Zero-Day Banking Fraud Detection with Generative AI

Modern banking systems, operating at high frequency, require a delicate balance between rapid fraud detection and the explainability demanded by regulations such as GDPR. Traditional models struggle to identify "zero-day" attacks due to extreme data class imbalance and the lack of historical precedents.

This paper presents a dual-path generative framework designed to overcome these limitations. The architecture separates real-time anomaly detection from offline adversarial training. A Variational Autoencoder (VAE) establishes a legitimate transaction manifold based on reconstruction error, ensuring an inference latency of less than 50ms. Simultaneously, an asynchronous Wasserstein GAN with Gradient Penalty (WGAN-GP) synthesizes high-entropy fraudulent scenarios to stress-test the detection boundaries.

To handle the non-differentiability of discrete banking data (e.g., Merchant Category Codes), a Gumbel-Softmax estimator was integrated. Furthermore, a trigger-based explainability mechanism is introduced, where SHAP (Shapley Additive Explanations) is activated only for high-uncertainty transactions, reconciling the computational cost of XAI with real-time throughput requirements.

For those evaluating on-premise deployments, there are trade-offs to consider between the initial infrastructure costs and the long-term benefits in terms of data control and regulatory compliance. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these trade-offs.