www.ijcrt.
org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
Smart-Pay Shield Using GAN
1Jemima D,2LathehaavanthanieM.,3Mr.P.Murugan 4Dr.J.Hemalatha,5Mr.C.Praveenkumar
1,2
UG student, Department of Computer Science and Engineering AAA College of Engineering and
Technology, Sivakasi.
5
Assistant Professor, Department of Computer Science and Engineering AAA College of Engineering and
Technology, Sivakasi.
4
Professor & Head, Department of Computer Science and Engineering AAA College of Engineering and
Technology, Sivakasi.
Abstract: The widespread adoption of digital payment systems, particularly the Unified Payments Interface
(UPI) in India, has led to a corresponding increase in fraudulent transactions, necessitating the development
of more intelligent and adaptive fraud detection mechanisms. Conventional systems based on static rules or
traditional machine learning models like Support Vector Machines (SVM) and Random Forests (RF) lack the
capability to effectively identify new, evolving fraud patterns due to their limited generalization ability and
inability to simulate unseen behavior. In this study, we introduce a novel fraud detection framework using
Generative Adversarial Networks (GANs), comprising a Generator that learns to synthesize realistic
fraudulent transaction data by capturing the statistical properties of known fraud cases, and a Discriminator
that is trained to distinguish between real and synthetic fraudulent data as well as legitimate transactions. This
adversarial learning process enhances the model’s ability to identify both common and rare forms of fraud,
including zero-day attacks. The system is trained on a diverse set of transaction-level features including
amount, timestamp, geolocation, merchant category, device fingerprinting, IP address history, and user
behavior profiles, enabling a comprehensive analysis of transaction context. Extensive experiments
demonstrate that the GAN-based model achieves higher precision, recall, and AUC scores compared to
baseline classifiers, particularly in imbalanced datasets where fraudulent transactions are rare. Furthermore,
the model's architecture supports integration into live fraud monitoring pipelines through RESTful APIs,
allowing for real-time detection with continuous model refinement using streaming data. In future work, we
aim to incorporate multimodal data such as biometric signals, voice commands (from voice-based UPI apps),
and social network behavior to further improve detection accuracy and contextual awareness. This research
highlights the potential of GANs as a powerful tool in financial cybersecurity, providing adaptive, scalable,
and intelligent protection against the dynamic landscape of UPI fraud.
INTRODUCTION
The growth of digital payment systems has transformed financial transactions worldwide. Among them, the
Unified Payments Interface (UPI) has transformed India, enabling seamless, real-time transactions between
many banks and platforms.Made simple and convenient, UPI has grown exponentiallysince its introduction,
with billions of transactions happeningevery month. With its extensive usage, however, there has been a sharp
spike in fraudulent transactions taking advantage of weaknesses in digital payment ecosystems. Phony
transactions, phishing, identity theft, and social engineering fraud on UPI users have become highly
sophisticated, causing severe threats to financial security and trust among users. Legacy fraud detection
techniques predominantly employ rule-based systems or classic machine learning models such as decision
trees, logistic regression, and support vector machines. These techniques are very reliant on historical data and
known patterns and are not as effective when faced with new or evolving patterns of fraud. The fraudsters will
always adjust their approach in order to get around existing security Identify applicable funding agency here.
