International Journal of Advances in Engineering and Management (IJAEM)

Volume 6, Issue 06 June 2024, pp: 98-100 www.ijaem.net ISSN: 2395-5252

UPI Fraud Detection Using Machine

Learning
Shabreshwari R M, Shafiya Mehrooz, Sidra Fatima, Tanmai
R B, Prof. Ganesh Manasali
Department of Electronics and Communication,
PDA College of Engineering, Kalabuaragi- 585103
----------------------------------------------------------------------------------------------------------------------------- ----------
Date of Submission: 28-05-2024 Date of Acceptance: 06-06-2024
----------------------------------------------------------------------------------------------------------------------------- ----------
ABSTRACT: Digital Fraud has become a threat commerce. Fraudulent transactions have become a
across all the sectors. Its pivotal for any growing problem in online banking, and fraud
organization now to have a dedicated focus to detection has always been challenging. Along with
detect and prevent fraud and increase their focus on UPI development, the pattern of UPI fraud has
Security. Machine learning algorithms have shown always been updated. Fraudsters do their best to
promise in analysing large volumes of transaction make it look legitimate, and UPI fraud has always
data to identify patterns and anomalies indicative of been updated. They try to learn how fraud detection
fraudulent transactions. Making use of a systems work and continue to stimulate these
heterogeneous dataset that includes both authentic systems, making fraud detection more complicated.
and fraudulent transactions, we utilize feature Therefore, researchers are constantly trying to find
engineering and data preparation techniques to new ways or improve the performance of the
identify significant trends. Using past data, the existing methods. People who commit fraud
chosen machine learning model is trained and its usually use security, control, and monitoring
ability to distinguish between real and fraudulent weaknesses in commercial applications to achieve
transactions are evaluated. The model takes into their goals. However, technology can be a tool to
account important features such transaction combat fraud. To prevent further possible fraud, it
amount, timestamp, payer and payee data, location, is important to detect the fraud right away after its
and device information. For a thorough analysis, occurrence.
time-based features are given more weight. The Fraud detection in banking is considered a
model is included into the UPI system for real-time binary classification problem in which data is
processing after the selected algorithm has been classified as legitimate or fraudulent. Because
adjusted to maximize performance. In order to banking data is large in volume and with datasets
ensure prompt intervention, alert systems are containing a large amount of transaction data,
implemented in tandem with the fraud detection manually reviewing and finding patterns for
model. fraudulent transactions is either impossible or takes
a long time. Therefore, machine learning-based
I. INTRODUCTION: algorithms play a pivotal role in fraud detection and
The landscape of financial transactions in prediction. Machine learning algorithms and high
India has undergone a remarkable transformation processing power increase the capability of
with the advent of digital payment systems, among handling large datasets and fraud detection in a
which the Unified Payments Interface (UPI) stands more efficient manner. Machine learning
as a hallmark of innovation and convenience. UPI, algorithms and deep learning also provide fast and
conceived and implemented by the National efficient solutions to real-time problems. In this
Payments Corporation of India (NPCI), has paper, we propose an efficient approach for
democratized financial inclusion by providing a detecting UPI fraud that has been evaluated on
seamless, interoperable, and instant platform for publicly available datasets and has used optimised
transferring funds between individuals, businesses, algorithms Random Forest, Light GBM, XGBoost,
and institutions. In recent years, there has been a Decision treeand logistic regression individually.
significant increase in the volume of financial An ideal fraud detection system should detect more
transactions due to the expansion of financial fraudulent cases, and the precision of detecting
institutions and the popularity of web-based e- fraudulent cases should be high, i.e., all results

DOI: 10.35629/5252-060698100 |Impact Factorvalue 6.18| ISO 9001: 2008 Certified Journal Page 98
 International Journal of Advances in Engineering and Management (IJAEM)
Volume 6, Issue 06 June 2024, pp: 98-100 www.ijaem.net ISSN: 2395-5252

should be correctly detected, which will lead to the learningalgorithm. The model is the “thing”
trust of customers in the bank, and on the other that is saved after running a machine learning
hand, the bank will not suffer losses due to algorithm on training data and represents the
incorrect detection.Thus this project embarks on a rules, numbers, and any other algorithm-
quest to develop a robust and adaptive UPI Fraud specific data structures required to make
Detection system capable of thwarting fraudulent predictions.
activities and safeguarding users' financial assets. 2. MODEL TRAINING: After model
II. OBJECTIVES: construction it is time for model training. In
1. To gather and pre-process a diverse UPI this phase, the model is trained using training
transaction dataset for fraud detection. data. At the end it will report the final accuracy
2. To identify pertinent features and create new of the model.
metrics to enhance model discrimination. 3. MODEL TESTING: During this phase a
3. To select and train machine learning second set of data is loaded. This data set has
algorithms for effective fraud detection. never been seen by the model and therefore it’s
4. To integrate the trained model into the UPI true accuracy will be verified.
system for real-time processing and implement 4. MODEL EVALUATION: Model Evaluation
alert mechanisms. is an integral part of the model development
5. To establish continuous monitoring, updates, process. It helps to find the best model that
and explore ensemble methods for model represents our data and how well the chosen
robustness. model will work in the future. Evaluating
model performance with the data used for
III. METHODOLOGY: training is not acceptable in data science
TRAINING MODEL: Supervised because it can easily generate overoptimistic
machine learning is one of the category of machine and overfitted models. There are two methods
learning where the model is trained by input data of evaluating models in data science, Hold-Out
and expected output data. For creating such model, and CrossValidation. To avoid overfitting, both
it is necessary to go through the following phases: methods use a test set (not seen by the model)
1. MODEL CONSTRUCTION: A model to evaluate model performance.
represents what was learned by a machine

