See discussions, stats, and author profiles for this publication at: [Link]

net/publication/376520239

Modern Deep Learning Techniques for Credit Card Fraud Detection: A Review
(2019 to 2023)

Research · December 2023

DOI: 10.13140/RG.2.2.32173.67043

CITATIONS READS

0 734

1 author:

Rajat Banger
Bournemouth University
1 PUBLICATION 0 CITATIONS

SEE PROFILE

All content following this page was uploaded by Rajat Banger on 15 December 2023.

The user has requested enhancement of the downloaded file.

 Modern Deep Learning Techniques for Credit Card
Fraud Detection: A Review (2019 to 2023)
Rajat Banger
dept. computing and Informatics
Bournemouth University
Bournemouth, United Kingdom
[Link]

Abstract— This paper presents a review of different deep 2021. Thus, considering the necessity of improvement, this
learning approaches for credit-card fraud detection, proposed paper provides the latest and in-depth review of the literature
between 2019 and 2023 inclusive. The paper discusses the dealing with CCFD, focused particularly on deep learning.
approaches and findings of the different literature from two The motive of this paper is to cover the gap in available review
journal databases: IEEE and Elsevier. The study discusses the literature on CCFD and to provide a detailed and up-to-date
importance of different techniques like data sampling, feature source of information to the researchers and readers, who are
extraction, and model architecture in credit card fraud interested in the advancement of deep learning in CCFD.
detection (CCFD). This review finds that unbalanced data is one
of the challenges in CCFD and discusses the data balancing This paper contributes to the following analysis of all the
approaches like Synthetic Minority Over-sampling Technique covered literature:
(SMOTE), SMOTE-Edited Nearest Neighbour (SMOTE-ENN),
Generative Adversarial Network (GAN), Variational Auto • Analysis of the problem the research is focused on.
Encoder (VAE) and GAN (VAE-GAN), Conditional VAE
• The information of the dataset used.
(CVAE) which help to deal this challenge. Also, this paper helps
to learn the implementation of feature extraction techniques to • The approach followed for implementation and testing
improve the performance of the CCFD model for instance, Full of the proposed model.
Centre Loss (FCL), 2-layer autoencoder and discusses various
architecture models like Attention-based models and • The evaluation methodology followed to check the
autoencoders to improve the CCFD. This review paper covers efficiency of the model.
the gap in the existing literature review on CCFD and discusses
the deep learning approaches that are not included in previous • Effectiveness of the proposed model.
research papers. The rest of the paper is organized in different sections. The
next section provides the methodology of the paper. The third
section presents the literature review, which is a major part of
Keywords— credit card fraud detection, deep learning, review, this paper. The fourth section is a discussion, which includes
neural network a detailed analysis of all the included literature. The fifth
section discusses the limitations of this paper and the last
section concludes the paper with the key findings and
I. INTRODUCTION suggestions for future research.
Credit card is a lucrative mode of payment as users get
easy and flexible payment options like buy now pay later, EMI
(Equated Monthly Instalments) payments, and other rewards
from the card providers. This along with some other benefits, II. METHODOLOGY
contributes to many active credit-card users. However, there The research methodology followed in this study is
are threats of frauds that can potentially lead users to financial represented in the flowchart shown in Fig 1.
losses. To check and prevent the credit card frauds, credit card
fraud detection (CCFD) has become important for financial
institutions. Credit card transactions are imbalanced as the
ratio of illegitimate to legitimate transactions is significantly
lower [1] and analysis of large volumes of transactions are two
of the challenges in CCFD. Researchers have suggested
various artificial intelligence (AI) approaches which can deal
with these challenges because of their high computation
power and ability to accurate predictions.
Al-Hashedi and Magalingam in their study [2] have
discussed the proposed fraud detection techniques for
different domains of the financial sector, between 2009 and
2019. Similarly, Hilal, Gadsden and Yawney in their study [3]
have covered different techniques for anomaly detection in
financial fraud. Both literatures are diverse towards various
domains of the financial sector and thus lack to provide a deep
and focused review on a particular domain. The authors,
Cherif et al. have discussed the work done on CCFD in their Fig. 1. Flowchart of Research Methodology followed for the research.
systematic review literature [4] but their study is limited to
machine learning and consists of journals between 2015 to

[Link]
 This research starts with the research definition and then artificially produced data may impact the precision of the
inclusion and exclusion criteria are set for the journals such model on real-world data. In another study [7], the GAN
that only published journals available on two popular model outperformed the SMOTE in terms of sensitivity but an
databases: IEEE and Elsevier, are considered. Another filter is increase in false positive rate is also observed which might be
related to the scope of the study, for that, the journals focused because of synthesized data. Both GAN and SMOTE are
on deep learning in CCFD between 2019 to 2023 inclusive are outperformed by another model, Variational Auto Encoder
included. The reason to bind the study under these criteria is (VAE) and Deep Neural Network (DNN) which is proposed
to focus on decent quality research that is latest and peer- by Tingfei, Guangquan and Kuihua in their study [8]. The
reviewed. Then the analysis of the included research paper is proposed model has performed 3 percent better than both
performed following the methodology of the paper. models when evaluated on confusion matrices, sensitivity and
specificity parameters using a European dataset. As the
III. LITERATURE REVIEW proposed model is a supervised model thus may perform
differently in real-world scenarios where the environment is
Credit card is an electronic mode of payment, and to make unsupervised. Another study [9] has addressed the same
it a reliable and secure payment environment, multiple challenge of unbalanced data, in which the author has
authorizations and fraud prevention system are included in its proposed an ensemble approach ESMOTE-GAN based on
process. In Fig 2. the general model of the CCFD system is SMOTE and GAN. Here SMOTE generates the initial
described. Here, the “Predictive Model” block is the point samples and the GAN model is trained on that to create a
where CCFD is performed which is a focus of all the studies synthesized dataset. The model has outperformed the other
included in this paper. models used for comparison and has improved the
Unbalanced data is a challenge while developing an AI performance by 3.2 percent in fraud detection rate with zero
model for CCFD. Esenogho et al. in their research [5] have false alarm rate, but this approach has a limitation that it leads
suggested a solution to this problem where they have proposed to overfitting of the model and the reason is the similarity of
a data sampling method, Long Short-Term Memory (LSTM) the generated samples to the minority class of the original
along with SMOTE-Edited Nearest Neighbour (SMOTE- dataset. This challenge is overcome by the research [10],
ENN). Here, SMOTE (Synthetic Minority Over-sampling where the authors have proposed data sampling by the
Technique) is responsible for the oversampling of the dataset SMOTE-ENN technique. This model is a stacking ensemble
and ENN performs sampling on that dataset, this technique model in which LSTM along GRU acts as a base learner and
also protects the overfitting of the model. The efficiency of MLP as a meta learner, which is trained on the dataset sampled
this technique is checked by comparing its performance with by the SMOTE-ENN approach and stratified 10-fold cross-
the performance of the model on the original dataset. Also, a validation technique is used to test the model on unseen data.
stratified 10-fold cross-validation technique is used to The model has outperformed other models and has achieved
evaluate the performance of the model on unseen data. The 100 percent sensitivity and 99.7 percent specificity. In paper
proposed model has outperformed the other models and has [11], another data sampling technique is proposed where the
achieved 0.996 sensitivity and 0.998 specificity, which challenge of unbalanced data is tackled by using CVAE which
reflects that the accuracy of the model is reliable. Also, the outperforms both SMOTE and Random Oversampling
model has outperformed other existing state-of-the-art models techniques. This study is evaluated using the FID analysis test
when compared using two datasets, the Taiwan credit card on MNIST and Fashion MNIST datasets. The research [12] by
dataset, and the German credit dataset. Almarshad, Gashgari, Ding et al. have introduced another data sampling approach,
and Alzahrani in their study [6] have proposed another data VAE-GAN in which fake data is produced by the model when
sampling approach, Generative Adversarial Network (GAN). it is trained on the minority data. This model has outperformed
In this technique, the minority data is given as input to the other data sampling techniques GAN, VAE, and SMOTE
GAN which generates more similar data and hence results in when compared with the implementation of synthesized
data balancing. The model has attained a 0.989 AUC value datasets on Logistic Regression, Decision Tree, Random
which is highest among all the compared models. Here, fine- Forest, Neural Network, and XGBoost models. There is
tuning on the produced dataset is performed to match the another study [13] where the GAN GAN-based model, MIMO
similarity of the generated data to the original data, but the ADV-O and TimeGAN ADV-O is proposed for data
sampling. In this study, the anomaly is detected by considering
the fraud as time-series and the proposed model has
successfully predicted the behaviour of the fraudster. The
details of this literature are described in TABLE I.
Misra et al. have advocated feature extraction technique in
their research [14], where they have proposed the use of a 2-
layer Autoencoder to extract lower dimension features from
the dataset and the new dataset is used to train three
classification models such as Multilayer Perceptron (MLP),
K-Nearest Neighbour (KNN), Logistic Regression (LR). The
proposed technique has outperformed the state-of-art data
sampling techniques. Another study [15] has also proposed a
similar model, Deep Auto-Encoder and Deep Neural Network
which has outperformed the PCA models. Cheng et al. in their
study [16] has proposed another feature extraction technique,
Spatial- Temporal Attention-based Neural network (STAGN).
Fig. 2. General model representing Credit Card Fraud Detection [16].
In this technique, the spatial aggregation collects the
 TABLE I. EVALUATION MATRIX FOR DEEP LEARNING ALGORITHMS BASED ON DATA SAMPLING APPROACH.

