Graph Neural Network-Based Real-Time Fraud Detection for UPI Transactions

Objective: To design and deploy a real-time fraud detection system for UPI transactions using graph machine learning to identify suspicious transaction patterns and prevent financial losses.

The surge in Unified Payments Interface (UPI) transactions across India has made digital payments seamless, but also exposed millions to sophisticated frauds such as mule accounts, phishing, and synthetic identity attacks. Financial institutions, payment gateways, and everyday users are at risk, with losses running into crores due to undetected fraud patterns embedded in massive transaction networks.

This project proposes a real-time fraud detection solution leveraging Graph Neural Networks (GNNs) to model UPI transaction flows as evolving graphs, capturing complex relational patterns between accounts, devices, and transactions. The system ingests anonymized transaction data, constructs a dynamic graph, and applies supervised GNNs to classify suspicious edges or nodes in real time, integrating seamlessly with UPI backends via a REST API.

Key features include: an end-to-end ML pipeline with scalable ETL for streaming UPI data, graph construction and feature engineering, GNN model training and inference (using public datasets such as PaySim or simulated UPI data), a Python Flask dashboard for live alerts, and explainability modules for flagged transactions. The working prototype demonstrates detection accuracy, latency, and integration with simulated UPI transaction streams.

This solution enables Indian banks and fintechs to proactively mitigate fraud, protect users, and reduce compliance risk, with potential scalability to millions of users and adaptation for other digital payment platforms. The project delivers deployable code, a dashboard, and a technical report, ready for industry adoption.