You are on page 1of 6

2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS)

Comparison among different CNN Architectures for


Signature Forgery Detection using Siamese Neural
Network
2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS) | 978-1-7281-8529-3/20/$31.00 ©2021 IEEE | DOI: 10.1109/ICCCIS51004.2021.9397114

Soumya Jain Meha Khanna Ankita Singh


Student, Information Technology Student, InformationTechnology Assistant Professor, Information
department department Technology department
IGDTUW IGDTUW IGDTUW
New Delhi, India New Delhi, India New Delhi, India
jainsoumya2105@gmail.com meha_khanna@outlook.com ankita@igdtuw.ac.in

Abstract—Signature is the most common way to indicate Signature forgeries can be classified into one of the
knowledge and acceptance of a document. As many documents following types:[7]
and contracts are now starting to use paperless electronic
formats, the term "signature" has been substantially 1. Blind Forgery (Random):
broadened. Whichever form it takes, the key importance of the A forger who has no access to the original signature
signature is identity authentication for managing security. may indulge in this type of forgery. The forged
Signatures being one of the most widely used methods for the signature in such cases shares little or no similarity with
same, play a crucial role in financial, legal, and social aspects the genuine signature.
of one's life. Thus, Signature forgery, that is falsely copying
another individual’s signature is an issue of utmost concern. 2. Trace-over Forgery (Unskilled):
The chances of two or more signatures made by the same The forger traces over the genuine signature which they
individual being identical are minimal, thus making signature have access to. This type of forgery is difficult to detect
forgery detection an arduous task. Our paper aims to apply the in photocopied or scanned documents.
state-of-the-art methodology, Siamese Neural Networks, on the
chosen data set, draw meaningful insights and perform a 3. Skilled (Practiced) Forgery:
comparative analysis between some variants of these neural In this case, the forger has one or more specimens of
networks to identify and authenticate handwritten signatures.
the genuine signature. The accuracy of the simulation
Keywords—signature, forgery, authentication, Siamese,
depends on how much the perpetrator practices before
convolutional neural networks the actual forgery. This type of forgery is the most
difficult to detect.
I. INTRODUCTION In signature forgery detection, classifying an original
As stated by the West's Encyclopedia [6]- “A signature signature is forged and vice versa are both problematic
is a mark or sign made by an individual on an instrument or situations. However, more emphasis should be on
document to signify knowledge, approval, acceptance, or classification of forged signatures asoriginal.Forgery should
obligation. Its purpose: to authenticate a writing, or provide not pass as a genuine signature as it poses a tremendous
notice of its source, and to bind the individual signing the security threat.
writing by the provisions contained in the document.”
II. RELATED WORK
The handwritten signature is one of the most common
forms of biometrics that has applications in day-to-day life. In [1] various notations for signature forgery detection
Despite being one of the earliest, most basic, and popularly are discussed. The Siamese network consists of two
accepted methods for identification and verification, identical convolutional networks that accept two signatures.
confirming the genuineness of a signature is particularly These images are taken from the rows of the dataset that
challenging. may be similar or different. These two convolutional neural
networks are then connected by a cost function. This cost
It is a myth that the authentic signatures of an individual function calculates a distance metric between the highest-
will be identical when done multiple times. Signing requires level feature representation on both the networks. Datasets
coordination of the eyes, fingers, arms, nerves, and muscles used:
with the brain. Other factors like emotional condition,
personality, age, health, and environment in which the 1.CEDAR signature database: 55 signers, 1320 genuine and
person signs, influence the signature. Thus, due to numerous 1320 forged signatures.
determinants involved, some elements may not appear the 2.GPDS300 signature corpus: 300 signers, 7200 genuine
same in each signature. and 9000 forged signatures.
What makes signature forgery detection even more 3.GPDS synthetic signature database :4000 signers,
difficult and crucial is the skill and precision, with which a 96000 genuine and 120,000 forged signatures.
forgery of the genuine signature is done.The focus of the
forger is usually on making an accurate copy of the genuine 4.The BHSig260 signature dataset: 260 signers(Bengali and
signature, rather than on signing fluently. Hindi), 6240 genuine and 7800 forged signatures.

ISBN: 978-1-7281-8529-3/21/$31.00 ©2021 IEEE 481

Authorized licensed use limited to: Birla Institute of Technology and Science. Downloaded on January 31,2022 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS)
In [14] a method based on Agglomerative Clustering is
The accuracies with the above databases were as follows: proposed for signature verification. It suggests that the
1. 100 clusters for similar images should be identical.
2. 76.83
III. PROPOSED METHODOLOGY
3. 77.76
4. 84.11 (Hindi) 86.11 (Bengali) A. Dataset
Signature_Verification_Dataset[9] from Kaggle having
The authors in [2] demonstrate a CNN in which an image signatures of Dutch users (both genuine and fraud) was
goes through 3 convolutional and max-pooling layers
chosen for the study. The dataset comprises of 2149 image
which are in an alternating fashion. The pooled feature map
from the last max-pooling layer is flattened and sent into the files and 2 .csv files.
fully connected layers. The dataset used in this research has
6000 signatures with 1000 genuine and 1000 forged per a) Image Files
subject. Total 3 unique signers There are 16 extra images,8 ● 69 unique signers
genuine and 8 forgeries per subject for holdout. ● Images of genuine and forged signatures for every
Shayekh Mohiuddin Ahmed Navid et al [3] suggest a person.
method in which VGG-19 which is used for forgery ● Total 2149 .pngfiles (train and test set included),
detection, is connected to a convolution layer with 256 out of which 1649 are unique. Train-Test split is
neurons, followed by layers having 128 and 64 neurons, 70:30.
respectively. The convolution-64 layer was then connected b) .csv files
to a fully connected layer of 512 outputs followed by layers
with 256, 128 outputs respectively and finally a dense layer ● train.csv has 887 unique values in the first column
with 2 outputs. and 1649 in the second.
● test.csv has 252 unique values in the first column
JiveshPoddara et al [4] suggests a method in which a
signatureimage is classified using both Crest Trough and and 500 in the second.
CNN methods.After classification, algorithms for forgery ● Both csv files have 3 columns. The first 2 columns
detection (Harris Algorithm followed by Surf Algorithm) contain entries for each original signature of an
are applied on the image. The dataset contains 1320 pictures individual and corresponding other signatures of
(not in grey scale), with 25 Unique pictures. the same person.
● The third column is a binary indicator of match (0)
ShalawMshir et al [5] proposes a method in which deep
convolutional neural networks (CNN) networks are or mismatch (1) of the signatures.This is how the
connected by a triple loss function. This function computes data is labelled for verification.
a similarity-scale that includes the distance between feature
representations on each of the two CNNs forming the
Siamese network.The dataset used is Kaggle dataset
contains 30 users, each having 15 signatures. This approach
is completely independent of the writer's approach.

Apart from the above 5 research works, the following


research papers were also studied: Fig 1: Example of images in dataset
JiaXin Ren et al [10] suggests a method in which features
are extracted from the processed image by calculating local
B. Pre-processing
structural patterns. Signature verification done by matching
extracted features against the features stored in the database.
Rajesh Kumar et al [11] proposes a method in which from
each pre-processed image, features are extracted. The
extracted vectors (for both images) are then combined in the
„pairing‟ module. Absolute difference is then calculated
between corresponding features.Two-class classification is
then used for verification purpose.
Hadjipanteli et al [12] proposes a method which uses grid-
based feature extraction. The evaluation is done using
binary grid masks. A WD verification scheme is followed
which is divided into 2 phases- training and testing.
Verification is done on a well-established dataset.
Fahmy et al [13] proposesa system that consistsof three
phases, namely-signal modelling, feature extraction, and
feature matching.
Fig 2: Steps in Preprocessing

482

Authorized licensed use limited to: Birla Institute of Technology and Science. Downloaded on January 31,2022 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS)
distance to an example of the same class and differentiates
The preprocessing (fig 2) required for this dataset was done that with the distance to positive examples. [8]
as follows
1. Preparing labels and image paths
2. Getting/fetching the image paths
3. Gray scaling the images- The images are converted Fig 4 Contrastive loss function
to single channel form, that is interpreted as
grayscale. E. Variations of CNN architecture used
4. Applying necessary transformations- Resizing of
all the images to a common size is necessary for A CNN consists of many kinds oflayers, starting at the
training the neural network. The images are pixels of raw images. Each layer performs a computation
upscaled/downscaled to a size of 100x100 pixels. and feeds the result to the next layer. The result is fed to a
5. Shuffling the images into random sets of 8. linear classifier.
This research proposes 3 CNN architectures that can be used
for signature forgery detection systems for optimal results as
C. Algorithm Used mentioned below:
Siamese convolutional neural network is a category (class) 1) CNN Architecture 1
of neural networks which requires two or more subnet In the first CNN architecture, as shown in fig 5, the 4 layers
functions. The sub-networks are connected with a in the LHS of fig 4 are repeated three times in the same
contrastive loss function that computes a similarity index order. In each successive convolution 2d layer, the number
based on the distance between features on each of the two of channels in the input signature image and number of
sides of the Siamese network. channels produced by convolution is increased.
While training this Siamese neural network, two signature Local response normalization (which is non-trainable) layer
images are given as input and the output is a dissimilarity square normalizes the pixel values within a local
index comparing the features of the two images, as shown in neighborhood in a feature map. This is followed by max
fig 3. pooling for reducing feature dimension. This is followed by
CNN model consists of three main components a fully connected layer with ReLU activation, followed by
a. Convolution layer- “scans” the image (pixel vector), dropout and 2 more fully connected(linear) layers.
extracts features and tries to judge its belongingness to
a label.
b. Pooling layer- used to reduce the number of parameters
and prevent overfitting, by down-sampling. It computes
the maximum of the width and height of the input.
c. Fully connected layer-Usually contains the final output
of the CNN. Neurons of this layer have
connections(full) to all activations in the preceding
layer.
Other layers might be added for normalization, activation,
flattening etc.

