A systematic review of literature on credit card cyber fraud detection using machine and deep learning

Research questions

This review attempts to summarise and analyse the ML and DL credit card cyber fraud detection algorithms from 2019 to 2021. The following research questions (RQs) are therefore posed:

RQ1: What ML/DL techniques are utilised in detection of credit card cyber fraud? This question aims to specify the ML/DL techniques that have been applied.

RQ2: What percentage of credit card cyber fraud detection articles discussed supervised, unsupervised, or semi-supervised techniques? This question seeks to determine the proportion of research articles that employ supervised, unsupervised, and semi-supervised credit cyber fraud detection techniques.

RQ3: What is the estimated overall performance and outcomes of ML/DL models? This question focuses on ML/DL model performance estimation and model results.

RQ4: What are the research trends, gaps, and potential future directions for cyber fraud detection in credit card? The question guides to uncover research trends, gaps in the existing literature, and future direction of credit card cyber fraud research.

Data sources and research strategy

After determining the research questions, we constructed the research as follows:

• – The main search terms are determined by the research questions.

• – Boolean operators (AND and OR) are used to restrict search results.

• – The search terms utilised for this review are related to detect cyber fraud in credit card and ML/DL techniques used for fraud detection.

The methodology incorporates the following electronic literature databases in order to obtain a comprehensive and broad coverage of the literature and to maximise the probability of discovering highly relevant articles:

• – Google Scholar—ACM—IEEE Xplore—SpringerLink—Web of Science—Scopus.

For the purpose of locating the most relevant article, particular Keywords were formulated into a search string. This string was divided into search units and Boolean operators were used to combine them. All of the mentioned resources have keyword-based search engines. We selected the following search string to retrieve the most relevant studies:

We include “artificial intelligence” OR “deep learning” OR “machine learning” thus that we can find studies that utilised any of these techniques. Additionally, we included the “credit card fraud” OR “card fraud” OR “card-fraud” OR “credit-fraud” OR “card cyber fraud” OR “transaction fraud” OR “payment fraud” OR “fraud detec*” OR “bank* fraud” OR “financ* fraud” term to concentrate on any fraud-related content so that we do not miss any relevant articles.

We conducted a search for the above string in six digital libraries. The research string is edited and converted into an appropriate search query input for each library. Table 1 provides the detailed search queries. We limited our review to journal articles, excluding conference article, books, and other publications. In December 2021, our search conducted for the years from 2019 to 2021. There were a total of 2,094 items retrieved from research libraries. Table 2 depicts the distribution of the items throughout the libraries. We identified 365 duplicate article. After eliminating the duplicates, we continued with the selection process based on the remaining 1,729 article. In addition to the automatic searches of digital libraries, snowballing mechanism was also used.

Study selection

We executed the above search strategy during December 2021 and identified 2,094 article. After removing duplicates (365 articles), the titles and abstracts of 1,729 unique citations were screened for eligibility. We screened the titles and abstracts for relevance. If the study’s relevance could not be verified due to insufficient abstract information or the absence of an abstract, the citation was assigned for full-text review. Thus we reviewed the full text of 281 studies. Disagreements on the included studies were resolved through discussion and consensus. The selected article were filtered to ensure that only relevant studies were included in our review. Then the article were exported to EndNote and grouped for each database and then exported to a literature review management software called Rayyan (Ouzzani et al., 2016) to facilitate the screening and selection process. To initiate the filtering and selection processes, duplicate articles gathered from multiple digital resources are eliminated. Then using inclusion and exclusion criteria, removed the irrelevant article. Using quality evaluation processes we included only the qualified article that offer the most effective answers to our study objectives. Using the collected article references, we searched for further related publications. Figure 1 displays the article selection process. The inclusion and exclusion criteria utilised for this review are detailed in Table 3. After the filtration process was completed, 181 article were observed for this study.