Proposed
Reference Title Year Dataset Evaluated Against Evaluation Method
Model
A Neural Network Ensemble with Support vector machine (SVM),
European_Dataset + LSTM with Sensitivity, Specificity and
[5] Feature Engineering for Improved 2022 multilayer perceptron (MLP),
Taiwan + German SMOTE-ENN AUC
Credit Card Fraud Detection decision tree, traditional AdaBoost.
Generative Adversarial Networks- Un-supervised Attentional Anomaly Correlational matrix,
Based Novel Approach for Fraud Detection-Fraud Detection Network Confusion matrices,
[6] 2023 European_Dataset GAN
Detection for the European (UAAD-FDNet), Random Forest Precision, Recall and F1-
Cardholders 2013 Dataset (RF), AdaBoost-LGBM. The score
Using generative adversarial
Sensitivity, Specificity,
networks for improving
[7] 2019 European_Dataset GAN SMOTE Precision, F-measure, and
classification effectiveness in
the accuracy.
Credit Card Fraud Detection
Using Variational Auto Encoding Confusion matrices,
[8] 2020 European_Dataset VAE-DNN GAN, SMOTE
in Credit Card Fraud Detection Sensitivity, Specificity
Ensemble Synthesized Minority
Oversampling-Based Generative
ESMOTE- KNN, CART, SVM, ANN, LR, RF, Confusion matrices and
[9] Adversarial Networks and 2023 European_Dataset
GAN and XGBoost. False alarm rate
Random Forest Algorithm for
Credit Card Fraud Detection
Sensitivity, Specificity,
A Deep Learning Ensemble with
European_Dataset + LSTM-GRU + AdaBoost, Random Forest, MLP, ROC (Receiver Operating
[10] Data Resampling for Credit Card 2023
Taiwan MLP LSTM, GRU Characteristic Curve) and
Fraud Detection
AUC (Area Under Curve)]
On oversampling imbalanced data
[11] with deep conditional generative 2021 MNIST CVAE SMOTE and Random oversampling FID analysis
models
GAN, VAE, SMOTE and Five
Credit Card Fraud Detection models are implemented using the
Based on Improved Variational sampled dataset Logistic Recall, F1-score, PR curve
[12] 2023 European_Dataset VAE-GAN
Autoencoder Generative Regression, Decision Tree, Random and ROC curve
Adversarial Network Forest, Neural Network, and
XGBoost
An Adversary Model of MIMO ADV-
Balanced Random Forest, Random
fraudsters’ behaviour to Improve O and Precision, Recall and F1-
[13] 2023 Worldline oversampling, SMOTE, K-Means
Oversampling in credit card fraud TimeGAN score
SMOTE, and CT-GAN
detection ADV-O

