International Journal On Engineering Technology and Sciences – IJETS

ISSN (P):2349-3968, ISSN (O): 2349-3976 Volume XI - Issue III, March - 2024

UPI FRAUD DETECTION USING MACHINE LEARNING

1
Axlin Jenes J F, 2Mahendran N, 3Siluvai Priyan S, 4Vivek Siddharath, 5Sajil L
1,2,3,4
Student, 5Assistant Professor,
Department of Information Technology,
Loyola Institute of Technology and Science, Nagercoil, Tamilnadu.

Abstract : The project focuses on the development of a machine learning model that can analyze UPI transaction data
in real-time to identify fraudulent activities. The primary objective is to create a system that enhances the security of
UPI transactions and reduces financial losses due to fraud. With the advancement of technology, today most of the
modern commerce is relying upon the online banking and cashless payments. Due to adaption of online payment
among businesses, the fraud cases are also increasing which cause financial losses to them. Fraudsters are inventing
new techniques to perform fraudulent transaction which seem legitimate. Hence, there is an urgent need to develop
fraud detection measures which can deal with these fraudsters on real time basis. Deep learning techniques have the
capability to detect these fraudulent transactions efficiently and has a huge scope in fraud detection. However, there are
many challenges faced by the researchers in online transactions fraud detection because the datasets are not publicly
available due to privacy issue of the financial institutions as customers data is sensitive and it can be misused and the
datasets which are available are imbalanced. This paper presents a review of deep learning techniques used for online
transactions fraud detection. It also provides the information about datasets used by the researchers and the results
achieved by them in their research work.

Index Terms - Fraud Detection, Feature Engineering, Cross-Validation, Emerging Technologies.

I. Introduction
patterns or anomalies. Although these systems have
The emergence of online banking has caused a
shown some degree of effectiveness, they have
paradigm shift in the current financial transaction
shortcomings in terms of accuracy, scalability, and
landscape, providing individuals and organisations
adaptability, especially when it comes to sophisticated
globally with unmatched ease and efficiency. But there
and ever-evolving fraud techniques.
are drawbacks to this digital transformation as well,
the most significant of which is the growing frequency A prevalent obstacle within the current framework
of fraudulent activity in online transactions. The swift is its dependence on static rules, which may not be
expansion of online banking services has made it able to identify subtle or evolving patterns that point to
easier for bad actors to take advantage of weaknesses, fraudulent activity. Furthermore, the rule-based
which puts the security and integrity of financial approach frequently finds it difficult to manage the
systems at serious risk. Furthermore, the COVID-19 subtleties and intrinsic complexity of transaction data,
pandemic's start has acted as a trigger, quickening the particularly given the ever-changing environment of
shift to remote operations and raising the possibility of online banking transactions. The current system's
fraudulent activity in the digital sphere. It is therefore vulnerability to false positives and false negatives is
more important than ever to create reliable fraud another drawback. Static rules have the potential to
detection systems in the face of the epidemic highlight unintentionally identify genuine transactions as
how important it is for both customers and financial fraudulent (false positives) or fail to identify actual
institutions to strengthen their defences against fraud. fraudulent activity (false negatives), which can result
The increasing trend of financial transactions occurring in inefficiencies, unhappy customers, and financial
on digital platforms underscores the need for advanced losses for financial institutions as well as consumers.
security measures and flexible approaches to protect Furthermore, particularly in the context of the Unified
the integrity of online banking systems. Payments Interface (UPI), the current systems would
find it difficult to handle the enormous volume and
Exiting System
variety of transaction data generated in real-time. The
Regarding the field of online banking fraud scalability and computing efficiency needed to
detection, the systems that are now in place primarily efficiently process and analyse large-scale transaction
depend on conventional techniques and rule-based datasets may be lacking in traditional methodologies.
methods. In order to identify potentially fraudulent
transactions, these systems frequently use static rules
and thresholds, usually based on established

33
 International Journal On Engineering Technology and Sciences – IJETS
ISSN (P):2349-3968, ISSN (O): 2349-3976 Volume XI - Issue III, March - 2024

