The rapid proliferation of Unified Payments Interface (UPI) as a primary mode of digital financial transactions in India has simultaneously escalated the incidence and sophistication of payment fraud. Traditional rule-based fraud detection mechanisms have proven inadequate against dynamic and evolving fraud patterns due to their static nature, high false positive rates, and poor real-time adaptability. This paper presents Smart UPI Fraud Detect — a novel, intelligent fraud detection framework that leverages a hybrid machine learning architecture integrating Deep Q-Networks (DQN) for adaptive reinforcement learning with eXtreme Gradient Boosting (XGBoost) for high-precision classification. The proposed system is trained on a comprehensive dataset of UPI transaction records encompassing features such as transaction amount, timestamp, sender/receiver identifiers, device fingerprints, geographic location, and behavioral patterns. Advanced preprocessing techniques including SMOTE (Synthetic Minority Over-sampling Technique) are employed to address severe class imbalance inherent in fraud datasets. Experimental evaluation on a Kaggle-sourced real-world dataset demonstrates an accuracy of 98.7%, precision of 97.9%, recall of 98.5%, F1-score of 98.2%, and a minimal false positive rate of 1.2%, significantly outperforming baseline models including Random Forest, SVM, LSTM, and standalone XGBoost. The system supports real-time classification, adaptive learning from new fraud patterns, and explainability through feature importance analysis. This work contributes a scalable, production-ready fraud detection solution for the Indian digital payments ecosystem.

Keywords: UPI Fraud Detection, Machine Learning, Deep Q-Network, XGBoost, SMOTE, Digital Payments, Anomaly Detection, Reinforcement Learning, Real-time Classification, Financial Security