![Proposed framework Small Card Data: Electronic commerce refers to the sale of electronic goods. A dataset from Kaggle [15] contains 3075 samples and 12 features, including cash and fraud status. The dataset, known as SCD, has fewer rows and columns. Tall Card Data: The paper uses Tall Card Data (TCD), an online database with 10 million samples and nine features. The dataset includes customer ID, gender, state, card balance, transactions, international transactions, credit line, and fraud risk [16]. Due to limited computational capacity and longer training durations. The "class imbalance" is](https://www.researchgate.net/profile/Srilatha-Komakula/publication/385380158/figure/fig1/AS:11431281287531550@1730301692412/Proposed-framework-Small-Card-Data-Electronic-commerce-refers-to-the-sale-of-electronic_Q320.jpg)
Content uploaded by Srilatha Komakula
Author content
All content in this area was uploaded by Srilatha Komakula on Oct 30, 2024
Content may be subject to copyright.
International Journal of
INTELLIGENT SYSTEMS AND APPLICATIONS IN
ENGINEERING
ISSN:2147-67992147-6799 www.ijisae.org
Original Research Paper
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2392
An Exploration of Deep Learning Algorithm for Fraud Detection using
Spark Platform
Mrs. Srilatha Komakula1*, Dr. M. Jagadeeshwar2
Submitted: 25/01/2024 Revised: 03/03/2024 Accepted: 11/03/2024
Abstract: Fraudulent activities pose an important threat in several areas, requiring robust and efficient mechanisms for detection. It is
critical to halt fraudulent transactions since they have a long-term influence on financial circumstances. Anomaly detection has several
essential applications for detecting fraud. This paper presents a novel fraud detection method using deep learning algorithms, combining
Convolutional Neural Networks (CNNs) for feature extraction from transaction data and Long Short-Term Memory (LSTM) networks for
capturing temporal dependencies in financial transactions, thereby enhancing robust and efficient detection mechanisms. This paper
proposes a new framework that combines Spark with a deep learning technique. However, it compares the performance of deep learning
approaches for credit card fraud detection with other machine learning algorithms, including CNN-LSTM, on three distinct financial
datasets. This paper also employs several machine learning algorithms for fraud detection, such as random forest, SVM, and KNN. Various
parameters are used in comparative analysis. Both the training and testing datasets achieved more than 96% accuracy. The text outlines the
creation of a high-performance deep learning model for detecting credit card fraud. The paper proposes hybrid attention to integrate current
time output with unit state, determining its weight, and optimizes accuracy using Adam optimization. It uses various machine learning
methods for a comparative study using the proposed deep architecture.
Index Terms: Fraud detection, deep learning, machine learning, online fraud, credit card frauds
1. Introduction
In the digital age, fraud detection has become a significant
concern. Artificial intelligence, particularly deep learning
algorithms, has become a key tool in this fight. Apache
Spark, a distributed computing platform, offers scalability
and speed for analysing massive datasets and identifying
fraudulent activities in real-time. Deep learning algorithms,
including neural nets, autoencoders, recurrent neural
networks, and convolutional neural networks, are used to
detect complex patterns and anomalies in transaction data.
Spark can parallelize the training and deployment of deep
learning models on massive datasets, ensuring no fraudulent
transactions are overlooked. Its real-time streaming
capabilities keep the detection engine running, analysing
transactions as they occur and preventing losses before they
materialize. Spark integrates with popular deep learning
frameworks like TensorFlow and PyTorch, making it an
ideal choice for detecting and preventing fraud in the digital
age. Responsibility involves assessing a person's
responsibility for their actions. The rise of new technology
has increased the opportunity for fraud, particularly in credit
card use. Financial losses from fraud affect institutions,
individual customers, and banks' reputations. Non-financial
losses, which are difficult to quantify in the short term but
become more apparent in the long run, are also a concern
[1]. The customer will no longer be able to rely on his bank
and will switch to a more reliable competitor. Since the
introduction of credit card payments, fraud prevention has
worked to prevent and detect fraudulent transactions. The
Address Verification System, Card Verification Method,
and Personal Identification Number are all fraud-prevention
technologies. In contemporary banking, several fraud
detection algorithms and machine learning technologies are
applied. Using classification algorithms, each transaction is
allocated to a risk category. Artificial intelligence
algorithms create the model on the basis of databases.