information of the location through graph-based features and This model has outperformed other models in terms of
temporal deals with the time of the transactions. This approach precision and recall on three three-month datasets and
has performed better than all other compared models. Li, Liu, average value as well when evaluated by taking the means of
and Jiang in their research [17], have proposed the Full Centre ten runs. In another research [20] LSTM based model,
Loss (FCL) and Deep Neural Network (DNN) model which ensemble-LSTM is proposed which has three phases: base
has outperformed all the compared models. The performance classifier training, middle vector generation and voting
of the model is evaluated using parameters like F1-score and classifier training. The proposed model has outperformed
AUC_PR whose values are 0.805 and 0.879 on the European other state-of-art models like LSTM, GRU, ensemble GRU
dataset and, 0.813 and 0.825 on the private dataset,
and Ensemble model in terms of accuracy but in terms of
respectively. The study has made a unique contribution by
efficiency, Ensemble GRU has outperformed all the models
discussing the importance of the loss function in feature
extraction and its impact on the model's performance. The along with the proposed model. A popular deep learning
comparison of the evaluation matrices of the above-discussed model Convolutional Neural Network (CNN) is proposed by
models is shown in TABLE II. Alarfaj et al. in their study [21]. The model with 20 layers has
outperformed the other models and has achieved an accuracy
Xie et al. in their research [18] have proposed a Time-aware of 99.72 percent. However, the implementation of the model
Attention-based Interactive Long Short-Term Memory (TAI- can be questioned as biased towards the proposed model
LSTM) model which captures the transactional behaviour of because, in the CNN model, a sampling layer is introduced to
the card owner. The performance of the model is tested deal with imbalanced data whereas no data balancing is
through the Holm One and Friedman test and it is evaluated introduced for machine learning models. Another targeted
on confusion matrices where it has outperformed the other question on the implementation of the model is the removal
competitive models. The approach of this model is based on of fraudulent transactions from training data during the
trade-time where the model extracts the information related sampling layer of the proposed model. RB and KR have also
to the time intervals of the transactions. Another time- proposed a similar model, Artificial Neural Network (ANN)
dependent research [19], has proposed Time-aware in their study [22]. The proposed model has achieved the
Historical-attention-based LSTM (TH-LSTM) model. The highest accuracy, which is 99.92 percent, but SVM has
proposed model tracks the behavioural changes over the outperformed the other models for precision and recall and
repetitive transactions in the past. Here, the dataset is divided has achieved 97.43 percent and 89.76 percent, respectively.
monthly, and the model is evaluated multiple times such as As precision and recall are important parameters in CCFD
one month will be a training set, and another is a testing set. this model may not perform well in the real environment.
 TABLE II. EVALUATION MATRIX FOR DEEP LEARNING MODELS BASED ON FEATURE EXTRACTION.

Evaluation
Reference Title Year Dataset Proposed Model Evaluated Against
Method
An Autoencoder-Based A combination of a 2-
Accuracy,
Model for Detecting layer Autoencoder Classification models, Different sampling
[14] 2020 European_Dataset Precision, Recall,
Fraudulent Credit Card with a classification techniques + classification models
F1-score
Transaction model
A novel combined approach
Recall, Precision,
based on deep Autoencoder Deep Auto-Encoder
[15] 2023 European_Dataset PCA and DNN F1-Score, AUC-
and deep classifiers for credit and DNN
ROC, AUC-PR
card fraud detection
Logistic Regression, GBDT (Gradient-
Graph Neural Network for Spatial-temporal Precision-recall
Boosting), MLP, Deep and wide, CNN-
[16] Fraud Detection via Spatial- 2020 Industrial attention-based neural curves,
max, Ada-BM, LSTM-seq, STAGN-
Temporal Attention network (STAGN) Sensitivity, AUC
notrans, STAN STAGN-notemp
Softmax Loss (SL), Large Margin Softmax
Deep Representation
Loss (LMSL), Angular Softmax Loss
Learning with Full Centre European_Dataset + F1 score and
[17] 2020 FCL with DNN (ASL), Large Margin Cosine Loss
Loss for Credit Card Fraud Private Company AUC_PR
(LMCL), Additive Angular Margin Loss
Detection
(AAML), Angle Centre Loss (ACL)

In the study [23] by Lei et al. a Distributed Deep Neural approach without sharing their data which improves data
Network Model (DDNN) is proposed for CCFD. The privacy and reduces data handling cost. In another study [24]
accuracy, precision, recall and F1-score of the model are based on DNN, a combination model of CatBoost and Deep
99.9422, 99.9649, 99.9772 and 99.9710, respectively. This Neural Networks is proposed where SMOTE is used for data
model has outperformed other models on all the parameters balancing and PCA is done for feature extraction. The model
except recall, where SVM has performed better and has is account-based thus all the users are separated into new and
achieved a 99.9947 percent recall value. Here, the idea of old categories such that if there is a fraud transaction available
parallel computing is introduced to train the model allowing in the history of any account every new transaction is also
multiple companies to work together via a distributed flagged as Fraud until there is no report of fraud in 120 days

TABLE III. EVALUATION MATRIX FOR DEEP LEARNING MODELS BASED ON NEURAL NETWORK ARCHITECTURE.

