Upi Fraud Shield: Real-Time Explainable Financial Fraud Detection and Proactive Early Warning Using Graph Neural Networks

Digital payment fraud is becoming more organized.Multiple transfers, devices, collusion between merchants and money-mule chains present patterns that can not be picked up by simple transaction models.This paper presents UPI Fraud Shield, a prototype for students to monitor fraud in real-time and explain why a transaction was flagged.UPI Fraud Shield uses table-based learning, time-based graph learning, explainable AI and a forward-looking warning layer.The system represents users, merchants and devices as a graph that is updated in a way over time.A LightGBM model employing boosting identifies behavioural signals such as speed of transfers and difference in amount compared to the common.At the moment a Graph Attention Network that looks into each transaction learns relationships between entities.A weighted ensemble picked by validation combines the two results.These are then passed to a policy layer that stores the model's risk separate from the final decision to ALLOW, REVIEW or BLOCK.Analysts get evidence for each alert via SHAP explanation graph context explanation and a what-if analysis that shows what would happen if conditions were to change.A React dashboard helps users monitor in time and investigate fraud rings.A Temporal Fraud Risk Prediction and Early-Warning Layer looks back 30 days to predict if an account or network will be part of fraud for the 7 days.This changes the system from reacting to fraud to warning about future risk.We tested on a synthetic data set of 9,060 transactions with a 5.00% fraud rate, split in time order into 70% training, 15% validation and 15% test.On the newest test slice (1,359 transactions), the validation-selected ensemble reaches 97.35% accuracy, 83.91% precision, 76.84% recall, 80.22% F1 score, 0.9072 ROC-AUC and 0.8103 PR-AUC.This is only marginally better than LightGBM alone (79.78% F1 score) and the difference is not statistically distinguishable, so the graph part is complementary but not the main driver.Median end-to-end scoring latency is 4.74 ms (P99 31.32 ms) on a CPU.Early-warning results are not reported here.