You are on page 1of 5

2022 IEEE 13th Control and System Graduate Research Colloquium (ICSGRC), 23 July 2022, Shah Alam, Selangor,

Malaysia

Intrusion Detection System: An Automatic Machine


Learning Algorithms Using Auto-WEKA
Venus W. Samawi Suhad A. Yousif Nadia M. G. Al-Saidi
Department of Computer Science Department of Computer Science Department of Applied Sciences
Isra University Al-Nahrain University University of Technology
Amman, Jordan Baghdad, Iraq Baghdad, Iraq
venus.samawi@iu.edu.jo suhad.a.yousif@nahrainuniv.edu.iq nadia.m.ghanim@uotechnology.edu.iq

Abstract— Network Intrusion Detection Systems (NIDS) are Anomaly-based IDSs utilize artificial intelligence and
essential to maintaining network security. This study is machine learning techniques to construct a trustworthy
concerned with developing an Intrusion Detection System (IDS) intrusion-detection model that detects suspicious activity
based on Automated Machine Learning (AutoML) to reduce (activities that are not defined as normal in the developed
false alarms and provide accurate NIDS. The proposed model is model). Machine learning-based IDSs are more generalized
developed using two machine learning (ML) software tools compared with signature-based systems. Nowadays, most
2022 IEEE 13th Control and System Graduate Research Colloquium (ICSGRC) | 978-1-6654-6806-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICSGRC55096.2022.9845166

(Weka and RapidMiner). To study the performance of various IDSs employ the rule-based approach in which few rules can
ML algorithms, four different classifiers are applied to the detect most intrusions [2,4, 5].
intrusion detection dataset, namely Naïve Bayes (NB),
Multilayer Perceptron (MLP), Random Forest (RF), and Much research concerning network intrusion detection
Sequential Minimal Optimization (SMO). Auto-WEKA is used focuses on anomaly-based methods. Various approaches are
to select the best classifier with its appropriate hyperparameters employed by researchers, including genetic algorithms [1, 6],
automatically. Auto-model implements the best classifier where fuzzy rule-based genetic classifiers are developed to
resulting from Auto-WEKA in RapidMiner to answer the detect intrusions, and data mining approaches [7,8]. Machine
question, which tool among them achieves the best accuracy ML and deep learning (DL) approaches are also employed to
with minimal effort (i.e., best-suited to non-expert developers)? develop NIDS [2,4, 5, 9, 10]. Most developed IDSs are trained
Finally, the performance of the classifiers (resulting from Weka, and tested using benchmark datasets, including KDD Cup'99,
Auto-WEKA, and the Auto Model in RapidMiner) is evaluated
NSL-KDD, and CSE-CIC-IDS2018 [11-13]. Although KDD
utilizing the NLS-KDD dataset. The experimental results show
that RF outperforms the other classifiers in terms of accuracy
Cup'99 and NSL KDD are old datasets, most researchers still
with adequate time consumption. It was also found that Auto- use them due to the availability of significant findings in the
WEKA is the preferable one since it automatically selects the literature [2]. The attack cases in KDD Cup'99 and NSL-KDD
best classifier (in terms of accuracy) with its appropriate datasets fall into four categories (DoS, R2l, U2r, and Probing)
hyperparameters with minimal effort. and normal cases. Some researchers choose to develop a
classifier to identify two classes (normal and attack) [2, 14,
15], and others choose to develop IDSs that can further
Keywords—Auto Machine Learning, Intrusion Detection, identify the attack type (i.e., a classifier that can identify five
WEKA, Auto-WEKA, RapidMiner, Machine Learning classes) [1, 4-6, 14].
I. INTRODUCTION When using machine learning algorithms, the main
problem is choosing the proper algorithm and setting its
With the massive growth of internet and communication
hyperparameters, which is difficult for non-expert users.
technologies and the enormous number of applications
AutoML tools such as Auto-WEKA and Auto Model in
available on the network, its security has become a significant
RapidMiner are perfect solutions that help non-expert choose
problem that needs to be resolved. Various tools and
the proper ML algorithm and set its hyperparameters. The
protection systems preserve network security within
main contribution of this paper is to develop an ML intrusion
cyberspace, including intrusion detection systems (IDS). The
detection model utilizing two ML tools (Weka and
(IDS) monitors hosts and network traffic to detect malicious
RapidMiner) to study the performance of four ML algorithms
activities or policy breaching. When suspicious activity is
(four Rule-based classifiers) using Weka: Naïve Bayes (NB),
detected, the IDS will generate an alert and report it to the host
Multilayer Perceptron (MLP), Random Forest (RF), and
or network administrator or centrally collect utilizing an event
Sequential minimal optimization (SMO). As Auto-WEKA
management system [1,2]. IDSs are classified into host
automatically chooses the most proper ML algorithm and
intrusion detection systems (HIDS), which protect endpoints
hyperparameters for a dataset, we will compare the accuracy
by comparing current and previous snapshots, and NIDS
of the Auto-WEKA recommended ML algorithm with the
monitors network traffic to detect network threats [3].
accuracy of the four ML algorithms utilizing Weka. Finally,
Technically, two methods are used to develop IDS: the algorithm with the best accuracy is implemented using the
signature-based and anomaly-based methods. The first detects Auto Model of RapidMinor, where its performance is
intrusions based on specific patterns (called signatures) that compared with Auto-WEKA. To evaluate the performance of
the malware uses. These patterns (signatures) are already the developed model, the NSL-KDD dataset is used. Based on
defined in the system. The signature-based intrusion detection the experimental results, we will answer three questions:
system cannot detect new attacks because their signatures are
unknown. Therefore, the systems should be updated when a 1) Which ML algorithm achieves the best accuracy?
new attack is discovered. Alternatively, anomaly-based 2) Does feature selection a crucial stage for the NSL KDD
methods can detect unknown malware attacks, making them dataset to achieve high accuracy IDS?
beneficial for addressing the rapid development of malware.