Data extraction

This process aims to analyse the final selection of article in order to collect the data required to answer the four research questions. Table 4 displays our data extraction form. In the final column of Table 4, the reason for extracting the corresponding data were given. We answered RQ1 and RQ2 using information regarding techniques and datasets. We used this information to group studies with comparable datasets and techniques. Extraction of each article’s discussion and findings was an aid in estimating the overall performance of approaches and answering RQ3. By extracting out the article’ objectives and conclusions, we are able to recognise trends, conduct gap analysis, determine future research, and provide a response to RQ4. As a result, in order to identify the gaps and define the next direction of future research should take, on the basis of the article’s objectives and conclusions, we conducted a summary analysis.

Distribution of chosen articles throughout the years

To explore the most recent techniques described in journals published in this field, limits were placed on publishing years. Our review selected article that were published from 2019 to 2021. In Fig. 2 we specified the distribution of article by year of publication. Since our study was completed in December 2021, it is important to note that article published after December 2021 were not included.

Publication type

In this review, we evaluated only journal publications. Table A1 displays the selected research articles published during the observation period.

Data synthesis results

This section examines the ultimately selected article (181 article). In order to provide a response to each of our four research questions, a synthesis of the data is performed. For RQ1: What types of ML/DL algorithms and datasets are used in credit card cyber fraud detection?

Cyber fraud detection techniques

In this part we address RQ1, which seeks to specify the ML/DL techniques used in detecting cyber fraud in credit card from 2019 to 2021.

Deep learning

Deep learning (DL) is subsection of ML uses data to teach computers how to perform tasks. The fundamental tenet of DL is that as we expand our NN and train them with new data, their performance continues to improve. The main advantage of DL over traditional ML is its higher performance on large datasets. The most frequently used DL algorithms in cybersecurity are feed forwards neural networks (FNNs), stacked autoencoders (SAE), and convolutional neural networks (CNNs). As shown in Fig. 3, DL techniques have been used in 34 reviewed articles. A total of 39 reviewed articles used combination of DL and ML techniques to detect fraud in credit card.

An artificial neural network (ANN) employs cognitive computing to aid in the development of machines capable of employing self-learning algorithms including pattern recognition, natural language processing, and data mining. ANN presents more accurate results because it learns from the patterns of authorized behaviour and thus distinguishes between ‘fraud’ and ‘non-fraud’ in credit card transaction. We explored 36 articles that used ANN in our review. In Agarwal (2021), ANN implemented for identity theft detection. The proposed model aims to use the different layers in a NN to determine the fraud transaction. The result shows that applying an ANN gives accuracy nearly equal to 100%. The result shows that ANN is best suited for determining if a transaction is fraudulent or not. New recent study applied ANN to detect fraud. The ANN technique has been used then compared with ML algorithms such as SVM, KNN. The result shows that ANN gives accuracy more than other ML algorithms, the suggested model is optimal for detecting credit card fraud (Asha & Suresh Kumar, 2021).

In Abdulsalami et al. (2019), back-propagation neural network (BPNN) and K-means are applied. The results indicate that the BPNN is more accurate than K-means algorithm. BPNN obtained accuracy of 79.9%. The results also indicate that K-means reduced prediction time provided it and advantage over BPNN. In Daliri (2020) harmony search algorithm with ANN (NNHS) are applied to improve fraud detection in banking system. The results show acceptable capability in fraud detection based on the information of customers. In Oumar & Augustin (2019) ANN with LR applied for fraud detection. Back-propagation has decreased the error function and enabled the model to discriminate between a fraudulent and a legitimate transaction. The suggested model is 99.48% accurate in its predictions and highly reliable.