The use of learning methods in fraud detection is a useful
tool since it allows for the finding of patterns in large-scale
datasets with many variables. The model is often a
parametric function that predicts the likelihood of
transaction fraud. Furthermore, fraudulent transactions often
coincide in both time and location. Similarly, [2]
investigated credit card fraud detection via individual.
Fraudulent transactions often coincide in time and location,
and credit card fraud detection can be improved by adding
“statistical features” resulting from a feature. Machine
learning (ML) is an AI subfield that trains processers to
predict future outcomes based on known patterns in existing
data [3], requiring attention to factors like quick reaction
time, “cost sensitivity," and “feature pre-processing.”.Many
studies have used ML models to solve a range of issues.
Deep learning (DL) algorithms have been utilised in a
variety of applications, such as data centres and mobile
1Research Scholar, Department of Computer Science, Chaitanya (Deemed
to be University), Warangal(Urban), Telangana, India
2Professor, Department of Computer Science. Chaitanya (Deemed to be
University), Warangal(Urban), Telangana, India
Corresponding Author Email: srilatha.kom@gmail.com
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2393
communications. Deep learning approaches have gained
interest due to their promising outcomes in computer vision,
NLP, and audio, but a small number of works have explored
their use in identifying credit card fraud (CCF) [4]. It
identifies CCF using a number of deep-learning techniques.
Figure 1 depicts a payment card authorization mechanism.
Biometric authentication is classified into three types:
physiological, behavioural, and combination authentication
[5], [6].
Fig 1: Payment card authorization process
This paper focuses on the performance of deep learning
techniques like CNN and LSTM for credit card fraud
classification. Classic machine learning techniques like
SVM, Decision Trees, and Logistic Regression are suitable
for small datasets. Deep learning models excel at detecting
complex patterns and anomalies in large datasets, leading to
higher accuracy in identifying fraudulent transactions
compared to rule-based or shallow learning algorithms. The
pre-built algorithms and tools in Spark MLlib facilitate
deploying deep learning models on Spark clusters,
simplifying the development and deployment process. The
paper presents a novel deep learning framework for financial
fraud detection using Spark, compares different machine
learning architectures, and presents a unique stacked-based
strategy for feature selection.
2. Background and Related Work
Financial institutions can navigate the digital landscape
effectively by utilizing deep learning's ability to understand
complex patterns and Spark's data processing and real-time
analysis skills. These algorithms ensure transaction validity
and secure the money. Convolutional neural networks,
popularised by AlexNet's ImageNet Challenge win, have
become increasingly used in computer vision [7, 8], with
more sophisticated deep networks like VGGNet improving
classification and recognition [9]. Deep residual networks
(ResNets) have transformed the area of deep learning,
notably for computer vision applications such as image
recognition and object identification. Their significance
arises from their capacity to solve a significant challenge in
training deep neural networks: the vanishing gradient issue
[10]. Similarly, in [11], the author said that in today's
technological age, particularly in online business and
banking, Mastercard transactions are fast rising. Mastercard
has evolved into a very useful piece of equipment for
Internet shopping. This increased usage creates significant
harm as well as fraud charges. It is critical to halt fraudulent
transactions since they have a long-term influence on
financial circumstances. Anomaly detection has several
essential applications for detecting fraud. Detecting fraud
concerning online transactions may be the most difficult task
for financial institutions and banks. As a result, financial
institutions and banks must have highly effective fraud
detection procedures to minimise their losses due to charge
card fraud transactions. Many researchers have discovered
various strategies to spot these scams while also taking them
down. Following the study of the dataset, the dependability
is 97% for LOF and 76% for IF. The author mentioned in
[12] that in today's technological age, particularly in Internet
commerce and banking, Mastercard transactions are fast
rising. Mastercard has evolved into a very useful piece of
equipment for Internet shopping. This increased usage
creates significant harm as well as fraud charges. It is critical
to halt fraudulent transactions since they have a long-term
influence on financial circumstances. Anomaly detection
has several essential applications for detecting fraud. This
platform identifies fraud relatively rapidly, resulting in less
loss and danger.
3. Proposed Modelling
This paper proposes a revolutionary framework that
combines Spark with a deep learning technique. Figure 2
illustrates the suggested framework. This paper also
employs several machine learning algorithms for fraud
detection, such as random forest, SVM, and KNN.
Dataset Description
Credit card transaction datasets are crucial for machine
learning models to identify fraudulent activity. They track
customer spending habits, merchant categories, and
geographic locations, helping understand customer
preferences and segmenting. Lenders use these datasets to
assess default risk and determine loan terms, while
borrowers' credit history and income are also considered.