If none, delete this. measures, and hence more adaptable, dynamic, and intelligent fraud detection systems are
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e291
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
required.Recent advances in deep learning have opened up new pos-sibilities for fraud detection, with neural
network models capa-ble of learning complex and non-linear patterns in large datasets. Generative Adversarial
Networks (GANs), invented by Goodfellow et al., have been incredibly successful in a broad range of
applications such as image generation, data augmentation, and anomaly detection. A GAN consists of two
neural networks—the Generator and the Discriminator—trained in parallel in a competitive setting. The
Generator attempts to create synthetic samples that are indistinguishable from real data, whereas the
Discriminator attempts to distinguish between real and synthetic data. This adversarial mechanism enables the
networks to learn complex data distributions, and hence GANs can be a promising solution for fraud
detectionapplications.For detection of UPI fraud, GANs can be used to address two critical problems: absence
of labeled data of fraudulent transactions and identification of novel patterns of fraud. Using
The generation of synthetic realistic fraudulent transactions,GANs can augment small datasets for improved
training of models. Moreover, the Discriminator network, trained in the identification of real vs. fraudulent
transactions, learns to detect subtle variations typical of real fraud. This paper presents a novel GAN-based
technique for UPI fraud detection. The primary objectives are to improve detection of existing and new fraud
patterns, reduce false positives and enable the system to be robust against the dynamic nature of fraud. The
technique involves data preprocessing to obtain useful features such as transaction value, timestamp, location,
merchant category, and device. A GAN is trained to generate synthetic fraudulent transactions, and the
Discriminator is tuned to mark transactions as authentic or fraudulent. Experimental outcomes demonstrate
that the suggested GAN-based model is better compared against traditional machine learning methods
regarding accuracy, recall, and F1-score. The model also exhibits robust performance even in the face of unseen
fraud patterns, and thus it can be implemented in real-world settings in banks. The rest of this paper is structured
as follows: Section II gives related work in fraud detection and GAN applications. Section III explains the
proposed methodology. Section IV gives experimental results and analysis. Finally, Section V concludes the
paper and future research directions.
2.Literature review:
[1] X. Zhang et al., “HOBA: A novel feature engineering methodology for credit card fraud detection
with a deep learning architecture,”Inf.Sci.,May2019.
This paper introduces HOBA, a new feature engineering approach that enhances the performance of deep
learning models for fraud detection. By using behavioral analysis and temporal features, HOBA captures
dynamic user patterns, significantly improving detection rates and reducing false positives. The proposed
deep learning model outperforms traditional classifiers in handling complex fraud patterns.
[2] N. Carneiro, G. Figueira, and M. Costa, “A data mining based system for credit-card fraud
detection,” Decis. Support Syst.,vol.95,pp.91–101,Mar.2017.
The authors present a data mining framework using decision trees, neural networks, and support vector
machines to detect fraudulent transactions. They highlight the importance of feature selection, data balancing,
and ensemble methods in improving model accuracy and generalization. Their results suggest that hybrid
models yield better detection precision in real-time settings.
[3] B. Lebichot et al., “Deep-learning domain adaptation techniques for credit cards fraud detection,”
INNS Big Data Deep Learn.Conf.,2019.
This study explores domain adaptation methods for fraud detection using deep learning. The paper addresses
the challenge of data distribution shifts between different institutions or regions and proposes techniques such
as adversarial adaptation and transfer learning to maintain performance across domains.
[4] H. John and S. Naaz, “Credit card fraud detection using local outlier factor and isolation forest,”
Int. J. Comput. Sci. Eng.,Sep.2019.
The authors use unsupervised anomaly detection techniques—Local Outlier Factor (LOF) and Isolation
Forest—for detecting fraudulent transactions in highly imbalanced datasets. Their model effectively
identifies rare and novel fraud patterns without relying on labeled data, making it suitable for dynamic
environments.
[5] C. Phua et al., “Communal analysis suspicion scoring for identity crime in streaming credit
applications,” Eur. J. Oper. Res.,Jun.2009.
This paper introduces a suspicion scoring technique based on communal analysis for real-time detection of
identity fraud in streaming credit applications. The system evaluates behavior deviations within a community
of users and flags abnormal activities using statistical and clustering techniques.
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e292
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
[6] R. Bolton and D. Hand, “Statistical fraud detection: A review,” Stat. Sci., Aug. 2002.
A comprehensive review of statistical techniques for fraud detection, including supervised, unsupervised, and
semi-supervised methods. The authors discuss the trade-offs between model complexity, interpretability, and
detection accuracy, emphasizing the need for adaptive methods to cope with evolving fraud tactics.
[7] P. A. Dal et al., “Credit card fraud detection: A realistic modeling and a novel learning strategy,”
IEEE Trans. Neural Netw.Learn.Syst.Sep.2017.
This work focuses on realistic fraud detection scenarios by simulating real-world challenges like delayed
feedback and imbalanced data. The authors introduce a learning strategy that combines sampling, cost-
sensitive learning, and incremental training to improve fraud recognition performance over time.
[8] S. Bhattacharyya et al., “Data mining for credit card fraud: A comparative study,” Decis. Support
Syst., Feb. 2011.
