ISSN 2581 – 3781

Available Online at www.ajcse.info

Asian Journal of Computer Science Engineering 2025;10(2):1-10

REVIEW ARTICLE

Analytical Review of Deep Learning Architectures in Financial Frameworks for

Fraud Identification of Credit Cards
Prithviraj Singh Rathore*
Department of Computer Sciences and Applications, Mandsaur University, Mandsaur, Madhya Pradesh, India
Received on: 12-03-2025; Revised on: 10-04-2025; Accepted on: 02-05-2025
ABSTRACT
The finance sector has made cybersecurity a crucial focus due to the exponential rise in cyberattacks
and digital financial crimes. One of the most significant issues is credit card fraud, which has grown
in frequency alongside the proliferation of both online and physical shopping. When faced with
increasingly complex fraud strategies, traditional rule-based and anomaly detection systems might be
resource-intensive and fail to detect them. The purpose of this research is to examine state-of-the-art deep
learning (DL) methods for efficient and precise detection of credit card fraud. Its main focus is on deep
neural architectures, including long short-term memory, convolutional neural networks, autoencoders,
and hybrid ensemble models. These methods demonstrate superior performance in capturing non-
linear relationships and temporal dependencies within transactional data. The use of data preprocessing
methods such as principal component analysis and the synthetic minority oversampling technique helps
with class imbalance and dimensionality reduction. Experimental results validate the effectiveness of
DL approaches in enhancing detection accuracy and robustness, thereby contributing to more resilient
financial security systems.
Key words: Anomaly detection, autoencoder, convolutional neural network, credit card fraud, deep
learning, fraud detection, GRU, long short-term memory, real-time analytics, xgboost

INTRODUCTION A credit card allows its holder to purchase goods

and services or withdraw cash within a specified
Internet banking is rapidly expanding, and financial credit limit, offering the advantage of deferred
institutions increasingly provide the public with payments to the next billing cycle.[3,4] However,
actual business facilities. Electronic payment systems this convenience also presents an opportunity
have grown vital in today’s competitive financial for fraudulent activities. Fraudsters can exploit
market due to the convenience they provide when vulnerabilities to perform unauthorized transactions,
purchasing goods and services. The introduction often bypassing existing security mechanisms and
of cashless transactions and the associated causing substantial financial losses in a short span
insurance against lost, damaged, or stolen items has without the cardholder’s knowledge.
revolutionized consumer convenience.[1,2] Credit and The crime of fraud is committed when one party
debit cards have further expanded this insurance. uses deceit to obtain money or property unfairly
By requesting confirmation from consumers before or unlawfully, even if the other party does not face
making a purchase using their credit cards, an extra immediate legal repercussions for their actions.[5-7]
layer of protection is incorporated. Credit card To mitigate such risks, financial systems employ
fraud, however, is an ongoing issue that hinders two primary countermeasures: Fraud prevention and
both customer and bank profits. Internet fraudsters fraud detection mechanisms. Fraud prevention refers
are always coming up with new techniques, to proactive techniques aimed at eliminating fraud
making it harder for financial institutions and other before it occurs, whereas fraud detection systems
organizations to detect and prevent their crimes. focus on identifying fraudulent transactions that
have already bypassed preventive controls. These
Address for correspondence: detection systems must evaluate every transaction,
Prithviraj Singh Rathore regardless of previous filters, to promptly recognize
E-mail: [email protected] anomalies and prevent further loss.

