at kitchens. The residential exposure varies from 45.92 to 13860
μw/ m2. The maximum value of residential exposure was 2797
μw/ m2.
Keywords—power density; health consequences; residential
exposure; dwellings
I. INTRODUCTION
In the recent years, wireless data transmission has recorded
an impressive technological development. Due to increased
use of mobile, the installation of new base stations in the
vicinity of dwellings increases the public exposure to
electromagnetic field. As the base stations are installed in the
residential areas, the residential exposure comes into the
picture. Because there will be the existence of EM field
everywhere in the dwellings. Each person from dwellings gets
exposed to EM field from cellular base station antennas.
Generally in urban areas the population is dense therefore
cell size is reduced to cover each and every part of the city.
This requires more no of base stations. To find out the likely
impact of the EMF exposure from cellular base stations,
different studies were carried out in the different areas of EMF
exposure measurement [1]-[5]. And possible consequences of
human exposure to EM fields from cellular base stations [6]-
[10]. More awareness is present in developed countries
regarding EM field exposure due to cellular base stations,
while less awareness in Asian countries. This paper presents
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Dr. Mahesh Chavan
Shivaji University: Department of Electronics Engineering
KIT’s College of Engineering
Kolhapur, India
maheshpiyu@gmail.com
work related to EM radiation emitted from cellular base
stations in close vicinity of dwellings. We investigated
different cases of residential human exposure to EM radiation
from base stations in the village of Pachgaon. Located in south
west side of Kolhapur. Base station antenna towers may be
located on ground, rooftops of residential buildings or on
commercial buildings or on hotel rooftops. We measured the
magnitude of EM field and power density in various parts of
the dwellings. The contribution of EM radiation from cellular
mobile base station antennas is higher as compare to other EM
field sources.
II. MATERIALS AND METHODS
In Kolhapur, there were different cellular systems such as 2G
and 3G systems like GSM, CDMA and UMTS which provides
mobile voice and data services. Each system uses different
frequency bands i.e. 800, 900, 1800 and 2100 MHz. all
systems provide voice, data and internet services. Here in
Kolhapur different service providers available are Reliance,
Airtel, Vodaphone, and Idea etc. To estimate the worst case
residential exposure, we measured EM field levels in different
parts of the dwellings. In our study, base station towers are
mounted on ground and average numbers of antennas on the
towers were around 16 to 20. All investigated dwellings were
situated around the base station antenna towers. The average
distance between base station antenna tower and dwellings
was varying from 10m to 50m.
EM field were measured in different parts of dwellings such
as hall, bedroom, kitchen and balcony and rooftops. In most of
the cases distance between dwellings and antenna tower were
around 20m. Considering the average height of the Indian
people the EM field exposure was measured at height 1.5m
with the help of three axis electromagnetic field meter model
KM-195 as shown in figure 1.0. This meter is having
frequency range 200MHz to 2.5 GHz which covers entire
frequency band used for cellular mobile communication.
Initially base station antenna sites were selected then EM
field and power density were measured in different parts of the
dwellings.
 1.00 pm 2902 1360.2
2.00pm 3045 1392.1
3.00 pm 3200 1371.3
4.00 pm 2871 1307.9
5.00 pm 2910 1397.3
6.00 pm 3747 1308.8
7.00 pm 2196 1288.9
8.00 pm 2549 1223.7

The magnitude of power density and electric field were high

at roof-top and balcony. While lower in kitchen and medium
Fig 1.0 Three axis field strength meter KM 195 in hall and bedroom. To provide better results for each
location measurements were repeated and averaged.
First it was measured in hall, kitchen, bedroom and roof To analyze measured data and interpreting the final results
terrace. Measurements were repeated several times in the from the point of view of residential exposure.
dwellings. Finally average was taken in each case. This was
repeated in somewhat around 50 homes. Electric field was TABLE II. DAILY VARIATIONS IN EMF EXPOSURE AT BEDROOM
measured in mv/m and power density in μw/ m2. Time
Power density µw/
m2 Electric field mv/m
Measurements of power density and electric field were taken
during the daytime from 8.00 a.m. to 8.00 p.m. All 8.00 am 393.6 543.7
measurements were carried out on sunny days. The readings 9.00 am 524.2 522.3
from base station were measured after every one hour from
8.00 a.m. to 8.00 p.m. to observe the fluctuations in the power 10.00 am 528.9 531.6
density. Here we have considered three cases which are as 11.00 am 511.7 514.2
follows 12.00 pm 406.1 516.8
 Case 1: EMF exposure measured in dwellings which
are situated on ground and 20 meters away from 1.00 pm 609.1 417.7
cellular mobile tower. 2.00pm 389.2 501.8

 Case 2: EMF exposure measured in dwellings on 3.00 pm 568.9 480.1

second floor on which cellular mobile tower is 4.00 pm 484.9 549.3
situated.
5.00 pm 579 529.3
 Case 2: EMF exposure measured in dwellings on 7 th
floor on which cellular mobile tower is situated. 6.00 pm 478.6 438.3
7.00 pm 526.6 455
III. RESULTS
8.00 pm 472.3 544.2
It is found that inside the dwellings the magnitude of power
density and electric field intensity was varying according to
IV. DISCUSSION
place. For better results power density and electric field was
Measured values of EM field levels for all investigated
measured near windows inside the dwellings. Table 1.0 and
cases, ranged between 45.92 to 13860 µw/ m2. The values of
2.0 shows daily variations in power density and electric field
power densities were higher near the base stations and it
at terrace and bedroom respectively.
decreases as we move away from the base station antennas.
TABLE I. DAILY VARIATION IN EMF EXPOSURE AT TERRACE Power density was varying according to direction of base
station. In our study it was found that power densities were
Location Terrace low in hall and kitchen, medium in bedrooms and high in
Home balcony and rooftops.
Date 25/05/2015
Time Power density µw/ m2 Electric field mv/m TABLE III. AVERAGED POWER DENSITIES AT DIFFERENT
PLACES
8.00 am 2974 1423.2 Location Average power density µw/ m2

9.00 am 2865 1383.9 Case 1 Case 2 Case 3

Hall 338.92 4312.84 338.92
10.00 am 2678 1388.5
Bedroom 373.37 1074.38 258.37
11.00 am 2796 1326.9 250.60 195.04 45.92
Kitchen
12.00 pm 2741 1349.2 Terrace 2597.90 11492.23 13860
Power density and electric field changes in accordance with Figure 3 show the graphical representation of table IV. It
number of voice calls, distance from base station antenna found that EMF exposure is grater for terrace and second
tower and height of the antenna. Table 3.0 gives the average floor and lower at higher floors in the dwellings.
power densities at different parts of the dwellings. All measured values of EMF exposure for all cases and
There was almost 70 % contribution is from RF fields from locations are 890 μw/m2, 4268.62 μw/m2, 3625.80 μw/m2
cellular mobile base stations. Indian regulations on human respectively for case1, case2 and case3. Measured power
exposure to EM fields are based on reference levels set by densities do not exceed the reference levels set by ICNIRP and
DoT (Department of Telecom). Table 4.0 represents averaged DoT.
electric field at different places in the dwellings.
V. CONCLUSION
Our study investigated the level of human exposure to EM
radiation emitted by cellular mobile antennas at dwellings.
Residential exposure was evaluated in dwellings in close
proximity with base station antenna towers. Measurements
showed that power density varies with height of antenna,
distance between dwellings and base stations and number of
antennas radiating in same direction. Also it was found that
intensity of power density was higher at roof-terraces, medium
at Hall and bedroom and lower in kitchen. The maximum
power density and electric field was recorded as 13860 μw/ m2
and 1224.5 mv/m. The reference level for general public
exposure was never exceeded in any location. The outcome of
our experimentation is that, human exposure in investigated
Fig. 2. Power densities at different places. dwellings situated in the proximity of base station antennas for
mobile phones complies with DoT guidelines and international
Figure 2 shows the graphical representation of table 3. It Standards.
shows that EMF exposure is higher at terrace and lower at
other places in the dwellings.
References
Table IV. Measured electric field at different places
Location Average electric field mv/ m [1] Institute of Electrical and Electronics Engineers. Safety levels with
respect to human exposure to radio frequency electromagnetic fields, 3
Case 1 Case 2 Case 3 kHz to 300 GHz. Report IEEE C95.1-1991 (Piscataway, NJ) (1992).
Hall 357.45 1275.12 375.45 [2] International Commission for Non-Ionising Radiation Protection.
Guidelines for limiting exposure to time-varying electric, magnetic and
Bedroom 375.17 636.42 312.09 electromagnetic fields (up to 300 GHz). Health Phys. 74, 494–522
(1998).
Kitchen 307.00 271.13 131.57 [3] International Commission for Non-Ionising Radiation Protection. Health
issues related to the use of hand-held radiotelephones and base
Terrace transmitters. Health Phys. 70, 587–593 (1996).
1266.25 2691.07 9266.23
[4] National Radiation Protection Board. Board statement on restrictions on
For the frequency bands 800, 900 1800 and 2100 MHz, the human exposure to static and timevarying electromagnetic fields. NRPB
reference levels are of 28, 43, 58 and 61 v/m respectively. Report 4(5) (Chilton, UK: NRPB) (1993).
[5] National Council on Radiation Protection and Measurements. Biological
effects and exposure criteria for radio frequency electromagnetic fields.
NCRP Report 86 (Bethesda, MD: NCRP) (1986).
[6] US Federal Communications Commission, Office of Engineering and
Technology. Evaluating compliance with FCCV-specified guidelines for
human exposure to radiofrequency electromagnetic fields. OET Bulletin
65 (1997).
[7] Cooper, J., Marx, B., Buhl, J. and Hombach, V. Determination of safety
distance limits for a human near a cellular base station antenna, adopting
the IEEE standard or ICNIRP guidelines. Biolelectromangetics 23, 429–
443 (2002).
[8] CENELEC standard EN 50383, ―Basic standard for calculation
and measurement of electromagnetic field strength and SAR
related to human exposure from radio base stations and fixed
terminal stations for wireless telecommunications system‖ (110
MHz – 40 GHz), CENELEC standard EN 50383
[9] RADIATION PROTECTION STANDARD ―Maximum Exposure
Fig. 3 Electric field at different places. Levels to Radiofrequency Fields — 3 kHz to 300 GHz‖ Radiation
Protection Series Publication No. 3 ARPANSA.
[10] CENELEC Standard EN50420, ―Basic Standard for the Evaluation of
Human Exposure to Electromagnetic Fields from Stand Alone Broadcast
Transmitter‖ (30 MHz–40 Ghz), CENELEC Standard EN50420, 2006.
[11] CENELEC standard, EN50492, ―Basic standard for In-situ measurement
of electromagnetic field strength related to human exposure in the
vicinity of base stations‖, CENELEC standard, EN50492, 2008.
[12] CENELEC Standard EN50400, ―Basic Standard to Demonstrate the
Compliance of Fixed Equipment For Radio Transmission (110 Mhz –
40 Ghz) Intended for Use in Wireless Telecommunication Networks
With the Basic Restrictions or the Reference Levels Related to
General Public Exposure to Radio Frequency Electromagnetic Fields,
When Put Into Service, CENELEC Standard EN50400, 2004.
[13] Byung Chan Kim, Jae-Hoon Yun, and Seong-Ook Park, Member, IEEE
―Uncertainty Estimation for Evaluating Human Exposure Levels to RF
Electromagnetic Fields from Cellular Base Stations‖ IEEE
TRANSACTIONS ON ELECTROMAGNETIC COMPATIBILITY,
VOL. 54, NO. 2, APRIL 2012.
[14] B. C. Kim, H. -D. Choi and S.-O. Park, ―Methods of Evaluating
human exposure to electromagnetic fields radiated from operating
base stations in Korea, ‖ Bio- electromagnetics, vol. 29, no. 7, pp.
579– 582, Oct. 2008.
[15] CENELEC EN50166, ―Human Exposure to Electromagnetic
Fields.High Frequency‖ (10 KHz to 300 GHz), CENELEC EN 50166-2,
Jan. 1995.
[16] S. I. Henderson and M. J. Bangay, ―Survey of RF exposure levels from
mobile telephone base stations in Australia,‖ Bioelectromagnetics, vol.
27, pp. 73– 76, 2006.
 Parallelization of Association Rule Mining:
Survey
Shivani Sharma∗ and Durga Toshniwal†
Dept. of Computer Science & Engineering
Indian Institute of Technology, Roorkee
Uttarakhand, India. 247667
∗ Email: shivani.vce@gmail.com
† Email: durgatoshniwal@gmail.com

Abstract—In todays big data era, all modern applications are
generating and collecting large amount of data. As a result,
data mining is encountering new challenges and opportunities
to make algorithms such that, this voluminous data can be
effectively and efficiently transformed into actionable knowl-
edge . Traditional algorithms were designed to run sequentially
over a single machine. But, as the volume of data increases
computational cost associated with its processing also increases.
This causes problems in analysing data on a single sequential
machine and instead of assisting in data analysis, the processor
serve more like a bottleneck. Parallel and distributed ap-
proaches improve the performance in terms of computational
cost as well as scalability but experience some limitations
during load balancing, data partitioning, job assignment, moni-
toring etc. MapReduce, a parallel programming model is a new
concept which provides seemingly unlimited computing power,
cheap storage as well as, can overcome above limitations. This
makes it a topic of upcoming research interest. A detailed
literature review of some existing methods is given along with
their pros and cons.
1. Introduction
With the advancement in technology, all modern appli-
cations are generating and collecting large amount of data.
Some typical examples include social networking, wireless
sensors network, computer networks etc. As a result, data
mining require more effective and efficient algorithms to
transform this huge data into actionable knowledge. Data
mining can be defined as a process of extracting hidden
patterns and predictive information from large volume of
data. Broadly this hidden knowledge can be extracted using
methods like, Association Rule Mining, clustering, sequence
analysis, classification or forecasting. Association rule min-
ing is one of the important technique used for extracting
interesting patterns and correlations among items in large
dataset [21]. There are two traditional approaches followed
called candidate-generation based (Apriori algorithm) and
Candidate-less (FP-growth algorithm).
The rule mining task can be divided into two computa-
tionally intensive subtasks i.e frequent itemsets generation
and association rule generation. It require the algorithms to
be very efficient and scalable. Traditional algorithms were
designed to run sequentially over a single machine. But, as
the volume of data increases computational cost associated
with its processing also increases. This causes problems in
analyzing data on a single sequential machine and instead
of assisting in data analysis, the processor serve more like
a bottleneck. To deal with upcoming issues, due to large
scale data, parallelization of mining algorithms becomes
inevitable. Parallelization of association rule mining tech-
niques can be done either by dividing and distributing data
over multiple nodes and generating association rules locally
followed by merging local results to obtain global asso-
ciation rules. This is called Data-parallelization. Second
approach is, parallelizing algorithm itself called Algorithm-
parallelization.
A detailed literature review of various existing methods
is given along with their pros and cons. Section 2 includes
some background of traditional association rule mining tech-
niques. Section 3 includes the discussion of some existing
research work done for parallelization of association rule
mining. Section 4 explores some of the existing techniques
for implementing association rule mining on MapReduce
with their respective advantages and limitations. Section 5
draws conclusion.
2. Association Rule Mining
Association rule mining can be defined as a process
of extracting correlations and associations among items in
large dataset [21]. Association rule is the implication of
the form X− > Y where X and Y are different items
in transactional or relational database. X− > Y holds
in dataset D, two properties: Support(s)- Total number of
transactions containing both X and Y (X union Y) out of
total present transactions. Equation1.
i.e.
support(X− > Y ) = P (XU Y ). (1)
and Confidence(c)- Percentage of transaction containing X,
given that it already has Y Equation2.
i.e.
conf idence(X− > Y ) = P (Y |X). (2)