Multilayer perceptron (MLP) is the most approach in ML because to its excellent accuracy in approximation nonlinear function. MLP comprises of three distinct layers. We explored 14 articles that used MLP in our review. In Alias, Ibrahim & Zin (2019), MLP and fifteen other types of supervised ML techniques are examined to determine the one with highest accuracy for detecting fraudulent transaction. The result shows that MLP generated the greatest detection accuracy of 15 algorithms, at 98%. Can et al. (2020) applied MLP and other ML techniques such as DT, RF, and NB. Regarding amount-based profiling, both MLP and classifiers demonstrated substantial improvements. In Faridpour & Moradi (2020), a novel ML-based model for detecting fraud in banking transaction utilising customer profile data is provided. In the proposed model, bank transactional data is utilised and an MLP with adjustable learning rate is trained to demonstrate the transaction authenticity, thus improving detection process. The suggested model surpasses SVM and LR. The accuracy of the proposed model is 0.9990.

Convolution neural network (CNN) is composed of multiple layers, output of which are used as inputs to layers that follow. ConvNET’s purpose is to reduce the input into a framework that is easier to comprehend, without sacrificing crucial information for making accurate predictions. CNN used in seven articles in the review. In Agarwal et al. (2021), DL techniques like CNN, BILSTM with ATTENTION layer have been used to detect and classify the illegitimate transactions. The CNN-Bi-LSTM-ATTENTION model detects the fraudulent class with high accuracy. Analysis shows that the model is adequate and yields an accuracy of 95%. The results demonstrate that the addition of an attention layer increases the performance of the model, allowing it to accurately discriminate between fraudulent and legitimate transactions. A CNN, NB, DT, and RF hybrid model is deployed in Aswathy & Samuel (2019), these algorithms are used as single models. Then these are used as hybrid models using majority voting technique. Adaptive boosting algorithm was used to boost the performance of classifiers.

DNNs, which provide potent tools for automatically producing high-level abstractions of complicated multimodal data, have recently garnered a great deal of interest from business and academics. DNNs learn features on their own, resulting in an increasingly accurate learning process. DNNs have been shown to be more efficient and accurate. Four studies employed DNN. In Arya & Sastry (2020), the proposed model is flexible to data disparity and resistant to hidden transaction patterns. Adaptive optimisation is recommended to improve fraud prediction. Result demonstrates its superiority over current other methods.

Credit card fraud detection using uncertainty-aware DL was implemented in Habibpour et al. (2021). It is vital to evaluate the uncertainty of DNN predictions. According to the study, there are three uncertainty quantification (UQ) techniques, ensemble, Monte Carlo dropout, and ensemble Monte Carlo dropout that can be used to quantify the level of uncertainty associated with predictions and produce a categorisation that is reliable. According to the findings, the ensemble method is superior at capturing the uncertainty related to predictions.

Deep convolution neural network (DCNN) applied in four articles. The DCNN technique can improve detection accuracy when a huge volume of data is involved. In Chen & Lai (2021), existing ML models, including LR, SVM, and RF, as well as auto-encoder and other DL models. Results show a detection accuracy of 99% was attained over a 45-s duration. Despite the vast quantity of data, the model provides enhanced detection. DL technique provides high accuracy and rapid pattern in detecting complex and unknown patterns. 1DCNN, 2DCNN, and DCNN have also been utilised to detect credit card cyber fraud in Cheng et al. (2020), Deepika & Senthil (2019), Nguyen et al. (2020).

A recurrent neural network, often known as an RNN, is a structure that used to remember previous input sequences. It is comprised of links between the internal nodes of a directed graph. Depending on the amount of their internal memory. RNN applied in seven articles in this review (Bandyopadhyay & Dutta, 2020; Chen & Lai, 2021; Forough & Momtazi, 2021; Hussein et al., 2021; Osegi & Jumbo, 2021; Sadgali, Sael & Benabbou, 2021; Zhang et al., 2021). In Bandyopadhyay & Dutta (2020), Implementing and applying RNN on synthetic dataset. The suggested model can detect fraudulent transaction with a 99.87% accuracy. The outcomes demonstrate that the approach is relevant and appropriate for detecting fraud. In Forough & Momtazi (2021), a deep RNN-based ensemble model and an ANN-based voting approach proposed. The ensemble model leverages a variety of RNN as the fundamental classifier and combines output using an FFNN as voting method. Classification employs a number of GRU or LSTM network. The outcomes indicate that the suggested model outperforms competing models. The proposed model is superior to existing models in this field. Bidirectional gated recurrent unit (BGRU) is applied in Sadgali, Sael & Benabbou (2021). Algorithms such as, GRU, LSTM, BRU, and SMOTE utilised in this model. BGRU obtained a high accuracy of 97.16%.