The paper presents a comparative analysis of several data mining algorithms—decision trees, neural
networks, SVMs, and Bayesian networks—on fraud detection tasks. It concludes that ensemble and hybrid
approaches outperform individual models, especially when paired with robust preprocessing techniques.
[9] N. Sethi and A. Gera, “A revived survey of various credit card fraud detection techniques,” Int. J.
Comput. Sci. Mobile Comput.Apr.2014.
A survey that categorizes fraud detection techniques into supervised, unsupervised, and hybrid approaches. It
reviews algorithms such as KNN, neural networks, genetic algorithms, and fuzzy logic, providing insights
into their strengths, limitations, and real-world applicability.
[10] A. O. Adewumi and A. A. Akinyelu, “A survey of machine-learning and nature-inspired based
credit card fraud detection techniques,”Int.J.Syst.AssuranceEng.Manage.,Nov.2017.
This survey explores both machine learning and bio-inspired methods (e.g., ant colony optimization, particle
swarm optimization) for fraud detection. It analyzes their effectiveness in feature optimization, decision
making, and adaptability to changing fraud behaviors.
[11] A. Dal Pozzolo et al., “Calibrating Probability with Undersampling for Unbalanced
Classification,” IEEE SSCI, 2015.
The paper proposes a calibration framework that improves classifier performance on imbalanced datasets via
strategic undersampling and probability calibration. The authors demonstrate that their approach yields better
precision-recall trade-offs in fraud detection tasks.
[12] A. Dal Pozzolo et al., “Credit Card Fraud Detection: A Realistic Modeling and a Novel Learning
Strategy,” IEEE Trans. NeuralNetw.Learn.Syst.,2018.
An extension of previous work, this paper offers an in-depth learning strategy combining incremental
learning, adaptive thresholds, and real-world constraints. It emphasizes the need for temporally aware models
and proposes continuous updates to maintain relevance.
[13] T. Nguyen et al., “Machine Learning Based Approaches for Detecting Fraudulent Transactions in
Online Payment Systems,”JISA,2021.
This study investigates the use of supervised machine learning methods—logistic regression, decision trees,
and neural networks—for online payment fraud detection. It evaluates different sampling and feature
selection techniques, concluding that balanced datasets and ensemble learning significantly boost
performance.
[14] R. Singh and S. Bharti, “Financial Fraud Detection Using Machine Learning,” Proc. ICCIDS,
2021.
This paper examines the application of machine learning algorithms—especially ensemble models and deep
neural networks—for detecting financial fraud. It highlights key challenges such as data imbalance, model
interpretability, and the need for real-time deployment frameworks.
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e293
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
3. PROPOSED METHODOLOGY
The main objective of this research is to create an intelligent system that can identify fraudulent UPI
transactions, including fraud types that haven’t been identified yet. The suggested model, in contrast to
conventional rule-based systems, makes use of GANs to create fictitious fraudulent transactions, improving
the discriminator’s capacity to recognize intricate fraudulent patterns. A GAN-based training mechanism, a
data preprocessing pipeline, and an evaluation system that compares performance to industry-standard fraud
detection benchmarks are all integrated into the framework.
Data preprocessing, feature extraction, GAN model training, and fraud classification using the trained
discriminator comprise the method’s structured flow.
3.1 Data Collection and Preprocessing
This research work is built upon the real-life UPI transacting dataset which has many other attributes. These
attributes include transaction amount, time stamp, ids of sender and receiver, location, device information used,
transaction status, fraud labels, etc. Because fraud datasets are naturally imbalanced - that is, they include a
small percentage of false transactions compared to a considerably larger crowd of legitimate transactions -
special data preprocessing strategies are employed. The initial preprocessing phase involves filling up missing
values using either mean or mode imputation, based on attribute type. Categorical attributes such as city, device
type, and transaction modes are usually encoded using one-hot encoding techniques. Continuous attributes like
transaction amount and interval time used in transactions are normalized since they need to remain in a standard
range; given that it also prevents the unnatural behavior of deep learning models during training, outliers with
high transaction amounts will be detected using statistical threshold techniques and removed from the training
data to avoid skewed model learning.
