Enhancing financial decision-making under cyber threats: a dual-branch framework integrating bayesian deep learning and explainable AI

Yuhe Wu, Yuran Chen, Zhuang Liu, Wayne Lin

Abstract: Amidst the increasingly severe landscape of financial and digital security, network attacks have become a major threat to user privacy and the integrity of financial systems. As attack methodologies evolve, traditional detection methods have proven insufficient. While modern deep learning models have enhanced detection capabilities, their inherent “black-box” nature and predictive uncertainty limit their application in high-stakes financial decision-making that demands high reliability. To address this challenge, this paper introduces BnetX, an innovative “dual-branch” framework that deeply integrates Bayesian deep learning (BDL) and explainable AI (XAI). Through Bayesian inference, BnetX is able to quantify the uncertainty of model predictions, thereby significantly enhancing the reliability of detection outcomes. Concurrently, the framework integrates XAI techniques such as SHAP and Grad-CAM to visualize key decision-driving features and improve the transparency of the model’s decision-making process. Comprehensive experiments on the KronoDroid dataset, which includes both static and dynamic features, demonstrate that the proposed BDNN model achieves an F1-score of 98.25%, significantly outperforming various baselines, including TabNet. Furthermore, cross-domain validation experiments on a financial network fraud dataset confirm the framework’s excellent generalization capability. The BnetX framework not only improves detection accuracy but also provides a complete decision information chain comprising “prediction, uncertainty, explanation, and the uncertainty of the explanation”. It offers a validated, reliable, and interpretable new paradigm for building and deploying trustworthy AI systems in high-risk financial environments.

External IDs:doi:10.1007/s10479-025-06973-2