ª
*&&&
Association rule mining aims to extract frequent item-
sets with support(s) >= Min sup and Finally generating
association rules from these frequent item set with and
confidence(c)>= Min conf . Here, Min sup and Min conf
are threshold values. Association rule mining techniques can
be divided into two major catagories- Candidate-generation
based techniques and Candidate-less techniques. Number of
algorithms exist under each catagory but importantly Apriori
algorithm is a candidate-generation based approach and FP-
Growth is a candidate-less approach.
Apriori Algorithm [21] uses the prior-knowledge in
terms of properties of frequent itemset. It is an iterative
approach where k-frequent itemsets are extracted using (k-
1)-frequent itemsets. The algorithm exhibits anti-monotonic
property called Apriori property [21], states that ” For a
frequent itemset all its non-empty subsets must also be
frequent.” Apriori can be seen as two step process: Fre-
quent itemset generation and Association rule generation.
Frequent itemset generation can be further divided into
two subtasks called Joining where, lk−1 candidate sets are
joined with itself to generate candidate k-itemsets (Ck ). and
Pruning where, all those candidate frequent itemsets with
support less than Min sup get pruned off. The working of
apriori can be seen in Figure 1.
Figure 1. Working Flow of Apriori Algorithm [21]
FP-Growth algorithm [20] is candidate-less approach
therefore, do not generate candidate itemset in between.
It uses tree as a data structure which is much compact
and frequent itemsets are mined directly from the tree.
FP-Tree is a composed and compressed representation of
large dataset. This approach require only two database
scans. First scan is required to obtain the support count
of each item in dataset. Infrequent items with support
less than Min sup are removed and others are arranged in
decreasing order of their support. Second scan over dataset
is made to construct FP-Tree. Finally this tree is used to
extract frequent itemsets directly using bottom-up strategy.
Both traditional approaches can be compared over several
parameters as shown in Table 2.
Traditional algorithms are designed to run sequentially
over single machine [20], [21]. With increase in data vol-
ume, computational intensiveness increases such that it be-
comes impossible for single machine to perform efficiently
[3].
3. Parallelization of Association Rule Mining
Association rule mining can be viewed as two step
process [3] [21]- frequent itemset generation and asso-
ciation rule generation. Being computationally intensive
tasks parallelization is next step required to be taken to
reduce the work load over one sequential machine and
distribute it among several nodes. Parallelization of associ-
ation rule mining can be achieved in two major ways: Data
parallelization- which divide and distribute data over mul-
tiple nodes and generate association rules locally followed
by merging local results to obtain global association rules.
Second is Algorithm parallelization where algorithmic tasks
are distributed over number of nodes. Based on the literature
review done, it can be concluded that most of the research
work is being done on data parallelization.
3.1. Parallelization of Apriori Algorithm
Apriori is most widely used frequent itemset generation
algorithm which, iteratively generate k candidate itemsets
from present (k − 1) frequent itemsets. Several interesting
and effective research work to parallelize the mining of
itemsets can be seen in [11], [12]. Initially Apriori was
implemented on multi-processors and further [1], [2], [13]
implemented them with distributed architecture. Summary
of parallel version of Apriori is defined in [8] as given
below: Count Distribution- can be seen as a most direct form
of parallelization for apriori algorithm. At each node global
candidate itemset as well as frequent itemsets are stored. For
respective local data present on each node, support count
of candidate itemset is calculated using apriori algorithm
and finally, these local results are exchanged among nodes
respectively. Candidate Distribution- here, each node donot
maintain global results but partitions the candidate itemsets
with respect to the partition of dataset. Each node locally
calculate support count of their own candidate itemsets.
Data Distribution- combines the above two approches to-
gether by partitioning both datasets as well as candidate
itemsets at the same time such that, each node can work
independently.
Parallel implementation of finding frequent itemsets fol-
lowed by sequential extraction of association rule depletes
the performance gained so far when the number of generated
frequent itemset is large. Hence, parallelization of second
phase is equally important to enhance the overall perfor-
mance in terms of scalability and efficiency. SeaRum [4] is
one of the initial attempt in our knowledge for developing
both phases of Apriori algorithm in distributed environment.
3.2. Parallelization of FP-Growth Algorithm
Multi-tree approach [10] was one of the initial works
done to parallelize FP-Growth algorithm. Broadly, they used
three steps to achieve parallelism which were followed in
sequence by parallel processing flows. Initially, horizontal
subset of data is analyzed. Secondly local FP-Tree is build
 Comparison of Apriori and FP-Growth
Parameters Apriori FP-Growth
Technique Candidate-based Candidate-less
Time Execution time is more as need to produce candidate It is much smaller than apriori. (Fast)
every time. (Slow)
Memory usage Due to large candidate generation require large Uses tree data structure and donot generate candidate
space. itemsets, require less memory.
No. of scans Multiple scans for generating candidate sets Only two scans are required.
Property Use apriori property, Prune step, Join step Conditional frequent pattern tree and base are con-
structed from database, saatisfying min sup value.

TABLE 1. C OMPARATIVE TABLE OF A PRIORI AND FP-G ROWTH [5]

in parallel and finally, on this local FP-Tree mining pro-
cess is carried out. From every processing flow candidate
patterns are obtained and then merged together. Further,
enhancement were made in merging algorithms using clus-
ter computing environment. Moreover, in [13] certain con-
straints were proposed for massive dataset, which should
be followed in parallel itemset extraction to obtain good
scalability. further improved itemset extraction with better
hardware resource exploitation on multi-core processors. It
proved to be a new innovation in the field which enhanced
the performance of FP-Growth algorithm via improving
the temporary locality for accessing the data at different
levels of memory. In [12] initial effort was made to address
cache-hint optimization using the above technique. Certain
applications were developed [11], implementing the above
for mining itemsets in parallel.
4.1. MapReduce
MapReduce [14] is a parallel programming model which
support distributed computing, required for mining large-
scale data. MapReduce composed of two procedure i.e Map
which basically performs sorting and filtering of data and
Reduce merges the map output to produce final result.
Mapreduce is a framework which can process the paralleliz-
able problems with massive data set by using a large number
of machines. Machines collectively called as cluster which
can take benefit of locality of data.

3.3. Limitations of Parallelization of Association

Rule Mining

Implementing sequence of algorithms for association

rule mining in distributed or parallel environment enhances
the performance in terms of scalability and computational
speed-up [20]. But, still persists various limitations while
implementing sequential algorithms in parallel environment
[3], [22] such as: Partitioning and Distribution, Job as-
signment and Monitoring, Load Balancing, Communication
Overhead. All these activities require one central control
which can moniter and maintain other distributed nodes,
this may be prone to failure, costly as well as uneasy
to handle for complex applications. Thus there is scope
to leverage parallelization in a better fashion for further
enhancing performance.
4. Association Rule Mining on MapReduce
Parallel programming model is a new concept which
provide higher computation power and robustness to failure.
MapReduce is one of the promising parallel programming
model discussed in further Sections.
Figure 2. One Phase of MapReduce Computation [14]
Firstly Mapping is done by dividing datasets into subsets
which are distributed as input to each processor where, code
of respective mining algorithm run to obtain local frequent
itemsets. Each processor produce (key/value list) as output
so that next step can use it for further processing. Secondly
Reduce step shuffle the output obtained and provide input
to reduce processors. Reduce code is executed to merge the
local result and obtaining global rules. Finally sorting is
performed to obtain final result. All these steps combatively
called phase of MapReduce as in Figure 2. One or more
phases may be required to obtain desired pattern or action-
able knowledge out of massive database.
Key benefits of MapReduce is that it automatically
handles failures as well as hides the complexity of fault-
tolerance from the programmer [8]. But support a particular
kind of schema for input file hence, do not offer much
space for improvements. However, the MapReduce frame-
work [18] is not suitable for the frequent itemset mining
algorithm like the Apriori algorithm which have intensive
iterated computation.
4.2. Parallelization of Apriori Algorithm on MapRe-
duce
Several versions of parallellized form of Apriori algo-
rithm to handle issues like improper load balancing, uneven
data distribution, high communication overhead have been
developed and discussed in further Sections
In K-MapReduce Apriori algorithm [23] method extracts
all frequent itemsets from the given dataset using k phases
of MapReduce. Phase ’i’ ran iteratively to obtain frequent
i-itemset where i= 1,2,3,4....k. Improved performance in
terms of scalability and computationally efficiency is ob-
tained. but one MapReduce phase need to wait and schedule
every time for next phase to start, which is pure burden to
the mining task.
The waiting time issue is resolved [7], by using only
a single phase to extract all frequent k-itemsets from the
given data set. But, the performance of method is not much
satisfactory .
In [9] MRApriori stratergy, introduced a new modified
version of Apriori algorithm and requires only two phases of
MapReduce to extract all possible frequent k-itemsets. The
MRApriori performs outstandingly in terms of all desired
deliverables in comparison to other approaches, as only
two phases are required for extracting all possible frequent
k-itemset from entire dataset. But in realistic scenerio, if
number of partial frequent k-itemset generated is huge than
then each node take insignificant amount of time to process
hence, perform inefficiently .
An enhanced variantIMRApriori [3] of MRApriori re-
duced the processing time in case frequent k-itemset gen-
erated is huge [3]. Efficient pruning technique is used for
removing number of partial k-itemsets which are infrequent,
to achieve the computational speed-up. Significant amount
of infrequent itemset get pruned of in phase one itself
which improves the efficiency by removing the case of huge
number of partial frequent itemsets.
4.3. Parallelization of FP-Growth on MapReduce
Initially, efforts were made to parallelize and speed up
Apriori algorithm because of its simplicity. It has been
shown that FP-Growth is much faster than Apriori hence,
logically parallel implementation of FP-Growth will provide
much faster speed up and low cost overhead. Initial parallel
FP-Growth was implemented across multiple-thread and
with shared memory in [10], [11]. But, this approach re-
stricts the percentage of computation that can be parallelized
as well as if dealing with huge dataset, requirement of
huge memory space, high communication cost were another
issue need to be handle. To deal with these limitations
distributed variant of algorithm was designed, which run
over the cluster of machines. [15] applied strategy to solve
the issues like communication overhead, and achieved satis-
factory performance over hundreds of node. But still issues
like scalability, automatic fault-tolerance were hindering the
desired performance.
In [2] Parallel FP-Growth(PFP) method parallelized
the FP-Growth over distributed machines such that each
machine executes independent mining tasks. Such partition
remove the computational dependency which in turn reduces
communication overhead between them. PFP divides the
dataset and save them on different P Nodes. Each part
is called ’Shard’. Parallel counting of support value cor-
responding to each item ’I’ is performed on each part,
using a MapReduce Pass and store these value in F-list.
All ’I’ items are divided into ’Q’ groups (G-List), each
given with unique group Id. Parallel FP-Growth uses another
MapReduce Pass where ’Map’ phase produces output as
Key/Value pair and Reducer take output of mappers in form
of group dependent transactions and again correspondingly
group them into shard. Each machine again assigned with
one or more shard to build a local FP-Tree and growth
them recursively. Finally, discover frequent patterns. PFP
is able to achieve ’Near-Linear’ Speed-up and is fairly
suited for query recommendation. But, it do not take load
balancing into consideration, hence limits if data is huge. It
also need multiple database scans as well as inter processor
communication is high.
An improved variant of PFP called Balanced
PFP(BPFP) [1] resolved the issue of load balancing
when dataset is massive. The algorithm proposed additional
estimate work load and dataset into several units on which
association mining is performed to obtain frequent itemsets.
Key Feature of BPFP is Partitioning Mining Task into
quit even subtasks, which in turn produce higher level of
parallelization. BPFP Performs well even with huge data
volume.
Much similar to PFP a stratergy called CPFP [16] was
introduced which improved in terms of inter process com-
munication, a limitation of PFP. The algorithm implements
the separation strategy which simply ask single visit to
local database. Hence, reduces the inter processor and I/O
overhead.
PIFP-Parallelized Incremental FP-Growth [17], a
MapReduce Based strategy developed basically for large-
scale data processing. The periodic updates of database and
change in threshold value require data mining process begin
from the beginning which is an overhead. Here, presents
a scheme which adapt dynamic database (streaming data)
as well as change in threshold value. The strategy realizes
effective and efficient data mining without any repeated
computation. The strategy divides the whole problem into
two parts: firstly updates of the FP-Tree and secondly
frequent itemset mining. Algorithm effectively reduce the
duplicated work and allow to improve the speed-up, by
adding more number of machines. But PIFP suffer from load
balancing problem which effect the overall performance.
 Summary Table
Technique Platform Algorithm Achieved Remark
improved
K-map Apriori [8] MapReduce Apriori Scalability and works in only K phase High waiting time between two
phases
MR-Apriori [9] MapReduce Apriori Scalability and works in only two phase Perform insignificantly if gener-
ation of k-frequent itemset is
huge, each node take insignificant
amount of time
IMR-Apriori [3] MapReduce Apriori Perform well and much scalable
PFP [2] Mapreduce FP-Growth Near-Linear speed-up Issue of load balancing, multiple
scanning of dB and I/O overhead
BPFP [1] MapReduce FP-Growth Near-Linear speed-up with load balancing Additional computational overhead
of load estimation
CPFP [16] MapReduce FP-Growth Efficient performance, load balancing and Not compatable for dynamic
reduced I/O overhead database or change in threshold
PIFP [17] MapReduce Fp-Growth Adopt incremental changes in database and Load balancing issues still exists
threshold and reduce duplicacy
SeaRum [4] MapReduce Apriori Parallelization of association rule extraction
phase and provide SaaS platform
PRAMA [6] MapReduce Fp-Growth/ Near-Linear speed-up, High scalability, re- Combines Random sampling and
Apriori duce duplicates, Extract rules Directly Parallelization
YAFIM [18] Spark RDD Apriori Faster computation Faster computation than Mapre-
duce
NIMBLE [19] NIMBLE FP-Growth/ Portable, support rapid prototyping Designed for fast and efficient im-
Apriori plementation of MLDM algorithms