© 2025, AJCSE. All Rights Reserved 1

 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

The proliferation of digital technologies, such as get their hands on cardholder data. One is phishing,
telephone banking, automated teller machines in which the criminal pretends to be a financial
(ATMs), and online credit systems, has intensified official in order to trick the user into giving over
the risk and impact of fraud, making manual sensitive information. Skimming is another kind
verification of each transaction infeasible due to of card fraud where the criminal uses an interface
cost and scalability constraints. In this context, to enter a system that scans cards directly at points
machine learning (ML) and, more recently, deep of sale or ATMs. Protecting customers’ money
learning (DL) have emerged as powerful tools to and personal information requires detecting
identify fraudulent behavior through automated credit card fraud. Automated systems and human
classification techniques.[8,9] By learning from investigation are the two primary methods for
historical patterns of fraud, these systems can identifying fraudulent actions.
classify and detect new fraudulent instances with
higher accuracy.
DL is an approach to ML that uses multilayered, Types and Characteristics of Credit Card
deep artificial neural networks (ANNs) to simulate Fraud
brain function. These networks consist of multiple Credit card fraud is complex; therefore, it is
hidden layers capable of learning complex, non- important to know what it is. Examples of common
linear representations of input data.[10] Particularly
types are:
well-suited to binary classification challenges
such as fraud detection is their capacity to
model high-level abstractions. With the rise of Card-not-present (CNP) fraud
sophisticated fraud strategies and the limitations One distinguishing feature of CNP fraud is that it
of traditional rule-based methods, DL provides is both a hybrid and an output crime. Frauds can be
a flexible and scalable answer. It uses massive perpetrated in either the real world or online, which
amounts of transaction data to identify unique and is why considers them a hybrid cybercrime.[12,13]
ever-changing instances of fraud. Since con artists can get victims’ credit card
details by physically scanning and skimming their
cards, CNP fraud might be considered a hybrid
Structure of the Paper
cybercrime. Furthermore, criminals may employ
This paper is organized in the following social engineering techniques to send unwanted
way: Section II explains the fundamentals of emails in an effort to trick people into divulging
identifying fraudulent charges on credit cards. their credit card information.
In Section III, it goes over the DL process and
the preprocessing approaches. Section IV delves
Lost or stolen card fraud
into various DL architectures that are used for
Crucially, fraudulent transactions represent the
fraud detection. A thorough literature review is
minority class of primary interest. However,
presented in Section V, and the study is concluded
their rarity often hampers the classifier’s
with important findings and directions for future
learning process, resulting in suboptimal model
research in Section VI.
performance.[14] When added together, these two
groups accounted for 68% of all instances of card
FUNDAMENTALS OF CREDIT CARD fraud in South Africa during 2007 and 2008.
FRAUD DETECTION
The crime of credit card fraud is when someone Counterfeit card fraud
else makes transactions using another person’s Once a criminal has a legitimate card number, they
card without their authorization. It encompasses might use it to make a counterfeit card. The data
all forms of bank card fraud, including those can thereafter be written onto the magnetic stripe
involving debit and credit cards. Even though of a new, blank card or entered by hand onto the
most of the transactions happen online, it can surface of a counterfeit plastic card.[15] A criminal
still use the physical card if it is lost or has it is can find all the information they need to make fake
stolen.[11] Fraudsters use a variety of techniques to cards with just a few clicks on the Internet. Credit

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 2
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

card counterfeiting equipment, including custom Challenges in Fraud Detection

embossing machines, tipping machines, decoding
software, and programs for encoding credit card There are some challenges of fraud:
magnetic stripes, may be found on numerous • Massive imbalance in the data: Since the
online marketplaces. number of valid transactions (around 0.2
percent) much outweighs the number of
fraudulent ones. Crucially, fraudulent
Application fraud transactions represent the minority class of
Applicants commit credit card application fraud primary interest. However, their rarity often
when they supply false information. Finding a hampers the classifier’s learning process,
fraud system that can detect suspicious applications resulting in suboptimal model performance.
is the solution to the issue of application fraud. • Unavailability of data sets: Since card
In the first case, known as duplicate applications, companies are unable to divulge these to
numerous applicants provide the same or nearly the general public or academics for reasons
identical information; in the second, known as related to privacy, secrecy, confidentiality, and
identity fraud, numerous applicants utilize the
the law. The vast number of ML algorithms
same or nearly identical information to submit
and the tremendous interest in this field are not
their applications.
enough to propel research forward; real data
are essential. The substitution of synthetic data
Behavioral fraud and the lack of or difficulty gaining access to
Conducting sales on a “cardholder present” basis research and its results were also mentioned.
after fraudulently obtaining the details of valid • Limited published work: The characteristics
cards is behavioral fraud. This category includes of the data or the parameters utilized by
sales conducted over the phone or online, where classifiers are typically not divulged by
the only information needed is the card details. researchers. Funding institution constraints
For behavioral fraud, a fraud scorecard that may be to blame. It is challenging to exchange
forecasts which clients would default is a useful ideas in fraud detection because of security
tool.[16,17] Some of the reasons why traditional and privacy issues, which, according to,
credit scorecards fail to identify consumers who greatly hinders the development of innovative
are likely to default include fraud. How it works: approaches for detecting credit card fraud.
Applying scoring to the prevention of fraud is • Adaptive and innovative fraudulent behavior:
identical to applying it to any other use case, Constant relearning of the new patterns is
including profit, default, or collection. Figure 1 necessary due to the frequent and ongoing
depicts the many forms of credit card fraud; evolution of fraud and normal profiles.
the score is calculated by averaging previous Furthermore, the computer could learn the
occurrences. One of two outcomes is possible: a wrong details about fraudulent transactions in
real or fake consumer. the past and treat them as real. It just takes one
slip-up for new frauds to emerge that look like
regular transactions.