978-1-6654-6806-0/22/$31.00 ©2022 IEEE 42


Authorized licensed use limited to: Staffordshire University. Downloaded on March 28,2024 at 16:28:02 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE 13th Control and System Graduate Research Colloquium (ICSGRC), 23 July 2022, Shah Alam, Selangor, Malaysia

3) Which tool among Weka, Auto-WEKA, and the Auto tools (Weka and RapidMiner). First, using Weka, four
Model in RapidMiner produces the highest accuracy different classifiers are applied to the NSL-KDD dataset (NB,
with minimal processing steps (i.e., which tool best suits MLP, SMO, and RF) to select the best classifier from the
non-expert developers). accuracy perspective. Then, Auto-WEKA is automatically
used to select the best classifier with its suitable
In this paper, AutoML is discussed in section two. Section hyperparameters. Finally, the best classifier resulting from
three describes the adopted methodology. Experimental Auto-WEKA is selected to be executed by the Auto Model in
results are assessed in section four. Finally, we concluded in RapidMiner. Auto Model in RapidMiner will automatically
section five. specify the most suitable hyperparameters for the executed
classifier. Finally, the performance of the classifiers (resulting
II. AUTO MACHINE LEARNINGAUTOML from Weka, Auto-WEKA, and the RapidMiner) is evaluated
Even though ML algorithms do not need humans to learn to choose the recommended ID classifier among the above.
and prepare the data implemented by the algorithms The suggested model mainly consists of two phases: the
themselves, finding the best algorithms and hyperparameters classification phase and the evaluation metric phase.
to get the best results requires a highly skilled specialist data
scientist. So, to get good results, the data scientist tries
different algorithms with different parameters to come up with
the best setting that gives the most efficient results. These
processes are not easy since they depend on specialist humans
(data scientists) with computer science, linear algebra,
mathematics, and statistics skills. Of course, the data scientist
with these skills is rare and called "data science unicorns".
Therefore, many organizations recruit multiple data scientists
and data analysts to perform the complete ML processes. One
of the ways to resolve this issue is by introducing AutoML, a
completely automated process that reduces human
interference to a minimum [16]. Different AutoML platforms
are introduced, some are commercial, like RapidMiner. Some
are open source and free coding, while others are free but
require some coding processes like Auto-WEKA, the subject
of this paper. Auto-WEKA is a model that implements the
selection among a set of defined algorithms and chooses the
best hyperparameter optimization in classification and Fig. 1. Auto Weka Model
regression models over Weka. Therefore Auto-WEKA
settings the hyperparameters for many methods and A) Classification Phase
recommends the best parameters with good generalization
using model-based optimization. This work proposes three different classification models
utilizing the software tools (Weka, Auto-WEKA, and the Auto
Unlike Weka, Auto-WEKA will automatically run and Model in RapidMiner). Weka includes a set of different ML
determine the best algorithms and hyperparameters. This algorithms used for data mining labor. It consists of many
determination happened intelligently using the "sequential tools for the preparation of data, feature extraction and
model-based algorithm configuration" (SMAC) tool [17,18]. selection, clustering, regression, visualization, and
The options that need to be set in Auto-WEKA are few and classification. In this paper, four Different ML classifiers are
usually could be left at their default values. Most users are utilized. These classifiers are:
interested in only the most essential two options:
1) Naive Bayes (NB): is a popular and straightforward
Time-Limit: this parameter is the time needed to specify supervised learning technique. The Bayesian algorithm
the best classifier and configuration. Therefore, better can tackle diagnostic and predictive problems in general.
results are achieved when increasing the time limit. Because the NB classifier is a probabilistic model, these
issues can be resolved by choosing the option with the
MEM-Limit: This option specifies the memory limit to run highest probability [19]. This probability is chosen using
the classifier and should be increased in case of an Eq.(1).
extensive dataset.
P (y | x) =P (y) .P(x|y) /P(x) . (1)
Fig. 1 shows the internal Experiment steps implemented
inside the Auto-WEKA wizard. All these steps are Here, P(x) refers to the prior probability of x, P(y) refers
implemented under the WEKA umbrella. to the prior probability of y, P(x|y) refers to the posterior
probability of x given y, and P (y |x) refers to the
III. PROPOSED ML BASED IDS MODEL probability that x belongs to class-y [20, 21].
Network intrusion detection systems are used to detect 2) Sequential Minimal Optimization (SMO): The support
malicious attacks and reacted by reporting the attack or vector machine (SVM) classifies binary and multinomial
network intrusion to the IT personnel as soon as it happens. data. Its objective is to find the best hyperplane for
Therefore, NIDS is essential to preserve network security. dividing the dataset into two groups based on the features
This study is concerned with developing an IDS based on [20, 22]. SOM is a new SVM learning algorithm
AutoML to reduce false alarms and provide accurate NIDS. conceptually simple, easy to implement, faster, and has
The suggested model is developed using two ML software superior scaling properties for challenging SVM issues