TABLE 2. S UMMARY TABLE F OR A BOVE D ISCUSSED T ECHNIQUES

4.4. Other Approaches
Beside MapReduce, many other research work using
different approaches can be seen, which also ultimately
improves the association rule mining process while using
parallel environment. A cloud based service modelSeaRuM
[4] efficiently extracts association rules from huge frequent-
items. SeaRuM run a number of distributed MapReduce jobs
performing different tasks in cloud. The architecture contain
following jobs Data acquisition, Data Preprocessing, Item
Frequency Computation, Itemset Mining, Rule Extraction,
Rule aggregation and Sorting.
A randomized approach [6], PARMA stands for -A Par-
allel Randomized Algorithm For Approximate Association
Rule Mining. The approach combines the two different
methods named Random Sampling and Parallelization for
extracting association rules out of massively huge amount
of dataset. The overall cost of association rule mining can
be divided into two components: scanning and Mining.The
scanning factor increases very rapidly and dominant mining
and make the whole process unsellable and complex. Hence,
PARMA combines both approaches in a novel fashion, it
mines small random samples in parallel. Each sample is
given as input to MapReduce function running in orthogo-
nal manner. lastly, filtering and aggregation of association
rule from each sample are collected. PARMA can compute
association rules directly and is not limited to frequent
itemset extraction. Significantly outperform and has near-
linear speed Up and good scalability. Able to achieve near
constant runtime because of scaling data and nodes together.
Another stratergy YAFIM(Yet Another Frequent Itemset
Mining) [18], used different approach for using parallel
Apriori algorithm over Spark RDD framework in-spite of
MapReduce. Spark is a specially designed for iterative and
interactive algorithms of data mining. It is basically in-
memory based parallel computing model means all data
is loaded in memory itself. Secondly it do not use fixed
two state model as in MapReduce but provide DAG based
data flow. These features allow speeding up the computation
significantly for the iterative algorithms like Apriori. YAFIM
outperforms in terms of fast computation in comparison to
MapReduce.
A new method called NIMBLE [19], aims to achieve
portability of programming code, for fast and efficient
implementation of the parallel data mining algorithms. The
tool kit allows to build Machine learning and Data Mining
algorithms around reusable building blocks parallely such
that they can be easily utilized by other programming model
as well. This helps to achieve inter-portability. NIMBLE
facilitates the processing of variety of data formats due
to its built-in support as well as simplify the custom
data-format implementations. Strategies for optimization
and abstraction incorporate hand in hand to deliver high
performance runtime. It can be seen as a infrastructure,
providing limited effort parallelization and Support for
rapid prototyping.

All discussed approaches in Section 4.1, 4.2 and 4.3 can
be summarized in the form of Table 2
5. Conclusion
Association rule mining aims to find the correlations
and associations from the data. Association rules can
be extracted by using either of the two primary ap-
proaches i.e. candidate-generation based (Apriori algorithm)
or candidate-less (FP-Growth). The process of association
rule mining can be further divided into two subtasks frequent
itemset generation and association rule generation. Frequent
itemset generation is computationally very intensive. Hence,
parallelization of traditional approaches becomes inevitable.
Association rule mining techniques can be parallelized in
two ways i.e. data parallelization or algorithmic paralleliza-
tion. Based on the literature review done it can be concluded
that most of the research work has been done using data
parallelization approach but, not much research work is
being done on algorithmic parallelization. Parallelization of
association rule mining techniques over distributed architec-
ture helps us to achieve scalability and computational speed-
up. But, there are many limitations such as data partitioning
problems, uneven data distribution, improper load balancing,
data communication overhead etc. MapReduce, a parallel
programming model is a new concept which seemingly
provides high computational power hence, promises a great
scope to evolve association rule mining techniques in par-
allel environment.
References
[1] Le Zhou, Zhiyong Zhong, Jin Chang, Junjie Li, Huang, J.Z. and
Shengzhong Feng, ”Balanced parallel FP-Growth with MapRe-
duce,” in IEEE Youth Conference on Information Computing and
Telecommunications (YC-ICT), pp.243-246, 2010.
[2] Haoyuan Li, Yi Wang, Dong Zhang, Ming Zhang and Edward
Chang, ”PFP: Parallel Fp- Growth For Query Recommendation,”
in Proceedings of ACM conference on Recommender systems
(RecSys), 2008.
[3] Farzanyar, Z. and Cercone, N., ”Efficient mining of frequent
itemsets in social network data based on MapReduce framework,”
in EEE/ACM International Conference on Advances in Social
Networks Analysis and Mining (ASONAM), pp.1183,1188, 25-28
Aug. 2013.
[4] Apiletti, D., Baralis, E., Cerquitelli, T., Chiusano, S. and Gri-
maudo, L., ”SeaRum: A Cloud-Based Service for Association Rule
Mining,” in 12th IEEE International Conference on Trust, Secu-
rity and Privacy in Computing and Communications (TrustCom),
pp.1283,1290, 16-18 July 2013.
[5] http://www.slideshare.net/zafarjcp/data-mining-association-rules-
basics.
[6] Riondato, Matteo, Justin A. DeBrabant, Rodrigo Fonseca and Eli
Upfal, ”PARMA: a parallel randomized algorithm for approximate
association rules mining in MapReduce,” In Proceedings of the
21st ACM international conference on Information and knowledge
management, pp. 85-94, 2012.
[7] Lingjuan Li and Min Zhang, ”The Strategy of Mining Association
Rule Based on Cloud Computing,” in International Conference
on Business Computing and Global Informatization (BCGIN),
pp.475-478, 29-31 July 2011.
[8] Xin Yue Yang, Liu, Zhen and Yan Fu, ”MapReduce as a program-
ming model for association rules algorithm on Hadoop,” in 3rd
International Conference on Information Sciences and Interaction
Sciences (ICIS), pp.99-102, 23-25 June 2010.
[9] Othman Yahya, Osman Hegazy and Ehab Ezat,”An Efficient Im-
plementation of Apriori Algorithm Based on Hadoop-Mapreduce
Model,” in Proccedings of the International Journal of Reviews in
Computing, Vol. 12, pp.59-67, 31st December 2012.
[10] Zaiane, O.R., El-Hajj, M. and Lu, P., ”Fast parallel association rule
mining without candidacy generation,” in Proceedings IEEE Inter-
national Conference on Data Mining ICDM, pp.665-668, 2001.
[11] L. Liu, E. Li, Y. Zhang, and Z. Tang, Optimization of frequent
itemset mining on multiple-core processor, in Proceedings of the
33rd international conference on Very large data bases VLDB., pp.
1275-1285, 2007.
[12] A. Ghoting, G. Buehrer, S. Parthasarathy, D. Kim; A. Nguyen, Y.-
K. Chen, and P. Dubey, Cache-conscious frequent pattern mining
on modern and emerging processors, in journal of the VLDB , vol.
16, pp. 77-96, 2007.
[13] M. El-Hajj and O. R. Zaane, Parallel bifold: Large-scale parallel
pattern mining with constraints, in journal of Distributed and
Parallel Databases, vol. 20, no. 3, pp. 225-243, 2006.
[14] J. Dean and S. Ghemawat, Mapreduce: simplified data processing
on large clusters, in Community of ACM, vol. 51, no. 1, pp. 107-
113, Jan. 2008.
[15] Aouad, Lamine M., Nhien-An Le-Khac, and Tahar M. Kechadi.
”Distributed frequent itemsets mining in heterogeneous platforms.”
in Journal of Engineering, Computing and Architecture Vol.1, no.
2, 2007.
[16] Wang Yong, Zhang Zhe and Wang Fang, ”A parallel algorithm of
association rules based on cloud computing,” in 8th International
ICST Conference on Communications and Networking in China
(CHINACOM), pp.415-419, 14-16 Aug. 2013.
[17] Xiaoting Wei, Yunlong Ma, Feng Zhang, Min Liu and Weim-
ing Shen, ”Incremental FP-Growth mining strategy for dynamic
threshold value and database based on MapReduce,” in Proceed-
ings of IEEE 18th International Conference on Computer Sup-
ported Cooperative Work in Design (CSCWD), pp.271-276, 21-23
May 2014.
[18] Hongjian Qiu, Rong Gu, Chunfeng Yuan and Yihua Huang,
”YAFIM: A Parallel Frequent Itemset Mining Algorithm with
Spark,” in IEEE International Parallel & Distributed Processing
Symposium Workshops (IPDPSW), pp.1664-1671, 19-23 May
2014.
[19] Ghoting, Amol and Kambadur, Prabhanjan and Pednault, Ed-
win and Kannan, Ramakrishnan, ” NIMBLE: A Toolkit for the
Implementation of Parallel Data Mining and Machine Learning
Algorithms on Mapreduce,” in Proceedings of the 17th ACM
SIGKDD International Conference on Knowledge Discovery and
Data Mining, pp. 334-342, 2011.
[20] J. Han, J. Pei, and Y. Yin, Mining frequent patterns without
candidate generation, in SIGMOD, pp. 1-12, 2000.
[21] Agrawal, Rakesh, and Ramakrishnan Srikant. ”Fast algorithms for
mining association rules.” In Proccedings of 20th international
conference on very large data bases, VLDB, vol.12-15, pp. 487-
499, 1994.
[22] Geng, Xia, and Zhi Yang. ”Data mining in cloud computing.” In
International Conference on Information Science and Computer
Applications (ISCA 2013), 2013.
[23] X. Y. Yang, Z. Liu and Y. Fu, MapReduce as a Programming
Model for Association Rules Algorithm on Hadoop, in Proceed-
ings 3rd International Conference on Information Sciences and
Interaction Sciences (ICIS), vol. 99, no. 102, pp. 23-25, 2010.
 Effectiveness of LSP Features for Text
Independent Speaker Identification
Sharada V Chougule1, Mahesh S. Chavan2,
Finolex Academy of Management and Technology, Ratnariri1,
KIT’s College of Engineering, Kolhapur2,
shardavchougule@gmail.com1, maheshpiyu@gmail.com2

