Provable Robust Saliency-based Explanations
This addresses the need for trustworthy AI explanations by improving stability against attacks, though it appears incremental as it builds on existing adversarial training and explanation methods.
The paper tackles the problem of making saliency-based explanations for machine learning models more stable against attacks by introducing a new metric for assessing stability of top-k salient features and a training method called R2ET with theoretical guarantees. The result shows R2ET achieves superior stability across various data and models.
To foster trust in machine learning models, explanations must be faithful and stable for consistent insights. Existing relevant works rely on the $\ell_p$ distance for stability assessment, which diverges from human perception. Besides, existing adversarial training (AT) associated with intensive computations may lead to an arms race. To address these challenges, we introduce a novel metric to assess the stability of top-$k$ salient features. We introduce R2ET which trains for stable explanation by efficient and effective regularizer, and analyze R2ET by multi-objective optimization to prove numerical and statistical stability of explanations. Moreover, theoretical connections between R2ET and certified robustness justify R2ET's stability in all attacks. Extensive experiments across various data modalities and model architectures show that R2ET achieves superior stability against stealthy attacks, and generalizes effectively across different explanation methods.