Intelligent and Optimized Fraud Detection Framework for UPI Transactions

The Digital Payment sector witnessed a transformation in the last couple of years, in part due to the increase in the use of UPI Transactions in India. However, this also led to an increase in vulnerable, sophisticated, and complicated fraud schemes. In this paper, we propose an integration of Seagull Optimization Algorithm (SOA) and LightGBM, which can achieve desirable results in terms of accurately forecasting fraudulent UPI Transactions. In our method, we capture the fraud indicators using detailed feature engineering, which includes temporal patterns, transaction behavioral analytics, and transaction anomalies. The SOA metaheuristic algorithm optimizes the hyperparameters of LightGBM and outperforms traditional optimization methods in accuracy and performance. We demonstrate the model results using an experimental dataset of 100,000 transactions, achieving 98.08 % cross-validated AUC, 98.02 % test AUC, and 94.95% accuracy at the optimal decision threshold of 0.30. The model achieves high recall of 99.98 % while maintaining precision of 80 %, effectively balancing false positives and false negatives. We also showcase the importance of critical fraud detection indicators: Device_ID, Customer_ID, IP_Address, and Merchant_ID, demonstrating this research's contribution in creating an effective, real-time, and reliable fraud detection model for UPI Transactions in Digital Payment Systems.