Title: Enhancing Neural Network Reliability Through Bayesian-Driven Navigation of Decision Boundaries

Abstract:

As the deployment of neural networks expands into safety-critical sectors, rigorous testing has become indispensable for assessing and enhancing their dependability. Current evaluation techniques, including both black-box and white-box approaches, typically rely on global mutation or coverage-driven strategies. However, these methods often fail to efficiently identify a wide range of model failures while maintaining closeness to the original data distribution and semantic integrity.

To overcome this challenge, we introduce BayesWarp, a novel testing framework designed to mutate input regions that are critical to decision-making, as identified through interpretable saliency methods. The framework employs an uncertainty-aware Bayesian Optimization strategy to adaptively steer the testing process. This approach facilitates the discovery of diverse failures without sacrificing distributional or semantic proximity to the source data.

Our evaluation across MNIST, CIFAR-10, and ImageNet datasets, utilizing six distinct neural network models, indicates that BayesWarp outperforms existing methods in several key areas: failure discovery, failure diversity, test case quality, and coverage of critical neurons, all within a constrained mutation budget. These findings underscore the enhanced effectiveness of BayesWarp in testing scenarios. Furthermore, we demonstrate that fine-tuning models using the generated failure cases results in improved overall model performance.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC