2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)

December 11–12, 2019, Amity University Dubai, UAE

Fraud Detection using Machine Learning and Deep

Learning
Pradheepan Raghavan Neamat El Gayar
School of Mathematical and Computer Sciences School of Mathematical and Computer Sciences
Heriot Watt University Heriot Watt University
Dubai, UAE Dubai, UAE
[email protected] [email protected]

Abstract— Frauds are known to be dynamic and have no total loss from the credit cards in the world was $21.84
patterns, hence they are not easy to identify. Fraudsters use billion and projected that in 2020 it would be $32 billion. [2]
recent technological advancements to their advantage. They
somehow bypass security checks, leading to the loss of millions In this paper, we will be looking into 3 data sets. They
of dollars. Analyzing and detecting unusual activities using are the European dataset [3], the Australian dataset [4] and
data mining techniques is one way of tracing fraudulent the German dataset [4]. In this work we aim to benchmark
transactions. transactions. This paper aims to benchmark different ML and DL techniques. An ensemble of the best 3
multiple machine learning methods such as k-nearest neighbor performing models is also applied the all 3 datasets. We
(KNN), random forest and support vector machines (SVM), present our conclusions based on an empirical study
while the deep learning methods such as autoencoders, comparing different ML and deep learning models.
convolutional neural networks (CNN), restricted boltzmann
machine (RBM) and deep belief networks (DBN). The datasets The paper is organized as follows. Section 2 summarizes
which will be used are the European (EU) Australian and the related work and the background of the models
German dataset. The Area Under the ROC Curve (AUC), implemented. Section 3 provides details on implementation
Matthews Correlation Coefficient (MCC) and Cost of failure and experimental setup. In section 4 we present and discuss
are the 3-evaluation metrics that would be used. the results. Conclusions and future work are summarized in
sections 5 and 6 respectively.
Keywords—credit card, fraud detection, machine learning,
deep learning, random forest, k nearest neighbor, support II. RELATED WORK
vector machine, autoencoder, restricted boltzmann machine, Tuyls et al. [11] outline several challenges regarding Fraud
deep belief networks, convolutional neural networks
Detection. First, the highly unbalanced datasets in this
I. INTRODUCTION application where only a small percentage of the available
data is fraud makes training efficient models quite difficult.
Ever since the introduction of credit cards and online Besides other problems arise from noisy data and
payments, many scammers have found ways to exploit overlapping patterns. Most importantly the dynamics of
people and steal their credit card information to use them for frauds keep changing and classification models need to
unauthorized purchases. This leads to a huge amount of capture and adapt to this change. As follows we review
fraudulent purchases every day. Banks and eCommerce
some of the most relevant studies that have applied machine
websites are trying to identify these fraudulent transactions
and stop them from happening again. With Machine learning learning and deep learning models in the area of fraud
and Deep Learning methods, they are trying to stop the detection.
fraudsters before the transaction is approved. A. A Comparative study on KNN and SVM
Machine learning is one of the hottest topics of this Zareapoor and Shamsolmoali [5] published a paper on
decade and a subset of Artificial Intelligence. More and more Fraud Detection using different techniques namely Naïve
companies are looking to invest in machine learning to Bayes, KNN, SVM and Bagging Ensemble Classifier. In
improve their services. Machine learning is a combination of their paper, they discuss the various concerns while handling
various computer algorithms and statistical modeling to this problem, such as there is a non-availability of real-world
allow the computer to perform tasks without hard coding. data which forces researches to work on faux data as the
The acquired model would be learning from the “training banks and other financial institutions don’t make their data
data”. Predictions can be made or actions can be performed public due to privacy concerns as it is sensitive data. Also,
from stored experiential knowledge. Deep learning models most of the times the data is highly unbalanced as the
are a part of machine learning techniques which involves number of fraudulent transactions is only 2% and 98% of the
Artificial Neural Networks. Convolutional neural networks, transactions are legitimate. They refer to the concerns of big
Deep Belief Network, Auto-encoders, Recurrent Neural data and the computation time it takes for larger datasets in
Network, and Restricted Boltzmann Machine are all various these cases. One of the major challenges which frequently
methods. A properly trained NN would have the capability to reported in many research papers is about the dynamic nature
capture unique relationships over the whole dataset. of fraud. The nature of fraud cannot be defined by one
particular situation or style. Hence, it needs machine learning
Credit card fraud is a form of fraud involving the use of algorithms to be updated regularly so that malicious attempts
fake or stolen credit card information and causing financial could be caught in real time.
harm to account holders or merchants involved. The total
number of credit card fraud in Single Euro Payments Area In their experiment, they used the data from the UCSD-
(SEPA) in 2016 was 1.8 Billion Euros out of the total 4.38 FICO competition where it contained 100,000 instances of
Trillion Euros transaction, which is 0.4% lower than the credit card transaction with 20 attributes from an e-
previous year[1]. In 2015, according to the Nelson report, the commerce website. Only 2293 instance/transactions were

