UPI Fraud Shield: Real-Time Explainable Financial Fraud Detection and Proactive Early Warning Using Graph Neural Networks (International Research Journal of Modernization in Engineering Technology and Science)

Summary

• UPI Fraud Shield is a prototype designed to monitor digital payment fraud in real time using tabular learning, graph neural networks, explainable AI, and an early-warning layer.p.1
• The evaluation dataset consists of 9,060 synthetic transactions with a 5.00% fraud rate, covering 500 users, 100 merchants, and 200 devices.p.2
• The hybrid detection layer combines LightGBM (0.85 weight) and a Graph Attention Network (0.15 weight), triggering an alert at a validation-selected threshold of 0.35.p.3
• On the test slice of 1,359 transactions, the ensemble model achieved 97.35% accuracy, 83.91% precision, 76.84% recall, and an F1 score of 80.22%.p.5
• Detection performance varies across fraud patterns, achieving 96.7% recall for velocity abuse and 100% for money-mule chains, but 0% for merchant collusion.p.5
• The prototype achieved a median end-to-end processing latency of 4.74 ms and a P99 latency of 31.32 ms on a CPU across 2,698 timed transactions.p.6

Key insights

1. LightGBM alone serves as the primary driver of fraud detection performance, achieving an F1-score of 79.78% on the chronological test slice.p.5
2. Integrating the Graph Attention Network yields only a marginal, statistically indistinguishable F1-score improvement from 79.78% to 80.22%.p.5
3. A policy layer separates predicted model risk into ALLOW, REVIEW, or BLOCK decisions to keep probabilistic risk scores distinct from operational actions.p.3
4. The proactive early-warning layer uses a 30-day lookback window to forecast fraudulent involvement over a 7-day future horizon without automatically suspending accounts.p.3
5. The detection layer suffered from zero recall on merchant collusion incidents, missing all 17 collusion transactions present in the test slice.p.5
6. Warm scoring latency remains low, but the first transaction of a new calendar day requires a graph snapshot rebuild that takes 367 ms at P50.p.6