43
Authorized licensed use limited to: Staffordshire University. Downloaded on March 28,2024 at 16:28:02 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE 13th Control and System Graduate Research Colloquium (ICSGRC), 23 July 2022, Shah Alam, Selangor, Malaysia

than the standard SVM training algorithm. It is a method studies that compare different IDS methods. Furthermore,
for resolving the quadratic programming (QP) problem NSL-KDD does not contain redundant records compared with
that emerges during support-vector machine training the original KDD'99 CUP dataset. In this case, the classifiers
(SVM). It was created in 1998 by John Platt. will not be biased toward the most frequent records, making
various methods' evaluation results consistent and comparable
3) Multilayer perceptron (MLP): A Neural Network
determines how related linear and nonlinear data are. [11,12].
The input layer, output layer, and hidden layers, which
are situated between the input and output layers to form
a cascaded network structure, make up the main structure
of the MLP classifier. Only one hidden layer is sufficient
to build a network structure with the computing
capability to solve all of the problems in the plurality. In
some cases, an additional hidden layer is added to
improve the accuracy of the findings; however, these
additional layers would increase the training data's
processing time [23].
4) Random Forest (RF): RF is introduced as one of the
ensembles of ML algorithms since many decision trees
are constructed when training the model, and the results
of predictions from all decision trees are collected to
make the final result. RF is one of the complex nonlinear
supervised ML used in regression and classification [24].
The Auto-weka is part of Weka, which could be applied
either via Weka GUI (the most straightforward technique
using the Auto-WEKA Panel (AWP), which permits Auto-
WEKA execution directly on a loaded dataset) or by choosing
Auto-Weka from standard classifier lists (AWC). In this case,
we need to select the "use training set" option, and then Auto-
WEKA implements a statistically accurate evaluation to select
the best classifier and automatically specifies the most suitable
hyperparameters for that classifier. Finally, the results are
achieved based on the 10-fold cross-validation method, which
is applied internally by AWC.
Finally, RapidMiner is another data mining platform that Fig. 2. The proposed model
includes data preparation and preprocessing, visualization,
validation, and optimization. It also comes with the necessary Fig. 3 illustrates the experimental results of the four ML
machine learning and deep learning technologies. In addition, algorithms implemented using WEKA. It was found that RF
RM includes an automatic model, the Auto Model (AMRM), outperforms the other ML algorithms with accuracy (99.67)
which provides a prediction (classification and regression), based on its design technique that combines many decision
clustering, and outlier detection. In this work, we are trees to decide the type of attack and decrease the risk of
concerned with the prediction model, which provides NB, DT, overfitting. Furthermore, the dataset is partitioned externally
RF, Deep learning, SVM, and GBT algorithms. The using 10-fold cross-validation.
parameters of each model, as well as feature engineering and
optimization methodologies, are chosen manually. Fig (2)