Traditional Detection Techniques

One of the first approaches brought in was a rule-
based system for detecting credit card fraud. Such
systems are based on a pre-worked-out body of
logical rules developed by domain experts to mark
anomalous transactions. They all have different
attributes, such as the amount of transactions,
time, location, or frequency, that they target. An
Figure 1: Types of credit card fraud example here is of a transaction that is exceeding a

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 3
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

specific amount in a foreign place may be marked rely on supervised models trained on massive
as suspicious.[18] Although rule-based systems are datasets of correctly tagged transactions.
simple, obvious, and easy to execute, they are There is a valid and a fraudulent category for
ineffective and cannot keep up with the changes every single transaction.
in the nature of fraud. They have a tendency to • Unsupervised learning: One way to train ML
give out high false-positive outcomes and cannot algorithms is through unsupervised learning,
be useful in representing complex or subtle fraud which involves using datasets with potentially
patterns. The limitations of rule-based systems in unclear target variables.[19] Finding the most
learning and development make them an incomplete important patterns in the data is the model’s
solution to the credit card fraud detection problem goal. Cluster segmentation and dimensionality
when applying DL. They are, therefore, becoming reduction are examples of unsupervised
augmented and even superseded by even more learning approaches.
data-driven methods, more importantly DL models, • Semi-supervised learning: Models can be
which can learn on its own to detect non-linear trained using unlabeled data by semi-supervised
relationships and shifting fraudulent patterns. learning, a type of learning that combines
Despite its shortcomings, the rule-based systems are supervised and unsupervised techniques. In
also useful in hybrid frameworks of fraud detection, this approach, the best way to represent data
where they help perform the initial screening before is determined by the unsupervised learning
performing a deeper-level inspection. attribute, and then the relationships within that
representation are analyzed by the directed
ML approaches learning attribute, giving rise to predictions.
ML is a technique that may be applied to many
different issues, particularly in fields that deal OVERVIEW OF DL IN FRAUD
with data analysis and processing. The imbalanced DETECTION
dataset can be resolved with the help of ML, which
can be categorized as supervised, unsupervised, or Figure 3 shows that DL, a subset of ML, makes use
reinforcement ML. The ability to automatically of neural networks to either generate predictions or
recognize patterns across massive amounts of convert incoming data into new representations.
data is what makes ML approaches so useful for Among the many architectures used in DL are
detecting and preventing fraud. Differentiating deep reinforcement learning, convolutional neural
between genuine and fraudulent activities networks (CNNs), RNNs, transformers, and deep
becomes easier with the right ML models in place. neural networks (DNNs). When applied to tasks such
Eventually, these smart systems might figure out as picture recognition, Natural Language Processing
how to cheat previously unknown frauds. Figure 2 (NLP), Computer Vision (CV), and speech
displays various ML methods. recognition, these DL architectures have achieved
• Supervised learning: Supervised learning trains results that are competitive with, and even better than,
ML systems using labeled or modifiable data those of human specialists. At the same time, RNNs
sets with known variable goals. Classification, are shining brightest when it comes to sequential data
inference, and regression are all instances of modeling, which includes things such as credit card
supervised learning. Most ML approaches transactions.[11] To address the vanishing gradients

