A Machine Learning Framework for Detecting Fraudulent Transactions in UPI Systems
: Frictionless peer-to-peer and merchant payments made possible by India's Unified Payments Interface (UPI) systems' rapid expansion. Nevertheless, the increase in transaction volume has also led to a rise in fraudulent activity, including as bot-driven transactions, phishing, and QR code manipulation. Rule-based algorithms and static machine learning classifiers, two traditional methods of detecting fraud, find it difficult to keep up with changing fraud strategies. In order to identify fraudulent UPI transactions, this study suggests a machine learning system that combines ensemble learning methods, deep learning architectures, and sophisticated feature engineering. For accuracy and interpretability, the system combines hybrid ensembles (CNN + Random Forest), transformer-based models for sequential transaction analysis, and graph neural networks (GNNs) for relational modeling. Preprocessing methods including feature scaling, categorical encoding, and data balancing algorithms like SMOTE are applied to a sizable, real-world UPI dataset. Through experiments, the suggested framework outperforms baseline models, obtaining a ROC-AUC of 0.996, 99.4% accuracy, and 98.7% precision. Furthermore, disclosure is enhanced by explainable AI (XAI) methods like SHAP, and sfraud detection across banks is protected by federated learning. This study increases UPI security against changing fraud tactics by providing a scalable, real-time, and interpretable architecture.