II. PROPOSED SYSTEM: Real-time Processing: Our suggested solution

allows for real-time fraud detection and response in
high-volume transaction environments like the Unified
Payments Interface (UPI) by utilising the parallel
processing powers of CNNs. CNNs streamline the
simultaneous processing of massive amounts of
transaction data, making it easier to identify fraudulent
activity quickly and take prompt action to prevent
losses. Despite the inherent complexity of CNNs, we
have given priority to interpretability and openness in
our suggested solution. We include methods for
illustrating and interpreting CNN decision-making
processes so that interested parties can comprehend the
reasoning behind reported transactions and develop
faith in the dependability of the system. Our
technology facilitates cooperation between users,
regulators, and internal auditors by improving
interpretability, which in turn promotes responsibility
We present a unique strategy that uses
and adherence to regulatory standards.
convolutional neural networks (CNNs) for increased
Continuous Monitoring and Evaluation: To
fraud detection in online banking transactions in order
evaluate the efficacy and performance of our suggested
to overcome the shortcomings of current fraud
system over time, it is equipped with mechanisms for
detection systems. Compared to conventional rule-
ongoing monitoring and assessment. We make sure
based approaches, our suggested solution has various
that the fraud detection system is continuously
advantages, including better accuracy, flexibility, and
optimised and improved by examining user feedback,
scalability. We support the use of convolutional neural
comparing model predictions to ground truth labels,
networks (CNNs) as the primary technology for online
and monitoring important performance metrics like
banking transaction fraud detection. Especially in
precision, recall, and F1-score. The overall
image analysis tasks, CNNs have shown impressive
effectiveness and dependability of the system are
skills in feature extraction and pattern detection. We
increased by this iterative process, which permits
intend to leverage CNNs' capacity to automatically
ongoing enhancement and adaptability to changing
learn hierarchical characteristics and identify complex
fraud environments.
patterns suggestive of fraudulent activity by applying
Advantages
them to transactional data. Important elements of the
▪ Enhanced accuracy
system we've suggested include: Adaptive Learning:
▪ Adaptability to changing patterns
Unlike static rule-based systems, our suggested
▪ Continuous improvement
approach makes use of CNN-enabled adaptive learning
▪ Comprehensive fraud detection
techniques. With the ability to dynamically modify
▪ Scalability and efficiency
their internal representations and adjust their
▪ Real time detection
parameters in response to incoming data, these neural
networks are able to react in real-time to evolving fraud III. Results and Discussion
trends and new threats. Its ability to adjust strengthens The evaluation of our implemented models has
the system's defences against changing fraud strategies provided valuable insights into their performance.
and guarantees steady advancement over time. Notably, the Feedforward Neural Network (FNN) and
Feature extraction and transformation: To extract Convolutional Neural Network (CNN) have exhibited
pertinent characteristics from transactional data and commendable accuracy in test sets, showcasing their
turn them into interpretable representations for fraud ability to accurately classify transactions. However,
detection, our suggested approach makes use of CNNs. the relatively low average precision scores for both
CNNs can identify subtle patterns and abnormalities models indicate potential challenges with false
that may escape conventional rule-based methods by positive rates. This emphasizes the significance of
automatically learning discriminative features from raw considering precision-recall trade-offs, especially in
transactional attributes. The method of feature the context of imbalanced datasets like ours.
extraction improves the accuracy with which the In contrast, the machine learning models, including
system can distinguish between authentic and Decision Tree, Naive Bayes, Logistic Regression with
fraudulent transactions. L1 and L2 regularization, and K-Nearest Neighbors
(KNN), have demonstrated robust overall performance

34
 International Journal On Engineering Technology and Sciences – IJETS
ISSN (P):2349-3968, ISSN (O): 2349-3976 Volume XI - Issue III, March - 2024

