A Hybrid Autoencoder-XGBoost Framework for High-Performance UPI Fraud Detection

This paper proposes a scalable hybrid autoencoder–XGBoost system for detecting unified payment interface fraud. The approach begins by training an autoencoder on legitimate transactions to learn normal behavior patterns, where reconstruction errors are derived and utilized as anomaly scores. These scores serve as engineered features in an XGBoost classifier for final fraud classification. The system is tested on a synthetic dataset of 2.68 million transactions. The findings show near-perfect performance, with accuracy, precision, recall, and F1-scores close to 1.0 and a receiver operating characteristic curve-area under curve (ROC–AUC) of 0.99999995. However, these results are influenced by deterministic fraud patterns in the simulated dataset, leading to near-separable classes with domain-driven balance features. Therefore, the performance should be interpreted as proof-of-concept under controlled synthetic conditions rather than absolute evidence of real-world effectiveness. The model demonstrates the potential of anomaly-aware feature enrichment for handling severely imbalanced data. Future work will focus on validation with real-world UPI data and adaptive learning upgrades.