Abstract— The speech features used for speaker recognition
should uniquely reflect characteristics of the speaker’s vocal tract
apparatus and contain negligible information about the linguistic
contents in the speech. Cepstral features such as Linear Predictive
Spectral Coefficients (LPCCs) and Mel Frequency Cepstral
Coefficients (MFCCs) are most commonly used features for
speaker recognition task, but found to be sensitive to noise and
distortion. Other complementary features used initially for speech
recognition can be found useful for speaker recognition task. In
this work, Line Spectral Pair (LSP) features (derived from
baseline linear predictive coefficients) are used for text
independent speaker identification. In LSP features, power
spectral density at any frequency tends to depend only on close to
the respective LSP. In contrast, for cepstral features, changes in
particular parameter affects the whole spectrum. The goal here is
to investigate the performance of line spectral pair (LSP) features
against conventional cepstral features in the presence of acoustic
disturbance. Experimentation is carried out using TIMIT and
NTIMIT dataset to analyze the performance in case of acoustic
and channel distortions. It is observed that the LSP features
perform equally well to conventional cepstral features on TIMIT
dataset and have showed enhanced identification results on
NTIMIT datasets .
Index Terms- Linear Predictive Cepstral Coefficients
(LPCCs), Mel Frequency Cepstral Coefficients (MFCCs), Line
Spectral Pair (LSP)
I. INTRODUCTION
Speech is a natural source of communication for human
being. Human brain and auditory system analyzes the speech
well for obtaining the information contents from it. In the
similar manner, computer analysis of speech is useful for
various purposes. It helps in observing real time speech
spectra, to get some important characteristics carried by
speech (utterance) as well as some features of the speaker.
Speech and speaker recognition by machines are main two
streams found useful in variety of applications in real world.
When the task is ‘who’ is talking (rather than ‘what’ is said), it
is referred to as ‘speaker recognition’ in general. Depending
on the nature of end decision, it is categorized as speaker
identification and speaker verification. Irrespective of the end
task, the requirement of speech analysis is to extract the
measures of ‘speaker’ characteristics, distinguishing one
person from the other. This procedure is called feature
extraction and is one of the important front end procedures in
speaker recognition techniques. There is growing need of
using speech as a person identity in number of day-to-day
applications. Person authentication prior to admission to a
secure facility or for the transaction over telephone, verifying
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
person’s voice (criminal) for forensic purpose or in audio
conferencing are few examples of the same [1].
Speaker recognition is usually carried out in two steps
namely enrolling (training) and testing (decision making). In
training stage, a set of speaker dependent features are derived
from the input speech, which is useful to form the model of the
particular speaker. In testing phase, the test speaker’s features
are compared with that of the stored database models and end
decision is made using some form of matching algorithm. A
variety of speech features useful for speaker recognition are
studied in past. These features are primarily based two main
factors: i) Physical nature of human speech production organs
ii) The socio-educational background the speaker. The
features based on the first category are named as low-level
(physical) features, whereas the features relating the second
factor are called behavioral (high-level) features [2]. Low
level features derived from shape of vocal tract and source
excitation of human speech production model [3], are easy to
compute and requires small data for analysis and extraction.
In comparison, high-level features are generally related to a
speaker’s learned habits and style (such as particular word
usage or idiolect and long term prosodic), are difficult to
extract and require larger speech data. The shape of vocal
tract and excitation parameters is supposed to be the most
unique characteristics of individual’s speech.
Cepstral based features separates the source (vocal folds)
and system (vocal tract) parameters from the speech signal.
Mel-frequency cepstral coefficients (MFCCs) and Linear
Predictive Cepstral Coefficients (LPCCs) are most common
examples of the same. Studies in [4, 5, 6] showed that the
source or vocal tract features alone are susceptible to additive
noise and channel distortions and gives decreased recognition
performance. Vocal source features such as spectral centroids,
spectral band energy, spectral bandwidth, spectral crest factor
and Shannon entropy are used in combination with MFCC
features for speaker identification on TIMIT dataset [7].
Features like Minimum Variance Distortionless Response
(MVDR) [8] and Mean Hilbert Envelope Coefficients
(MHEC) [9] are proposed for improving the performance of
speaker recognition in noisy conditions. Front end processing
based on autoregressive models followed by modulation
filtering process is proposed for robust feature extraction in
noisy conditions [10].It is desirable that the features used to
represent the speaker specific characteristics must be robust
enough for various forms of added noise and undesired
distortions in the surrounding.
 In this work, we have used the LPC derived features (LSP
and LAR), which are less complicated to extract and basically
used for speech coding and recognition tasks. In the present
paper, we compare the performance of LSP and LAR features
to that of cepstral features for speech from telephone channel.
The paper is organized as follows: LP and LSP representation
is described in section 2 and 3 respectively. Description of the
database used and steps performed in experimental set up is
given in section 4 and 5. Results are discussed in section 6 and
conclusions are given in section 7..
II. LP ANALYSIS
Feature extraction converts the raw speech signal into a
compact and effective representation which is more suitable
to process than the original signal. For speaker recognition,
these features should uniquely represent the speaker specific
characteristics, ignoring linguistic contents. The source-filter
model of speech production is useful for speaker recognition
by investigating the source (vocal cords) and system (vocal
tract) components of the speech signal. The resonant
frequencies of the vocal tract are called formant frequencies
(or formants), which changes with different vocal tract
configurations. The peaks of the spectrum of the vocal tract
response correspond approximately to its formants
[3].Typically, the low frequency formants are phoneme
dependent and high frequency formants are speaker
dependent.
In LP model, knowledge of p samples is used to predict the
nth sample. The predicted value of nth speech sample is given
by:
sˆ(n)   a(k ) s (n  k )
p
(1)
k 1
In equation (1),
a(k ) p
k  1 are the prediction coefficients,
called model parameters.
Thus any signal can be defined by linear predictor and its
prediction error as:
s ( n )    a ( k ) s ( n  k )  e( n )
p
(2)
k 1
The transfer function for pth order all pole filter given by:
G 1 (3)
H ( z)  
A( z ) 1   a z
p
k
k
k 1
where G is the gain scaling factor for the present input and
A(z) is the pth order inverse filter.
III. LINE SPECTRAL PAIR REPRESENTATION
LP coefficients carries some speaker specific vocal tract
information such as resonant frequencies and bandwidth. The
instability of all pole filter (due to change in coefficients) and
poor interpolation, found LPC less useful for speaker
recognition in case of noise and distortions.
Itakura [11] introduced LSP as an alternative kind of LPC
spectral representation, which are having more stable and
efficient coefficients representing the vocal tract parameters.
LSP decomposes the pth order linear predictor A(z) into
symmetrical and anti-symmetrical part denoted by P(z) and
Q(z) respectively. The two polynomials P(z)and Q(z) are
given by the equations:
P( z )  A( z )  z  ( p 1) 1
A( z ) (4)
Q( z )  A( z )  z  ( p 1)
A( z )1
(5)
The linear predictor A(z) can be expressed in terms of
P(z) and Q(z)as:
1 (6)
A( z )  ( P( z )  Q( z ))
2
These roots are located on unit circle (including zeros at
z=1 and z=-1), with all complex zeroes having its conjugate in
z-plane. The LSP parameters are expressed as zeros (or roots)
of P(z) and Q(z) which occur in symmetrical pair, hence
named as Line Spectral Pairs (LSP). The frequencies with
respect to location of pair of complex conjugate zeroes is
called Line Spectral Pair Frequencies (LSF). The inverse
filter A(z) is a minimum phase filter as all its zeros lie inside
the unit circle. LSP collectively describe the two resonance
conditions arising from the interconnected tube model of
vocal tract [12]. P(z) and Q(z) correspond to vocal tract with
glottis closed and open respectively. If we denote the set of
complex roots as θ(k), then line spectral frequencies are
determined from (4) and (5) as:
Re{ (k )} (7)
 (k )  tan (1
)
Im{ (k )}
The advantage of LSP features is that, any power spectral
density (PSD) at particular frequency tends to depend on the
close (nearest) line spectral pair (or line spectral frequency).
In contrast, for cepstral features, the entire power spectrum
gets affected by changes in a particular parameter. This
property of LSP is called localization property, in which any
modification in PSD, can cause local change in LSP or LSF
[13]. As distortions or noise affect on the signal locally, LSP
also gets changed in the spectrum locally.
IV. DATABASE
TIMIT database: TIMIT dataset allows the examination of
speaker identification performance under almost ideal (clean)
conditions [14]. It is a noise-free speech corpus recorded
using high quality microphone sampled at 16 kHz. This
database consists of 630 speakers, with 10 utterances spoken
by each speaker. We have used first 7 utterances and last 3
utterances for testing. 597 speakers are chosen (randomly) out
of 630. So there are 4179 utterances for training and 1791
utterances for testing.
NTIMIT Database: NTIMIT dataset is having the same
speech from the TIMIT corpus, recorded over local and long
distance telephone loops. The speech here is degraded due to
use of carbon button transduction and actual telephone line
conditions [14]. This database consists of 630 speakers, with
10 utterances spoken by each speaker. We have used subset of
110 and 597 speakers out of 630 speakers from both the
datasets. First 7 (total 4179) utterance are used for training
and remaining 3 utterances (total 1791) for testing.

V. SPEAKER IDENTIFICATION EXPERIMENTS

Following are the main stage in the Text independent speaker
identification system build for experimental analysis:

1) Pre-processing- Speech signal have more energy in low

frequency than for high frequencies, called as spectral tilt
[3]. Also high frequency portion (which carries speaker
specific information) of the speech signal is suppressed
during sound production. To boost the high frequency
portion of speech spectrum, pre-emphasis is done ,which
is simple high pass filtering of speech signal. A Fig.1 PIA using TIMIT Database
pre-emphasis factor of 0.97 is mostly used.
2) Feature Extraction- Framing and windowing of the
speech signal is carried out (using hamming window of
size 25 ms and frame overlap of 10 ms). Linear
prediction coefficients are derived using Levinson
Durbin recursion, which are further transformed to Line
spectral pairs (to get LSF). Performance of LSP is
compared with conventional LPCC and MFCC features.
3) Pattern formation and Matching- Model of each speaker
is formed using clustering based LBG algorithm. The
quantization distortion with Euclidean distance measure
is used as matching function. As the goal here is to
analyze the performance at feature level, all other system
parameters are kept static.
4) Identification /Score calculation- For the closed-set, text Fig.2 PIA using NTIMIT Database
independent speaker identification system, identification
accuracy as defined in [15] is given in terms of
percentage of identification accuracy (PIA) as:

(8)

VI. RESULTS AND DISCUSSION

Performance analysis of closed-set, text independent speaker
identification system is carried out using four feature sets
LPC, LPCC, MFCC and LSP respectively on both TIMIT and
NTIMIT dataset. As speech in the TIMIT dataset is clean and
undistorted, all the four features give 100 percent accuracy
(except LPC, having 99.6 percent accuracy). The comparison
is shown in fig.1. Fig.2 shows the performance analysis of
features using NTIMIT dataset for training as well as testing.
As expected, as the speech in this dataset is somewhat
degraded, the percentage accuracy drops as compared to
TIMIT dataset. In this work, all the features are used in their
original (baseline) form, without use of any signal
enhancement or noise removal technique. It is observed (from
fig.2) performance of LSP is as good as MFCC for the subset
of 110 speakers. But the LSP features shows improved
performance for the second subset of 597 speakers. Though
cepstral coefficient features like MFCCs and LPCCs found to
be widely used in speaker recognition applications in clinical
environment, their performance is limited in various real
world conditions. In such cases LSP can be used as
complementary features to cepstral features. The localization
property of LSP is found to useful in enhancing the accuracy
in case of the degraded speech.
Fig.3 Effect of prediction order on speaker identification
The order of prediction coefficients (in case of LPC,LPCC
and LSP) also have significant effect on the correct
identification rate. Fig.3 shows the effect of prediction order
on PAR with LPC, LPCC and LSP feature set respectively. It
is observed that LPC is showing better performance for order
of 8, and the PAR falls down as the prediction order increased
further. For the case of LPCC, there is improvement in PAR
with increase in order, whereas LSP features get less affected
by the change in the order. These results are obtained on a
subset of NTIMIT dataset having 110 speakers.
VII. CONCLUSION
In this work, study and analysis of line spectral pair (LSP)
features is carried out on TIMIT and NTIMIT database. The
performance of LPC features, cepstral features and LSP
features is analogous for clean speech, whereas it drops down
in case of noisy and degraded speech. In case of large
population (597 speakers with 10 utterances), LSP features
showed improved accuracy as compared to other features for
NTIMIT dataset. One of the case of telephone channel
distortions (NTIMIT) is considered in the present work.
Further work can be extended by studying the performance of
the features by adding various artificial noise of different SNR
towards the recognition accuracy of speaker identification
system. Use of speech enhancement techniques or front end
modification may further improve the percentage of correct
speaker identification.

ACKNOLEDEMENT
The authors would like to thank TIFR Mumbai and Dr.
Samudravijaya K. for providing the speech database.