with varying precision scores. Particularly noteworthy treatment of missing data played a crucial role in
is Logistic Regression with L1 regularization (LR1), getting the dataset ready for efficient model training.
which stands out by achieving the best overall The selection of algorithms was an important factor
accuracy among all algorithms. This remarkable in guaranteeing the effectiveness of the system in
accuracy underscores LR1's effectiveness in correctly detecting fraudulent actions. We made sure our system
identifying fraud cases with minimal false positives. could detect a wide range of fraudulent patterns by
Such precision is of paramount importance in fraud implementing a diverse ensemble of machine learning
detection, where minimizing false positives is crucial algorithms, from more sophisticated techniques like
to avoid inconveniencing legitimate users. These convolutional neural networks to more conventional
results provide actionable insights for refining fraud methods like logistic regression and decision trees.
detection strategies in real-world applications. For With the contributions of each algorithm, a
instance, the outstanding precision achieved by LR1, comprehensive fraud detection system that could
along with the perfect precision scores attained by handle a variety of fraud scenarios was created.
Decision Tree, Naive Bayes, and KNN models, The system architecture was crucial in
suggests their suitability for deployment in scenarios guaranteeing scalability, efficiency, and usability
where false positives must be minimized. On the other during the development and deployment stages. By
hand, while FNN and CNN models demonstrate high using technologies like Flask, HTML, CSS, Python,
accuracy, further optimization or consideration of and other programming languages, we were able to
additional factors may be necessary to address smoothly integrate the fraud detection system into the
potential challenges with false positives. current financial infrastructure, increasing its usability
Moreover, in the context of user interface design, and accessibility for both stakeholders and end users.
presenting these insights in a comprehensible manner The seamless deployment and functioning of the
becomes essential. A user-friendly interface should system in real-world settings were made possible by
provide clear visualizations and summaries of the this harmonious technology integration.
model performance, emphasizing the trade-offs The fraud detection system is a major improvement
between accuracy and precision. This aids decision- in protecting financial transactions from fraudulent
makers in selecting models aligned with the specific activity, as we can see when we consider the results of
requirements of the application. Additionally, our work. It is crucial to preserving the integrity and
incorporating user feedback mechanisms into the reliability of financial systems because of its capacity to
interface can enhance the adaptability of the system, precisely detect fraudulent transactions while reducing
allowing for iterative improvements based on real- false positives. But our adventure doesn't end here.
world usage and evolving fraud patterns. Overall, the Maintaining protection and security in the ever-
integration of insightful model evaluations and a user- changing world of financial transactions will require
friendly interface forms a cohesive strategy for constant system improvement and refinement to meet
refining and deploying effective fraud detection changing fraud trends and new security risks. We are
systems in practical applications. prepared to meet upcoming challenges and protect the
integrity of financial systems around the globe with a
IV. Conclusion dedication to innovation and vigilance.
We have gone through several phases of data
collection, pre-processing, algorithm selection, and REFERENCES
system implementation in this extensive effort to create
1. ALESKEROV E, FREISLEBEN, B., and, RAO B (1997)
a strong fraud detection system. The end result is a CARDWATCH: A neural network-based database mining
solution that has the potential to greatly improve the system for credit card fraud detection. In Conference (pp.
security and dependability of financial transactions. We 220–226). IEEE, Piscataway, NJ
started our trip by obtaining a large and comprehensive
dataset that included complex transactional 2. Sahin M (2017) Understanding Telephony Fraud as an
Essential Step to Better Fight it [Thesis]. École Doctorale
information, which served as the foundation for our
Informatique, Télécommunication et Électronique, Paris
later investigations.
Setting up a strong basis for later model training 3. Abdallah A, Maarof MA, Zainal A (2016) Fraud detection
required the first stage of data pre-processing. We system: A survey. J Netw Comput Appl 68:90–113
carefully addressed issues, including the dataset's
4. ANDREWS PP, PETERSON MB (eds) (1990) Criminal
unequal class distribution, using calculated pre-
Intelligence Analysis. Palmer Enterprises, Loomis, CA
processing techniques to reduce biases and guarantee
the stability of our models. Methods like 5. ARTÍS M, AyUSO M, GUILLÉN M (1999) Modeling
standardization based feature scaling and careful different types of automobile insurance fraud behavior in the
Spanish market. Insurance Math Econ 24:67–81

35
 International Journal On Engineering Technology and Sciences – IJETS
ISSN (P):2349-3968, ISSN (O): 2349-3976 Volume XI - Issue III, March - 2024

6. BARAO MI, TAWN JA (1999) Extremal analysis of short

series with outliers: Sea-levels and athletics records. Appl
Stat 48:469–487

7. BLUNT G, HAND DJ (2000) The UK credit card market.

Technical report, Department of Mathematics, Imperial
College, London.

8. BOLTON RJ, HAND DJ (2001) Unsupervised profiling

methods for fraud detection. In Conference on Credit Scoring
and Credit Control 7, Edinburgh, UK, 5–7 Sept

9. Phua C, Lee V, Smith K, Gayler R (2010) A

comprehensive survey of data mining-based fraud detection
research. https://doi.org/10.48550/ARXIV.1009.6119

10. Summers SL, Sweeney JT (1998) Fraudulently misstated

financial statements and insider trading: An empirical
analysis. 73(1):131– 146https://www.jstor.org/stable/248345.

36