Proposed
Reference Title Year Dataset Evaluated Against Evaluation Method
Model
Time-Aware Attention-Based
LSTM, GRU, Bi-RNN, TAH-
Gated Network for Credit Card
[18] 2022 China TAI-LSTM RNN, TLSTM, HAInt-LSTM, Confusion Matrices
Fraud Detection by Extracting
Ug-RNN, STA-NN
Transactional Behavior

STA-NN, A-LSTM, HOBA,

Learning Transactional Behavioral
European_Dataset + A-CNN, LSTM, GRU, Bi-
[19] Representations for Credit Card 2022 TH-LSTM Confusion Matrices
China RNN, UGRNN, HAINT-
Fraud Detection
LSTM, T-LSTM, TAH-LSTM

Ensemble of deep sequential

European_Dataset + Ensemble- LSTM, GRU, ensemble GRU Recall, Precision, and F1, AUC-
[20] models for credit card fraud 2021
Brazilian LSTM and ensemble model ROC, AUC-PR
detection
Decision Tree, KNN
Credit Card Fraud Detection Using
CNN with algorithm, Logistic
[21] State-of-the-art Machine Learning 2022 European_Dataset Confusion matrices
twenty layers Regression, SVM, Random
and Deep Learning Algorithms
Forest Tree and XG Boost.
Credit card fraud detection using Confusion matrices, Precision,
[22] 2021 European_Dataset ANN SVM, K-NN
artificial neural network Recall, Accuracy
Accuracy, Precision, Recall, F1-
A distributed deep neural network CDNN model, Logistic
score, Receiver operating
[23] model for credit card fraud 2023 European_Dataset DDNN Regression, SVM, Decision
characteristic curve (ROC), Area
detection Tree
under ROC curve (AUC)
A Proposed Model for Card Fraud CatBoost and
ROC-AUC and Precision-Recall
[24] Detection Based on CatBoost and 2022 Vesta Neural XGBoost and LightGBM
Curve
Deep Neural Network Networks

Champion-challenger analysis for Feed-Forward K-S statistics and AUROC,

[25] credit card fraud detection: Hybrid 2019 South_korea Neural Hybrid Ensemble Learning precision, recall, and cost
ensemble and deep learning Network reduction rate.
and after that, it automatically gets converted into not a comparative study to check the performance of the
legitimate account. The accuracy of the model is validated proposed model, Hierarchical Temporal Memory based on the
through 10-fold cross-validation and the model is compared Cortical Learning Algorithms (HTM-CLA), against ANN
against XGBoost and LightGBM models on ROC-AUC and trained by the Simulated Annealing technique (SA-ANN) and
Precision-Recall curve. CatBoost has achieved 0.974 AUC Long Short-Term Memory ANN (LSTM-ANN). According
and DNN has achieved 0.84 AUC. The ideology of the model to the results, the proposed model is outperformed by SA-
is interesting, but the proposed model is not straightforward as ANN which has obtained an accuracy of 85.71 percent
both models are implemented separately and give different whereas LSTM-ANN is the least performer.
outputs. Instead of a combined model, it depicts the
comparison of two models. One more study [25] has This review of twenty-four pieces of literature helps readers
suggested a neural network-based model, the Feed-forward to develop knowledge about the effectiveness of different
Neural network model which has outperformed the Hybrid deep learning models in CCFD and about the implementation
Ensemble Learning Model on all parameters like recall process to develop an efficient model. Also, this has helped to
(transaction), recall (card), and cost reduction rate by 3.8 understand the importance of data sampling techniques and
percent, 2.1 percent and 5.5, respectively. The techniques of feature extraction methods. This review has compared various
similar approaches which will help researchers and
all the neural network-based proposed models for CCFD are
available in TABLE III. practitioners to opt for a better approach.