REFERENCES

[1] Douglas O’Shaughnessy, Speech Communications Human and

Machine, University Press, 2001.
[2] Tomi Kinenen, Haizhou Li, An overview of text independent
speaker recognition: From features to supervectors, Speech
Communication, 2009.
[3] Rabiner, L. and Schafer, R., Digital Processing of Speech
Signals. Prentice Hall, Inc., Englewood Cliffs, New Jersey,
1978.
[4] Bojan Kotnik, Damjan Vlaj, Zdravko Kacic, Bogomir Horvat,
Robust MFCC feature extraction procedure using efficient
additive and convolutive noise reduction procedures,
ICSLP'02 Proceedings, pp. 445 - 448, 2002.
[5] Gugnason J., Brookes M., Voice source cepstrum coefficients
for speaker identification , ICASSP 2008, pp.4821-4824.
[6] Tomi Kinenen, Pavo Alku, On separating source and vocal
tract information in telephony speaker verification, ICASSP
2009.
[7] Danoush Hosseinzadeh and Sridhar Krishnan, Combining
vocal source and MFCC features for enhanced speaker
recognition performance using GMMs, MMSP-2007, pp.
365-368.
[8] Seyed Omid Sadjadi and John H L Hansen, Hilbert envelope
based features for robust speaker identification under
reverberant mismatched conditions, ICASSP 2011.
[9] Qin Jin and Thomas Fang Zheng, Overview of front end
features for robust speaker recognition, APSIPA ASC 2011.
[10] Sriram Ganapathy, Sri Harish Mallidi and Hynek Hermanshy,
Robust feature extraction using modulation filtering of
autoregressive models, IEEE/ACM transactions on Audio,
Speech and Language Processing, vol.22, no.8, pp.1285-1295,
August 2014.
[11] Itakura F., Line spectrum representation of linear predictive
coefficients of speech signals, J. Acoustic Society Am. 57, 537
(A), 1975.
[12] T. Backstrom and C.Magi, Properties of line spectral pair
polynomials: AReview, Signal Process.,vol.86, no.11, pp.
3286-3298, 2006.
[13] K.K. Paliwal, On the use of line spectral frequency parameters
for speech recognition, Digital signal processing, 2,
80-87,1992.
[14] Douglas A Raynolds, Automatic speaker recognition using
Gaussian mixture speaker models,The LINCON Laboratory
Journal, vol.8 no.2, PP.173-192, 1995.
[15] Raynold D. and Rose R. , Robust text independent speaker
identification using Gaussian mixture speaker models, IEEE
Trans. On Speech and Audio Processing 3 , January 1995.
 Analysing Road Accident Data Using Association
Rule Mining
(Analysing RTA data using ARM)
Sachin Kumar
Centre for Transportation Systems (CTRANS)
Indian Institute of Technology Roorkee
Uttarakhand, India
sachinagnihotri16@gmail.com
Abstract—Road accident is one of the crucial areas of
research in India. A variety of research has been done on data
collected through police records covering a limited portion of
highways. The analysis of such data can only reveal information
regarding that portion only; but accidents are scattered not only
on highways but also on local roads. A different source of road
accident data in India is Emergency Management research
Institute (EMRI) which serves and keeps track of every accident
record on every type of road and cover information of entire
State’s road accidents. In this paper, we have used data mining
techniques to analyze the data provided by EMRI in which we
first cluster the accident data and further association rule mining
technique is applied to identify circumstances in which an
accident may occur for each cluster. The results can be utilized to
put some accident prevention efforts in the areas identified for
different categories of accidents to overcome the number of
accidents.
Keywords—Data Mining; Road Accidents; Accident Analysis;
Association Rule Mining.
I. INTRODUCTION
Road accident is one of the undesirable events that are
uncertain and unpredictable. Road accident is one of the major
causes of unnatural deaths, disability and property damage. In
a report by MORTH, 2014 [1], it is stated that 0.4 million
accidents are reported each year in India that makes India as
one of the countries with larger accident rate. The report stated
that there is a minor decrease in accident rate from year 2012
to 2013 but it does not guarantee that this decrease will exist
for future years also. However, fruitful research is required in
order to identify circumstances of accident occurrence.
In India, when an accident took place, it is recorded by
the concerned police officer of that region’s police station.
Police stations only cover the accidents which have happened
in their territories. Raj V. P. [2] discussed that in India, the
method used for collecting, compiling and recording the
accident data needs lot of improvements. He states that the
reports which are prepared at accident sites are very basic and
non-analytical for research purpose. Various researches [3-5]
have been done on this data using statistical techniques such
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Durga Toshniwal
Computer Science & Engineering Department
Indian Institute of Technology Roorkee
Uttarakhand, India
durgatoshniwal@gmail.com
as Poisson models, Negative Binomial models etc. As
previously mentioned that this data was collected from police
stations, it contains accident records that happened on limited
portion of highways under their territories. Hence their results
only explore the characteristics of accidents on some portions
of highways.
However, authors from other countries have used quality
data for accidents and analyzed those using statistical
techniques. Poisson models [6-7] and negative binomial
(NB) models [8-10] have been used extensively to
identify the relationship between traffic accidents and the
causative factors. It has been widely recognized that Poisson
models outperform the standard regression models in handling
the nonnegative, random and discrete features of crash counts
[11-12].
Although Poisson models perform better, the constraint of
the mean being equal to the variance in Poisson models is
often violated by over-dispersed accident data. As an
alternative, Negative Binomial (NM) models can be used to
accommodate this over-dispersion by incorporating an
additional independently distributed error term. However, with
the assumption of independent observations, both the Poisson
models and the NB models not suitable to handle the
heterogeneous nature of road accident data.
There are various regression analysis based models i.e.
linear regression models, negative binomial regression models
and Poisson regression models which are the popular methods
in road accident data analysis as these methods can find the
factors associated with a road accident. Such information can
be very useful for the purpose of preventing road accidents by
taking preventive measure on the locations of accident
occurrence. Although these techniques are found useful in
analyzing road accident data and identifying associated
factors, [13] finds in his study that using these techniques in
road accident data analysis may lead to some problems if the
accident data have higher dimensions. These problems are
exponential growth in the parameters with respect to growth in
the number of variables and invalid results of statistical tests
because of sparse data in contingency tables. One of the basic
requirements of regression models is that they have certain
model specific assumptions and certain relationships between II. DATA SET DESCRIPTION
independent and dependent variables which when not The data for this research has been obtained from GVK-
followed may lead to errors and provides wrong outcomes EMRI, Dehradun. The data set contains 15574 road accidents
[14]. Hence, a different technique for analyzing road accidents records from Jan 2009 – Dec 2014 that occurred in Dehradun,
data is required which can handle the heterogeneous nature of Uttarakhand. Previous researches in India on accident analysis
road accidents and provide outcomes that can lead to some have been done on the data collected from police stations that
solutions to prevent accidents rate. cover a limited portion of some highways. The attributes of
Data mining is a promising area that can be defined as a the data are mostly categorical in nature. A brief description of
set of techniques such as clustering, classification, association the data is given in the Table1. Data transformation has been
rule mining etc. which can be used to analyze large amount of applied to some attributes such as victim age has been
data and extract hidden patterns. Data mining is widely used in converted to nominal variable and assigned nominal values
several studies related to transportation systems such as road such as child for age group (1-12), teen(13-19), Young(20-30),
accident data analysis, road pavement data analysis, road Adult(30-45), senior (45+) to strengthen the quality of data.
roughness data analysis etc. Similarly, other attributes have been transformed into
In India, the technology used for data collection related to appropriate form for better analysis of the data.
transportation systems is not up to mark as compared to other
countries i.e. UK, US and European countries. Various studies
Table 1: Road Accident Attributes
in India [3-4] shows that accident data for analysis were
collected on manual basis and hence this data is for short Attribute Name Type Values
duration and use of data mining techniques cannot extract
Victims injured Ratio 1, 2,>2
much information from this data as data mining requires large
data to extract meaningful results. The analysis of road Children, Teen, Young,
Victim Age Nominal
accident data collected from police station records which Adult, Senior
covers the limited number of accidents that had occurred in Victim Gender Binary Male/Female
the territories of the concerned police stations. The studies Victim economic White card, Pink card,
from other countries have used data mining to analyze road Nominal
status Yellow card
accident data and extract useful results [14-18].
In India, there is a need for different source of accident Morning, Forenoon,
Time of accident Nominal
data so that we can make use of data mining. 108 is an Afternoon, Evening, Night
emergency number that handle all kind of emergency Day Nominal Weekday, Weekend
situations such as pregnancy related emergency, heart stroke,
road accidents and many more. This emergency ambulance Winter, Summer, Rainy,
Month Nominal
service has saved more than one million lives with over Spring
22,000 emergencies served per day. This service is running in Region Nominal 1,……,120
several states of the country and keeps track of record of every
emergency served statewide. The main advantage of EMRI Lighting on road Binary Yes/No
service data is that it covers road accidents from every region
of the district and hence it provides a wide coverage of road Roadway Feature Nominal Intersection, Bend, Slope
accident locations.
Accident
As this data is not publically available, no research work Nominal Critical/Non-critical
has been reported yet on this data. The quality of any research Severity
is primarily depends on the quality of data set used. In this Hospital, Colony, Market,
paper, we have analyzed the road accident data to identify the Area Around Nominal Forest, Hills, Vegetation
correlation between different set of attributes in the data which land
will help to understand different circumstances for different
type of accidents. Our approach mainly uses clustering to Road Type Nominal Highway/Local
divide the data into different groups and further applying the
association rule mining technique to get the descriptive
analysis each cluster. Further, its results can be used to
overcome the accident rate by improving traffic policies and III. METHODOLGY
ensuring traffic safety on roads.
The organization of this paper is given as follows. In A. Data Preprocessing
Section2, a brief description of the data set used s given. In Data preprocessing [19] is the primary task to be done
Section3, methodology used to analyze the road accident data prior to analysis to get the data ready for analysis. As good
is discussed. Section 4 presents the results and discussion data can only provide good results, data preprocessing
section and finally we concluded in section5 with some becomes necessary prior to analysis. In this part of our study,
suggestions. we removed those tuples from the data which were having
multiple values missing. As this data is a real world data,
manual filling information can change the results. Data
transformation has been applied on some attributes such as
age, traffic density so that they can be interpreted easily. After
preprocessing of data set, 9640 road accidents are found
suitable for further analysis.
B. Clustering Algorithm
There is a variety of clustering algorithms [20] such as
partition based clustering and density based clustering that can
be used for clustering of datasets. The clustering algorithm is a
type of unsupervised learning algorithm which aims to group
the data into clusters or groups. The objects in a one cluster
are more similar to each other and objects in different clusters
are dissimilar to each other. In our road accident data set, we
have more number of categorical attributes; hence we need an
algorithm that is best suited for categorical data. K-modes
algorithm is a clustering algorithm that can be used for
categorical data. The K-modes algorithm uses a distance
function to find the similarity among objects which is defined
below:
Given a data set D, the distance between two objects X &
Y, where X & Y are described by N categorical variables, can
be computed as
d (X, Y) = ∑(δ(Xi, Yi)) for i=1 to n (1)
Where,
0, Xi = Yi
δ(Xi, Yi) =
(2)
1, Xi ≠ Yi
In the above equations, Xi and Yi are the values of object
X and Y for attribute i. This distance measure is often referred
as simple matching dissimilarity measure. Here, we are
providing a brief description of the K-modes clustering
algorithm.
K-mode clustering procedure:
In order to cluster the data set D into k cluster, k-modes
clustering algorithm perform the following steps:
1. Initially select k random objects as cluster centers or
modes
2. Find the distance between every object and the
cluster center using distance measure defined in
equation1
3. Assign each object to the cluster whose distance with
the object is minimum
4. Select a new center or mode for every cluster and
compare it with the previous value of center or mode;
if the values are different, continue with step 2.
C. Association Rule Mining
Association rule mining [20] is a very popular data mining
technique that extracts interesting correlations among different
set of attributes in the data. This technique is mainly based on
market basket analysis of market data whose focus is to
identify the goods which a customer will prefer to buy
together and more frequently. Association rule mining
generates different set of rules that define the correlation
among different set of attributes in the data set. Various
interesting measures such as support and confidence are there
which identifies the strong rules. In general, a support value
indicates the frequency of occurrence of a rule in the data set.
On the other hand, confidence defines the reliability of a rule.
Hence, association rules with high confidence and support
value are of main interest.
Given a data set D of n transactions where each transaction
TЄ D. Let I= {I1,I2,..In} is a set of items. An item set A will
occur in T if and only if A⊆T. AB is and association rule,
provided that A⊂I, B⊂I and A∩B= Ø. In Association rule
mining, support, confidence and lift metrics are well known
interesting measure which are used in choosing strong
association rules.
IV. RESULTS AND DISCUSSION
A. Cluster Analysis
R-statistical software was used for k-mode clustering. We
used Akaike Information Criteria (AIC) [21], Bayesian
Information Criterion (BIC) [22] and Consistent AIC (CAIC)
[23] to identify the number of clusters in the data set. A total
of 15 models were generated from 1 cluster to 15 clusters.
Figure 1 illustrates the evolution of BIC, AIC and CAIC for
all the 15 models generated. We can see a fall in the values of
AIC, BIC and CAIC with respect to an increase in the number
of clusters. Based on the Figure 1, we select the model with 6
clusters as there is no improvement shown in the values of
AIC, BIC and CAIC. K-modes clustering algorithm was
applied to obtain the six clusters. These six clusters obtained
by K-modes algorithm is described in following subsection.
B. Cluster Description
Cluster 1 involves 69% of two wheeler accidents happened
in populated areas such as markets, hospitals, local colonies
across highways and local roads. The road features associated
with these two wheeler accidents were intersections and
bends. Two wheeler accidents which occurred on intersections
and bends on highways involved one injured only. Two
wheeler accidents at local roads are mostly involved two or
more injured victims.
Cluster 2 consists of two wheeler accidents that occurred
less populated areas such highways that go through a hill area,
forest area or vegetation area. Accidents with two victims are
mainly involved in this cluster.