Figure 2: Machine learning Figure 3: Deep neural network

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 4
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

problem, the long short-term memory (LSTM) was SMOTE

developed.[20] While RNNs perform well as models The SMOTE strategy employs an oversampling
for sequence data, their training becomes challenging procedure with the main objective of increasing
when gradients explode or disappear. The Gated the minority class’s case count, leading to the
Recurrent Unit (GRU) can maintain data and perform formation of a new synthetic sample.[22] A new
LSTM-like tasks without an additional memory unit, synthetic sample can be generated from any
despite its development in. To improve the accuracy point along the line that starts with the minority
of their state forecasting algorithms, bidirectional class samples’ nearest neighbors. The experiment
versions of these networks use data from both the employs the SMOTE class of the imbalanced-
past and the future. learn package to perform the over-sampling. The
total number of occurrences of the minority class
after resampling divided by the total number
General DL Workflow for Fraud Detection of occurrences of the majority class yields the
The typical workflow for implementing DL in sampling ratio.
fraud detection involves several key stages:
Principal component analysis (PCA)
Data collection and integration Data scaling commonly employs principal
The fraud detection system is built upon the records component analysis. A new set of uncorrelated
of financial institutions’ transactions. The input variables, or PCs, is formed that sequentially
transactional data are retrieved from the transaction maximizes variance to reduce the number of
database to execute the fraud detection process.[21] correlated variables from n to m. The newly
The following will be displayed if the database has n introduced variables are linear combinations of
bits of transactional data, which it will refer to as K: their older, less important counterparts. The goal
is to account for as much variation in the original
K = {H1,H2……Hn}: 1≤l≤n (1)
data as can be feasibly explained using the first
Database (K), transaction data (H), and the total principal component.
number of records relevant to transactions are
shown in the first equation. The preprocessing step
DL Models for Fraud Detection
takes the input data H1, selected from the database,
to move on with processing, the detection method. There are some models of DL for fraud detection
discussed below:
Data preprocessing
The data that H1 collected from the database are Recurrent neural networks (RNNs) and LSTM
to be preprocessed by the preprocessing module. The fraud detection sector has adopted RNN-based
Data cleaning, normalization, and encoding should DL techniques because these algorithms are often
precede training. considered to be among the most accurate for
sequence analysis tasks.[23] Rapid neural networks
Model selection and training (RNNs) can assess the changing temporal
Prioritize CNNs, LSTMs, or hybrid models as behavior of various bank accounts by mimicking
your DL architectures of choice. Train them with the sequential reliance between subsequent credit
transaction data from the past. card transactions.
LSTM artificial RNN [Figure 4]. To forecast
a transaction label from a sequence of prior
Validation and testing transactions, LSTM neural networks use
Evaluate model performance using labeled test feedback connections between hidden units
data to assess generalizability. linked with discrete time steps and, unlike
traditional feedforward networks, learn sequence
Data Preprocessing Techniques dependencies over time.[24] To address the problem
of regular RNNs training with vanishing and
Below are some techniques of data preprocessing: exploding gradients, the LSTM neural network

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 5
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

was created. A promising use of LSTM networks is best possible solution for any given problem.
the detection of fraudulent credit card transactions. Its building blocks are neurons arranged in a
hierarchical fashion. A DNN is an ANN that uses a
network of linked neurons to simulate the way the
Autoencoders
human brain works. A multitude of mathematical
An autoencoder is an unsupervised learning model computations are performed by these neurons
that attempts to reconstruct its input, denoted as x, using data supplied from the activation function
through a compressed representation. The output x ̂ is and the input layer. With the help of a feedforward
a reconstruction of the original input x. Autoencoders multilayer perceptron, a specific kind of three-
learn by encoding the input into a lower-dimensional layer neural network, the researchers were able
latent space and then decoding it back to its original to organize the data in this specific case. This
form. The encoder maps the input to the hidden network only processes input in one way. For
layer (latent representation), and the decoder mathematical computations in the hidden layer and
reconstructs the input from this representation. output layer, a differentiable sigmoid activation
Using low-dimensional, nonlinear hidden layers, function is utilized. It tackles non-linearly
the autoencoder captures essential features of the separable problems using the sigmoid activation
input data. However, due to the bottleneck structure function.[26] To train itself, the network makes
which limits the information throughput the model use of the backpropagation learning method, a
must learn efficient compression of the input. The supervised learning technique. The network was
structure of the autoencoder, including the hidden trained using an online financial dataset.
layer, is illustrated in Figure 5. MLP, as depicted in Figure 6, consists of
four nodes that take in a vector of four inputs
(x1,x2,x3,x4). Divide all transactions into two
DNNs/Multilayer Perceptrons (MLPs) groups: suspicious (S) and non-suspicious (-S). It
is from the MLP’s output that it determines if a
The unique architecture of DNNs enables them
transaction is suspicious (S) or not (-S).
to learn and adapt, ultimately leading to the