TABLE IV discusses the approach of a few papers where IV. Discussion

system design is unique. One of these research [26] has
proposed an enhanced Deep Belief Network (DBN) with CCFD is an area which demands continuous development
Homogeneity Oriented Behaviour Analysis (HOBA). Here, and adaptation to emerging frauds, as fraudsters constantly
the HOBA technique is implemented to analyse the past innovate new tactics and strategies to exploit vulnerabilities in
behaviour of the customers which is used for feature financial systems. Thus, it is an attractive area of study for
engineering. In this approach, unauthorized transactions and researchers. As shown in Fig 3, the graph of the yearly
high-risk authorized transactions are rejected at the contribution of researchers in CCFD with deep learning is in
authorization step of the model and the alert has been shared inclining form for the past 5 years such that the highest
with the monitoring department through the alert management contribution made is in the year 2023 which individually sums
module. The model is trained offline and tested online on real to the eight submitted papers on IEEE and Elsevier only.
data. The approach has two set limits of false positive rate, According to the inclusion criteria of this study, the total
first, if the value is below 1 percent the model is good, and the share of published research papers on IEEE is 54 percent and
maximum acceptable value is 3 percent. HOBA-DBN model on Elsevier is 46 percent as illustrated in Fig 4. Analysis of
has outperformed other models with an F1-score 0.577 and datasets used in the reviewed literature depicts that all the
58.33 percent recall. The accuracy of the model is validated datasets are highly imbalanced and consist of a small number
through 10-fold cross-validation. In another research [27], of illegitimate transactions. From the bar chart in Fig. 5, it is
Habibpour has proposed an approach to find uncertainties in observed that ten studies are conducted on European_Dataset
the prediction of the CCFD model and for these three only, which belongs to Université Libre de Bruxelles (ULB)
techniques of uncertainty quantification Monte Carlo and has a total of 284,807 transactions out of which 492 are
Dropout, Ensemble, and Ensemble Monte Carlo Dropout are fraudulent transactions. Five studies are there which have used
compared using confusion matrices, where two different at least two datasets in their study out of which one is
confusion matrices, certain and uncertain categories are European_Dataset, many of them have conducted their search
considered. Here, both MCD and ensemble models have by using real-world data. The datasets from financial
resulted in improvement in the performance of the DNN institutions of China and Vesta are also used in two studies
model. Osegi and Jumbo in their research [28] have conducted each.

TABLE IV. THE EVALUATION MATRIX FOR DEEP LEARNING MODELS IS CLASSIFIED BASED ON THE MODEL ARCHITECTURE.

Reference Title Year Dataset Proposed Model Evaluated Against Evaluation Method
HOBA: A novel feature engineering
Accuracy, Precision,
methodology for credit card fraud BPNN+RFM, SVM+RFM and
[26] 2021 China HOBA F1-score, Recall, False
detection with a deep learning RF+RFM
Positive Rate
architecture
Monte Carlo dropout, Two different
Uncertainty-aware credit card fraud ensemble, and confusion matrices:
[27] 2023 Vesta compared with each other
detection using deep learning ensemble Monte certain and uncertain
Carlo dropout categories

Comparative analysis of credit card AN ANN trained by the Simulated

fraud detection in Simulated Annealing Australian- Annealing technique (SA-ANN),
[28] 2021 HTM-CLA Accuracy
trained Artificial Neural Network and German Long Short-Term Memory ANN
Hierarchical Temporal Memory (LSTM-ANN)
 are sensitivity, specificity and AUC which provide great
insights into the performance of the model.

V. LIMITATIONS:
The study is limited to Deep learning models proposed in
the last five years (2019-2023) for CCFD and considers
research from only two editorial journals (IEEE and Elsevier)
which are published in the English language and are peer-
reviewed.

VI. CONCLUSION:
Data sampling techniques provide a solution to the
problem of unbalanced data but the challenge of overfitting
Fig. 3. Graph representing the number of literatures published annually on
credit card fraud detection with deep learning. the model has arisen. From the study of models based on
feature extraction techniques, it is observed that the overfitting
of the model can be avoided, and these models have also
performed better than models based on data sampling
techniques. After Evaluating the performance of the neural
network models, it is observed that the combination of neural
networks with other techniques helps to increase the
performance of the model. This paper has also disclosed the
importance of features like time and location for behavioural
analysis of transactions which is important for proper training
of the model. This paper paves the path for future studies the
learnings from this paper can be utilized for the development
of a better CCFD model. One suggestion for future studies is
to explore the scope of distributed deep learning to implement
different data sampling techniques at different nodes of
parallel computing.

