Real-time adaptive graph-temporal fraud detection (RAGTFD) for UPI transactions

With over 650 million transactions processed every day, the Unified Payments Interface has revolutionized digital payments in India, outperforming international systems like Visa. However, this growth has fueled sophisticated frauds, including tampered quick-response codes, phishing links, and behavioral manipulation, which challenge traditional rule-based and static machine learning systems. This paper proposes the Real-time Adaptive Graph-Temporal Fraud Detection framework, a multi-layered system for robust, real-time fraud detection in Unified Payments Interface transactions. The framework integrates Long Short-Term Memory networks for sequential behavior analysis, Graph Neural Networks for relational pattern detection, and adaptive thresholding for dynamic risk assessment. Explainable artificial intelligence modules, using Shapley Additive Explanations and Local Interpretable Model-Agnostic Explanations, provide transparent justifications for flagged transactions. Deployed in a custom mobile application, the framework enables real-time transaction screening and quick-response code validation, supported by a Streamlit dashboard for offline trend analysis.