The present paper aims to evaluate the effectiveness of
classification models in detecting CCF using three distinct
datasets: "European Card Data (ECD), Small Card Data
(SCD), and Tall Card Data (TCD)." The datasets are
skewed, with few fraud incidents compared to ordinary
transactions. The datasets are labeled with '0' indicating no
fraud and '1' indicating fraud.
European Card Data: The “machine Learning Group of
Université Libre de Bruxelles” received a dataset from
Kaggle [14], containing 284,807 European cardholder
transaction data from September 2013. Only 492 of the
samples were fraudulent, accounting for 0.172% of the total.
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2394
Due to transactional details' sensitivity and client
information privacy. The dataset pertains to electronic commerce, which involves the purchase and sale of
electronic items.
Fig 1: Proposed framework
Small Card Data: Electronic commerce refers to the sale of
electronic goods. A dataset from Kaggle [15] contains 3075
samples and 12 features, including cash and fraud status.
The dataset, known as SCD, has fewer rows and columns.
Tall Card Data: The paper uses Tall Card Data (TCD), an
online database with 10 million samples and nine features.
The dataset includes customer ID, gender, state, card
balance, transactions, international transactions, credit line,
and fraud risk [16]. Due to limited computational capacity
and longer training durations. The “class imbalance” is
maintained, with 28,000 fraud instances out of 500,000 total
samples. The study aims to analyze the data using a small
amount due to restricted computational capacity and longer
training durations associated with classifiers.
Data Pre-Processing
Data preprocessing is crucial for developing effective
machine learning models, especially for detecting credit
card fraud. In this study, three datasets (ECD, SCD, and
TCD) are inspected manually and statistically to provide
updated input for classifiers. The goal is to generate the best
possible output, which can be impacted by missing data,
categorical features, variable scale, and high dimensionality.
For ECD, all features were numeric, no missing values were
found, and no features were removed. SCD's categorical
features were transformed to numbers and the 'Transaction
Date' feature was eliminated due to the large number of
missing values. TCD was used in a smaller proportion of the
dataset, with all variables numeric and the 'custID' feature
removed. Correlation is found during data exploration to
determine the dependability of variables. This helps
eliminate data features with similar behavior, reducing data
dimensions and allowing for quicker training and improved
classification outcomes.
The SCD dataset shows no negative correlation between
features and no consistent pattern as shown in Figure 3.
Larger transaction amounts are associated with a higher
average daily transaction number. The number of
chargebacks per day, amount, and frequency are highly
connected, making the high-risk country feature crucial for
fraud classification. Due to the dataset's small size, no
dimensionality reduction is performed.
Dataset
sample
Decision
module
Deep
learning
Customer
transaction
database
Computing
algorithm
output
Bank alert
Allow
transaction
Incoming
transaction
Legitimate
transaction
Fraudulent
transaction
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2395
Table 1: Data exploration
Fig 3: Correlation Heatmap of the SCD dataset
Data Scaling and Standardization
Machine learning involves crucial steps such as scaling and
standardization to prepare data for efficient model training.
These steps are essential when dealing with varied sizes of
features in datasets. In the ECD dataset, the features
'Amount' and 'V1' have mean values. To make the input
more intelligible, approaches like scaling and
standardization aggregate features to nearly the same scale.
In this paper, the “StandardScaler” and Python's “sklearn”
are employed to convert each feature into a mean and
standard deviation.
Test, Train and Validation Split
In machine learning, the test, train, and validation split is a
fundamental technique for evaluating the performance of a
model. It involves dividing the data into three sets:
Electronic commerce refers to the sale of electronic goods.
For training 60% data is used, Test data is a randomly
chosen portion of 20%. Validation of 20% used to validate
the classifier, preventing overfitting and enhancing model
performance. This technique is crucial for any machine
learning project to build accurate and generalizable models.
Apache Spark3
Apache Spark is known for its distributed computing
capabilities, allowing it to process large volumes of data in
parallel across a cluster of machines. This can significantly
improve the speed of data processing, enabling faster fraud
detection. Spark 3 introduces enhancements to Spark SQL
and DataFrames, making it easier to express complex data
manipulations and queries. This can simplify the
preprocessing and analysis of credit card transaction data,
leading to more effective fraud detection algorithms. MLlib,
Spark's machine learning library, continues to evolve with
each Spark release. Spark 3 may include improvements to
existing machine learning algorithms, making it easier to
develop and deploy more accurate models. Adaptive query
execution, introduced in Spark 3, can dynamically optimize
query plans based on runtime statistics. This can lead to
more efficient processing of complex analytical queries,
which is beneficial for tasks like feature engineering in fraud
detection.