Fig. 4. Chart representing the share of a particular database in selected REFERENCES

literature.
[1] W. Ning, S. Chen, S. Lei, and X. Liao, “AMWSPLAdaboost Credit
Card Fraud Detection Method Based on Enhanced Base Classifier
Diversity,” IEEE Access, vol. 11, pp. 66488–66496, 2023, doi:
10.1109/ACCESS.2023.3290957.
[2] K. G. Al-Hashedi and P. Magalingam, “Financial fraud detection
applying data mining techniques: A comprehensive review from 2009
to 2019,” Comput Sci Rev, vol. 40, p. 100402, May 2021, doi:
10.1016/[Link].2021.100402.
[3] W. Hilal, S. A. Gadsden, and J. Yawney, “Financial Fraud: A Review
of Anomaly Detection Techniques and Recent Advances,” Expert Syst
Appl, vol. 193, p. 116429, May 2022, doi:
10.1016/[Link].2021.116429.
[4] A. Cherif, A. Badhib, H. Ammar, S. Alshehri, M. Kalkatawi, and A.
Imine, “Credit card fraud detection in the era of disruptive
technologies: A systematic review,” Journal of King Saud University -
Computer and Information Sciences, vol. 35, no. 1, pp. 145–174, Jan.
2023, doi: 10.1016/[Link].2022.11.008.
[5] E. Esenogho, I. D. Mienye, T. G. Swart, K. Aruleba, and G. Obaido,
Fig. 5. Graph for comparison of datasets used in the reviewed literature. “A Neural Network Ensemble With Feature Engineering for Improved
Credit Card Fraud Detection,” IEEE Access, vol. 10, pp. 16400–16407,
The comparison of the reviewed literatures discusses the 2022, doi: 10.1109/ACCESS.2022.3148298.
performance of various deep learning models in tackling the [6] F. A. Almarshad, G. A. Gashgari, and A. I. A. Alzahrani, “Generative
Adversarial Networks-Based Novel Approach for Fraud Detection for
imbalanced data issue for CCFD with the help of different the European Cardholders 2013 Dataset,” IEEE Access, vol. 11, pp.
sampling techniques like SMOTE, ENN, GAN, and VAE- 107348–107368, 2023, doi: 10.1109/ACCESS.2023.3320072.
GAN. It also highlights the importance of feature extraction [7] U. Fiore, A. De Santis, F. Perla, P. Zanetti, and F. Palmieri, “Using
techniques such as FCL, Autoencoders and TAI-LSTM in generative adversarial networks for improving classification
extracting key features for the CCFD model. This is also effectiveness in credit card fraud detection,” Inf Sci (N Y), vol. 479,
observed that the selection of evaluation matrices plays a key pp. 448–455, Apr. 2019, doi: 10.1016/[Link].2017.12.030.
role in testing the performance of the model for example in the [8] H. Tingfei, C. Guangquan, and H. Kuihua, “Using Variational Auto
Encoding in Credit Card Fraud Detection,” IEEE Access, vol. 8, pp.
case of the European dataset the legitimate transactions are 149841–149853, 2020, doi: 10.1109/ACCESS.2020.3015600.
more than 99 percent thus selection of accuracy matrices for
[9] F. A. Ghaleb, F. Saeed, M. Al-Sarem, S. N. Qasem, and T. Al-
this is not an optimal choice as the performance will not be Hadhrami, “Ensemble Synthesized Minority Oversampling-Based
critically evaluated. The commonly used evaluation matrices Generative Adversarial Networks and Random Forest Algorithm for
 Credit Card Fraud Detection,” IEEE Access, vol. 11, pp. 89694–89710, [19] Y. Xie, G. Liu, C. Yan, C. Jiang, M. Zhou, and M. Li, “Learning
2023, doi: 10.1109/ACCESS.2023.3306621. Transactional Behavioral Representations for Credit Card Fraud
[10] I. D. Mienye and Y. Sun, “A Deep Learning Ensemble With Data Detection,” IEEE Trans Neural Netw Learn Syst, pp. 1–14, 2022, doi:
Resampling for Credit Card Fraud Detection,” IEEE Access, vol. 11, 10.1109/TNNLS.2022.3208967.
pp. 30628–30638, 2023, doi: 10.1109/ACCESS.2023.3262020. [20] J. Forough and S. Momtazi, “Ensemble of deep sequential models for