978-1-7281-3778-0/19/$31.00 ©2019 IEEE

334
Authorized licensed use limited to: Ontario Tech University. Downloaded on May 10,2025 at 22:45:24 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

fraudulent which makes the ratio of legitimate to fraud as structure 30-2-30, 30-10-30, 30-20-10-2-10-20-30 and 30-
100:3. 25-20-10-20-25-30, basically two with one hidden layer and
two with 5 hidden layers. All this was built in python using
The methodology they used was splitting the datasets into Tensorflow with learning rate 0.01 and training the network
4 different sets which contained 20% fraud, 15% fraud, 10% for 100 epochs. For evaluating the performance, he used
fraud and 3% fraud. In their experiment, they realized that recall-at-k and precision-at-k where k is the number of
the accuracy or error rate would not be the best evaluation instances.
measure in this case. Hence, they used True Positive Rate
(fraud catching rate), True Negative Rate, False Positive With k at 1000, single hidden layer AutoEncoder worked
Rate (false alarm rate) and False Negative rate. These four much better compared to the stacked multilayered
values help in reflecting the performance more than the AutoEncoder. However, as k increased the stacked model
accuracy and the error rate. They used a 10-cross-fold produced much better results over single layer model.
validation for their experiments. The final result showed that
KNN had a much better false alarm rate and fraud catching D. Using Restricted Boltzmann for Fraud Detections
rate as compared to SVM and Naïve Bayes Classifier on all Restricted Boltzmann machines (RBM) can be used for
four sub-datasets. data reconstruction in an unsupervised learning setting [9].
Pumsirirat and Yan’s Paper [9] used Keras to implement this
B. Random Forest in Fraud Detection high-level NN. They used H2O package to find out the Mean
Randhawa et al’s paper [6] on Credit card fraud detection Squared Error, Root Mean Square Error and Variable
using AdaBoost and majority voting explores many different importance of the attributes in each dataset. Keras was used
machine learning algorithms such as Naïve Bayes, Random to get the Area Under the Curve and Confusion Matrixes for
Forest, Gradient Boosted Tree, etc. In this paper, they use each case.
“Majority Voting” for combining two or more algorithms.
They found that RBM works much better in producing
The study also investigates AdaBoost ensemble model and
reports that AdaBoost is very sensitive to anomalies and AUC and Accuracy for larger datasets (European) compared
outliers. to the smaller datasets (German and Australian). This may be
due to the fact that with smaller datasets, it’s harder to
They use the RapidMiner as an implementation Software recognize the fraud as the data is too sparse. However, with
and the experiments are conducted using the South-East Asia larger datasets its easier to detect “not-fraud” as there is a lot
region credit card data. This dataset is a highly imbalanced of data to learn/train from.
data set with less than 1% fraud transaction. All of the
classifiers were evaluated using 10-fold cross validation to E. Using CNN for for detecting Suspicious Activity
reduce the bias. The classifier performance is evaluated using In Chouiekha and El Haj’s paper [10], Convolutional