K-Nearest Neighbor (KNN)
KNN is used as classification for storing training data points
and classifies new points based on their similarity to these
points. It calculates the distance to all training points using
a distance metric, assigns the most frequent class among its
nearest neighbors, and finds the K training points. KNN is
known for its ease of use, adaptability to non-linear
problems, and sensitivity to local data structure. However,
it's crucial to consider its computational cost, sensitivity to
irrelevant features, and the right K value for specific tasks.
Support Vector Machine (SVM)
SVM is a classification algorithm that learns from labeled
data to classify new unseen data points. It aims to find the
decision boundary that maximizes separation between
classes, leading to better generalization and robustness.
Training support vector machines (SVMs) may be
computationally costly, particularly for large data sets,
despite their effective performance in many circumstances.
Also, getting the parameters just right is key, particularly for
picking the kernel and the regularisation parameter, and
effectiveness in non-linear problems. However, it's crucial
to consider computational cost and interpretability
challenges when selecting an algorithm for a specific task.
CNN Model
CNN is a type of network that comprises Convolution,
activation, and pooling layer.
Convolutional Layer. This is the fundamental process
which comprises numerous convolution units, the
parameters of which are improved via a backpropagation
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2396
method. Convolution operations are used to extract distinct
characteristics from the input. Increasingly network layers
may repeatedly extract more sophisticated information from
low-level data. It's a small matrix of numbers that's used to
perform convolution, a mathematical operation that
balances or transforms an image based on the values in the
kernel.
where Zl+1 Z (i, j) signifies “feature matrix”; f signifies
“Kernel sixe”.
Pooling Layer. A pooling layer is a crucial part of
convolutional neural networks (CNNs) that reduces feature
maps' spatial dimensions while maintaining crucial
information. It is used to control overfitting and add spatial
invariance. The output matrix is transferred to the pooling
layer for feature extraction and filtering.
LSTM stands for Recurrent Neural Network (RNN).
Backpropagation based on gradients reduces loss in neural
networks. Standard NNs cannot remember existing
information and must begin learning from scratch. RNNs are
neural networks with memory. As the gradient progresses
back in RNN, the weight update decreases. RNN creates a
short-term memory network, as the final layers lose learning
capacity and recall early occurrences. LSTMs address this
issue by providing a cell state as the network's memory and
gates in each phase to preserve vital information while
rejecting irrelevant information. There are forget, input, and
output gates used. Forget, input, and output gates decide
what to keep, add, and conceal. The hidden state recalls the
model's previous inputs but has short-term memory, while
the cell state remembers essential information from the
beginning of the series. Figure 4 depicts the whole LSTM
cell flow [13]. Each dotted box signifies a step. The
following are the LSTM conversion functions:
where σ is in range of [0, 1], tanh signifies a “hyperbolic
tangent function” in range of [−1, 1].
Figure 4: An LSTM Cell
CNN-LSTM.
The paper employs CNN and LSTM to address the challenge
of learning information from variable-length sequence
input. The improved CNN structure is comparable to
ConvNets, with nine layers, six convolutional and three
connected. RNNs, on the other hand, struggle with gradient
issues as input sequence duration increases. LSTM
calculates the need for new inputs and information loss from
cell states using a gate mechanism in each unit. It also
handles the retrieval of characteristics at word and character
levels. The model is structured similarly for Char-Channel
and Word-Channel channels, distinguished by embedding
granularity. The convolution kernel and input sequence are
sent to the CNN segment, which calculates the convolution
result.
Fig 5: Structure of ConvNets model.
The LSTM technique has been consolidated into LSTM (x).
Neural networks for convolution uses both short term and
long term memory, can be structured in series or parallel.
The series structure is commonly used, but it suffers from
information loss due to data compression during
convolution. The parallel system is selected in the place of
series to capture all channels' structures simultaneously.
This allows for the full use of LSTM's advantages in
capturing time-series features in neural networks.
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2397
Ve and Vg are word-level embedding vectors, and “two
channel outputs” are combined to form a hidden layer
output.
The hidden layer's results are sent to the “fully connected
layer”, followed by the softmax.
Hybrid Attention. The “dynamic adaptive weight”, a
crucial component, is determined using Equations (10) and
(11) to calculate the weight score ω.
where represents LSTM output, ct represents LSTM
status and b indicates bias. Subsequent, the score ω can be
written as in
where x indicates sequence length.