Fig 5:CNN Architecture 1

2) CNN Architecture 2
In the second CNN architecture (fig 6), results are more
optimized. It uses batch normalization as opposed to local
Fig 3: Siamese Neural Network Architecture response normalization used in architecture 1.
Batch Normalization is a trainable layer and is carried out
for each pixel across all activations in a batch. The kernel
D. Contrastive loss size for all max pool 2d layers is taken to be 2.
Contrastive loss(in fig 4) maps vectors that model the
similarity of the inputs. Contrastive loss takes the output of
the network for a negative example and calculates its

483

Authorized licensed use limited to: Birla Institute of Technology and Science. Downloaded on January 31,2022 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS)
Over the past years, use of various deep learning algorithms
like CNNs (VGG-19 and others), Harris and Surf algorithms
have been used. Other implementations also included
algorithms like clustering [9]. However, use of Siamese
CNN for forgery detection remains state of the art.
This approach uses Siamese Networks. Architecture 1 is like
those used in [1] and [5]. Changes to architecture 1 have
been made to get architectures 2 and 3, which optimize the
training loss further and overcome the shortcomings of the
initial implementation.

Unique features of this implementation-


1. Contrastive Loss Function
2. Similarity Index as metric
3. Optimized architectures 2 &3 uses batch normalization
instead of local response normalization.
IV. RESULTS AND EVALUATION
Fig 6: CNN Architecture 2
A. Results Obtained
3) CNN Architecture 3 Table 2: Training loss values for 3 CNN Architectures
In CNN architecture 3, the layers were kept the same as that
of architecture 2. The number of channels in the input
signature image and those produced by convolution was CNN Number of Training loss
Architecture epochs
increased. Increasing the output channels results in typical
down sampling to trade off spatial resolution for greater
Architecture 1 50 1.04
channel depth.
Architecture 2 50 0.56
F. Comparing with related work in the field
Architecture 3 50 0.43

Table 1: Comparison of Related Work


The three variants of Siamese networks created using three
Paper Layers Activation Evaluation different CNN architectures were trained on the chosen
Function(s) Metric(s) dataset for 50 epochs each.
Siamese Convolution, ReLU Accuracy
CNN [1] pooling, * Local The values of contrastive loss obtained in each case are
dropout, and Response shown in table 2.
fully Normalisatio
connected n. It was observed that the value is high for the first
layers.
architecture but decreases considerably in architecture 2 on
CNN Convolution, ReLU and Accuracy and making the following changes:
[2] pooling, SoftMax. Entropy Loss 1. Replacing local response normalization by batch
dropout, and
fully normalization(trainable).
connected 2. Decreasing the number of layers in the architecture.
layers.
VGG-19 VGG19 ReLU, Accuracy, Further changes in architecture 2 lead to down sampling
[3] (except last Sigmoid and Binary Cross-
layer) SoftMax. entropy,
resulting in a considerable drop in training loss for
followed by 3 Precision, architecture 3.
convolution Recall, F1 To test and evaluate the model‟s working, dissimilarity
and 4 dense score. index (using Euclidean distance) was calculated on a test
layers.
set.
CNN, Convolution, ReLU and Accuracy
Harris max pooling, Sigmoid
& Surf and fully Two examples of the result obtained are shown in Fig 7.
[4] connected Each figure displays the actual label of the image (on the
layers. right side) and the calculated dissimilarity.
Siamese Convolution, ReLU Triple Loss In the first image, the image is of an original signature and
CNN [5] pooling, and Function
fully thus, the dissimilarity is low.
connected However, in the second image because of high dissimilarity
layers. (0.72) between the 2 signatures, the second signature is a
forged one.