the Matthews Correlation Coefficient (MCC). Neural Networks (CNN) are used for Fraud Detection. A
database was created with 18000 artificial images of 300
MCC helps in measuring the efficiency of the 2-class customers’ activity during 60 days. They used Customer
problem by taking TP, TN, FP and FN rates into account. In Details Records in such a way that long conversation or an
their experiment, they found that Random Forest had one of unusual number of vouchers used would be detected. CNN
the best MCC rates at 0.990 compared to other methods such is applied to the images to detect fraudulent activity. 50% of
as SVM, gradient boosted trees, etc. By using AdaBoost, the data set was used for training, 25% for validation and
Random Forest with achieved 100% accuracy and MCC rate 25% of for testing. Images have been rescaled to improve
to be 1. In this study the generalization ability of the classifier performance. The proposed Deep CNN(DCNN)
developed model should be carefully examined as it is not contained 7 layers with 3 Convolutional layers, 2 pooling
clear how good it would perform on unknown data. layers, 1 full connected layer and finally 1 SoftMax
In general, the study reports that hybrid approaches regression layer.
produce more reliable results compared to single classifier Results were evaluated using accuracy. Deep CNN’s
models. performance is compared against SVM, Random Forest and
C. Detecting Fraud using AutoEncoders based on Gradient Boosting Classifier (GBC). The results show that
Reconstruction Error DCNN outperforms SVM by 5%, Random Forest by 10%
and GBC by 3%. Deep CNN was found to train almost twice
Tom Sweers in his bachelor thesis, [7] describes faster than the rest of the methods.
AutoEncoders as an effective neural network which can
encode the data as it would learn to decode it as well. In this F. Neural Networks vs Bayesian Belief Network
approach the Autoencoders are trained to non-anomaly The study in [11] compares Bayesian Belief Network
points, introduced to the anomaly points to classify it as (BBN) against Artificial Neural Networks. BBN are found to
‘fraud’ or ‘no fraud’ according to the reconstruction error better detect fraud cases and to use less training time. ANN
which is expected to be high in the case of anomalies that the however had faster prediction times when applied in real
system has not been trained on. Here, any value above the time.
upper bound value or threshold could be considered an
anomaly. This is a method which was also used in Z. Chen el G. Summary and Motivation
at’s paper [8] on Autoencoder-based network anomaly Various Machine Learning techniques have been used to
detection. detect frauds such as SVMs, KNN, K-Means, Random
Chen reported that stacked AutoEncoders worked better Forest, Naïve Bayes and many more. Most of them deal with
than for one hidden layer AutoEncoders in detecting unbalanced data sets where the amount of fraud is very low.
anomalies. The AutoEncoder network he used had the The main evaluation metrics for fraud detection are True
Positive Rate, False Negative Rate and Matthews Correlation

335
Authorized licensed use limited to: Ontario Tech University. Downloaded on May 10,2025 at 22:45:24 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