Long short-term memory (LSTM) is helpful technique to predict fraud because of the history knowledge it contains and the link that exists between prediction outputs and historical input. LSTM architecture enables sequence prediction problems to be learned through long-term reliance. LSTM and BiLSTM applied in eight articles (Agarwal et al., 2021; Alghofaili, Albattah & Rassam, 2020; Benchaji, Douzi & El Ouahidi, 2021; Cheon et al., 2021; Forough & Momtazi, 2021; Nguyen et al., 2020; Osegi & Jumbo, 2021; Sadgali, Sael & Benabbou, 2021). In Alghofaili, Albattah & Rassam (2020), a new model developed to improve both the present detection techniques and the detection accuracy in light of huge data. Findings demonstrated that LSTM performed perfectly, achieving 99.95% accuracy. Benchaji, Douzi & El Ouahidi (2021) recommended a model with the purpose of recording the previous purchasing behaviour of card holders. The results show that LSTM model obtained a high level of performance and accuracy.

DL based hybrid approach of detecting fraudulent transactions applied in Cheon et al. (2021). The new model includes a Bi-LSTM-autoencoder with isolation forest. This model proposed a detection rate of 87% for fraudulent transactions. The suggested model scored the highest mark. This model has the potential to be employed as an effective method for detecting fraud.

Deep belief network (DBN) applied in one article (Zhang et al., 2021). The new model utilised DBN and advanced feature engineering base on a Homogeneity-oriented behaviour analysis (HOBA). Results indicate that suggested model is effective and capable to identify fraud. DBN classifier with HOBA achieves a performance that is superior to that of the standard models.

Boltzmann machine (RBM) comprises of visible and hidden layers linked by symmetric weights. The neurones in the visible layer correspond to the X inputs, whilst the responses of the neurones H in hidden layer reflect the eventuality distribution of the inputs. RBM appeared in three articles in the review (Niu, Wang & Yang, 2019; Suthan, 2021; Suvarna & Kowshalya, 2020). In Niu, Wang & Yang (2019), supervised and unsupervised techniques have been applied. XGB and RF as a supervised technique obtain the best performance with AUROC is 0.961. RBM provides the best performance among unsupervised techniques. Results indicate that supervised models outperform the unsupervised models. Because of the problem of inadequate annotation and data imbalance, unsupervised techniques remain promising for credit card fraud detection.

A generative network (GAN) is comprised of two feed forward neural network, a Generate and a Discriminator, competing each other. The G produces new candidates while the D evaluates the quality. Each of the two networks is typically a DNN with multiple layers interconnected. GAN appeared in seven articles (Ba, 2019; Fiore et al., 2019; Tingfei, Guangquan & Kuihua, 2020; Hwang & Kim, 2020; Niu, Wang & Yang, 2019; Wu, Cui & Welsch, 2020; Veigas, Regulagadda & Kokatnoor, 2021). In Ba (2019), GANs employed as an oversampling technique. The findings indicate that Wasserstein-GAN is reliable during training and creates accurate fraudulent transactions comparing with other GANs. In Fiore et al. (2019), GAN employed to enhance the effectiveness of classification. A model for addressing the problem of class imbalance is described. GAN trained to generate minority class instances, then combined with training data to create an augmented training set to enhance performance. The results indicate that a classifier trained on expanded data outperforms its original equivalent.

The input-output mapping between the encoding and decoding phases is discovered by the autoencoder (AE). The input is mapped by the encoder to the hidden layer, and the input is rebuilt by the decoder using the hidden layer as the output layer. AE appeared in 18 articles in this review. AE mentioned in 18 articles within this review. In Misra et al. (2020), autoencoder model for cyber fraud detection is applied. Two-stage model with an autoencoder that coverts the transaction characteristics to a lower-dimensional feature vector at the first step. A classifier is then fed these feature vectors in a subsequent step. Results show that the suggested model outperform other models.