3.2 Feature Engineering
Crucial in improving the predictive capability of the model, feature engineering transforms raw transaction
data into derived attributes that give insights into user behavior and possible anomalies. Some features may
include transaction frequency within certain time intervals, average transaction amount per user, score of device
diversity that indicates the number of devices that can be linked to a single account, a time gap in between two
consecutive transactions. Geographic movement, or the movement from one location to another while carrying
out consecutive transactions, is also taken into consideration since abrupt changes imply possible compromised
accounts. Each of the built features is assumed to enhance the ability of normal and abnormal transaction
pattern differentiation.
3.3GAN Model Architecture
In the system proposed, Generative Adversarial Networks play a central role. The Generative Adversarial
Network consists of two main parts: the Generator and the Discriminator.The generator tries to create synthetic
samples of fraudulent transactions that look like real fraudulent records, while the discriminator tries to
differentiate between real and fake transactions.The generator accepts a random noise vector as input,typically
sampled from a uniform or normal distribution, and maps it through several hidden layers to produce output
that resembles the feature vector of a transaction. ReLU and Tanh activation functions are employed in hidden
layers and output layers, respectively, to ensure a non-linearity and bound range of output. On the other side,
the Discriminator receives transaction feature vectors, which can be either real or synthetic, and
predictstheirauthenticity. The architecture consists of multiple dense layers using LeakyReLU activations that
culminate in a final output neuron with a sigmoid activation to perform binary classification.Both networks
continuously improve themselves by adversarial training, which leads to a Generator that can create
convincingly realistic fraudulent transactions and a Discriminator able to classify fraud effectively.
3.4 Training Strategy
Training GAN involves an iterative process in which the generator and discriminator are trained alternately.
Initially, the discriminator is trained on a mixed batch of real and synthetic transactions to ensure its
performance on correct classifications. The generator is then updated according to the feedback from the
discriminator, with the desired objective of producing samples misclassified by the discriminator as
real.Batches of 64 transactions are typically used to train the models under consideration. Adam optimizer is
also used for both networks, at a learning rate of 0.0002 for stable training. Label smoothing is also incorporated
while training the Discriminator to prevent becoming overconfident in its predictions, which may, in turn,
destabilize the GAN’s learning process.The loss functions for both networks are binary cross entropy losses.
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e294
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
The objective of the generator here is to minimize the capability of the discriminator in distinguishing between
real and generated samples; in contrast, the discriminator wants to maximize this ability.
3.5 Model Evaluation
When training completes, the Discriminator is now used as a classifier for fraud. The test it is submitted to
involves the detection of real transactions not seen during training. The evaluation metrics include Accuracy,
Precision, Recall, F1-Score, and the ROC AUC score. The amount of transactions that were flagged to be
fraudulent and were in fact fraudulent. The fraudsters that were in fact fraudulent and got detected.The F1-
Score basically turns Precision and Recall into a single measure of performance. In essence, ROC-AUC checks
the ability of the model to separate between the legits and the frauds across various classification thresholds.
The GAN approach is compared to older models like SVMs,Random Forests, and Logistic Regression. The
outcomes from all of the experiments show the GANs having the most prominent recall and F1 results,
confirming their ability to detect fraud, particularly for new patterns not contained within training data.
3.6 Handling Class Imbalance
A significant advantage of using GANs is their ability to tackle class imbalance. Since it is well-known that
fraud transactions make up a very small population in any data set,model training tends to be biased toward
the majority class (legitimate transactions). Using the Generator, we generate synthetic fraudulent transactions
to augment the minority class so that we end up with a better-balanced dataset without traditional oversampling
techniques like SMOTE. This augmentation aspect can lead to improved model performance, especially recall,
which is critical for any fraud detection system.
3.7 Deployment Considerations
The challenge of taking the GAN-based fraud detection model to production carries with it the need to factor
operational efficiency and flexibility into the design. As fraud patterns are subject to change over time, it is
imperative that systems for continual learning and periodic retraining model implementation are in place for
gaging with incoming new data.In addition, inference time needs to be optimized to accommodate fraud
detection in real time. Lightweight instances of a trained Discriminator can be crafted for speedy prediction
purposes. Moreover, in order to assure system trustworthiness, explainable AI approaches, possibly making
use of SHAP values or LIME, can be integrated for the interpretation of why a given transaction has been
classified as fraud. Also, security measures must be taken into account to safeguard the system itself against
adversarial attacks, with the threat of manipulative actors attempting to alter input features to throw the
detection mechanism off their trail.