Coefficient. Some studies recommend using Neural Scikit-Learn, and Tensorflow. Rstudio was occasionally used
Networks as a solution for unbalanced datasets. for data cleaning.
In Tom Sweers’ thesis [7], his methodology of For K-nearest neighbor, we use cross-validation on the
introducing Autoencoders to normal data and detecting fraud training data set to determine the best value of the neighbors
based on reconstruction error is unique. In Pumsirirat et al’s K for each data set. The best K for each data set is then used
paper [9], we can see that AUC and Confusion matrix can be to conduct further analysis on the entire data set.
used to evaluate the models. We also see that how deep
learning fails when there is less instances in a dataset For Support vector machines and Random Forest we
(Australian and German). Working with smaller dataset use a grid based search approach to find the best parameter
failed to achieve good prediction scores. This shows that not for each model. We use the Python method GridSearchCV,
always deep learning would able to solve problems for with the parameters depicted in figures 1 and 2 for the SVM
smaller dataset. While Chouiekha et al. [10] report that and the Random forest, respectively.
DCNN and outperforms SVMs, Random Forest and Gradient
Boosted Classifier, Tuyls et al. [11], confirm that Artificial
Neural Networks have a much faster fraud catching process
compared to Bayesian Belief Networks. Fig. 1. Parameters for SVM
In this study we present an empirical comparison of
various machine learning and deep learning models inspired The best parameters are used to evaluate the models with
by the previous studies. Our intention is to investigate the the entire data.
performance of various models using data sets with different
sizes, complexities and characteristics. The goal is to come
up with recommendations on best practices of picking the
most suitable model for a fraud detection application given
the data of particular description.
In particular we compare the performance of SVM,
KNN and Random forest to Deep Learning methods such as Fig. 2. Parameters for Random Forest
AutoEncoders, RBM, DBN and CNN. Since we will be
working with three different datasets, we would be able to
see to what extent these machine learning techniques hold The basic idea behind the Autoencoders is that it could
valid. We also consider fine-tuning methods such PCA, reconstruct its input. So here for fraud detection, we train the
reduction of features, cleaning data, hyperparameters etc. to autoencoders only on the normal transaction. While running
improve classifier performance. the experiment on the test data, it would produce
reconstruction errors for each of the instances. The
Furthermore, we investigate combining the 3 best
expectation is that normal transactions would produce lower
performing models in an ensemble using majority voting. In
reconstruction errors while fraud transactions/instances
the next section we describe the data used, the experimental
would produce higher values. A certain threshold value is
setup and the main learning models.
set, such that if reconstruction error is above this threshold
III. IMPLEMENTATION AND ANALYSIS then the particular instance/transaction is fraudulent.
Otherwise the transaction is considered. In our experiments
A. Data Sets we test different threshold values and present the related
In this study our experiments will be conducted on 3 data results.
sets. The European Dataset contains the transactions made Similar to Autoencoders, the restricted Boltzmann
by credit card users in 2 days in September 2013 [3]. All of machine (RBM) produces free energy that is again tested
the fields except time and amount have been PCA against a threshold to determine normal vrs fraudulent
transformed. It contains only 492 fraud instances out of transactions. The RBM model used here is created by
284,807 instances. Weiman Wang for fraud detection. [12]
Both the Australian Dataset and the German Dataset [4] For deep belief network, we use an adapted version of
are acquired from the UCI ML repository. The data sets were the model by AlbertUP, which is implemented in Tensorflow
anonymized such that no personal information is provided. for supervised and unsupervised pattern recognition
The Australian contains 383 normal instances and 307 problems [13].
fraud instances. In the same time the German data set With CNN, we convert the dataset into a 2D array
contains 1000 instances; of which 700 of them are normal instead of a 1D array. The data passes through a series of
instances and 300 of them are fraud instances. The European convolutional layers and max-pooling layers followed by a
dataset is considerably larger than the Australian and the layer to flatten the data. Finally, the data is classified at
German datasets. SoftMax layer. Fig 3 shows the architecture of the CNN we
Our experiments try to investigate the effectiveness of use.
different machine learning and deep learning models using
data set with varying size and complexities.
B. Experimental Setup
All of the implementation were conducted in Python and
using various libraries such as NumPy, Pandas, Keras,

336
Authorized licensed use limited to: Ontario Tech University. Downloaded on May 10,2025 at 22:45:24 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

IV. RESULTS
As follows we present the results obtained from the
experiments on the European Dataset, the Australian Dataset
and the German data set.
Fig. 3. CNN Architecture
A. European Dataset
We test our models using cross-validation and choose the
top three performing models and combine them using TABLE I. EUROPEAN DATASET RESULTS
majoring voting. The diagram below depicts the basic
Method
structure of the model.
MCC AUC Cost of Failure
RBM 0.176 0.9109 227360
Autoencoders 0.2315 0.8943 127220
Random Forest 0.7947 0.8507 30340
CNN 0.8096 0.8764 25700
SVM 0.8145 0.9004 21220
KNN 0.8354 0.8887 22660
Ensemble (KNN, 0.8226 0.8964 21740
SVM and CNN)