The dynamic adaptive weight is used to weight the output
vector ci, as illustrated in
4. Results and Discussions
In this section, various machine learning techniques are
tested for performance metrics on the dataset. A deep
learning model is constructed with various ratios of training
and testing sets. The study emphasises the importance of
selecting appropriate grading techniques to assess the
correctness of grading. The Confusion Matrix is used as a
classification metric, displaying the generated data based on
four possible combinations of expected and actual values:
false positive (FP), true positive (TP), false negative (FN),
and true negative (TN).
Fig 5: Confusion matrix. (Source: compiled by authors)
Accuracy measures the model's predictions, focusing on the
fraction of transactions correctly classified as fraudulent or
genuine. Precision measures the frequency of fraud
detection, with high accuracy resulting in low false alarm
rates. Recall measures the percentage of genuine fraud
instances accurately classified as fraudulent, indicating the
model's ability to detect true fraud cases. The F1 score,
combining accuracy and recall, is beneficial for detecting
fraud accurately and minimizing false alarms, ensuring a
more comfortable experience for real customers.
(14)
(15)
(16)
(17)
Area under the Curve (AUC): The area under the ROC
curve is known as AUC, while the probability curve is
known as ROC. To assess the model's outcomes, the worst
AUC is 0, and the greatest AUC is near 1. Tables 2 and 3
reveal the results of using CNN-LSTM with varied training
and testing split ratios. Figures 7 and 8 show the distribution
of fraud scores for split ratios of 60-40 (Case 1), 70-30 (Case
2), and 80-20 (Case 3).
Table 2: Comparing different data splits on train dataset
Train
data
Test data
Training
accuracy
Time
elapsed
(sec)
(Case 1)
60
40
0.9774
14.21
(Case 2)
70
30
0.9687
16.57
(Case 3)
80
20
0.9537
18.54
True
Positive
(TP)
False
Negative
(FN)
True
Positive
(TP)
True
Negative
(TN)
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2398
Figure 7: Fraud score distribution of training for split ratio
Table 3: Comparing different data splits on test dataset
Train data
Test
data
Testing
accuracy
Time
elapsed
(sec)
(Case 1) 60
40
0.9621
0.35
(Case 2) 70
30
0.9547
0.38
(Case 3) 80
20
0.9432
0.42
Figure 8: Fraud score distribution of testing for split ratio
Table 4: ML and Proposed method comparison table on the
basis of performance measures
Classifiers
Accur
acy
Specificity
Precision
F1-
score
SVM
94.8
95.7
87.2
88.4
KNN
94.1
97.8
63.4
78.5
CNN-LSTM
(Proposed)
98.5
98.7
98.9
99.1
Figure 9 shows that proposed CNN-LSTM is having highest
accuracy of 98.5%, specificity of 98.7%, precision 98.9%
and F1-score of 99.1% which is evident that the proposed
CNN-LSTM outperformed as related to other two ML
techniques.
Fig 9: Performance comparison of SVM, KNN and CNN-
LSTM proposed method
5. Conclusion
Deep learning algorithms and Spark's scalability are
powerful tools for combating credit card fraud. However,
challenges persist, and ongoing research and development
efforts are improving their capabilities. The main objective
of fraud detection systems is to accurately predict fraud
instances and reduce false-positive cases. This paper
examines various methods for detecting fraud in card
transactions using an unbalanced dataset. The under-
sampling approach balances the dataset during pre-
processing, while CNN-LSTM measures fake observants
and combines probabilities to find alerts. The model
employs a ranking technique to prioritize alerts and correct
class imbalances. CNN-LSTM outperforms all algorithms
with an F1-Score of 99.1%. Future work should focus on
techniques that enhance model interpretability and build
trust in decision-making. Regular monitoring and updating
of data and models are essential to maintain effectiveness.
Combining deep learning with other machine learning
techniques can enhance overall performance. By addressing
these challenges and leveraging the strengths of deep
learning and Spark, organizations can significantly improve
their ability to detect and prevent credit card fraud,
protecting both customers and financial interests.
References
[1] Burhanudin, Burhanudin. (2022). Examining the effect
of service value and reputation on customer loyalty with
trust and electronic word of mouth as mediation. Jurnal
Aplikasi Manajemen. 20.
10.21776/ub.jam.2022.020.03.05.