In Wu, Cui & Welsch (2020), dual autoencoders generative adversarial networks (DAEGAN) is employed for the imbalanced classification problem. The new model trains GAN to duplicate fraudulent transaction for autoencoder training. To create two sets of features, two autoencoders encode the samples. The new model outperforms several classification algorithms. Due to extremely skewed class distributions, credit card datasets present classification situations that are unbalanced. To address this difficulty. New model proposes in Tingfei, Guangquan & Kuihua (2020) employing oversampling technique based on variational automatic coding (VAE) in combination with DL techniques. Results demonstrate that the VAE model outperforms synthetic minority oversampling strategies and conventional DNN methods. In addition, it performs better than previous oversampling techniques based on GAN models.

Percentage of articles that address supervised, unsupervised, or semi-supervised in credit fraud detection?

This section answers RQ2 which attempts to show the proportion of gathered research article that employ supervised, unsupervised, or semi-supervised techniques. We examined credit card fraud detection techniques described in research article. According to Fig. 4, 74% of the chosen article utilised the supervised technique. Consequently, supervised technique is the most commonly employed in the reviewed article. In contrast, 12% utilised unsupervised techniques, and 12% utilised both supervised and unsupervised techniques. A total of 2% of reviewed article utilised semi-supervised learning. Additionally, 1% utilised reinforcement learning. Supervised and unsupervised learning have been implemented in 2019, 2020, and 2021. While semi-supervised learning only implemented three times in 2021. In the same manner, reinforcement learning has only been utilised in 2021. Compared to supervised and unsupervised learning, semi-supervised learning and reinforcement learning were not embraced by a large number of researchers. The ML/DL techniques type of each study article is listed in Table C1 for more information. The proportion of supervised, unsupervised, and semi-supervised is showed in Fig. 4.

Overall performance estimation of ML/DL model in credit fraud detection

This section addresses RQ3, which concerns the estimate of ML/DL model performance. Accuracy of estimation is the primary performance indicator for ML/DL models. This question focuses on the following features of estimating accuracy; performance metric, accuracy value, and dataset. As the construction of ML/DL models is dependent on the dataset, we examined the data sources of ML/DL models in the reviewed article. In addition, we found a number of datasets utilised in the experiments of associated article. This review articles employs two sets of datasets; real-word data set and synthetic dataset. The dataset utilised most frequently in the reviewed article is a real-word dataset. In addition, 154 research article employed real-world datasets, eight utilised synthetic datasets, and 19 did not specify the dataset source.

Evaluation metrics were used to calculate ML/DL model performance. Confusion matrix provides output matrix that characterises the model’s overall effectiveness. ML/DL model’s accuracy is compared using confusion matrix sensitivity and specificity, F-score, precision, receiver operating characteristic (ROC), and area under precision recall area (AUPR).

In this review, a number of different performance indicators have been used in addition to accuracy. As shown in Table C1, we found 177 article that clearly presented the performance metrics of the proposed models. Four article did not mention the performance metrics. We discovered that 177 of reviewed article mentioned the performance indicators of their suggested models. However, four reviewed article did not mention the performance metrics. In this review, accuracy, recall, precision, and F-score were often employed as performance indicators. Accuracy is the proportion of test set records that were properly categorised transaction to fraudulent or non-fraudulent. The ration of true positives to all positives is referred to as precision. The proportion of fraudulent transactions that we correctly detected as fraudulent compared to the total number of fraudulent transactions would be the precision. Recall is percentage of all correctly classified predictions made by an algorithm. In addition, the value of F1 provides a single score that is proportionate to both recall and precision. Full two-dimensional area under the entire ROC curve is measured by AUC. One of the best indicators for analysing the effectiveness of credit card fraud detection is the ROC curve. The classification’s quality is measured by MCC. Because it covers true positive, true negatives, false positive, and false negatives, it is a balanced metric. MCC utilised in 13 reviewed article.

