A Deep Hybrid Learning Approach for Fraud Detection in Digital Payment Systems Using Bi-LSTM and Ensemble Boosting

The emergence of electronic payment systems such as UPI and mobile wallets has simplified transactions but has also exposed the system to smart fraud attacks. Legacy fraud detection systems which are based on rule-based systems are unable to detect new threats in real-time. This work evaluates a comparative fraud detection framework using Bi-LSTM networks and ensemble boosting techniques (XGBoost and LightGBM) to enhance detection accuracy and scalability in electronic financial systems. Bi-LSTM captures sequential interdependencies between transactions, and ensemble models detect static behavioral features with good accuracy. Thorough preprocessing and feature engineering—temporal feature extraction and anomaly-sensitive transformation—enhanced model performance. Experimental results indicate that XGBoost performed the best with the best test accuracy (98.99%) and ROC-AUC (0.9997), while LightGBM and Bi-LSTM performed closely. The union of deep temporal learning and explainable ensemble techniques presents an efficient solution to real-time fraud detection, where accuracy and explanation are a prerequisite to deployment in the financial sector.