Hybrid and Ensemble Deep Models

A hybrid ensemble model that combines the
capabilities of CNNs, LSTMs, and transformers,
and is built on the XGBoost meta-learner.
Figure 4: Long short-term memory
Incorporating the supplementary features of the
two models helps strengthen and improve credit
card fraud detection systems.[27] To understand the
spatial patterns and localized characteristics in
the transactional data, CNNs are utilized as base

Figure 5: Autoencoder with hidden layers Figure 6: Multilayer perceptron

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 6
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

learners. As a foundational learning mechanism, sparseness of fraud cases. Their ANN model beats
LSTMs are used to grasp the sequential patterns all other models in the credit card fraud detection
and temporal dependencies included in transaction arena in terms of accuracy and durability, with a
sequences. LSTMs can detect suspicious patterns record of 99.99%.[28]
of transactions that would not be obvious from a Elmangoush et al. investigate the matter and create
single transaction. On the other hand, transformers a dependable model to detect cases of credit card
are used as foundational learners to record intricate fraud. A synthetic minority oversampling approach
interdependencies and interactions in transaction has been implemented to tackle the issue of class
sequences. imbalance. Developing a model for credit card
The suggested hybrid ensemble model makes use identification using sequential DL methods was the
of XGBoost as its meta-learner feature. XGBoost subsequent stage. By enhancing feature extraction
is an effective gradient-boosting technique that and representation using data supplied by SMOTE,
captures intricate patterns and works wonders this approach tackles the issue of insufficient features.
with tabular data. A hybrid ensemble model uses They compared the proposed model to existing
a stacking technique to aggregate predictions literature in the field and looked at its accuracy,
from the CNN, LSTM, and transformer base detection rate, and f-measure. The outcomes
learners, with XGBoost serving as the meta- demonstrate that the proposed model surpasses the
learner. Pictured in Figure 7 are the output vectors most advanced models in this domain.[29]
PCNN, PLSTM, and PTransformer, which are the results of Mohammad and Logeshwaran proposed a learning
each base model’s prediction on the validation set strategy for real-time credit card fraud detection.
following its independent training on the training Their framework leverages deep learning (DL)
set. The XGBoost meta-learner trains on the techniques to enable fast and accurate identification
ensemble of predictions from each base learner of fraudulent transactions. By utilizing advanced
and uses these outputs as input features to make DL algorithms, the system analyses real-time
the final classification decision. transactional data to detect anomalies indicative
of fraud. Trained through a data-driven approach,
the DL-based model can dynamically adapt to
LITERATURE REVIEW
emerging fraud patterns, enhancing its detection
Credit card fraud detection research on four capabilities over time. This adaptability allows it
DL architectures, ANN, CNN, autoencoder, to maintain high accuracy in identifying fraudulent
and sequential models, is compiled here. Class behaviour. Moreover, the framework supports real-
imbalance, sparse features, and real-time detection time analytics on large volumes of data, making it
are some of the main issues that are tackled. well-suited for financial institutions that process
A comparative summary of recent approaches, high transaction volumes. Notably, the architecture
performance metrics, and future directions is offers significantly higher throughput for a
provided in Table 1 for a concise overview. given latency compared to traditional methods,
Yadlapalli et al., the major focus of this study is potentially reducing response time and minimizing
to build an ANN model and compare it with other financial losses due to fraudulent activities..[30]
famous ML models that can detect credit card Satti et al., the investigation’s objective is to
theft and help users stay secure. The reliability establish a connection between these frauds.
and the resilience of the ANN model were ensured The ML literature mentions a number of credit
by fine-tuning it to handle class imbalance and card recognition methods, including XG Boost,

Figure 7: Proposed stacking ensemble approach

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 7
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