summarize the suggested ML-based-IDS illustrated above. Accuracy
B) Evaluation Metric (Accuracy)
RF
Various evaluation metrics are used based on values of:
NB
True Positive (TP) and True Negative (TN): Correct class
SMO
prediction (negative or positive).
MLP
False Positive (FP) and False Negative (FN): Class
Misclassification (swap negative and positive classes). 80 85 90 95 100
MLP SMO NB RF
The evaluation metric used in this work is accuracy (as in 1).
Accuracy % 97.3 97.14 87.2 99.67
2
Fig. 3. The ML-Base IDS: Applying Weka
IV. EXPERIMENTALNTAL RESULT: ASSESSMENT
In this work, Auto-WEKA is used in two different ways
This section illustrates the experimental results of the (AWP, AWC). Both ways choose the RF algorithm as the best
suggested ML-based-IDS. The NSL-KDD dataset is used to classifier. Fig 4. Show that AWP outperforms AWC random
evaluate the suggested ML-based-IDS's performance. NLS- forest algorithm (99.98, 99.40, respectively) since AWP is
KDD is an efficient benchmark due to the availability of directly applied to the loaded dataset (i.e. the accuracy is
significant findings in the literature besides its usability in

44
Authorized licensed use limited to: Staffordshire University. Downloaded on March 28,2024 at 16:28:02 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE 13th Control and System Graduate Research Colloquium (ICSGRC), 23 July 2022, Shah Alam, Selangor, Malaysia

calculated based on the training and testing portions of the Comparing the developed model results to the one
data set, which increases accuracy). In contrast, AWC applied presented in [14] proves the importance of using AutoML
a statistically rigorous evaluation internally (10-fold cross- algorithms (Auto Weka) since the author(s) tried several ML
validation) where the accuracy indicates the average accuracy algorithms with their difficulty in setting all preprocessing
of the 10-folds. The results are achieved when the time limit steps on the dataset and manually setting the hyperparameters
is set to 100. To answer the question, do all AutoML behaves (trial and error) to get the best accuracy. While the AutoML
equally when applied to the same dataset? we compare the internally chooses the best algorithm and hyperparameters
result of Auto-WEKA with another AutoML (Auto-model in without the interference of the data scientist.
Rapid Miner) using the best algorithm (RF). AMRM comes
with minimum accuracy (93.9) due to dataset partitioning System Perfomance
depending on the hold-out method, with 60% for training and
40% for testing). Accuracy % Time (Secs.)
The time consumption of the used ML algorithms is 100%
illustrated in Fig. 5. Although the NB algorithm is 90%
accomplished quickly due to its simplicity, the accuracy is 80%
70%

SECONDS
very low compared with the other algorithms. The RF is the 60%
best algorithm in terms of accuracy with adequate training 50%
time, whereas MLP is time-consuming, although it 40%
30%
outperforms NB and SMO in terms of accuracy. The time 20%
consumption of MLP is due to the time spent in the training 10%
phase, which depends on the dataset used and the number of 0%
RMR AWP AWC
MLP SMO NB RF
samples used to train the MLP. Finally, Fig. 6. shows the F RF RF
relation between accuracy and time for each suggested Time (Secs.) 44041 99 2.08 74.34 25.4 27 5
classification model. Accuracy % 97.3 97.14 87.2 99.67 93.9 99.98 99.4