Table I summarizes the results obtained on the European

Dataset. In particular the table lists the Matthew Correlation
Coefficient (MCC) and the Area Under the Curve measure
(AUC) for a set of machine learning models.
RBM and AE have high false positives (false alarm rate)
and therefore perform poorly with respect to MCC and cost.
Random Forest has good AUC and MCC values. CNN,
SVM, and KNN have the best performance in terms of MCC
and AUC. We can see the SVM has the least in terms of cost
Fig. 4. Majority Voting Based Model Structure of failure while Autoencoders and RBM have the highest.
Random forest, although producing good results, is poor in
terms of the cost. The top three performing models for this
C. Evalulation Metrics data set are SVM, KNN and CNN.
As follows we describe the main evaluation criteria we The majority voting classifier is constructed from the top
consider in this study. 3 performing models. The ensemble method performs better
than SVM and CNN individually, however it has a similar
The Matthews Correlation Coefficient is a measure to cost to SVM. However, SVM has a better AUC value. Here,
evaluate the quality of a two class/binary classifier. It was the recommendation would be to choose SVM instead of the
proposed by Brain W. Matthews in 1975. The coefficient ensemble if the company is looking to reduce the cost as
returns +1 for a perfect prediction, while a value zero much as possible as the ensemble method would take longer
indicates a random prediction. Matthews Correlation is also time in terms of both training and testing while SVM has the
known as the phi coefficient. Davide Chicco mentions that least in terms of testing and training.
MCC is a much better measure than accuracy and F1 score
as the other two can be misleading because they do not B. Australian Dataset
consider all four values of the confusion matrix [14].
TABLE II. AUSTRALIAN DATASET RESULTS
ROC curve is the receiver operation characteristic. It
helps in determining the precision of the model because of Method
the imbalance in the dataset. ROC curve is basically plotting MCC AUC Cost of Failure
of TPR on the x axis against FPR on the y axis. Sometimes RBM 0.15 0.5546 24600
two ROC curves may have the similar Area under the curve Autoencoders 0.2318 0.6174 12220
(AUC), then in that case, we need to look further into the CNN 0.6408 0.8227 6430
finer details such as the Cost of failure. Random Forest 0.684 0.8416 4700
KNN 0.6905 0.8425 6460
The idea behind the cost of failure is that each of the DBN 0.6999 0.8441 6790
False Negatives (Frauds detected as Normal) would have a
SVM 0.7085 0.8551 3380
cost of $1000 and False Positives (Normal instances detected
Ensemble1 (KNN, 0.7144 0.8573 5290
as fraud) would have a cost of $100 to the company/entity. SVM, DBN)
We are using this method to evaluate the top three models Ensemble2 (KNN, 0.7281 0.8655 3470
because sometimes all three of them have very similar MCC SVM, Random
and AUC values. Similarly, the cost of the resulting Forest)
ensemble classifiers is also calculated.

337
Authorized licensed use limited to: Ontario Tech University. Downloaded on May 10,2025 at 22:45:24 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