Table 1: Literature review based on deep learning architectures for credit card fraud in financial systems
Reference Study on Approach Key findings Challenges Future direction
Yadlapalli et al. (2025) ANN vs. ML ANN optimized for Achieved 99.99% Sparse fraud patterns; Extend to hybrid DL
models class imbalance accuracy; outperformed imbalanced data models for further
traditional ML models improvement
Elmangoush et al. (2024) Sequential DL with SMOTE + Sequential 0.99924 Accuracy and Class imbalance and Explore advanced
SMOTE DL for better feature 0.75976 F‑measure; feature insufficiency feature engineering and
extraction superior to existing real‑time detection
models
Mohammad and Detecting digital DL with real‑time High throughput and Latency vs. accuracy Incorporate edge
Logeshwaran (2024) fraud in real‑time analytics adaptability to new trade‑offs computing for ultra‑low
frauds latency detection
Satti et al. (2024) Comparative Compared DL models AUC of up to 98.1%, DL models still lag in Combine ensemble and
analysis of ML/DL with XGBoost, SVM, F1‑score of 85.71, and generalization DL for robust detection
RF, etc. accuracy of 99.9%
Bharath et al. (2023) Autoencoder‑based ILSDFD using Adapts to new fraud Evolving fraud Employ continuous
DL autoencoder and patterns with high patterns and feature learning and adaptive
feature selection performance drift training
Chaudhari et al. (2023) DL + Face Two‑step verification Enhanced verification Privacy and hardware Improve facial
detection using face + DL mechanism with DL dependency verification robustness
integration with biometric fusion
Gambo et al. (2022) CNN + ADASYN CNN with ADASYN Achieved 0.9982 Class imbalance and Optimize CNN
Sampling accuracy, 0.9965 CNN overfitting architecture and use
precision, 0.9999 recall explainable AI for
transparency
ANN: Artificial neural network, ML: Machine learning, DL: Deep learning, CNN: Convolutional neural network, AUC: Area under the curve, AI: Artificial intelligence,
ADASYN: Adaptive synthetic

Decision Tree (DT), Random Forest (RF), Support activity. However, the dynamic nature of consumer
Vector Machine (SVM), Logistic Regression (LR), behavior and the evolving tactics of fraudsters
and extreme learning method. Ultramodern deep pose significant challenges to traditional ML
literacy algorithms are still needed toward lower models. To address these limitations, the authors
fraud losses due to their low delicacy. Exercising proposed a deep learning (DL)-based framework
the most recent developments in deep literacy capable of analyzing both historical and real-time
algorithms has been primarily ideal in order to transactional data to detect anomalies. DL models
achieve this. Toward achieving effective results, demonstrated superior performance compared to
a relative analysis of machine literacy and deep conventional ML methods due to their ability to
literacy ways was conducted.[31] extract complex features and adapt to new fraud
Bharath et al., the proliferation of online shopping patterns. Their proposed solution includes a
and payment systems has been linked to an increase two-stage verification system that combines DL
in financial fraud, particularly the fraudulent use of techniques with facial recognition technology to
credit cards. As a result, instantly implementing enhance the accuracy and robustness of credit
systems that can detect credit card theft is essential. card fraud detection.[33]
People all across the world are increasingly making Gambo et al. investigate the skyrocketing growth
purchases using stolen credit cards, and this situation of online purchases made through e-commerce
needs to be addressed immediately. Limiting this platforms, highlighting the pervasiveness of credit
kind of behavior is good for consumers because card payments. Consumers and banks lose billions
the system’s main purpose is to stop users from of dollars annually because of CCF strategies that
being charged for services and goods they did are always evolving to match the speed of industry
not authorize. Extreme caution is required when transformation. It is critical to successfully
selecting fraudulent transaction attributes for use in distinguish fraudulent transactions from the
ML-based credit card fraud detection.[32] many lawful ones. To address the imbalance in
Chaudhari et al. explored various machine learning the dataset, this study suggests a CNN model for
(ML) techniques for classifying fraudulent credit credit card fraud detection that is based on the
card transactions. These approaches are based adaptive synthetic sampling technique.[34]
on the assumption that patterns in historical The literature review is summarized in Table 1,
transaction data can help identify suspicious which highlights the emphasis, methodology,

