A Dual-Path Generative Framework for Zero-Day Fraud Detection in Banking Systems

Researchers Nasim Abdirahman Ismail and Enis Karaarslan have introduced a novel AI architecture designed to tackle one of banking's toughest challenges: detecting never-before-seen 'zero-day' fraud in real-time while meeting strict regulatory explainability requirements. Their Dual-Path Generative Framework fundamentally decouples the detection pipeline, using a Variational Autoencoder (VAE) to establish a baseline of legitimate transactions and flag anomalies in under 50 milliseconds. This speed is critical for high-frequency environments where delayed decisions mean lost funds.

In parallel, an asynchronous Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) continuously stress-tests the system offline by synthesizing high-entropy, realistic fraudulent scenarios. A key technical innovation is the integration of a Gumbel-Softmax estimator to handle non-differentiable, discrete banking data like Merchant Category Codes. To reconcile performance with compliance, the framework features a trigger-based explainability system where computationally intensive SHAP (Shapley Additive Explanations) analysis is activated only for high-uncertainty transactions flagged by the VAE, preserving throughput.

The paper, published on arXiv, addresses the core trade-off between the low-latency demands of real-time fraud prevention and the computational burden of Explainable AI (XAI) required by regulations like GDPR. By separating the lightning-fast detection path from the resource-intensive adversarial training and on-demand explanation path, the framework offers a practical blueprint for deploying advanced generative AI in heavily regulated, performance-critical financial systems.