3.9 System Architecture
The UPI fraud detection system architecture using GANs is designed to control the flow of transactional
data from ingestion fraud classification. It aims to accommodate massive-scale financial datasets whilst
guaranteeing the accurate and real-time detection of fraudulent transactions. The architecture consists of several
logically connected components working together toward that end. It starts with the data ingestion pipeline that
collects transactional data from various sources such as UPI servers, banking APIs, and payment gateways.
This data generally consists of important fields like transaction IDs, timestamps, user identification, transaction
amounts,geography information, device information, and labels for whether the transaction is deemed to be
fraudulent or legitimate. Once data are ingested, a preprocessing module cleans and transforms the raw input
data into an accepted format for machine learning. This may include several steps toward the completion of
data preprocessing, such as handling missing values, normalization of numerical features, and encoding of
categorical data through either one-hot or label encoding. It also has some feature engineering work, such as
generating behavioral features like transaction frequency, device diversity, and average transaction amount per
user. These newly derived features help capture the anomalous patterns caused by fraud activities. This data-
prepared dataset is then sent into the GAN training module. The GANs lie in the heart of the architecture,
which is composed of two neural networks: a generator and a discriminator.The generator produces synthetic
fraudulent transaction data from learning the distribution of actual fraudulent records. The generator is fed
random noise to output transaction-like feature vectors. The discriminator, on the other hand, is determining
whether a particular transaction sample is real or fake. Both networks carry out a competitive training routine
in which the generator tries to fool the discriminator, and the discriminator tries getting better at classifying.
Thus, the training enables the discriminator to learn complicated patterns of frauds that are otherwise difficult
for traditional methods to detect. At the end of the training process, the discriminator is redefined and adopted
as the actual fraud detection model for new transactions going for real-time classification. In live
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e295
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
implementation, the incoming UPI transactions pass through the trained Discriminator model, which outputs a
probability saying that the transaction is fraudulent. Transactions above a given threshold may be marked for
review or automatically blocked based on policy. To ensure that the model remains efficient with time, an
evaluation and monitoring system keeps track of key performance metrics such as precision, recall, F1-score,
and ROC-AUC, while also detecting concept drift—changes in the character of transaction patterns that may
indicate the necessity to retrain the model with fresh data. The overall architecture is made scalable all the way
down.
Fig. 1. System Architecture
4.Results And Discussion
The experimental results obtained from the UPI transactional data on the implementation of the GAN-based
proposed fraud detection system are listed here. The evaluation methodology followed standard classification
measures and compare it with the traditional machine learning models. The discussion, in turn, explains the
strength and weaknesses, desirability of applicability, etc., for the practical implementation.
4.1 Experimental setup
The experimental works were performed on a dataset of anonymized UPI transactions. The dataset reported
both legit transactions and fraudulent ones, although the distribution among the classes was heavily
imbalanced. The GAN model was implemented in TensorFlow and trained in a very powerful environment
with GPU support. Classical machine-learning models such as Logistic Regression, Random Forest, and
Support Vector Machines were used for benchmarking. These models were trained and tested under the same
conditions.
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e296
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
4.2 Evaluation metrics
The models for fraud detection were evaluated using the following metrics:
Accuracy: Measures the overall correctness of the model.
Precision: Indicates how many of the detected frauds were actually fraudulent.
Recall: Measures how many actual frauds were correctly detected.
F1-Score: Harmonic mean of Precision and Recall, offering a balanced performance view.
ROC-AUC Score: Reflects the model’s ability to distinguish between classes across different thresholds.
These metrics are particularly important in fraud detection, where identifying fraudulent transactions (high
recall) without
producing excessive false alarms (high precision) is critical.
4.2 Performance Comparison
GANs outperformed traditional machine learning models,particularly in the detection of unseen or rare
types of fraud. Logistic Regression and SVMs achieved fair levels of accuracy but failed miserably in recalling
a lot of fraud cases. Random Forests, still not good enough at detecting new kinds of fraud,did a bit better than
these two. The actual GAN Discriminator registered an F1 score 0.91, Recall of 0.94, and Precision of 0.88,
all well ahead of the other methods. This is paired with an ROC-AUC score of 0.96, clearly distinguishing
between the fraudulent and legal transactions. This enhanced performance can be attributed to the augmentation
of the rare class using synthetic data generated by the GAN’s Generator, subsequently aiding the
Discriminator in building its capacity in recognizing varying fraud patterns. This resolution of the class-
imbalance problem proved to be better than the classic oversampling techniques like SMOTE.