Table II summarizes the results obtained on the similar to KNN. For the German dataset it had the lowest
Australian Dataset. We can see that RBM and AE have the cost as well.
worst performance of all the methods. SVM, DBN, and
KNN have the best in terms of AUC and MCC. Random Table IV summarizes the frequency of the individual
Forest and CNN also have good values compared to the models ranking at the top 3 performing models for all the
others. We choose two ensemble models. The first ensemble datasets. SVM consistently was among the best performing
model (Ensemble 1) the is composed of 3 classifiers: the models of all data sets. KNN also produces good results with
KNN, DBN, and SVM. A second ensemble (Ensemble 2) is both the large dataset and the smaller datasets.
build using the models with the least cost of failures: KNN, TABLE IV. TOP PERFORMING MODELS
SVM and Random Forest. RBM and AE have the maximum
cost of failure as they again have a lot of false positives due Method
to the threshold. Number of times in Top 3
Support Vector Machines 3 Times
Table II shows that using Ensemble 1 (KNN, SVM, K-Nearest Neighbors 2 Times
DBN) has improved the MCC and AUC performance Convolutional Neural Networks 2 Times
compared to single SVM and the other methods. However, Random Forest 2 Times
the cost is higher than the random forest and SVM. This is Deep Belief Network 1 Time
because KNN and DBN have a high cost of failure values,
which likely influenced the classification. Ensemble 2 (KNN,
SVM, Random forest) on the other hand, which was based V. CONCLUSION
on combining classifiers with the least cost of failure
methods, classifier achieved a higher MCC, AUC, and lower Research related to Fraud Detection has been around for
cost value. Here we can see that combining methods have over 20 years now and has used various methods from
provided better results overall and Ensemble 2 is the best manual checking to customer end authentication. Machine
method of all as it has the best MCC, AUC values and the learning models have also had wide successes in this area.
lower cost. Deep learning models have been recently adopted in many
applications enabled by the rise in higher computation power
and cheap computing cost.
C. German Dataset
This paper provides an empirical investigation comparing
various machine learning and deep learning models on
different data sets for the detection of fraudulent transaction.
TABLE III. GERMAN DATASET RESULTS The main aim of this study is to find insights of which
Method methods would best suitable for which type of datasets. As
MCC AUC Cost of Failure nowadays, many companies are investing in new techniques
RBM 0.0984 0.5524 14160 to improve their business this paper could potentially help
Autoencoders 0.139 0.5614 22640 practitioners and companies to better understand how
KNN 0.2487 0.6047 21100 different methods work on certain types of datasets.
DBN 0.2725 0.5873 23640 Our study reveals that to detect fraud, the best methods
Random Forest 0.2912 0.6437 16970 with larger datasets would be using SVMs, potentially
SVM 0.4038 0.6857 16400 combined with CNNs to get a more reliable performance.
CNN 0.4291 0.7056 14220 For the smaller datasets, ensemble approaches of SVM,
Ensemble (SVM, 0.4439 0.7011 15620 Random Forest and KNNs can provide good enhancements.
CNN, Random Convolutional Neural Networks (CNN) usually, outperforms
Forest)
other deep learning methods such as Autoencoders, RBM
and DBN.
Table III summarizes the results obtained for the German A limitation of this study is however that it only deals
Dataset. Examining the results of the table we can see that with detecting fraud in a supervised learning context.
SVM, Random Forest and CNN are the best models in terms Although supervised learning methods such as CNN, KNN,
of performance (AUC and MCC values). Random forest, Random Forest seem attractive and produce good results,
CNN and SVM also have a better cost of failure than other they do not work well for dynamic environments. Fraud
models. Therefore, we use an ensemble of these three models patterns typically change over time and would be hard to
to build the majority voting classifier. catch. New data sets would need to be collected and machine
In general, the results depicted in tables I, II and III for learning models need to be retrained.
the three data set investigated show that combing the top- Autoencoders provide a good solution in that case as they
performing models outperform single models. Ensemble are only trained on normal (i.e. non-fraudulent) traffic.
improvement is more obvious on the smaller data sets (the Fraudulent transactions are detected as deviation from the
German dataset and the Australian dataset). For the normal patterns. Although training of Autoencoders is
European dataset, it produces results a little less than SVM. initially quite costly, it can be useful for the labelling of data
Random Forest works best for smaller datasets. sets. Once enough data is labelled, it can be used to retrain or
Convolutional Neural Networks was found to be the best build other supervised models.
deep learning method as it produces good results for both
European and German dataset, while its performance on the
Australian dataset was the 4th best and it cost of failure was

338
Authorized licensed use limited to: Ontario Tech University. Downloaded on May 10,2025 at 22:45:24 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

