DPI-Guardian: A Hybrid Graph-Temporal Framework for Real-Time Detection of Authorized Push Payment Fraud in Unified Payment Interfaces

Abstract

The proliferation of Digital Public Infrastructure (DPI), exemplified by India’s Unified Payments Interface (UPI), has revolutionized financial inclusion but has concurrently precipitated a surge in Authorized Push Payment (APP) fraud. Unlike unauthorized access, APP fraud involves victims socially engineered into voluntarily authorizing transfers, rendering traditional credential-based security ineffective. This paper introduces DPI-Guardian, a novel real-time anomaly detection framework designed to distinguish fraudulent intent from legitimate identity. We propose a hybrid architecture integrating Long Short-Term Memory (LSTM) networks for temporal behavioural analysis and Graph Neural Networks (GNN) for beneficiary relationship mapping. The study addresses critical gaps in current literature regarding latency constraints and social engineering indicators. The proposed framework targets a detection latency of <50ms to maintain UPI Service Level Agreements (SLAs) while significantly improving recall rates for coercion-based fraud.