cv

UPI FRAUD DETECTION USING

MACHINE LEARNING

Team Members
1. SAKSHAM BHATNAGAR -1DS21CS181
2. ADITYA SHUKLA -1DS21CS023
3. SHAAN SHARMA -1DS21CS196
4. PRAJWAL NP - 1DS21CS148

Under the Guidance of

Prof. Sandhya K.
Assistant Professor, Dept of Computer Science and Engineering

Department of Computer Science & Engineering, DSCE

 cv
ABSTRACT
Mobile payment systems have become a popular mainstream method, but
their growth has also increased fraud risks, impacting consumers and the
online economy. Machine learning offers a robust solution by detecting fraud in
transaction data, outperforming traditional statistical methods. This study
evaluated seven algorithms, with Convolutional Neural Networks (CNN)
achieving the highest accuracy. The project envisions expanding into a Fraud
Detection as a Service (FDaaS) platform for SMEs and e-commerce, offering
secure, API-based integration. Future plans include real-time detection, hybrid
models for better accuracy, explainable AI for transparency, and blockchain
integration for enhanced security, aiming to create a scalable and secure digital
payment ecosystem.

Department of Computer Science & Engineering, DSCE

 INTRODUCTION

The rise in mobile payments has increased online transactions, attracting fraud
that harms consumers and stunts online economy growth. Traditional fraud
detection methods struggle to identify fraud patterns, but big data and machine

cv
learning provide better solutions by analysing large datasets more effectively.
Algorithms like Convolutional Neural Networks (CNN), K-Nearest Neighbours
(KNN), Support Vector Machine (SVM), Logistic Regression, Naive Bayes,
Decision Tree, and Random Forest improve fraud detection by leveraging their
unique strengths. By comparing these algorithms, the goal is to create a robust
system for detecting fraudulent UPI transactions with high accuracy and
minimal false results, enhancing digital transaction security and contributing to
financial cybersecurity.

Department of Computer Science & Engineering, DSCE

 APPLICATIONS

• Real-Time Fraud Detection: Detects and flags fraudulent UPI transactions

instantly, minimizing losses.
• Automated Transaction Monitoring: Continuously monitors transactions

cv
for anomalies, reducing manual checks and errors.
• Fraud Detection as a Service (FDaaS): Cloud-based platform for SMEs to
access fraud detection tools without costly infrastructure.
• API-Based Fraud Detection: Provides an API for fintech companies to
integrate fraud detection into their systems.
• Customer Support for Fraud: Offers dedicated support to handle fraud
cases, recover funds, and secure accounts.

Department of Computer Science & Engineering, DSCE

 SHORT LITERATURE SURVEY

The base papers introduced a feature-based convolutional neural network model

for transaction fraud detection, showing better stability and classification than
other models. However, due to imbalanced data, achieving high accuracy

cv
remains challenging. To improve this, various supervised machine learning
algorithms, such as Decision Trees, Naive Bayes, Logistic Regression, and
CNN, are used and compared for accuracy.
The reference papers highlight the rise in credit card fraud and the need for
better fraud classification by banks. Machine learning methods, especially
Random Oversampling, achieved a precision and accuracy score of 0.99. The
study suggests using data sampling to balance data and optimize model
performance for fraud detection.

Department of Computer Science & Engineering, DSCE

 CHALLENGES IN THE
CURRENT WORK
• Data Quality: Difficulty accessing high-quality, labelled data due to privacy
issues and regulations, leading to potential inaccuracies.

cv
• Imbalanced Data: Few fraudulent transactions cause models to favour non-
fraudulent predictions, increasing false negatives.
• Changing Fraud Tactics: Constantly evolving fraud methods require
models to adapt quickly to remain effective.
• Feature Selection: Identifying effective features for distinguishing fraud
requires specialized knowledge.
• Real-Time Processing: Need for rapid, accurate analysis of transactions in
real-time without delays.

Department of Computer Science & Engineering, DSCE

 PROBLEM STATEMENT

Current fraud detection systems often rely on traditional rule-based approaches,

which struggle to identify evolving and complex fraud patterns. These methods
can lead to a high number of false positives (legitimate transactions flagged as

cv
fraud) and false negatives (fraudulent transactions not detected), negatively
impacting user experience and security. Additionally, the large volume of UPI
transactions makes manual monitoring impractical and inefficient.
Challenges include:

•Lack of Real-World Data: Sensitive financial information limits access to

data needed for effective experimentation and model training.