Figure 1: Cluster selection criteria
Cluster 3 consists of all accidents in which a vehicle falls
down from height. Most of these accident cases were critical
where area around was hill and involves more than two
victims injured. The main road feature responsible for these
accidents was blind turn on road. The main duration of these
accidents were morning time of 4:00 am to 6:00 am on hilly
roads and 8:00 pm to 4:00 am in other roads.
Cluster 4 consists of accidents involving multiple vehicle
accidents and divider/fixed object hit cases. These accidents
are mostly happened in night time around 11:00 pm to 4:00am
which are critical accidents whereas accidents at other time are
non-critical. These accidents have same tendency on both
highway and local roads. Intersections are the main places
where these accidents have occurred.
Cluster 5 involves pedestrian hit accidents. Local roads
involve more pedestrian hit accidents than highway roads.
Mostly these pedestrian hit accidents have occured in market,
colony, near hospitals, and other populated areas. Pedestrian
hit accidents at night time were critical whereas at day time
these accidents have non-critical accidents.
Cluster 6 consists of the accidents involving vehicle roll-
over cases. Vehicle roll-over cases were found at bends and
slopes on highways. It has been observed that 80% of these
accidents have happened on the forest and vegetation areas on
local roads.
All these clusters were further analyzed using association
rule mining to find the correlation among different attributes
in the data. Each cluster obtained was analyzed using
association rule mining algorithm i.e. Apriori algorithm using
WEKA 3.6 tool. For this purpose, a minimum support of 30%
was taken as threshold to identify strong rules. Although
various rules are generated for every cluster only certain rules
with strong confidence and lift value were taken into
consideration. The information extracted from these rules is
discussed in the following subsection.
C. Description of Association Rules
Apriori algorithm [24] has been applied on every cluster
using WEKA3.6 to generate association rules. Although
several association rules are generated for each cluster type,
we are summarizing the information conveyed by association
rules for each cluster.
1. Association rules for cluster1
The rules show that two wheeler accidents are mainly
occurs on specific road segments like intersections on
community areas like colony, markets. Intersections in
colonies near highways are more prone to two wheeler
accidents than colonies on local roads. Also market areas are
more likely to have two wheeler accidents with two or more
injured victims. For example, a rule {Road_type = “highway”
AND Road_feature = “intersection” AND Area_around =
“colony”}  {Victim_injured=”1”} specifies that in cluster1
if an accident occurs at intersections in colonies near highways
with two wheeler accidents; then the chances for 1 victim
injury is high. Another rule {Road_type = “local” AND
Road_feature = “intersection” AND Area_around = “market”}
 {Victim_injured=”>2”} indicates that if an accident occurs
at intersections in markets on local roads then that accident
will involve more than 2 victims injured.
2. Association rules for cluster2
The rules indicate that forest area and vegetation area that
are aside of certain highways are dangerous for two wheeler
accidents as sudden bend, slope at night time can cause
imbalance of driver and may cause accidents. Association rule
{Road_type =”highway” AND Area_around = ”vegetation”
AND Time =”Night”}  {Accident_severity =”critical”}
suggests that highways that goes through vegetation land at
night time are dangerous for two wheeler accidents and most
these accidents are critical.
3. Association rule for cluster 3
The rules show that most of the vehicle-fall from height
accidents involved more than 2 injured. It is found that
vehicles falling from height on hilly highways are severe
accident where more than two injured persons are there. The
reason might be the vehicle type is four-wheeler or similar
category which transports more than 2 persons at a time. Also,
it shows that mostly vehicles fall from height from hill
location are due to a bend on road.A rule {{Road_type
=”highway” AND Area_around =”hill” AND Road_feature
=”bend AND Victim_injured=”>2”}  {Accident_severity
=”critical”} indicates that an accident at bend on hilly
highways with more than two victims injured are critical
accidents. It indicates that vehicles that can transport more
than two persons such as car, taxi and bus are more prone to
fall height accidents at bends on hilly highways. So the drivers
of such vehicles should drive carefully on hilly highways to
prevent accidents as these accidents are critical accidents in
these regions.
4. Association rules for cluster 4
The rules indicate that multi-vehicular and fixed
object/divider hit accidents are mostly occurred at night time
on highway roads. Intersections on highways are another
cause for such type of accidents which are very difficult to
observe at night duration. A rule {Road_feature =
“intersection” AND Time = “night” AND Accident_severity =
”critical”}  {Victim_injured = ”>2”} implies that accidents
in this clusters involves more than two victims injured if
accidents happens in the night at intersections are critical
accidents. Rules for this clusters indicates that accidents have
same tendency in local as well as highway roads. So one
should carefully drives at nights and be careful at intersections
otherwise the critical accident can occur.
5. Association rule for cluster 5
The rules show that local colonies and markets on local
roads are the major places of pedestrian hit accidents. A rule
{Road_type=”local” AND Area_around=”market” AND
Victim_injured=”1”}  {Age=”Young”} indicates that if a
pedestrian hit accidents in market on local roads involves one
injury then the age of the victim involved in the accident will
be young. Another rule {Road_type=”local” AND
Area_around=”colony” AND Victim_injured= “2”} 
{Age=”child”} indicates that if an accident in colony on local
roads involves two victims injures then the age group involved
in the accident will be child group. The rules for this cluster
suggest that markets, colony and hospital area are the major
places where pedestrian hit cases have occurred. Hence
pedestrian should be careful in these locations.
6. Association rules for cluster 6
The rules indicate that vehicle roll-over accidents are
occurred at night in forest area and roads near vegetation
areas. A slope in forest road and a bend on road is the road
feature involved in these accidents. A forest road is more
prone to vehicle roll-over accidents in night time. The rule
{Area_around = ”forest” AND Time = ”night”} 
{Accident_severity=”critical”} indicates that roll-over
accidents at night time in forest area is critical accidents.
Another rule {Area_around = ”vegetation” AND
Road_feature = ”bend”}  {Victim_injured = “>2”} indicates
that bend on roads near vegetation areas are dangerous for
vehicle roll-over accidents and these accidents involves more
than two victims injured in accidents.
V. CONCLUSION AND SUGGESTIONS
In this paper, we used association rule mining to analyze
accident patterns for different types of accidents on the road
network of Dehradun district in Uttarakhand state of India.
Our approach used 9640 accidents records after preprocessing
that have occurred on Dehradun district road network during
2009 to 2014. These accident records are based on the
information collected by 108 ambulance service running in
Uttarakhand state that contains about 17 variables from which
13 variables were found suitable for analysis. We used K-
modes clustering that identified 6 clusters which are mainly
based on accident type. Further, association rule mining
technique is applied on these clusters to identify some
interesting rules that can help in understanding the
circumstances of accidents in different clusters. Although,
certain important rules were identified but more interesting
rules can be found if more information is available with the
data. As previously mentioned, in a country like India, data
collection process for traffic and road accidents are not as per
mark. This data provides information that was not available
with the previous data collection process by police officers.
We understand that 108 service is a lifesaving system
rather an accident data collection system. More staff can be
provided by government to collect more information like
speed of the vehicle at accident time, weather information,
road structure information and etc. in order to enhance the
research on road accidents in India, so that researchers can
find much better results that can help in preventing accidents
in India. The results can be utilized by the concern officers of
traffic and road safety department of India to put some
accident prevention efforts in the areas identified for different
categories of accidents to overcome the accidents.
ACKNOWLEDGMENT
The authors acknowledge the GVK-EMRI Dehradun to
provide the data for our research work.
REFERENCES
[1] MORTH, 2014. “Road Accidents in India 2013”, New Delhi: Ministry
of Road Transport and Highways Transport Research Wing,
Government of India. August 2014.
[2] R. V. Ponnaluri, “ Road traffic crashes and risk groups in India: analysis,
interpretations, and prevention strategies, IATSS Research, 35(2), 2012.
[3] M. Parida, S. S. Jain, and C. N. Kumar, “Road traffic crash prediction
on national highways” Indian Highways, Indian Road Congress, 40(6),
2012.
[4] C. N. Kumar, M. Parida, and S. S. Jain, “Poisson family regression
techniques for prediction of crash counts using Bayesian inference”
Procedia-Social and Behavioral Sciences, 104, 2013.
[5] R. Bandyopadhyaya, and S. Mitra, “Modelling Severity Level in Multi-
vehicle Collision on Indian Highways”, Procedia – Social and
Behavioral Sciences, 104, 1011-1019, 2013.
[6] B. Jones, L. Janssen, and F. Mannering, “Analysis of the Frequency and
Duration of Freeway Accidents in Seattle”, Accident Analysis and
Prevention, Elsevier, vol. 23, 1991.
[7] S. P. Miaou, and H. Lum, “Modeling Vehicle Accidents and Highway
Geometric Design Relationships”, Accident Analysis and Prevention,
Elsevier, vol. 25, 1993.
[8] S. P. Miaou, “The Relationship between Truck Accidents and
Geometric Design of Road Sections–Poisson versus Negative
Binomial Regressions”, Accident Analysis and Prevention, Elsevier,
vol. 26, 1994.
[9] M. Poch, and F. Mannering, “Negative Binomial Analysis of
Intersection-Accident Frequencies”, Journal of Transportation
Engineering, vol. 122, 1996.
[10] M. A. Abdel-Aty, and A. E. Radwan, “Modeling Traffic Accident
Occurrence and Involvement”, Accident Analysis and Prevention,
Elsevier, vol. 32, 2000.
[11] S. C. Joshua, and N. J., Garber, “Estimating Truck Accident Rate and
Involvements using Linear and Poisson Regression Models”,
Transportation Planning and Technology, vol. 15, 1990.
[12] M. J. Maher, and I. Summersgill, “A Comprehensive Methodology for
the Fitting of Predictive Accident Models”, Accident Analysis and
Prevention, Elsevier, vol. 28, 1996.
[13] W. Chen, and P. Jovanis, “Method of Identifying Factors Contributing to
Driver-Injury Severity in Traffic Crashes”, Transportation Research
Record-1717, 2002.
[14] L. Y. Chang, and W. C. Chen, “Data Mining of Tree based Models to
Analyze Freeway Accident Frequency”, Journal of Safety Research,
Elsevier, vol. 36, 2005.
[15] J. Abellan, G. Lopez, and J. Ona, “Analyis of Traffic Accident
Severity using Decision Rules via Decision Trees”, Expert System with
Applications, Elsevier, vol. 40, 2013.
[16] B. Depaire, G. Wets, and K. Vanhoof, “Traffic Accident Segmentation
by means of Latent Class Clustering”, Accident Analysis and
Prevention, Elsevier, vol. 40, 2008.
[17] V. Rovsek, M. Batista, and B. Bogunovic, “Identifying the Key Risk
Factors of Traffic Accident Injury Severity on Slovenian roads using a
Non-parametric Classification Tree”, Transport, Taylor and Francis.
2014.
[18] T. Kashani, A. S. Mohaymany, and A. Rajbari, “A Data Mining
Approach to Identify Key Factors of Traffic Injury Severity”, Promet-
Traffic & Transportation, vol. 23, 2011.
[19] P. N. Tan, M. Steinbach, and V. Kumar, “Introduction to data mining”.
Pearson Addison-Wesley, 2006.
[20] J. Han, and M. Kamber, “Data Mining: Concepts and Techniques”.
Morgan Kaufmann Publishers, USA, 2001.
[21] H. Akaike, “Factor analysis and AIC”. Psychome 52, 317–332, 1987.
[22] A. E. Raftery, “A note on Bayes factors for log-linear contingency table
models with vague prior information”. Journal of the Royal Statistical
Society, Series B 48, 249–250, 1986.
[23] C. Fraley, A. E. Raftery, “How many clusters? Which clustering
method? Answers via model-based cluster analysis”, The Computer
Journal 41, 578–588, 1998.
[24] R. Agrawal, R. Srikant, “Fast Algorithms for Mining Association Rules
in Large Databases”, Proceedings of the 20th International Conference
on Very Large Data Bases, p. 487-499, 1994.
 Distribution Methods of Web Pages and Testing
Neelaksh Sheel
Uttarakhand Technical University
Dehradun, India
sheelneelaksh@gmail.com
Abstract—Publishing the contents through internet we need
a medium i.e Web browser. Hyper Text Markup Language is
very helpful for writing the web contents and by the help of web
browsers we publish our page on net. It is invented as an
application so that people can take advantage to view immobile
web pages in succession. Technologies improves and the web sites
gain more advance features and have dynamic web applications
with the exchanging of contents from other websites, similarly
browsers features have turn out to be multi-principal operating
surroundings with shared resources with commonly trustable
web site principals. Internet Explorer, Google Chrome, Mozila
Firefox and many more have a multiple key principal operating
system assembly that gives a facility of web browser. The
constrained control to supervise the protection of all machine
resources for the collected web page is based on the defined rules.
An intranet is not a public network with the surfing limitation to
one organization. It uses the same technical principles and
procedure as the international public internet. Internet is the
most apparent opportunity for publishing the web pages but we
may need to limit the sharing of our web pages to a local internet
inside the group, instead of making this web page availability for
public. We can share out our web pages on local memory of
personal computer or server. This paper covers all the publishing
options and offers guidance for designing web page and the
distribution methods. It also covers the basics testing methods of
published web pages.
Keywords— Web Browsers; Web Pages; Web Server; Web
Page; Testing
I. INTRODUCTION
Publishing the web page through web browsers have in
progress to be a multiple primary operational place where key
is a web site [11]. Similarly multi-functional operating system,
modern proposals [2, 3, 6, 11, 12] and browsers like Internet
Explorer-8,9 [8] and Firefox 3rd version [4] supports
generalized concept for two way communication e.g. Post
Message and protection of frames for web developers. On the
other hand no existing browsers, including new architectures
like internet explorer, Google Chrome and Opera Mini have a
multi-functional operating system creation that gives a
browser-based operating system called Browser Kernel. The
restricted control manage the protection and sharing of system
resources in all respect with browser principles [5,7,9].
Kernel of browser directly interacts with the elementary part
of operating system and depicts a set of structure entitled for
principal of browser. We depict the separation boundary
across the distinct browser rules defined by the foundation
policy [1, 10] like protocol, domain-name and port using the
sandboxed operating system processes. Due to sandboxed
978-1-4673-9354-6/15/$31.00 ©2015 IEEE
Kunwar Singh Vaisla
Bipin Tripathi Kumaon Institute of Technology
Dwarahat, India
vaislaks@gmail.com
processes are unable to work together with the primary
method and have to use system calls grant by the Kernel of
browser. Primary process links with the other processes will
communicate with each other using inter-process
communications with the support of browser kernel. Other
than Opera Mini, Kernel of browsers offers the same shield to
hook up contents as to the standard network content. These
types of multi-principal operating system creation for a
browser fetch considerable security and dependability benefits
to the whole browser scheme: the negotiation or failure of
rules affects the Browser Kernel.
A group of instructions for retrieving, traversing and
presenting the information resources on the web is a computer
program known as web browser. Source of information is
acknowledged by Uniform Resource Locator or Uniform
Resource Identifiers. Source of information can be a shared
web page, videos, images or any other type of content. Hyper-
linking in resource provides the facility to the users to
navigate the used browsers to relate with actual resources.
Web browsers include Nexus (1990, first web browser by Tim
Burners-Lee), Erwise (1st web browser with graphical user
interface), Mosaic (1993, Marc Andreessen), Netscape
Navigator (1994), Microsoft Internet Explorer (1995), Opera
(1996), Mozilla Firefox (1998, later 2011), Apple’s Safari
(January-2003), Google Chrome (September 2008), and
Internet Explorer 8 (December 2011) [13].
Web browsers Netscape Navigator and Mosaic are simple
software applications that provide HTML code sustain
bookmarks and processed input. As website creates, so it
fulfils the requests of web browser. Now a day’s browsers are
more advance, supports more than one type of web codes like
PHP, XHTML, HTML 5 and active JavaScript. Secure
websites use encryption in the code. Advanced web browsers
permit the web developers to developed well designed
interactive websites e.g. Ajax facilitate browsers to robustly
upload updated information on the web without reload the
contents of webpage. Advancement in CSS permits the
browser to display the visual effects in responsive website or
web design. Cookies in the browser allow to remembering
your settings for specific webpage in a website. Sometimes
website works very well in one browser, but would not
function in another browser efficiently. Therefore, it is elegant
to install number of multiple browsers to seeing the best
output of your page. Even browsers are mostly intended to
utilize the World Wide Web, and have the authority to access
the information to be made available by our selected web
servers in local area networks [14].
 II. WEB PAGES
A. Setting of internet web site / web pages
After creating the HTML page we require to upload it on web
server (where computer have internet connection and have the
permission to send web contents on demand) for publishing it
on World Wide Web. If we are running individual web server,
then it’s just a matter of copying the files in suitable folder
available in computer drive. Most of the developers outsource
the web server to the professional Internet Service Provider
(ISP) for hosting their web site. All the ISP’s allocate space on
the server to publish personal web site after paying extra cost,
whenever we pay some extra amount then ISP helps us to gain
the attraction on web pages by including a number of
multimedia files.
Geocities (www.geocities.com), Tripod (www.tripod.com),
and Angelfire (www.angelfire.com) are very popular for
publishing the web contents, with the condition that they
incorporate advertisement of their choosing on the hosted web
pages. Our web site web address will also contain the name of
ISP (host).
http://domain.registration.com is a web site on which we can
check the uniqueness of our domain name, If the selected
name that isn’t already registered then permit the internet
service provider to upload this unregistered name to registered
on the web [15].
B. To upload web pages on web server
To publish the web pages through internet, web server uses
specific internet protocols for information exchange called
Hyper Text Transfer Protocol (HTTP). We need the software
to upload our webpage like Filezilla that uses the
communication standard File Transfer Protocol (FTP). The
FTP is a benchmark that uses to transfer the file on the web
server. Firstly server adjust these files and then broadcast
these HTML files for everyone who request for these using the
hyper Text Transfer Protocol (HTTP) with URL [15,16].
Web hosting service (ISP) company requires four significant
piece of information before putting web pages on our web site:
 Account Name ( user name or user ID)
 password
 The FTP address for web site.
 The directory folder where our web page files store.
Sometimes place these files in root directory, but
regularly we use subdirectory named www for the
community, or domain name.
Next we want do decide which best software with best facility
can upload web pages to the web server and maintain it [15],
like Netscape Navigator, Microsoft Internet Explorer,
Microsoft front Page or CuteFTP etc.
III. PUBLISHING WEB PAGES
A. Publishing pages with Netscape Navigator
Netscape Navigator web browser follows the steps to upload
the web pages are:
1. First of all put the path of web directory in the browser
Netscape Navigator’s locality box eg.
ftp://myname:mypassword@siteaddress.net/home/web/w
herever/, Now enter the account name in user name text
box and your personal password in password txtt box for
admittance the website instead of myname and
mypassword, FTP address for web site instead of
siteaddress.net, and the file name where web pages
reside as an alternative of /home/web/wherever/ [17].
2. Select the files and then upload.
3. Select the file you want to upload and click to open it.
Wait for transfer the file data uploading.
4. Now test the web page by double clicking the HTML
uploaded file on the FTP directory in Netscape Navigator
Window, if it works properly [23] then our contents will
be on the web.
Fig. 1. Starting of new blank web page [22]
In Netscape Navigator Composer Toolbar follow:
1. Select Blank Page to start from FILE menu “Fig. 1”.
2. Select web page from Template to attach with Netscape's
navigation server and choose template for transform data
on the web. Now click on the File drop down menu and
select Edit Page option.
3. Select a Web Page from magic box to hook up
Netscape's navigation server and respond some simple
questions to create and finalize the web page.
B. Description to publish the web pages
To publish the web pages that you have formed in Netscape
navigator Composer, click on "Publish" option given in the
toolbar. The dialog box will open for publishing the web page
“Fig. 2”.
The FTP location and user name will already be shown as we
have already configuration earlier. Gives the password or we
can leave it blank. For this blank password Netscape navigator
will enquire later. If we are uploading a new web page, then
the web page title and HTML Filename boxes can be put as
blank. Generally HTML Filename for "home page" would be
index.html. The web page title should have the name whatever
we want -- this title of web page will come into view in the
title bar of any browser window at the time of webpage
retrieval [16, 17, 18].