In addition, 30 of the 181 studies employed only a single performance metric, with the majority of these article using only accuracy (24) article, MCC (five) article, and execution time (one) article. Using single performance metric is insufficient for determining the quality of ML/DL model. However, article such as 43 and 74 utilised more than five performance indicators to represent the performance of their ML/DL model. In addition, a number of reviewed article give computational performance measurements as well as performance metrics. The length of time the model took to complete the assigned task is called execution time. To ascertain how long the model takes to detect fraud, the execution time is calculated. As a result, we guarantee that the model successfully achieves its goal. Execution time employed in Alghofaili, Albattah & Rassam (2020), Devi, Thangavel & Anbhazhagan (2019), Singh, Ranjan & Tiwari (2021). The loss rate function compares actual and expected training output to speed up learning. Loss rate employed in article (Alghofaili, Albattah & Rassam, 2020). Test of the effect of cost sensitive wrapping of Bayes minimal risk (BMR) applied in article (Almhaithawi, Jafar & Aljnidi, 2020) as a cost-saving measure. Balanced accuracy (BCR) combines the matrices of sensitivity and specificity to produce a balanced outcome. BCR presented in article (Layek, 2020). In (Arun & Venkatachalapathy, 2020) Kappa assesse the predication performance of the classifier model. Few article (Arya & Sastry, 2020; Bandyopadhyay et al., 2021; Bandyopadhyay & Dutta, 2020; Benchaji, Douzi & El Ouahidi, 2021; Rezapour, 2019) introduced mean square error (MSE) assessment metrics, mean absolute error (MAE), and root mean square error (RMSE). Table C1 shows the proposed ML/DL model along with performance and datasets.

Trend of research

To answer RQ4, we examine the trend of the reviewed article. In addition, we compare the models created over the three years to determine and evaluate which techniques recently garnered more attention. This also assist, to identify the gaps so that future research will be able to address them in their own work. First, we examined the distribution of the chosen article by the publication year. In year 2019 (47 articles), 2020 (70 articles), and 2021 (64 articles). Significant difference existed between the years 2019 and 2020, the number of published articles for credit card fraud detection increased (23 articles). However, there was no notable difference between 2020 and 2021 (six articles). Fig. 2 demonstrates this comparison.

In response to RQ1, we demonstrated that 110 distinct ML models, 34 distinct DL models, and 39 models that combine ML and DL have been utilised by researchers. RF, LR, and SVM are the most commonly employed ML approaches. ANN, AUE, and LSTM are the most utilised DL approaches. In addition, we observed increased interest in combining ML and DL models.

In our review, we count the various learning-based credit card cyber fraud detection techniques applied in the reviewed article to answer RQ2. From this review we found that the most common technique among the reviewed article is the use of supervised algorithm. Supervised algorithms applied in 74% of the reviewed article. A total of 12% of the reviewed article utilised unsupervised techniques. A total of 12% used supervised and unsupervised techniques. A total of 2% applied semi-supervised technique. A total of 1% used reinforcement technique. For the RQ3, we listed the performance metrics that each research article applied. We discovered that 24 out of 181 reviewed article utilised accuracy as their only key performance metric. We also found a number of datasets that utilised in the reviewed article. Majority of the reviewed article using real-world datasets. A total of 154 research article applied real-world data, eight article used synthetic data, and 19 did not mention the source.

In RQ4, we identified research gaps by investigating unexplored or infrequently studied algorithms. In addition, we found supervised learning as the most prevalent learning technique and SMOTE as the most prevalent oversampling technique. The majority of researchers focused on supervised techniques such as LR, RF, SVM, and NN.

