Sentinel-UPI: A Graph Neural Network Approach for Real-Time Fraud Detection in High-Volume Unified Payments Interface (UPI) Transactions

DOI

https://www.doi.org/10.59256/ijrtmr.20260601008

Published In

Volume 6, Issue 1 (2026)

Published Online

January-February 2026

Pages

56-64

Article Type

Original Article

Check for updates

The Unified Payments Interface (UPI) now drives India's digital economy. In FY 2023-24 alone, it processed over 131 billion transactions. However, this massive growth has created new opportunities for financial cybercrime. Banks traditionally rely on rule-based engines or standard machine learning models to stop fraud. These older systems are struggling. They treat transactions as isolated events and miss the hidden connections between accounts. Fraudsters use this blind spot to build money mule networks and circular trading schemes. To solve this problem, we built Sentinel-UPI. It is a real-time fraud detection framework powered by Graph Attention Networks (GAT). Instead of looking at single transactions, Sentinel-UPI analyzes the entire transaction neighborhood. It assigns attention weights to high-risk nodes to find bad actors quickly. We tested our model using a modified PaySim dataset that mimics real Indian UPI traffic. The results were highly positive. Sentinel-UPI achieved an F1-score of 96.4% and an AUC-ROC of 0.98. It outperformed standard baselines like XGBoost and Random Forest by a wide margin. Our model, processes transactions in under 20 milliseconds. This speed perfectly meets the strict Service Level Agreements (SLAs) needed for live UPI networks.