Fig. 2. Publishing of web page
After filling all the information click OK button and Netscape
navigator will initiate to send your web page information to
the defined web server. Successful uploading of data will
permit us to inspect our web page in any ordinary browser and
available this web page for public on internet who so ever will
knows the proper URL of our web site.
Once the web page has been available, now we can carry on to
make changes in the Composer window of Netscape navigator.
For visualize the alteration in effective manner upload the web
page again and again after saving the changes in web page file.
Fig. 3. Page properties of Netscape Navigator
If come back to edit web page new version afterwards, then
same procedure will be followed by the user to upload the
page. In fact we can do this [18] with any web page that would
feel to upload from initial point “Fig. 3”.
C. Publishing web page with Microsoft Internet Explorer
Microsoft internet explorer is the most commonly used
browser for the internet user. Microsoft Explorer file manager
comes with the every operating system launched by Microsoft
(every version). When we go for FTP address in the address
bar of internet explorer, explorer have the option to cut, paste,
rename and delete the file or directory subfolders and folders
from the web server as if it is available on the same hard disk.
Follow these instructions to upload the web page you have
created and save in memory, so that it becomes visible on web.
1. Click on Microsoft internet explorer. Type proper path
where the page is currently located and press enter.
Always keep in mind that you have to type the path
directly into the address bar.
2. Double Click the file which we want to upload on the
web. Select more than one file with holding ctrl key.
Must sure to include graphics file that need to go with the
HTML file (if any) [15].
3. Choose edit and then copy.
4. Now as the format opens type the user name and
password with web site address in the address bar e.g.
ftp://myname:mypassword@siteaddress.net/home/web/w
herever/ Type the account name and password for
seeing the web site on admin side. Internet and the
directory where the web pages go on the server as a
substitute of /home/web/wherever and then press enter
key.
5. Select edit and then paste. The web file will relocate from
your secondary memory to the web server.
D. Publishing Web Page with Microsoft front page
More sophisticated program such as Microsoft Front Page
2000 turns to publish entire web site in single step. It’s an
automated process. Either we are having one or N number of
web pages, put them all online by selecting the specified file,
and publishes on the web. Firstly specify the location to
publish your web page, fill the FTP address of web hosting
service.
Again enter the account name and password. Front page
automatically uploads all the pages linked together, that we
create or last update. Except Microsoft Front Page,
Macromedia Dream Weaver, Adobe pageMill, and NetObjects
Fusion also recommend automated uploading of web pages
and web site management features “Fig. 4”.
E. Publishing a Microsoft FrontPage web site to an A2
Hosting Server
After creating a web site in FrontPage, use FTP to publish it to
the A2 Hosting server associated with account name.
FrontPage extensions are proprietary and only work with
Microsoft IIS (Internet Information Server). A2 Hosting
servers do not support FrontPage extensions. To publish a
FrontPage web site to an A2 Hosting server, follow these steps
[19, 20]:
1. Click on FrontPage and open web site project files/folder.
2. Click on Publish Site given in file menu. Isolated web site
properties dialog box appears.
3. In remote web server type domain name and click on FTP.
4. On Remote Web site locality text box, type
ftp://ftp.example.com, replace example.com with your
described domain name.
5. In the FTP directory text box, type public_html. The
public_html path assumes to upload files to the web site’s
document root directory. If subdirectory available then add
it to the end of the path.
6. Select the Use Passive FTP check box.
7. Click OK. FrontPage asks for username and password.
8. In the Name text box, type A2 Hosting account (cPanel)
username. In caswe of forgotten the cPanel username or
password, use the A2 Hosting Customer Portal to view it.
9. In the Password text box, type A2 Hosting account (cPanel)
password.
10. Press OK. FrontPage displays files on your local computer
as well as on the isolated remote web server.
11. To publish all revised web pages, click Local to remote
web server.
12. Click on Publish Web site. After FrontPage finishes
uploading the files to your A2 Hosting account, you can
view them in any web browser.
Fig. 4. Web Page with MS front Page [23]
F. Publishing web pages with cuteFTP:
“Fig. 5,” shows the most popular cuteFTP program for
windows freely available on different websites to download.
Transferring contents to a web server have following steps.
1. Makes the awareness of cuteFTP utilization and address
of web page with account name and password. Select a
group for web site in cuteFTP website Manager Window,
click on add website to access and edit host dialog box.
Fig. 5. CuteFTP default dialog box
2. Procedure to fill in each of the item:
(i) Label in website indicate name which will use to
denote our own website. No one can see this label
name, so we can enter any label name on this place.
(ii) Host address will be the FTP address. Web server use
this FTP address to send our web page contents on
the web. This will generally start FTP. The ISP
communicates the correct host address to the web
server.
(iii) Internet service providers that maintain the web
server generate user ID and password. Be conscious
that cuteFTP can remember password, it means
anyone other than administrator can modify our web
pages.
(iv) Unidentified user can change setting for downloading
the web files from public FTP services. There will be
no need of user ID’s or password at this moment. So
also generate the login type for security reasons.
(v) Auto detection should be configured. It will
automatically send HTML files and other contents
files to the server using some other protocol. Images
and other non text materials should be completely
compatible with all other types of computers.
(vi) Host type should also be set to Auto-detect, unless
you have trouble in connection to the server. In that
case, you need to find out what type of computers
you are connecting to and pick the corresponding
Host Type.
(vii) For initiating the remote directory, enter the name of
primary directory and folder on the web server. The
ISP will communicate us for that directory name.
(viii) We can leave remote directory.
(ix) Filter and local directory filter both should be blank.
This is where we would enter wildcard such as
*.htm*. All other files, such as .gif and .jpg, would
then be ignored.
(x) For the starting directory, enter the proper path where
the web pages saves.
(xi) We do not need to change any setting on the advance
tab unless we face the problems with our internet
connection. If it happens, ISP will help to figure out
the best settings.
3. After click OK, make sure that internet service is proper
and connection with web server is perfect. Mostly web
server issues a message publically, who are connected
with them. Some FTP software does not pay any
attention for this message, but CuteFTP have this facility.
4. Once makes the connection with server, two lists of files
will open. Left pane shows the list of files present on
your computer, while the right hand side pane shows the
list of files present on the server side computer. Select
the HTML and other files for uploading on server from
the left hand side window. After the select command,
click the upload button from the toolbar.[15]
5. After uploading the files view the web page on the server
and client with the help of different browsers.
6. After uploading the modifying files on the server, choose
FTP, and disconnect the connection “Fig. 6,7,8”.
IV. TESTING WEB PAGE
Every time we transfer the web pages to memory disk, web
sites or intranet server. Test every uploaded web page contents
thoroughly, preferably through the 28.8 Kbps modem
connection. Some other precautions are as follows:
1. Test the web site on the latest version of internet browser
like Microsoft Internet Explorer, Mozilla Firefox,
 Google Chrome and many others like DOS Lynux or
Opera Mini. Testing with older version of browser isn’t a
bad idea since many internet surfers still use outdated
version of browser and some web pages will not appear
properly on these older browsers.
Fig. 6. CuteFTP Site Manager
Fig. 7. CuteFTP Hosting Dialog Box
Fig. 8. CuteFTP Hosting default setting Dialog Box
2. Make sure the testing computer is set to a 16- colour
video mode or, at most a 256-colour mode. Also try
several different brightness settings on your monitor and
change the colour balance of monitor to make sure your
pages still look okay on display that are more blue or red
than yours.
3. Set the computer on 800X600 resolutions for web page
testing, but regulate the volume of browser window on
640X480 pixels. Check the every web page on
maximum and minimum size of the browser window on
the given resolution. If pages quality is good at these
resolutions then they will also looks fine at large
resolution. Test the web pages at 1024X768 or
1600X1200 resolution shows perfection.
4. Before testing the web page, switch off the auto
uploading of images by browser, to take the exclusive
review of the graphics. Also ensure the ALT tag
messages, upload the image by clicking the load image
button and review updated page again consciously.
5. Set the font size by the diversity of font size facility
available in the browser for each page. It ensures the
drop layout content of each section in a web page.
6. Begin with home pages and sequentially go through with
each link. After each link back or home button should be
available in every web page.
7. Wait for loading the full contents of web page, also
check that all the uploaded images are appear on the
defined section of the page.
8. If the website is complex, then make the checklist of
testing for every web page to ensure the best quality.
9. Check the loading time of the page. Use 28.8 Kbps
modem, if possible when associated with dissimilar ISP’s
that running our web server. Also major the loading time
on 14.4 Kbps modem and compare them, it must be just
double the time taken by 28.8 Kbps modem.
If web page pass this entire test successfully, then we can be
certain that the page will looks great on every web browser
and on every computer connected with the internet in this
world.
V. CONCLUSION & FUTURE WORK
This paper provides the basic information we need to prefer
among the most ordinary content delivery methods used for
uploading the web contents. It also provides the step by step
information to uploading the web page on the interanet and
then internet by the use of different browsers and software. It
also describes the process of inserting web pages contents and
images files on a web server workstation by using commonly
available file transfer protocol software. Finally it offers a
checklist to test the web pages on VDU in all respect, once
contents are uploaded on the web.
In the designing of the web pages we have to use the HTML,
XHTML, XML, PHP, java script etc. Testing of proper
delivery of contents on client side, response time of server and
many more on different parameters is left for future work.
REFERENCES
[1] Helen J. Wang, Chris Grier, Alexander Moshchuk, Samuel T. King,
Piali Choudhury, Herman Venter Microsoft Research, University of
Illinois at Urbana-Champaign, University of Washington “The Multi-
Principal OS Construction of the Gazelle Web Browser” MSR Technical
Report MSR-TR-2009-16
[2] D. Crockford. JSONRequest. http://www.json.org/jsonrequest.html.
[3] D. Crockford. The Module Tag: “A Proposed Solution to the Mashup [17] http://www.ccsf.edu/Pub/Fac/composer.html
Security Problem”. http://www.json.org/module.html.
[18] http://en.wikipedia.org/wiki/Microsoft_FrontPage#mediaviewer/File:Mi
[4] Firefox3fordevelopers,2008.”https://developer.mozilla.org/en/Firefox_3 crosoft-FrontPage-screenshot.png
_for_developers”.
[19] http://www.a2hosting.com/kb/getting-started-guide/publishing-your-
[5] C. Grier, S. Tang, and S. T. King. “Secure web browsing with the OP web-site/publishing-your-web-site-with-microsoft-frontpage
web browser”. In Proceedings of the 2008 IEEE Symposium on Securiy
and Privacy, 2008. [20] http://support.lanset.net/pages/conf_2.html

[6] W. H. A. T. W. Group. Web Applications 1.0, February 2007. [21] http://penta2.ufrgs.br/edutools/composer/tutcomposer6.html

http://www.whatwg.org/specs/web-apps/current-work/. [22] http://www.dollarhost.com/tutorials/FrontPage/
[7] What’s New in Internet Explorer 8, 2008. http://msdn.microsoft.com/en- [23] http://support7.qnx.com/download/download/26848/HTML5_Developer
us/library/cc288472.aspx. s_Guide.pdf
[8] IEBlog: IE8 Security Part V: Comprehensive Protection, 2008.
http://blogs.msdn.com/ie/archive/2008/07/02/ie8-security-part-v-
comprehensive-protection.aspx.
[9] C. Reis and S. D. Gribble. “Isolating web programs in modern browser
architectures”. In Proceedings of Eurosys, 2009.
[10] J.Ruderman. The Same Origin Policy.
http://www.mozilla.org/projects/security/components/same-origin.html.
[11] H. J. Wang, X. Fan, J. Howell, and C. Jackson. “Protection and
Communication Abstractions in MashupOS”. In ACM Symposium on
Operating System Principles, October 2007.
[12] W3C XMLHttpRequest Level
http://dev.w3.org/2006/webapi/XMLHttpRequest-2/.
[13] www.wikeipedia.com/web_browser
[14] http://www.techterms.com/definition/web_browser
[15] Teach Yourself HTML 4 by Dick Oliver
[16] http://math.hws.edu/TMCM/java/labs/WebPublishingLab.html
ICCCS-2015 Author Index

Author Index

Abderrahmane, Daif 78
Abuna, Felix 87
Adewumi, Adewole 85, 86
Agarwal, Akash 17
Agba, Basile 17
Ahmed, Abd El-Aziz 64
Ahmed, Mahmood 64
Akingbesote, Alaba 46
Akinola, Ayotuyi 46
Annamalai, Senthamarai Selvan 38
Appavoo, Perianen 69
Ashley-Dejo, Ebunoluwa 26, 28
Asowata, Osamede 5

Baichoo, Sunilduth 73, 77

Bala, Rajni 9
Bansal, Mohit 52
Barbot, Jean-Pierre 20
Betz, Bourke 87
Bhekisipho, Twala 47
Binns, Richard 55
Biolek, Dalibor 65, 66
Biolkova, Viera 65, 66
Bokhoree, Chandradeo 45
Boodoo-Jahangeer, Bibi Nazmeen 73, 77
Bukusi, David 87
Busawon, Krishna 20, 55

Catherine, Clarel 12
Catherine, Pierre 15, 16
Chavan, Mahesh 90, 92
Cheema, Amarjeet 27
Chen, Chien-Hung 21
Cherutich, Peter 87
Chetty, Naganna 75
Chien, Wu-Fan 30
Chikudo, Admiral 68
Chiu, Chien-Ching 21
Chougule, Sharada 92
Christo, Pienaar 5
Chuku, Peter 70
Coonjah, Irfaan 12, 15, 16
Curum, Brita 19
ICCCS-2015 Author Index

Dasgupta, Ranjan 10, 62

Dash, Ratnakar 22
Datcu, Octaviana 20
Dhawale, Chitra 88, 89
Djonon Tsague, Hippolyte 47
Dookhitram, Kumar 45
Dookhun, Preetamsing 69
Dunbar, Matthew 87

El Jamiy, Fatima 78
Engelbrecht, Klarissa 33
Esan, Omobayo 28, 36

Farquhar, Carey 87
Flores, Denys A. 40
Foogooa, Ravi 45

Garai, Chandan 10, 62

Gill, Komal 35
Gomes, Teena 61
Gooria, Vandanah 76
Gopee, Ajit Kumar 68
Gulati, Ravi 7
Gungea, Meera 69
Gunness, Sandhya 25
Gupta, Amit 14