[2] Gurav, Prof & Badhe, Mrs & Nagtilak, Mrs &
Sonawane, Mr & Agarwal, Mr. (2022). Credit Card
Fraud Detection. International Journal for Research in
0.9774 0.9687 0.9537
14.21 16.57 18.54
0
5
10
15
20
Case 1 Case 2 Case 3
Value
Split data case
Training accuracy
0.9621 0.9547 0.9432
0.35 0.38 0.42
0
0.2
0.4
0.6
0.8
1
1.2
Case 1 Case 2 Case 3
Value
Split data case
Training accuracy
0
20
40
60
80
100
Accuracy Specificity Precision F1-score
Performance level in %
Metric
SVM
KNN
CNN-LSTM (Proposed)
International Journal of Intelligent Systems and Applications in Engineering IJISAE, 2024, 12(3), 2392–2399 | 2399
Applied Science and Engineering Technology. 10.
2009-2012. 10.22214/ijraset.2022.41594.
[3] Y. Abakarim, M. Lahby, and A. Attioui, ‘‘An efficient
real time model for credit card fraud detection based on
deep learning,’’ in Proc. 12th Int. Conf. Intell. Systems:
Theories Appl., Oct. 2018, pp. 1–7, doi:
10.1145/3289402.3289530.
[4] V. Arora, R. S. Leekha, K. Lee, and A. Kataria,
‘‘Facilitating user authorization from imbalanced data
logs of credit cards using artificial intelligence,’’
Mobile Inf. Syst., vol. 2020, pp. 1–13, Oct. 2020, doi:
10.1155/2020/8885269.
[5] A. O. Balogun, S. Basri, S. J. Abdulkadir, and A. S.
Hashim, ‘‘Performance analysis of feature selection
methods in software defect prediction: A search method
approach,’’ Appl. Sci., vol. 9, no. 13, p. 2764, Jul. 2019,
doi: 10.3390/app9132764.
[6] B. Bandaranayake, ‘‘Fraud and corruption control at
education system evel: A case study of the Victorian
department of education and early childhood
development in Australia,’’ J. Cases Educ. Leadership,
vol. 17, no. 4, pp. 34–53, Dec. 2014, doi:
10.1177/1555458914549669.
[7] Krizhevsky A, Sutskever I, Hinton G E (2012)
ImageNet classification with deep convolutional neural
networks. In: Paper presented at the twenty-sixth annual
conference on neural information processing systems,
Lake Tahoe, Nevada, 3–6 December 2012, pp 1097–
1105
[8] Szegedy C, Liu W, Jia Y, Sermanet, P, Reed S (2015)
Going deeper with convolutions. In: Paper presented at
the IEEE conference on computer vision and pattern
recognition, Boston, Massachusetts, 7–12 June 2015,
pp 1–9.
[9] Simonyan K, Zisserman A (2015) Very deep
convolutional networks for large-scale image
recognition. In: Paper presented at the international
conference on learning representations, San Diego,
California, 7–9 May 2015, pp 1–14
[10] Andrew G, Menglong Zhu, Bo Chen, Dmitry
Kalenichenko (2017) MobileNets: efficient
convolutional neural
[11] H. John and S. Naaz, “Credit card fraud detection using
local outlier factor & isolation forest,” International
Journal of Computer Sciences and Engineering, vol. 7,
no. 4, pp. 1060–1064, 2019.
[12] W. F. Yu and N. Wang, “Research on credit card fraud
detection model based on distance sum,” in IJCAI
international joint conference on artificial intelligence,
pp. 353–356, Hainan, China, 2009.
[13] M. Phi, Illustrated Guide to LSTM’s and GRU’s: A step
by step explanation, Towards Data Science, 2018.
Accessed on: Nov. 23, 2020. [Online]. Available at:
https://towardsdatascience.com/illustrated-guideto-
lstms-and-gru-s-a-step-by-step-explanation-
44e9eb85bf21
[14] J. Desjardins, How much data is generated each day?,
World Economic Forum, April 17, 2019. Accessed on:
Nov. 18, 2020. [Online]. Available:
https://www.weforum.org/agenda/2019/04/how-much-
data-is-generatedeach-day-cf4bddf29f/
[15] A. O. Adewumi and A. A. Akinyelu, "A survey of
machine-learning and nature-inspired based credit card
fraud detection techniques." International Journal of
System Assurance Engineering and Management 8, no.
2 (2017): 937-953.
[16] T. T. Nguyen and V. J. Reddi, "Deep reinforcement
learning for cyber security," arXiv preprint
arXiv:1906.05799 (2019).