[11] V. A. Fajardo et al., “On oversampling imbalanced data with deep credit card fraud detection,” Appl Soft Comput, vol. 99, p. 106883, Feb.
conditional generative models,” Expert Syst Appl, vol. 169, p. 114463, 2021, doi: 10.1016/[Link].2020.106883.
May 2021, doi: 10.1016/[Link].2020.114463. [21] F. K. Alarfaj, I. Malik, H. U. Khan, N. Almusallam, M. Ramzan, and
[12] Y. Ding, W. Kang, J. Feng, B. Peng, and A. Yang, “Credit Card Fraud M. Ahmed, “Credit Card Fraud Detection Using State-of-the-Art
Detection Based on Improved Variational Autoencoder Generative Machine Learning and Deep Learning Algorithms,” IEEE Access, vol.
Adversarial Network,” IEEE Access, vol. 11, pp. 83680–83691, 2023, 10, pp. 39700–39715, 2022, doi: 10.1109/ACCESS.2022.3166891.
doi: 10.1109/ACCESS.2023.3302339. [22] A. RB and S. K. KR, “Credit card fraud detection using artificial neural
[13] D. Lunghi, G. M. Paldino, O. Caelen, and G. Bontempi, “An adversary network,” Global Transitions Proceedings, vol. 2, no. 1, pp. 35–41, Jun.
model of fraudsters’ behaviour to improve oversampling in credit card 2021, doi: 10.1016/[Link].2021.01.006.
fraud detection,” IEEE Access, pp. 1–1, 2023, doi: [23] Y.-T. Lei, C.-Q. Ma, Y.-S. Ren, X.-Q. Chen, S. Narayan, and A. N. Q.
10.1109/ACCESS.2023.3337635. Huynh, “A distributed deep neural network model for credit card fraud
[14] S. Misra, S. Thakur, M. Ghosh, and S. K. Saha, “An Autoencoder detection,” Financ Res Lett, vol. 58, p. 104547, Dec. 2023, doi:
Based Model for Detecting Fraudulent Credit Card Transaction,” 10.1016/[Link].2023.104547.
Procedia Comput Sci, vol. 167, pp. 254–262, 2020, doi: [24] N. Nguyen et al., “A Proposed Model for Card Fraud Detection Based
10.1016/[Link].2020.03.219. on CatBoost and Deep Neural Network,” IEEE Access, vol. 10, pp.
[15] H. Fanai and H. Abbasimehr, “A novel combined approach based on 96852–96861, 2022, doi: 10.1109/ACCESS.2022.3205416.
deep Autoencoder and deep classifiers for credit card fraud detection,” [25] E. Kim et al., “Champion-challenger analysis for credit card fraud
Expert Syst Appl, vol. 217, p. 119562, May 2023, doi: detection: Hybrid ensemble and deep learning,” Expert Syst Appl, vol.
10.1016/[Link].2023.119562. 128, pp. 214–224, Aug. 2019, doi: 10.1016/[Link].2019.03.042.
[16] D. Cheng, X. Wang, Y. Zhang, and L. Zhang, “Graph Neural Network [26] X. Zhang, Y. Han, W. Xu, and Q. Wang, “HOBA: A novel feature
for Fraud Detection via Spatial-Temporal Attention,” IEEE Trans engineering methodology for credit card fraud detection with a deep
Knowl Data Eng, vol. 34, no. 8, pp. 3800–3813, Aug. 2022, doi: learning architecture,” Inf Sci (N Y), vol. 557, pp. 302–316, May 2021,
10.1109/TKDE.2020.3025588. doi: 10.1016/[Link].2019.05.023.
[17] Z. Li, G. Liu, and C. Jiang, “Deep Representation Learning With Full [27] M. Habibpour et al., “Uncertainty-aware credit card fraud detection
Center Loss for Credit Card Fraud Detection,” IEEE Trans Comput Soc using deep learning,” Eng Appl Artif Intell, vol. 123, p. 106248, Aug.
Syst, vol. 7, no. 2, pp. 569–579, Apr. 2020, doi: 2023, doi: 10.1016/[Link].2023.106248.
10.1109/TCSS.2020.2970805. [28] E. N. Osegi and E. F. Jumbo, “Comparative analysis of credit card
[18] Y. Xie, G. Liu, C. Yan, C. Jiang, and M. Zhou, “Time-Aware fraud detection in Simulated Annealing trained Artificial Neural
Attention-Based Gated Network for Credit Card Fraud Detection by Network and Hierarchical Temporal Memory,” Machine Learning with
Extracting Transactional Behaviors,” IEEE Trans Comput Soc Syst, Applications, vol. 6, p. 100080, Dec. 2021, doi:
vol. 10, no. 3, pp. 1004–1016, Jun. 2023, doi: 10.1016/[Link].2021.100080.
10.1109/TCSS.2022.3158318.

View publication stats