BE-RAGAN: A Bayesian Ensemble GAN Framework with Black- Scholes Risk Feature Integration for Scalable Financial Fraud Detection

Abstract

Fraud detection in financial transactions remains a critical challenge due to evolving fraud strategies, large-scale datasets, and the need for high detection accuracy with minimal false alarms. This paper proposes the Bayesian Ensemble Risk-Aware Generative Adversarial Network (BE-RAGAN), a hybrid and scalable fraud detection framework that integrates Black-Scholes feature engineering, Variational Autoencoders (VAE), Nyström approximation-based Gaussian Processes (GP), Random Projection Trees (RPTree), and Gated Recurrent Units (GRU) with Bayesian Reliability Fusion. The framework is evaluated on the Kaggle Synthetic Financial Datasets for Fraud Detection, which contains 6 million highly imbalanced transactions. Comparative experiments demonstrate that the Bayesian Ensemble Risk- Aware GAN outperforms baseline models including DL Ensemble and UAAD-FDNet variants. BE- RAGAN achieves a sensitivity of 0.970, specificity of 0.984, AUC of 0.968, precision of 0.987, F1-score of 0.968, and recall of 0.867, surpassing the performance of competing methods across all key metrics. These results confirm the robustness, adaptability, and scalability of BE-RAGAN for large-scale and real-time fraud detection. In addition, the framework enhances transparency through Bayesian reliability-based confidence scores, supporting interpretability in fraud risk assessment.