Combination techniques that employ multiple algorithms are becoming increasingly prevalent in the detection of cyber fraud. Detecting cyber fraud in credit card increasingly involves the use of DL. DL techniques utilised 34 times in the reviewed article, whereas 39 of the reviewed article applied a combination of DL and ML techniques for credit card cyber fraud detection. DL is advantageous for fraud detection since it solves the difficulty of recognising unexpected and sophisticated fraud patterns. Moreover, as the number of fraud cases to be recognised is relatively limited, DL may be effective. DL have garnered the most attention and had the most success in combating cyberthreats recently. Due to its ability to minimise overfitting and discover underlying fraud tendencies. Moreover, the capacity to handle massive datasets.

For supervised learning algorithms to predict future credit card transaction, each observation must have a label. Given that there is no classification for these observations, this could be a problem when trying to identify fraudulent transactions. Additionally, since fraudsters constantly alter their behaviour, it is challenging to develop a supervised learning model for a given transaction. The normal class is often the only one that unsupervised algorithms need labels for, and they can predict future observations based on deviations from the normal data. Future research should give more attention to unsupervised and semi supervised techniques, which can yield new insights. In addition, paying more attention to DL techniques such as CNN, RNN, and LSTM, we recommend that further research may be conducted on ML techniques, especially semi-supervised and unsupervised techniques in order to improve ML model performance. In addition, performing additional research on DL techniques is needed. As a result of the unavailability of a balanced dataset and the shortage of datasets, financial institutions are encouraged to make the essential dataset available, so that research outputs will be more effective and qualitative.

To detect cyber fraud in credit card, supervised, unsupervised, and semi-supervised ML/DL techniques applied in the reviewed article. Figure 4 displays that 74% of the reviewed article utilised supervised techniques. As a result, it is the most common technique used in the reviewed article. In addition, according to the reviewed article, classification and regression techniques been always of interest. On the other hand, 12% of selected articles applied unsupervised techniques, 12% of selected articles applied both supervised and unsupervised techniques, while 2% articles applied semi supervised techniques, and 1% articles applied reinforcement learning. A growing trend in this field is the use of ensemble techniques that capitalise on the benefits of several classification methods. The use of ensemble methods increased in 2020 and 2021 comparing with 2019. The other interesting finding is that DL approaches have attracted considerable interest during 2019 to 2021. The number of research articles that used DL techniques as single technique or combined with other ML techniques in 2019 is 15 articles, in 2020, 30 articles, and in 2021, 28 articles. It appears that the popularity of DL algorithms has increased.

The countries that published research on utilising ML/DL techniques to detect credit card cyber fraud is growing over time. In 2021, Ghana, Romania, Taiwan, and Vietnam are among the new countries that made an effort in detecting cyber fraud. India is the pioneer when it comes to the publication of ML/DL studies. Figure 5 depicts the number of article published by country and year (2019, 2020, and 2021).

Gap analysis and the future direction

The most effective way for determining the approaches that are most appropriate for this research problem is to categorise the ML/DL algorithms used in detecting cyber fraud in credit card. Additionally, it is beneficial to determine why particular tactics were chosen. Supervised algorithms have always been of interest, as 74% of the reviewed articles have been used supervised algorithms, with the most commonly used being RF then LR then SVM. Unsupervised learning algorithms also applied in 12% articles with the most commonly used being Isolation forest. However, it is interesting that only 12% of the 181 reviewed studies utilised unsupervised learning techniques. Semi-supervised approach employed in 2% of the reviewed articles. It appears that semi-supervised and unsupervised learning techniques may be researched further. According to reviewed articles (Choubey & Gautam, 2020; More et al., 2021; Muaz, Jayabalan & Thiruchelvam, 2020; Shirgave et al., 2019), unsupervised or semi-supervised learning techniques such as one-SVM, isolation forest, and K-means clustering should be utilised more in credit card fraud detection.