4.3 Results Visualization
Moreover, the application benefit of our design was validated further by t-SNE plots to visualize how well
the model of separating fraudulent and legitimate transactions in the feature space. The class boundaries created
by the GAN trained models were clearer than those of the baseline models. Confusion matrices showed a high
true positive rate and a low false negative rate for the GAN model, which is vital for fraudulent prevention in
real-time settings.
5.CONCLUSION
The unified payments interface (UPI) being embraced as a means for online trading has transformed how
financial services work in India as well as developing nations. Of course, this increase is also accompanied
with a steep rise in cyber crimes-including those that take advantage of flaws in verification of an electronic
transaction or behavioral monitoring systems. This paper presented a new approach to fraudulent detection in
UPI systems based on using the capabilities of Generative Adversarial Networks (GAN). It would accomplish
the targeted functions of finding known fraudulent patterns, thus discovering unaccessible and previously
unknown fraudulent behaviors that traditional models would not detect easily. The architecture thus designed
is to manage the all encompassing transaction data from inception and preprocessing to model training,
deployment, and even real-time inference. The generator part of the GAN was anticipated to learn from actual
fraud data then create convincing fake fraud transactions; while the discriminator was taught to discern between
real and generated transactions. This taught the Discriminator to improve its ability in fraud detection
constantly through this antagonistic learning paradigm. The analysis results indicate that not only Artificial
Neural Network classifiers, even other traditional classification algorithms such as Logistic Regression,
Support Vector Machines, and Random Forests, perform better than GAN in terms of effectiveness. The GAN-
based Discriminator also achieved high precision and recall scores along with an F1 score and AUC-ROC
value to reflect a trustable balance between accuracy of fraud detection and minimizing false positives. One of
the core strengths, however, of this system is in the capability of the GAN architecture to tackle the core
problem of class imbalance, which is the most significant challenge facing fraud detection, as generally, a lot
more legitimate transactions occur than fraudulent ones. Unlike oversampling techniques which risk
introducing noise, our developed meaningful synthetic samples to enhance minority class richness, thus
improving generalization of the model to diverse fraud strategies. The model was validated through
visualization techniques such as t-SNE and confusion matrices, which showed excellent differentiation
capability for normal and anomalous transactions. The results might seem encouraging, but a few limitations
exist and future directions could be noted. The GAN model requires high computation costs, along with
extensive tuning by a chief optimizer while it trains to preempt mode collapse or to prevent overfitting.
Furthermore, similar to many deep learning systems, the decision-making mode of the model is devoid of
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e297
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
interpretability, which is vital for any financial application for regulatory compliance and user trust. This
model, hence, should be construed to dwell on future works related to the incorporation of explainable AI
(XAI) frameworks so as to make the model’s decisions more transparent. More so, real-time performance
optimization and testing under live transaction streams would further validate the practicality of the system for
deployment.
6.FUTURE WORK
Despite the encouraging results demonstrated by the GAN based UPI fraud detection system developed
herein, there exists a formidable scope for improvement pertaining to its operational applicability and
robustness. Explainable AI (XAI) stands out among the most crucial aspects that require attention. As it stands
now, the Discriminator operates like a black box classifier; therefore, combining imparting interpretability
using methods like SHAP or LIME would allow stakeholders to understand why specific transactions are
flagged, thus improving transparency and garnering user trust. Also crucial is optimizing the system for real-
time deployment. While the system might function adequately in offline environments, its computational
complexity seems to act as a stumbling block for production environments that would impose considerable
latency constraints on its use. Techniques such as model compression, pruning, or quantization may help reduce
inference times and promote responsiveness. More sophisticated measures to counter fraud attempts can be
accomplished through the integration of multi-modal data such as user behavior, device fingerprinting, and
location history. In addition to these, temporal and contextual signals would allow the model to learn more
subtle patterns of fraudulent behavior. Continuous learning mechanisms should be studied, in which feedback
from confirmed fraud cases is used to gradually update the model. This enables the model to maintain its
performance against changes in fraud strategies without complete retraining. Finally, implementing the model
using a federated learning framework considering privacy and data security would allow joint learning across
institutions without sharing sensitive customer data to stay compliant with data protection regulations. Thus,
improving interpretability, providing real-time capability, reasoning with multiple data sources, supporting
adaptive learning, and ensuring privacy protection are all vital future work directions for making this system
effective and deployable in real-world UPI ecosystems.