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 8
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

important findings, obstacles, and suggested decision trees and support vector machines. In: IMECS
future paths of each study. 2011 - International Multi Conference of Engineers and
Computer Scientists. Vol. 1. Berlin: Springer; 2011.
p. 442-7.
CONCLUSION AND FUTURE WORK 6. Wawge SJ. A survey on the identification of credit
card fraud using machine learning with precision,
The growing number of online transactions and the performance, and challenges. Int J Innov Sci Res
Technol 2025;10:3345-52
ever-changing methods used by cybercriminals
7. Mantha G. Transforming the insurance industry with
have made the detection of credit card fraud an sales force: Enhancing customer engagement and
increasingly pressing issue. There have been operational efficiency. North Am J Eng Res 2024;5:3-4.
several applications of DL approaches because 8. Singh V. Predicting loan default risk in P2P lending
of their efficacy and reliability. Credit card platforms: A study of lending club borrowers. Int J Sci
Res 2023;12:2255-60.
fraud employing DL models has recently been
9. Chaudhari B, Verma SC. Synergizing generative AI and
discovered to be significantly more common. machine learning for financial credit risk forecasting
To find credit card scams more effectively than and code auditing. Int J Sci Res Comput Sci Eng Inf
traditional rule-based and classical ML methods, Technol 2025;11:2882-93.
this study shows that DL methods such as LSTM, 10. Abakarim Y, Lahby M, Attioui A. An Efficient Real
Time Model For Credit Card Fraud Detection Based On
CNN, Autoencoders, and hybrid ensemble
Deep Learning. In: Proceedings of the 12th International
models are the best. These models greatly Conference on Intelligent Systems: Theories and
improve the accuracy of detection and cut down Applications. New York, NY, USA: ACM; 2018. p. 1-7.
on false positives by learning complicated, non- 11. Mienye ID, Jere N. Deep learning for credit card fraud
linear patterns and temporal relationships from detection: A review of algorithms, challenges, and
transactional data. Due to the dynamic nature of solutions. IEEE Access 2024;12:96893-910.
12. Akdemir N, Yenal S. Card-not-present fraud
financial fraud tactics, DL frameworks provide a victimization: A routine activities approach to
scalable and adaptable means of preventing real- understand the risk factors. Güvenlik Bilim Derg
time fraud in online banking systems. 2020;9:243-68.
The main goal of future work will be to create real- 13. Chaudhari B, Verma SC. Achieving high-speed data
time scam detection systems that use federated consistency in financial microservices platforms using
NoSQL using Nosql (Mongodb, Redis) technologies.
learning to make sure that institutions can train Int J Adv Res Sci Commun Technol 2024;4:750-9.
models in a way that respects privacy. Furthermore, 14. Mishra A. Ai-powered cybersecurity framework for
by integrating explainable artificial intelligence, secure data transmission in Iot network. Int J Adv Eng
model transparency will be enhanced, which will Manag 2025;7:5-13.
help human analysts comprehend and have faith 15. Burns P, Stanley A. Fraud Management in the Credit
Card Industry. Philadelphia, PA: Ten Independence
in automated fraud choices. Expanding datasets to
Mall; 2002. p. 1-16.
include multi-channel transaction data and applying 16. Ramanujam B. Statistical in sights in to anti-
graph neural networks to uncover relational fraud money laundering : Analyzing large-scale financial
patterns are also promising directions for advancing transactions. Int J Eng Res Technol 2025;14:1-6.
fraud detection capabilities. 17. Chatterjee P, Das A. Adaptive financial recommendation
systems using generative ai and multimodal data.
J Knowl Learn Sci Technol 2025;4:112-20.
REFERENCES 18. Pillai V. System and method for intelligent detection
and notification of anomalies in financial and insurance
1. Unogwu OJ, Filali Y. Fraud detection and identification data using machine learning. Pat Off J 2025.
in credit card based on machine learning techniques. 19. Kali H. Optimizing credit card fraud transactions
Wasit J Comput Math Sci 2023;2:16-22. identification and classification in banking industry
2. Wawge SJ. Evaluating machine learning and deep using machine learning algorithms. Int J Recent Technol
learning models for housing price prediction: A review. Sci Manag 2024;9:85-96.
Int J Adv Res Sci Commun Technol 2025;5:367-77. 20. Garg S. AI, blockchain and financial services:
3. Dornadula VN, Geetha S. Credit card fraud detection Unlocking new possibilities. Int J Innov Res Creat
using machine learning algorithms. Procedia Comput Technol 2022;8:1-4.
Sci 2019;165:631-41. 21. Kolli CS, Tatavarthi UD. Fraud detection in bank
4. Majumder RQ. A review of anomaly identification in transaction with wrapper model and Harris water
finance frauds using machine learning systems. Int J optimization-based deep recurrent neural network.
Adv Res Sci Commun Technol 2025;5:101-10. Kybernetes 2021;50:1731-50.
5. Sahin Y, Duman E. Detecting credit card fraud by 22. Karthika J, Senthilselvi A. Smart credit card fraud