In the three years, DL techniques have been examined increasingly frequently. Utilising DL to get greater accuracy and efficient performance. By applying DL techniques, new fraudulent patterns can be recognised and system can respond flexibly to complex data patterns. Thus, for efficient credit card fraud detection, researchers are encouraged to conduct additional study on DL techniques. Several studies such as (Benchaji, Douzi & El Ouahidi, 2021; Jonnalagadda, Gupta & Sen, 2019; Kalid et al., 2020) suggested further study of DL techniques for detection in credit card. Moreover, as each ML/DL technique has its own limitations, it is necessary to consider combining the ML and DL algorithms for promising detection results. Several article such as (Agarwal, 2021; Dang et al., 2021; Gamini et al., 2021; Kalid et al., 2020; Singh & Jain, 2019) suggested combinations of DL methods and traditional ML methods to cyber fraud detection from an unbalanced data and enhance the accuracy.

Several reviewed article cited the lack of the dataset as the limitation of their work. According to Meenu et al. (2020), the research outcomes will be more effective and of higher quality if the financial institutions make the crucial data set of various fraudulent actions available. As a result, one of the key problems in many studies is the lack of data. Limitations on the availability of the data could be overcome if there is a vital data set of diverse fraudulent activities across nations. Maniraj et al. (2019) noted that when dataset size increase, algorithm precision also increases. It appears that adding additional data will undoubtedly increase the model’s ability to detect fraud and decrease the number of false positives. The banks themselves must formally support this. The study (Seera et al., 2021) proposed conducting further evaluation of their generated model with real data from diverse regions.

Additionally, the datasets are significantly skewed, which is a problem. Numerous studies attempted to develop a model that could perform properly with data that is highly skewed. Several articles (Balne, Singh & Yada, 2020; Ojugo & Nwankwo, 2021; Shekar & Ramakrisha, 2021; Voican, 2021; Vengatesan et al., 2020), unbalanced data was applied, and balancing the dataset using sampling techniques such as oversampling or undersampling is left as future work. Several articles (Ahirwar, Sharma & Bano, 2020; Almhaithawi, Jafar & Aljnidi, 2020; Manlangit, Azam & Shanmugam, 2019) applied oversampling techniques.

Undersampling techniques have been applied in several article (Amusan et al., 2021; Ata & Hazim, 2020; Muaz, Jayabalan & Thiruchelvam, 2020; Rezapour, 2019; Zhang, Bhandari & Black, 2020). In Amusan et al. (2021), a random undersampling technique was used, and the study recommended that other balancing data techniques be explored. One reviewed article (Ata & Hazim, 2020) applied an undersampling technique. However, the study recommends adopting the suggested model by using massive dataset instead of using sampling technique. In addition, some articles such as Trisanto et al. (2021) and Singh, Ranjan & Tiwari (2021) applied undersampling techniques and oversampling techniques.

Oversampling technique such as SMOTE, ADASYN, DBSMOTE, and SMOTEEN have been used. Undersampling techniques such as random undersampling (RUS) has been applied. In light of this, future studies should consider applying alternative oversampling techniques, such as borderline-SMOTE and borderline oversampling with SVM, as well as undersampling techniques. In addition to fraud location, an algorithm to determine the timing of the fraud is required (Alghofaili, Albattah & Rassam, 2020; Chen & Lai, 2021). In addition, an algorithm can be developed to predict fraudulent transactions in a real-time and deploying the service on various cloud platforms to make it easily accessible and reliable (Ingole et al., 2021).

Limitation of the review

Our review is restricted to journal article published in 2019, 2020, and 2021 that apply ML/DL techniques. By using our methodology in the early stages, we eliminated several irrelevant article. This assured that the selected article met the requirements for our review. Even though we searched the most prominent digital libraries for the article, there may be more digital libraries having relevant research article that were not included for this study. The snowballing method used to include relevant article that excluded during automatic searching in order to address this limitation. In addition, as it is probable that while looking for the keywords, we would have missed some synonyms. Hence, we also analysed the search terms and keywords for recognised collection of research works. We restricted our search to only English-language articles. This creates a language bias, as there may be article in this field of study written in other languages.

Appendix A

Appendix B

Appendix C