Fig. 6. ML_Based IDS: System Performance


Accuracy
V. CONCLUSION
AWCRF In this study, the problem of IDS is tackled using machine
AWPRF learning algorithms. Four algorithms are applied to the ID
dataset using WEKA, namely NB, MLP, SMO, and RF. The
RMRF experimental results show that RF outperforms the other
classifiers in terms of accuracy with reasonable time
RF consumption. Furthermore, when Auto-WEKA is applied on
90 92 94 96 98 100 the NSL-KDD, the best accuracy is achieved with RF in
AWP and AWC (99.98 and 99.67, respectively). It is
RF RMRF AWPRF AWCRF essential to mention that the (99.98) accuracy is reached when
Series1 99.67 93.9 99.98 99.4
the whole dataset is used in training, which is not preferable
since it is crucial to develop IDS that could recognize a new
type of intrusions. On the other hand, the Auto model in
Fig. 4. The ML-Base IDS: Applying WEKA, Auto_WEKA, and AMRM RapidMiner uses the hold-out method in dataset partitioning,
while AWC uses automatic 10-fold-cross validation.
Therefore, the recommended approach is AWC since its
accuracy is high and can deal with new intrusion types.
TIME CONSUMPTION
Finally, by comparing the accuracy of AutoWEKA with
previous studies, we conclude that by using AutoML, high
95 accuracy can be reached without the need to perform
complicated preprocessing steps and test various ML
75 algorithms and manual settings of the hyperparameters
SECONDS

(based on trial and error) to improve system accuracy.


55
REFERENCES
35
[1] Naqshbandi, A., M., S., Samawi W. V. & Herik V., J. Intrusion
15 detection by a dynamic length rule, AWERProcedia Information
Technology & Computer Science. [Online].2013, 3, pp 1460-1470.
-5 Available from: http://www.world-education-center.org/index.php/P-
RMR AWP AWC ITCS
MLP SMO NB RF
F RF RF
[2] Ahmad, Z., Shahid Khan, A., Wai Shiang, C., Abdullah, J., & Ahmad,
Time (Secs.) 44041 99 2.08 74.34 25.4 27 5 F. (2021). Network intrusion detection system: A systematic study of
machine learning and deep learning approaches. Transactions on
Emerging Telecommunications Technologies, 32(1), e4150
Fig. 5. Time Consumption for the Developed Models
[3] Kumar, B. S., Ch, T., Raju, R. S. P., Ratnakar, M., Baba, S. D., &
Sudhakar, N. (2013). Intrusion detection system-types and prevention.

45
Authorized licensed use limited to: Staffordshire University. Downloaded on March 28,2024 at 16:28:02 UTC from IEEE Xplore. Restrictions apply.
2022 IEEE 13th Control and System Graduate Research Colloquium (ICSGRC), 23 July 2022, Shah Alam, Selangor, Malaysia