•Data Noise: Errors in data, such as incorrect dates, reduce the accuracy and
effectiveness of fraud detection models, regardless of the size of the training
set.

Department of Computer Science & Engineering, DSCE

 OBJECTIVE

• Minimize Errors: Reduce false positives and negatives to improve fraud

detection accuracy.

cv
• Improve Accuracy: Use algorithms like CNN, KNN, Logistic Regression,
Naive Bayes, Decision Tree, and Random Forest for better detection.

• Enhance Security: Boost UPI transaction security and user trust by

protecting financial data.

• Adapt to Fraud: Develop models that can adapt to new and emerging fraud
tactics.

Department of Computer Science & Engineering, DSCE

PROPOSED SOLUTION

cv

Department of Computer Science & Engineering, DSCE

 DEMONSTRATION WORKFLOW
1. Login (Authentication): Secures access by verifying user credentials,
ensuring data confidentiality and integrity.

2. Upload Dataset (Data Preprocessing): Enables users to upload transaction

cv
data, preprocesses it by cleaning, handling missing values, and formatting for
machine learning analysis.

3. ML Model (CNN) Training/Testing: Trains CNN using historical data to detect

fraud patterns and tests its accuracy on new transactions.

4. Web Server: Manages communication between the user interface and

backend, ensuring efficient request handling and system scalability.

Department of Computer Science & Engineering, DSCE

 DEMONSTRATION WORKFLOW

5. User Interface (Transaction Testing): Simplifies interaction, allowing users to

upload data, test transactions, and view results intuitively.

cv
6. Transaction Validation: Analyzes transactions with the trained CNN model,
flagging fraudulent activities for further investigation.

7. Analysis (Results Reporting): Visualizes system performance using charts,

showing fraud detection metrics and model accuracy.

Department of Computer Science & Engineering, DSCE

DEMONSTRATION WORKFLOW

UI Design:
•Login Page: Simple with username and password fields.

cv
•Dashboard: Clean layout with access to Model Training, Verification, and
Analytics.

•Verification Form: Structured input fields with dropdowns, date pickers, and
a “Verify” button.

•Results Section: Clear display of verification results and factors.

•Analytics Section: Interactive bar graphs and pie charts for fraud trends.

Department of Computer Science & Engineering, DSCE

 MODULES

Login
Upload
Preview
Training/Testing
Check
Result (Valid Transaction)
Check
Fraud (Fraud Transaction)
Analysis
Accuracy Plot
 LITERATURE SURVEY
No. Author(s) Title Year Key Findings

Global consumer Explored global consumer

1 Nielsen, A. C. attitude towards 2005 attitudes, emphasizing online
online shopping shopping trends.

cv
Discussed practical applications
Hand, D. J., Blunt,
Data mining for fun of data mining, including fraud
2 G., Kelly, M. G., & 2000
and profit detection and profitability
Adams, N. M.
improvements.

Statistics for general Provided statistics on card

3 Epay News and online card 2007 fraud trends and prevention
fraud methods.

Department of Computer Science & Engineering, DSCE

 LITERATURE SURVEY
No. Author(s) Title Year Key Findings

UPI transaction fraud

Ghosh, S., & Reilly, D. Proposed a neural network model for
4 detection with a 1994
L. neural-network fraud detection in UPI transactions.

cv
Parallel granular Introduced parallel granular
Syeda, M., Zhang, Y. networks for fast UPI
5 2002 networks for fast and efficient fraud
Q., & Pan, Y. transaction fraud
detection.
detection

Stolfo, S. J., Fan, D.

W., Lee, W., UPI transaction fraud Presented meta-learning approaches
6 detection using meta- 1997 for UPI fraud detection with initial
Prodromidis, A. L., &
Chan, P. K. learning results.