Hamza, Nermin 64
Hans, Robert 63
Hefny, Hesham 64
Ho, Min-Hui 21
Horvat, Matija 24
Hosanee, Yeeshtdevisingh 79
Hosseinzadeh, Shohreh 71
Hyrynsalmi, Sami 71

Ierache, Jorge 60

Jaufeerally-Fakim, Yasmina 68
Jena, Sanjay Kumar 18
Jones, Brian 48
Juddoo, Suraj 13
Jutton, Teenah 69

Ka, Selvaradjou 38
Kekana, Johannes 63
Khedo, Kavi Kumar 19, 42
Kim, Hyoungjun 44
ICCCS-2015 Author Index

Kolka, Zdenek 65, 66

Koochakpour, Kaban 37
Koonjal, Premanand 76
Ku, Tsun 30
Kumar, Binay Binod 52
Kumar, Sachin 93

L. K. Cheung, Jonathan 74
Labeau, Fabrice 17
Laha, Sumit 82, 83
Langueh, Kokou 20
Le Roux, Petra 49
Lee, Jaehee 44
Lee, Kyungho 44
Leppänen, Ville 71
Lerato, Masupha 28
Liao, Shu-Han 21
Limthong, Kriangkrai 50
Liu, Han-Wen 30
Loock, Marianne 49

M. Abdou, Mohamed 56
Macharia, Paul 87
Maharaj, Manoj 3
Mahlobogwane, Zanele 39
Maiti, Sumana 10
Majhi, Banshidhar 18, 22
Mansur, Vidya 58
Manyere, Peter 51
Masupha, Lerato 36
Medhi, Nabajyoti 6
Migabo, Emmanuel 57
Mishra, Anurag 9
Misra, Sanjay 85, 86, 88, 89
Mocktoolah, Asslinah 42
Mohamed, Hossam 56
Mohapatra, Ramesh Kumar 18
Montenegro, Carlos W. 40
Muhammad, Najam Ul Islam 1
Muhongya, Kambale Vanty 3
Mujahid, Umar 1
Murthy, Dr. B K 27
Mvelase, Promise 67

Nayak, Deepak Ranjan 22

Nel, Andre 51
Ngobeni, Matimu 39
Ngwira, Selelman 31, 32
ICCCS-2015 Author Index

Ngwira, Seleman 29, 36

Njorege, Anne 87
Ntuli, Elesa 29
Nundoo-Ghoorah, Suniti 76
Nundran, Kevin 74

Ojo, Samuel 31, 32

Olwal, Thomas 57, 70
Ostrek, Mirela 24

Pal, Manjish 6
Panchoo, Dr. Shireen 79
Panchoo, Shireen 61
Patel, Charmy 7
Paupiah, Pravin Selukoto 11
Peeroo, Swaleha 48
Peng, Yong 53
Phate, Thato 57

Rajpal, Ankit 9
Ramjug-Ballgobin, Rajeshree 54, 55
Ras, Dirk 72
Rauti, Sampsa 71
Refaie, Rasha 64
Renke, Amar 90
Richomme, Morgan 74
Robberts, Michelle 33
Roy, Rinita 82
Rughooputh, H C S 55

Sack, Pablo 60
Sambai, Betsy 87
Samy, Martin 48
Saxena, Deepika 84
Saxena, Shilpi 84
Sayed Hassen, Sayed Zahiruddeen 54
Schoeman, Ruaan 5
Schwalke, Udo 59
Sharma, Anju 35
Sharma, Shivani 91
Sheel, Neelaksh 94
Shettar, Rajashekar 58
Singh, Omesh 23
Singh, Upasana 23, 25, 41
Singha Roy, Sayantan 62
Singhal, Rekha 8
Skvorc, Dejan 24
Sm, Ngwira 28
ICCCS-2015 Author Index

Sohoraye, Mrinal 76
Sonawane, M.S. 88
Sotenga, Prosper 70
Soyjaudah, K. M. S. 12
Soyjaudah, K.M.Sunjiv 15, 16
Spies, Chel-Mari 81
Spies, Jan 80
Srivastav, Vinita 89
Srivastava, Praveen 27
Srivastava, Siddharth 27
Stone, Kyle 80
Sudarshan, Sithu D 75
Suddul, Geerish 68, 74
Sungkur, Roopesh 41

Tarokh, Mohammad Jafar 37

Thombre, Avinash 2
Tiru, Banty 34
Toshniwal, Durga 91, 93
Trivedi, Aditya 2
Tsai, Cheng-Hung 30

Ul-Ain, Qurat 1

Vaisla, Kunwar Singh 94

Vaisla, Kunwer Singh 75
van der Poll, John 43
Van Der Poll, John Andrew 33
van Tonder, Joubert 43
Veerapen, Sonia 54
Venter, H.S. 72

Wang, Jidong 4
Wang, Xiaoyi 53

Yao, Yuangang 53
Yoon, Hyunsik 44

Zhan, Zheng 53
Zhang, Xinyang 4
Zhao, Xianghui 53
Zheng, Gang 20
Zuva, Tranos 28, 29, 31, 32, 36
ICCCS-2015 Keyword Index

Keyword Index

6LoWPAN 17

8-connected neighbor 82
802.11n 70

ABCD matrices 34
Access Categories(ACs) 70
Accident Analysis 93
Activation value 89
Ad hoc cloud mobile 46
Adaptive Field-Effect Transistor 59
Adaptive Neuro Fuzzy 54
Algorithm indexing 51
Android 24, 86
ANFIS 37
anomaly detection 50
Architecture 78
artificial noise injection 2
assessment 41
assistant node 4
Association Rule Mining 62, 93
attributes reduction 75
Automotive 11

Backscatter communication system 2

base station 29
Benchmark 8
BER 9
betweeness centrality 3
big data 25
Big Data 8, 13, 78
BigData 62
Bioinformatics 68
biomass concentration 55
Biometric 73
Biometrics 31, 32, 77
biometrics 36
bioreactor 55
Blended learning 61
Blind Watermarking 9
Blood Bank Management System 27
BPNN 18
browser 94
Bus Topology 34
ICCCS-2015 Keyword Index

categorization of privacy challenges 42

CBAR 62
Challenges facing mobile workers 33
channel capacity 21
Chaotic systems 20
chronic kidney disease 75
Classification 14, 88
classification 75
closeness centrality 3
cloud computing 85
Cloud computing 43, 67, 72, 84
CloudSim 84
Clustering 14, 62
cobit 40
cognitive load 19
cognitive radio 29
Cognitive Radio 38
combinatorial optimization 83
Complex Trigonometric Hyperbolic Functions 34
Computer forensi 11
Computer forensics 72
confidence interval 7
Confidentiality 87
Congestion control 38
connected components analysis 52
Context 26
context-aware mechanisms 19
Context-aware Recommender systems. 26
context-awareness 19
continued system usage 81
controller 55
cookies 94
correlation 7
Cost Per Click 30
Critical Infrastructure 44
CRM 78
CryptDB 64
cuckoo search algorithm 83
Customer engagement 48
Cyber defense 60
Cyber warfare 60
Cyberspace 60

Data cleansing 13
Data Mining 14, 91, 93
Data mining 56, 75
data mining 25
ICCCS-2015 Keyword Index

Data profiling 13
Data quality 13
Data quality dimensions 13
Data quality metrics 13
Data quality rules 13
Database 77
Database security 64
DDoS 10
design aspects 81
Development Platform 68
Differential power analysis 47
Digital forensics 72
Digital signal operations 58
digital television 29
Discrete Cosine S-Transform 18
Discrete wavelet transform (DWT) 22
diversification 71

E-assessment 31, 32
e-assessments 25
e-CRM 78
e-health 63
e-tutoring 49
Ear 77
education 19
EM Radiation 90
Embedding algorithm 82
emulator 66
Energy Efficiency 39
Enforcing Mechanism 27
engage 79
Enhanced Distributed Channel Access (EDCA) 70
enterprise resource planning (ERP) systems 23
Enterprise systems 43
ERP modules 23
ERP procurement 23
ERP selection 23
ethical tracking 80
evaluation framework 23
EXIF metadata 24
Extreme Learning Machine 9

Face Recognition 73, 77

Face recognition 36
Facebook 48, 76
Feature Vector 10
Feed Forward Neural Network 89
filter design 57
ICCCS-2015 Keyword Index

Forensic monitoring 72
FPGA 1
Framework. 33
free convection 56
ftp 94
Full-duplex 57
Fusion 31
Fuzzy 54
Fuzzy-PID 54

Game Based Learning 41

gate 47
Genectic Algorithm 82
Generation Rate Constraint 54
geo-location spectrum database 29
geolocation 74
geolocation API 74
Gephi 3
Global Feature 73
Global Warming 61
governance 40
Governor Deadband 54
Graphical User Interface(GUI) 79
Green cloud computing 35
Green ICT 45
Grocery stores 48
GUISET 46

Hadoop 62
Half-duplex 57
health 63
heat source/sink. 56
Hidden units 89
hierarchical wireless sensor networks 4
High-K dielectric 47
higher order thinking skills (HOTS) 25
Homomorphic encryption 64
hosting 94
Hybrid Feedback 28
hybrid method 37
Hypervisor 72

ICT 40
identity development 49
Image Steganography 82
Immersion 20
Implicit Feedback 28
impulsive noise environment 17
ICCCS-2015 Keyword Index

Informatic Security 60
information security awareness 49
Innovative video technology 69
inter-user interference 57
Internet of Things 71
Interval arithmetic 58
Intrusion detection and prevention system 35
IoT 71
IP Spoofing 10
ISA100.11a 17
iso 40
IT 40
IT security 44

k-Coverage 6
K-Means 14
Kernel principal component analysis (KPCA) 22
key management 4
Keywords: 26
Knowledge discovery 56
Knowledge management 43

law 76
LBP 73
learning analytics 25
Learning mathematics 69
Least squares support vector machine (LS-SVM) 22
licensed number plate 52
Line Spectral Pair (LSP) 92
Linear Predictive Cepstral Coefficients (LPCCs) 92
linear regression 7
live video streaming 24
Load Frequency Control 54
Local Feature 73
location privacy challenges 42
location-based 63

machine learning 50
Magnetic resonance imaging (MRI) 22
management 40
Map reduce 91
MapReduce 62
massively-parallel analog computations 65
Maturity models 45
Mauritius 76
Mel Frequency Cepstral Coefficients (MFCCs) 92
memcapacitor 66
meminductor 66
ICCCS-2015 Keyword Index

memristor 65, 66
Metrics 60
mHealth 81, 87
micro-services 74
Microblog 53
MIMO 57
MIMO-WLAN 21
mixed strategy 57
MJPEG 24
MLP 88
MNIST dataset 18
Mobile ad hoc networks 38
mobile application 63
Mobile application 86
mobile applications 74
Mobile Devices 41
Mobile knowledge workers 33
mobile learning 19, 49
mobile technology 63
Mobile worker perceptions 33
Mobile worker requirements 33
Mobile workforce 33
model 40, 65
MONOMI 64
Moores Law 59
Multimodal Biometrics 32
multiple timeline 50
Mutual Authentication 1

NABH 27
NACO 27
Naı̈ve Bayes 14
nearest neighbor 88
Neural network 88
neural network 52
Neural Network 59
Nigerian education 85
Normalized Correlation (NC) 9
NSA 10

obfuscation 71
Object-Oriented Programming(OOP) 79
Observability Singularity 20
observer 55
OCR 88
Online Sequential 9
Online social network 3
OpenSSH 12, 16
ICCCS-2015 Keyword Index

OpenVPN 12, 15
orientation angles 5

Pallelization 91
Parallelism 51
PCA 73
Peak Signal to Noise Ratio 82
performance evaluation 36
performance measurements 7
Performance testing 7
Personal of Interest analysis 30
physical layer secrecy 2
Polycrystalline 5
popularity 53
porous medium 56
portable device 29
Power Line Communication 34
practices 40
prediction 53, 75
Preprocessing 14
Privacy 87
privacy 71, 76, 80
privacy solutions 42
Professional Donor 27
proficiency level. 79
Protocol to Access White Spaces 29
proximity based social networking 42
PSNR 9
PSO and BPN methods 37
public concern 90

Q format representation 58
Quality of Service (QoS) 70
quantum algorithm 83
quantum-inspired cuckoo search algorithm 83
query processing 64

RDBMS 8
Recommender system 26
Recommender Systems. 28
Reconfigurable Logic 59
Relationship marketing 48
reliable communication 17
Residential Exposure 90
retweet 53
Reuleaux Tetrahedron 6
RFID 1
Road Accidents 93
ICCCS-2015 Keyword Index

rotation matrix 24
Rough sets 56
Rule extraction 89
Rule Induction 56

SASI 1
Secure data transmission 20
Secure indexes 64
Security 35, 87
security 4, 42, 71
Security Controls 60
self-interference 57
Sensor Nodes 39
Serious games 41
server mobile machine 46
service computing 67
service delivery models 85
service selection 46
SFSA 88
simulated annealing 83
single CCI ray-tracing approach 21
Sixsoid 6
smart card. 47
SMMEs 46, 67
social computing 49
Social media 48
Social Networks 30
Social Personal Analysis 30
Socio-constructivism 61
Software Agent 68
software tools 79
Spectrum Sensing 38
SPICE 65
SSAR data 51
SSIM 9
staff availability 80
staff tracking 80
standard 40
Star Topology 34
student feedback 41
suitability 85
Supply chain 43
Sustainable Development 45
SWOT analysis 85

Tasks scheduling 84
Taxi booking 86
TCP tunnel 15
ICCCS-2015 Keyword Index

Testing 94
Throughput 70
tilt 5
traditional health practitioner 63
Transmission Line 34
Transport Protocol 38
transportation 86
Transportation company 43
Travelling salesman problem 83
Tunneling 12, 16
TV whitespace 29

UDP tunnel 15
Ultralightweight 1
Unknown malicious code detection model 44

variable permeability 56
vehicle security 11
video rotation 24
Virtual Enterprise 67
Virtual Machine 84
visualization 3
VPN 12, 15, 16

web 80
web API 74
weighting technique 50
WEKA tool 14
White space device 29
wireless sensor networks 2
Wireless Sensor Networks 6, 17, 39
WirelessHART 17
Workflow 84
WSR. 57

XOR 82

YouTube Education 69

Zigbee 17
 Sponsored by:

MAURITIUS SUBSECTION

ISBN 978-1-4673-9353-9

IEEE Catalog Number CFP15CSG-DVD