(IJCSIT) International Journal of Computer Science and Information [14] Wendi, C.; Khan, S.; Anajemba, J.H.; Mittal, M.; Alenezi, M.; Alazab,
Technologies, Vol. 4 (1), 2013, 77 - 82 M. The Use of Ensemble Models for Multiple Class and Binary Class
[4] Lew J, Shah DA, Pati S, et al. Analyzing machine learning workloads Classification for Improving Intrusion Detection
using a detailed GPU simulator. Paper presented at: Proceedings of the Systems. Sensors 2020, 20, 2559. https://doi.org/10.3390/s20092559
IEEE International Symposium on Performance Analysis of Systems [15] Ali, A., Jillani, F., Zaheer, R., Karim, A., Alharbi, Y. O., Alsaffar, M.,
and Software (ISPASS). Madison, WI, USA: IEEE; 2019:151-152. & Alhamazani, K. (2022). Practically Implementation of Information
[5] Saranya, T., Sridevi, S., Deisy, C., Chung, T. D., & Khan, M. A. Loss: Sensitivity, Risk by Different Feature Selection
(2020). Performance analysis of machine learning algorithms in Techniques. IEEE Access, 10, 27643-27654.
intrusion detection system: A review. Procedia Computer [16] Mustafa, A. and Rahimi Azghadi, M., 2021. Automated machine
Science, 171, 1251-1260. learning for healthcare and clinical notes analysis. Computers, 10(2),
[6] S. M. A. Naqshbandi and V. W. Samawi, "One-rule Genetic-Fuzzy p.24.
classifier," 2012 IEEE International Conference on Computer Science [17] Kotthoff, L., Thornton, C. and Hutter, F., 2017. User guide for auto-
and Automation Engineering (CSAE), 2012, pp. 204-208, doi: WEKA version 2.6. Dept. Comput. Sci., Univ. British Columbia,
10.1109/CSAE.2012.6272759. BETA lab, Vancouver, BC, Canada, Tech. Rep, 2, pp.1-15.
[7] G. Meena and R. R. Choudhary, "A review paper on IDS classification [18] Kotthoff, L., Thornton, C., Hoos, H.H., Hutter, F., Leyton-Brown, K.
using KDD 99 and NSL KDD dataset in WEKA," 2017 International (2019). Auto-WEKA: Automatic Model Selection and Hyperparameter
Conference on Computer, Communications and Electronics Optimization in WEKA. In: Hutter, F., Kotthoff, L., Vanschoren, J.
(Comptelix), 2017, pp. 553-558, doi: (eds) Automated Machine Learning. The Springer Series on
10.1109/COMPTELIX.2017.8004032. Challenges in Machine Learning. Springer, Cham.
[8] F. Salo, M. Injadat, A. B. Nassif, A. Shami and A. Essex, "Data Mining https://doi.org/10.1007/978-3-030-05318-5_4.
Techniques in Intrusion Detection Systems: A Systematic Literature [19] Yousif, Suhad A., Hussam Y. Abdul-Wahed, and Nadia MG Al-Saidi.
Review," in IEEE Access, vol. 6, pp. 56046-56058, 2018, doi: "Extracting a new fractal and semi-variance attributes for texture
10.1109/ACCESS.2018.2872784. images." AIP Conference Proceedings. Vol. 2183. No. 1. AIP
[9] P. Mishra, V. Varadharajan, U. Tupakula and E. S. Pilli, "A Detailed Publishing LLC, 2019.
Investigation and Analysis of Using Machine Learning Techniques for [20] Ahmed Y. Al-Obaidi and Venus W. Samawi. "Opinion Mining:
Intrusion Detection," in IEEE Communications Surveys & Tutorials, Analysis of Comments Written In Arabic Colloquial", Proceeding of
vol. 21, no. 1, pp. 686-728, Firstquarter 2019, doi: the conference of World Congress on Engineering and Computer
10.1109/COMST.2018.2847722. Science (WCECS 16), vol. 1, pp. 470-475, 21–23 October 2016
[10] Karatas G,Demir O, Sahingoz OK. Increasing the performance of [21] Yousif, Suhad A., Sultani, Zainab. N., & Samawi, Venus. W.
machine learning-based IDSs on an imbalanced and up-to-date dataset. “Utilizing Arabic WordNet Relations in Arabic Text Classification:
IEEE Access. 2020;8:32150-32162. New Feature Selection Methods”. IAENG International Journal of
https://doi.org/10.1109/ACCESS.2020.2973219. Computer Science, 46(4), (2019).
[11] Information and Computer Science," KDD Cup 99 dataset,", [22] Yousif, A. Y.; Younis, S. M.; Hussein, S. A.; Al-Saidi, N. M. G..
University of California, 1999, an intelligent computing for diagnosing covid-19 using available
http://kdd.ics.uci.edu/databases/kddcup99/kddcup99.html blood tests International Journal of Innovative Computing, Information
[12] M. Tavallaee, E. Bagheri, W. Lu, and A. Ghorbani, "A Detailed and Control ; 18(1):57-72, 2022.
Analysis of the KDD CUP 99 Data Set," Submitted to Second IEEE [23] Jehad, R., & Yousif, Suhad A., “Classification of fake news using
Symposium on Computational Intelligence for Security and Defense multi-layer perceptron”. In AIP Conference Proceedings Vol. 2334
Applications (CISDA), 2009. https://www.unb.ca/cic/datasets/nsl.html No. 1 p. 070004, AIP Publishing LLC, 2021.
[13] A collaborative project between the Communications Security [24] Jehad, R. ;Yousif, Suhad A. “Fake News Classification Using Random
Establishment (CSE) & the Canadian Institute for Cybersecurity (CIC), Forest and Decision Tree (J48)” . Al-Nahrain Journal of
"CSE-CIC-IDS2018", 2018, https://www.unb.ca/cic/datasets/ids- Science; 23(4), 49-55, 2020.
2018.html

46
Authorized licensed use limited to: Staffordshire University. Downloaded on March 28,2024 at 16:28:02 UTC from IEEE Xplore. Restrictions apply.

You might also like