Department of Computer Science & Engineering, DSCE

 LITERATURE SURVEY
No. Author(s) Title Year Key Findings

Cost-based modeling Shared results of fraud and intrusion

7 Stolfo, S. J., et al. for fraud and intrusion 2000 detection using cost-sensitive
detection models.

8
Aleskerov, E.,
Freisleben, B., &
Rao, B.
cv
CARDWATCH neural
network-based
database mining
system
1997
Proposed CARDWATCH, a neural
network-based fraud detection
system.

Developed a neural classifier with

Kim, M. J., & Kim, T. Neural classifier with
9 2002 fraud density mapping for UPI fraud
S. fraud density map
detection.

Web services-based Suggested a web services-based

10 Chiu, C., & Tsai, C. collaborative scheme 2004 collaborative approach for UPI fraud
for fraud detection detection.

Department of Computer Science & Engineering, DSCE

 BASE PAPER EXPLANATION

Framework and Key Contributions:

•EARN Ensembler: Combines feature vectors from Ensemble Autoencoder

cv
and ResNet to capture both high-dimensional and low-dimensional features in
financial data. This versatile tool reveals essential data patterns through
advanced feature extraction methods.
•Ensemble Learning Feature Vector: Uses multiple feature extraction
methods to create a unified, comprehensive feature representation, enhancing
the model’s ability to detect patterns in financial transactions.
•RXT-J Classification Model: Handles big data effectively with exceptional
accuracy, using ResNeXt-GRU (RXT) and hyperparameter tuning via the Jaya
Algorithm (JA) to distinguish between normal and fraudulent transactions.

Department of Computer Science & Engineering, DSCE

 BASE PAPER EXPLANATION
Challenges in Fraud Detection:

•Lack of Public Data Sets: Limited access to publicly available credit card
transaction data due to privacy concerns, restricting research opportunities.
cv
•Cost Sensitivity: False positives and negatives in fraud detection involve
significant costs, with false positives leading to administrative expenses and
false negatives resulting in financial losses.
•Data Imbalance: Fraud detection datasets are often skewed, with many
legitimate transactions and few fraudulent ones, making accurate classification
challenging. Balancing the data through preprocessing is necessary.
•Feature Space Dimensionality: Credit card transaction data includes diverse
features, requiring dimensionality reduction techniques like PCA and deep
representation methods to optimize learning and data representation.

Department of Computer Science & Engineering, DSCE

 RESULTS
1. Model Performance:
CNN achieved the highest accuracy, outperforming models like KNN, Logistic
Regression, and Decision Tree.
Effectively reduced false positives and negatives.

2. Fraud Detection Accuracy:

Reliable detection of fraudulent transactions with real-time validation.

3. User-Friendly Interface:
Simplified testing and instant result display with clear insights.

4. Visualization Tools:
Bar charts, line graphs, and pie charts highlighted fraud trends and patterns.

5. Analysis:
Time-based insights identified peak fraud hours.
Merchant categories with higher fraud risks were highlighted.
Scalability across platforms like UPI was demonstrated.
 CONCLUSION
1. Effectiveness of CNN for Fraud Detection:
CNN outperformed other models like KNN, Logistic Regression, and Decision
Tree in detecting fraud, due to its ability to capture complex patterns.
2. Improved Accuracy and Reliability:
The system minimized false positives/negatives, ensuring accurate fraud
detection with minimal disruption to legitimate users.
3. Reduction in Financial Losses:
By accurately flagging fraudulent transactions, the system reduced financial
losses for both users and organizations.
4. User-Centric Design:
The intuitive interface made transaction testing and analysis easy and
efficient for users.
5. Actionable Insights:
Visualization tools like pie charts and graphs helped identify fraud trends
and improve decision-making.
6. Comprehensive Dataset Utilization:
The system utilized a diverse dataset, considering key factors like transaction
time, category, amount, and location for accurate predictions.
 FUTURE SCOPE
1. Real-Time Detection: Enhance speed and scalability for high-volume
transactions.
2. Hybrid Models: Combine CNN with advanced techniques for better
accuracy.
3. Explainable AI: Offer clear reasons behind fraud detection decisions.
4. Adaptability: Continuous learning to handle evolving fraud patterns.
5. FDaaS Platform: Enable integration with financial ecosystems for scalable
solutions.
6. Global Accessibility: Multi-language support and cloud-based deployment.
END