A Dual-Path Generative Framework for Zero-Day Fraud Detection in Banking Systems
arXiv:2603.13237v1 Announce Type: new Abstract: High-frequency banking environments face a critical trade-off between low-latency fraud detection and the regulatory explainability demanded by GDPR. Traditional rule-based and discriminative models struggle with "zero-day" attacks due to extreme class imbalance and the lack of historical precedents. This paper proposes a Dual-Path Generative Framework that decouples real-time anomaly detection from offline adversarial training. The architecture employs a Variational Autoencoder (VAE) to establish a legitimate transaction manifold based on reconstruction error, ensuring <50ms inference latency. In parallel, an asynchronous Wasserstein GAN with Gradient Penalty (WGAN-GP) synthesizes high-entropy fraudulent scenarios to stress-test the detection boundaries. Crucially, to address the non-differentiability of discrete banking data (e.g., Merchant Category Codes), we integrate a Gumbel-Softmax estimator. Furthermore, we introduce a trigger-ba
arXiv:2603.13237v1 Announce Type: new Abstract: High-frequency banking environments face a critical trade-off between low-latency fraud detection and the regulatory explainability demanded by GDPR. Traditional rule-based and discriminative models struggle with "zero-day" attacks due to extreme class imbalance and the lack of historical precedents. This paper proposes a Dual-Path Generative Framework that decouples real-time anomaly detection from offline adversarial training. The architecture employs a Variational Autoencoder (VAE) to establish a legitimate transaction manifold based on reconstruction error, ensuring <50ms inference latency. In parallel, an asynchronous Wasserstein GAN with Gradient Penalty (WGAN-GP) synthesizes high-entropy fraudulent scenarios to stress-test the detection boundaries. Crucially, to address the non-differentiability of discrete banking data (e.g., Merchant Category Codes), we integrate a Gumbel-Softmax estimator. Furthermore, we introduce a trigger-based explainability mechanism where SHAP (Shapley Additive Explanations) is activated only for high-uncertainty transactions, reconciling the computational cost of XAI with real-time throughput requirements.
Executive Summary
This paper introduces a novel Dual-Path Generative Framework designed to address the dual challenges of zero-day fraud detection in high-frequency banking systems: maintaining sub-50ms inference latency for real-time anomaly detection while enabling regulatory explainability under GDPR. The architecture effectively decouples anomaly detection from adversarial training via a VAE-based legitimate transaction manifold and an asynchronous WGAN-GP for synthetic fraud scenario generation. The inclusion of a Gumbel-Softmax estimator for discrete banking data and a trigger-based SHAP explainability mechanism demonstrates a thoughtful mitigation of technical constraints. The framework balances performance, compliance, and transparency in a novel way.
Key Points
- ▸ Decoupling real-time anomaly detection from adversarial training via dual-path architecture
- ▸ Utilization of VAE for legitimate transaction manifold modeling with sub-50ms latency
- ▸ Integration of Gumbel-Softmax and SHAP explainability to address discrete data and computational cost challenges
Merits
Innovative Dual Architecture
The dual-path framework provides a structured separation between detection and training, enabling optimized performance and compliance simultaneously.
Technical Integration
Smart use of Gumbel-Softmax for discrete data non-differentiability and SHAP for selective explainability shows strong engineering insight and adaptability.
Demerits
Complexity of Implementation
Coordinating asynchronous adversarial training with real-time detection may introduce operational complexity in deployment environments.
Assumption Dependency
Performance may be contingent on accurate VAE manifold modeling and WGAN-GP synthetic fraud diversity, which could be difficult to validate in real-world scenarios.
Expert Commentary
The Dual-Path Generative Framework represents a significant advancement in the intersection of fraud detection, explainability, and real-time processing. The authors have successfully navigated a complex trilemma—latency, accuracy, and compliance—by leveraging generative modeling in a targeted, modular fashion. The use of WGAN-GP for adversarial stress-testing is particularly commendable, as it introduces a dynamic threat modeling component that traditional rule-based systems cannot replicate. Furthermore, the selective deployment of SHAP via trigger mechanisms reflects a pragmatic understanding of XAI’s computational burden. This work contributes meaningfully to the broader discourse on adaptive security in finance and sets a new benchmark for balancing operational efficiency with regulatory obligations. One potential avenue for future research might explore scalability across distributed banking ecosystems or integration with federated learning for collaborative anomaly detection.
Recommendations
- ✓ 1. Conduct benchmarking against existing zero-day detection models using real transaction logs to validate latency and accuracy claims.
- ✓ 2. Evaluate the framework’s adaptability in multi-institutional banking networks via pilot deployments with shared anomaly data.