6.REFERENCES
[1] X. Zhang, Y. Han, W. Xu, and Q. Wang, “HOBA: A novel feature engineering methodology for credit card
fraud detection with a deep learning architecture,” Information Sciences, May 2019. [Accessed: Jan. 8, 2019].
[2] N. Carneiro, G. Figueira, and M. Costa, “A data mining based system for credit-card fraud detection,”
Decision Support Sytems, vol. 95, pp. 91–101, Mar. 2017.
[3] B. Lebichot, Y.-A. Le Borgne, L. He-Guelton, F. Oblé, and G. Bontempi, “Deep-learning domain
adaptation techniques for credit cards fraud detection,” in Proc. INNS Big Data Deep Learn. Conf., Genoa,
Italy, 2019, pp. 78–88.
[4] H. John and S. Naaz, “Credit card fraud detection using local outlier factor and isolation forest,” Int. J.
Comput. Sci. Eng., vol. 7, no. 4, pp. 1060–1064, Sep. 2019.
[5] C. Phua, R. Gayler, V. Lee, and K. Smith-Miles, “On the communal analysis suspicion scoring for identity
crime in streaming credit applications,” Eur. J. Oper. Res., vol. 195, no. 2, pp. 595–612, Jun. 2009.
[6] R. Bolton and D. Hand, “Statistical fraud detection: A review,” Statistical Science, vol. 17, no. 3, pp. 235–
249, Aug. 2002.
[7] P. A. Dal, G. Boracchi, O. Caelen, C. Alippi, and G. Bontempi, “Credit card fraud detection: A realistic
modeling and a novel learning strategy,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, pp. 3784–3797,
Sep. 2017.
[8] S. Bhattacharyya, S. Jha, K. Tharakunnel, and J. C. Westland, “Data mining for credit card fraud: A
comparative study,” Decision Support Systems, vol. 50, no. 3, pp. 602–613, Feb. 2011.
[9] N. Sethi and A. Gera, “A revived survey of various credit card fraud detection techniques,” Int. J. Comput.
Sci. Mobile Comput., vol. 3, no. 4, pp. 780–791, Apr. 2014.
[10] A. O. Adewumi and A. A. Akinyelu, “A survey of machine-learning and nature-inspired based credit card
fraud detection techniques,” Int. J. Syst. Assurance Eng. Manage., vol. 8, no. S2, pp. 937–953, Nov. 2017.
[11] A. Dal Pozzolo, O. Caelen, R. Johnson, and G. Bontempi, “Calibrating probability with undersampling
for unbalanced classification,” in Proc. IEEE Symp. Series Comput. Intell., 2015, pp. 159–166. doi:
10.1109/SSCI.2015.33.
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e298
www.ijcrt.org © 2025 IJCRT | Volume 13, Issue 5 May 2025 | ISSN: 2320-2882
[12] A. Dal Pozzolo, O. Caelen, R. A. Johnson, and G. Bontempi, “Credit card fraud detection: A realistic
modeling and a novel learning strategy,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, pp. 3784–3797,
2018. doi: 10.1109/TNNLS.2017.2736643.
[13] T. Nguyen, D. Tran, and W. Cao, “Machine learning based approaches for detecting fraudulent
transactions in online payment systems,” J. Inf. Secur. Appl., vol. 63, p. 103016, 2021. doi:
10.1016/j.jisa.2021.103016.
[14] R. Singh and S. Bharti, “Financial fraud detection using machine learning,” in Proc. Int. Conf. Comput.
Intell. Data Sci. (ICCIDS), 2021, pp. 345–350. doi: 10.1016/j.procs.2021.01.045.
IJCRT2505489 International Journal of Creative Research Thoughts (IJCRT) www.ijcrt.org e299