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 9
 Rathore: Analytical Review of Deep Learning Architectures in Financial Frameworks for Fraud Identification of Credit
Cards

detection system based on dilated convolutional neural Technologies, Signal and Image Processing (ATSIP);
network with sampling technique. Multimed Tools 2024. p. 455-8.
Appl 2023;82:31691-708. 30. Mohammad R, Logeshwaran J. Real-Time Credit
23. Benchaji I, Douzi S, El Ouahidi B. Credit card fraud Card Fraud Detection using Deep Learning Based
detection model based on LSTM recurrent neural Framework. In: 2024 2nd International Conference
networks. J Adv Inf Technol 2021;12:113-8. on Advances in Computation, Communication
24. Janumpally BK. Architecting serverless payment and Information Technology (ICAICCIT); 2024.
gateways: A systematic analysis of scale, security, and p. 496-500.
performance trade-offs. Int J Res Comput Appl Inf 31. Satti P, Tamilkodi R, Maduri N, Ratalu S,
Technol 2025;8:1186-201. Saritha K, Santharaju N. Credit Card Fraud Detection
25. Al-Shabi MA. Credit card fraud detection using Using Machine Learning and Amp Deep Learning
autoencoder model in unbalanced datasets. J Adv Math Algorithms. In: 2024 2nd International Conference
Comput Sci 2019;33:1-16. on Signal Processing, Communication, Power and
26. Mbunge E, Makuyana R, Chirara N, Chingosho A. Embedded System (SCOPES). United States: IEEE;
Fraud detection in E-transactions using deep neural 2024. p. 1-5.
networks-a case of financial institutions in Zimbabwe. 32. Bharath AP, Rajendran N, Devi SD, Saravanakumar S.
Int J Sci Res 2015;6:2319-7064. Experimental Evaluation of Smart Credit Card Fraud
27. Chatterjee P. Proactive infrastructure reliability: AI- Detection System using Intelligent Learning Scheme.
powered predictive maintenance for financial ecosystem In: 2023 International Conference on Innovative
resilience. J Artif Intell Gen Sci 2024;7:291-303. Computing, Intelligent Communication and Smart
28. Yadlapalli P, Srivatsal P, Polimera N, Srinivas M. Electrical Systems (ICSES); 2023. p. 1-6.
Credit Card Fraud Detection using Machine Learning 33. Chaudhari A, Sontakke A, Tendulkar S, Zambare S,
Algorithms and Artificial Neural Network. In: 2025 Yewale M. Detecting Credit Card Frauds Using Deep
International Conference on Artificial Intelligence and Learning and Face Detection. In: 2023 7th International
Data Engineering (AIDE). United States: IEEE; 2025. Conference On Computing, Communication, Control
p. 539-42. And Automation (ICCUBEA); 2023. p. 1-6.
29. Elmangoush AM, Hassan HO, Fadhl AA, Alshrif MA. 34. Gambo ML, Zainal A, Kassim MN. A Convolutional
Credit Card Fraud Detection Using Synthetic Minority Neural Network Model for Credit Card Fraud Detection.
Oversampling Technique and Deep Learning Technique. In: 2022 International Conference on Data Science and
In: 2024 IEEE 7th International Conference on Advanced Its Applications (ICoDSA); 2022. p. 198-202.

AJCSE/Apr-Jun-2025/Vol 10/Issue 2 10