484

Authorized licensed use limited to: Birla Institute of Technology and Science. Downloaded on January 31,2022 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS)
In this problem of handwritten signature forgery detection,
the positive class (1) is the forged signature class, and the
negative class (0) is the genuine signature class.
In addition to considering True Positives and True
Negatives, precision captures the False Positives. On the
contrary, recall takes into consideration the False Negatives.
False Positive in this case would be a genuine signature
being classified as a forged one.False Negative would mean
a forged signature being classified as a genuine signature.
Though both the negatives correspond to erroneous
situations, a forged signatureclassified as an original
signature would have severe consequences.
Fig 7: Dissimilarity Index
Thus, in the problem at hand, recall is given preference over
precision.To further validate the classification, F1 score is
B. Evaluation Metrics
also considered.
In addition to Euclidean Distance and Contrastive Loss, the
following metrics were used for model evaluation: V. CONCLUSION AND FUTURE WORK
1. Precision
2. Recall This paper analyses and compares previous work done in
3. F1 Score the field of forgery detection in signatures. The proposed
system performed an efficient signature forgery detection
For all 3 CNN architectures, 3 different values of method using a fully connected Siamese Neural Network
dissimilarity index (Euclidean distance) threshold, i.e., 0.4, architecture, which is state of the art.This methodology
0.5 and 0.6 were tested. consists of data acquisition, pre-processing, extraction using
There is always a tradeoff between precision and recall CNN and evaluation through contrastive loss.High value of
scores. In this specific case, that is, forgery detection, high recall (0.92, close to 1) was achieved in this implementation
recall is preferable over high precision. and Contrastive loss was minimized (0.43).
To further find abalance between the two, F1 score is also
considered. It was observed that threshold value 0.4 gave Signature forgery detection system can have many
the best results. The results have been shown in fig 8. applications, like in security management, enforcement, and
various other business purposes. The proposed system is
inexpensive in forgery detection at runtime.

As a future work, state of the art parameter coefficients can


be used to increase the deviation between forged and real
signatures which will eventually result in better accuracy.

VI. REFERENCES
[1] SounakDeya, AnjanDuttaa, J. Ignacio Toledoa, Suman K.Ghosha,
JosepLladosa, UmapadaPalb 2. “SigNet: Convolutional Siamese
Network for Writer Independent Offline Signature Verification” in
Elsevier 2017
[2] Jerome Gideon S, Anurag Kandulna, Aron Abhishek Kujur, DyanA,
KumudhaRaimond “Handwritten Signature forgery detection using
CNN” in 8th International Conference on Advances in Computing
and Communication (ICACC-2018)
[3] Shayekh Mohiuddin Ahmed Navid, Shamima Haque Priya, Nabiul
Hoque Khandakar, Zannatul Ferdous, AkmBahalul Haque 2.
“Signature Verification Using Convolutional Neural Network” in
2019 IEEE International Conference on Robotics, Automation,
Artificial-intelligence and Internet-of-Things (RAAICON)
[4] JiveshPoddara, VinantiParikha, Santosh Kumar Bharti
“Offline Signature Recognition and Forgery Detection using Deep
Learning” in International Journal of Computer Sciences and
Engineering.
[5] ShawlawMshir, Mehmet Kaya 2. “Signature Recognition Using
Machine Learning” in 2020 8th International Symposium on Digital
Forensics and Security (ISDFS)
[6] West's encyclopedia
[7] https://towardsdatascience.com/signature-fraud-detection-an-
advanced-analytics-approach-10c810cda26e
[8] https://towardsdatascience.com/contrastive-loss-explaned-
159f2d4a87ec
[9] https://www.kaggle.com/robinreni/signature-verification-dataset
Fig 8: Comparing metrics

485

Authorized licensed use limited to: Birla Institute of Technology and Science. Downloaded on January 31,2022 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.
2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS)
[10] Jing Wen, MoHan Chen, and JiaXin Ren“Off-Line Signature
Verification Based on Local Structural Pattern Distribution Features”
in Pattern Recognition, 2014, Volume 484, Springer.
[11] Rajesh Kumar, J.D. Sharma,Bhabatosh Chanda “Writer-independent
off-line signature verification using surroundedness feature” in 2011
Elsevier.
[12] A. Hadjipanteli*, E. N. Zois and A. Nassiopoulos
“Signature Verification Using Young's Lattice Grid Modeling”, in
JOURNAL OF Engineering Science and Technology Review, 2015.
[13] Fahmy, M. M. (2010). “Online handwritten signatureverification
system based on DWT features extraction and neural network
classification.” Ain Shams Engineering Journal, Pattern Recognition,
2014, Volume484.
[14] Varun Pandya “Offline Signature Verification using Clustering
Technique” in the International Conferenceon Pattern Recognition and
Artificial Intelligence - PRAI 2019.

486

Authorized licensed use limited to: Birla Institute of Technology and Science. Downloaded on January 31,2022 at 07:16:53 UTC from IEEE Xplore. Restrictions apply.

You might also like