VI. FUTURE WORK [3] Credit card fraud detection anonymized credit card transaction labeled
as fraudulent or genuine [Online]. Available:
The idea of implementing neural networks for fraud https://www.kaggle.com/mlg-ulb/creditcardfraud.
detection is worth further exploration. The main problem [4] Dheeru Dua and Casey Graff. UCI Machine Learning Repository.
with obtaining good NN models is the availability of 2017 [Online]. Available: http://archive.ics.uci.edu/ml/datasets/
appropriate training data. We are currently looking into more [5] Masoumeh Zareapoora, Pourya Shamsolmoalia. “Application of
data sets and availability of larger sized fraud transactions to Credit Card Fraud Detection: Based on Bagging Ensemble Classifier”
be able to train larger models. International Conference on Intelligent Computing, Communication
& Convergence (ICCC-2015). Procedia Computer Science 48 pp 679
We are also looking into other models of deep learning – 686. 2015.
networks. In particular Generative Adversarial Networks [6] Kuldeep Randhawa, Chu Kiong Loo, Manjeevan Seera, Chee Peng
(GANs) have been used for anomaly detection in the field of Lim, Asoke K. Nandi. “Credit card fraud detection using AdaBoost
and majority voting”. IEEE Access (Volume: 6), pp 14277 – 14284.
cybersecurity to detect cyber-attacks. Some research work 2018.
has used GANs to simultaneously train both the generator [7] Tom Sweers. “Autoencoding Credit Card Fraud”. Bachelor Thesis,
and discriminator on healthy eyes data to identify anomalies Radboud University. June 2018.
in the human eyes. This could be extended to the application [8] Z. Chen, C. K. Yeo, B. S. Lee, and C. T. Lau. “Autoencoder based
of fraud detection. network anomaly detection.” Wireless Telecommunications
Symposium, pp 1-5. 2018.
Fuzzy systems for anomaly/outlier detection can also be [9] Apapan Pumsirirat, Liu Yan. “Credit Card Fraud Detection using
paired with a neural network to achieve better results and to Deep Learning based on Auto-Encoder and Restricted Boltzmann
provide more interpretable models. Genetic programming Machine”. International Journal of Advanced Computer Science and
can also be used to provide optimization for neural network Applications, Vol. 9, No. 1, 2, pp 18-25. 2018.
parameters and architecture. [10] Alae Chouiekha, EL Hassane Ibn EL Haj. “ConvNets for Fraud
Detection analysis”. Procedia Computer Science 127, pp.133–138.
In general fraudsters keep getting intelligent and it will 2018.
remain challenging to catch them. Therefore, the [11] S. Maes, K. Tuyls, B. Vanschoenwinkel, B. Manderick. “Credit Card
development of hybrid, dynamic and adaptable systems for Fraud Detection Using Bayesian and Neural Networks”. 2002.
fraud detection will pose many open areas of research in the [Online]. Available:
https://www.researchgate.net/publication/2524707_Credit_Card_Frau
future. d_Detection_Using_Bayesian_and_Neural_Networks.
REFERENCES [12] Weiman Wang, Restricted Boltzmann Machine. GitHub. Aug 2017.
[Online] Available: https://github.com/aaxwaz/Fraud-detection-using-
[1] European Central Bank, “Fifth report on card fraud, September deep-learning/blob/master/rbm/rbm.py.
2018,” 26 September 2018. [Online], Available: [13] Albertbup, “A Python implementation of Deep Belief Networks built
https://www.ecb.europa.eu/pub/cardfraud/html/ecb.cardfraudreport20 upon NumPy and TensorFlow with scikit-learn compatibility”.
1809.en.html. GitHub. October 2018. [Online] Available:
[2] N. Report, "The Nilson Report," The Nilson Report, 2016. [Online]. https://github.com/albertbup/deep-belief-network.
Available: www.nilsonreport.com. [14] D. Chicco, "Ten quick tips for machine learning in computational
biology". BioData Mining, December 2017, pp 1–17.

339
Authorized licensed use limited to: Ontario Tech University. Downloaded on May 10,2025 at 22:45:24 UTC from IEEE Xplore. Restrictions apply.