IV. SYSTEM ARCHITECTURE:

Fig 1: System Architecture of UPI fraud detection

DOI: 10.35629/5252-060698100 |Impact Factorvalue 6.18| ISO 9001: 2008 Certified Journal Page 99
 International Journal of Advances in Engineering and Management (IJAEM)
Volume 6, Issue 06 June 2024, pp: 98-100 www.ijaem.net ISSN: 2395-5252

1. Dataset: Start with a labelled dataset making our project easy to understand by using a
containing information about UPI transaction kind of AI that can explain how it works. By
with input parameters. Each transaction is including these things in our future plans, we hope
labelled as genuine (Real) or fraudulent to keep our project working well in the always-
(fraud). changing world of technology.
2. Input: Pre-process the dataset to ready it for
input into the machine learning model. This VI. CONCLUSION:
involves handling missing data, encoding In conclusion, developing and putting into
categorical variables, normalizing numerical action a UPI fraud detection system is a crucial step
features, and other necessary preprocessing to make Unified Payments Interface transactions
steps. safer and more trustworthy. As digital payments
3. Feature Selection: Identify relevant features become more widespread, it's essential to safeguard
that contribute to distinguishing between users and financial institutions from fraud. The
genuine and fraudulent transactions. Feature proposed system, using advanced machine
selection improves the model's performance by learning, aims not only to spot unusual transactions
focusing on the most informative attributes, but also to adapt to new fraud patterns. By
using techniques like correlation analysis or analysing past transaction data, the system intends
recursive feature elimination. to identify subtle signs of fraud, strengthening the
4. Training the Model: Split the dataset into overall security of UPI transactions. The method
training and testing sets. Train selected we plan to follow involves collecting and preparing
machine learning algorithms (Random Forest, data, creating a model, integrating it into the
XGBoost classifier, Logistic regression, system, and continuously monitoring its
Decision Tree, GBM) on the training set. performance. We will prioritize ethical
During training, the model learns patterns that considerations, user privacy, and compliance with
distinguish between genuine and fraudulent regulations to ensure the responsible and legal use
transactions. of the system. If successful, the UPI fraud detection
5. Evaluation: Evaluate the performance of each system is expected to bring about better accuracy,
model using the testing set. Common metrics real-time monitoring, adaptability to new threats,
for fraud detection, such as precision, recall, increased user trust, and reduced financial losses.
F1 score, and accuracy, are considered. Choose The final output will include detailed
the algorithm with the highest accuracy for documentation, user manuals, alert systems, and
further use. seamless integration with the existing UPI
6. Selection of Best Model: Based on evaluation framework.
results, select the algorithm with the highest
accuracy for fraud detection. REFERENCES:
7. Prediction: Use the trained models to predict [1] Mahbuba Yesmin Turaba et al. “Fraud
whether new, unseen transactions are genuine Detection During Financial Transactions
or fraudulent. Input the features of a new Using Machine Learning and Deep Learning
transaction into each model, and the model will Techniques” in IEEE Oct 2022.
output a prediction. [2] Seyedeh Khadijeh Hashemi et al., “Fraud
Detection in Banking Data by Machine
V. FUTURE SCOPE: Learning Techniques”, in IEEE Dec2022.
The future scope of UPI Fraud Detection [3] Mr. Sunil S Mhamane and Mr. L.M.RJ Lobo
System Using Machine learning has a lot of “Internet Banking Fraud Detection Using
potential to get better. We're excited about trying HMM”, in IEEE July 2012.
out new and improved techniques in computer [4] Pradheepan Raghavan and Neamat El Gayar
learning, using behaviour details. These “Fraud Detection using Machine Learning
improvements will not only make our system better and Deep Learning”, in IEEE Feb 2020.
but also let us work with people from around the [5] G.Jaculine Priya and Dr.S.Saradha “Fraud
world who have different skills.Looking ahead, we Detection and Prevention Using Machine
know it's important to keep up with the latest Learning Algorithms: A Review”, in IEEE
technology. As students working on this project, 2021.
our main goal is to make our system work with the
new changes in quantum computing while making
sure it's clear and private. We also want to focus on

DOI: 10.35629/5252-060698100 |Impact Factorvalue 6.18| ISO 9001: 2008 Certified Journal Page 100