Robust Explanation for Free or At the Cost of Faithfulness

Devoted to interpreting the explicit behaviors of machine learning models, explanation methods can identify implicit characteristics of models to improve trustworthiness. However, explanation methods are shown as vulnerable to adversarial perturbations, implying security concerns in highstakes domains. In this paper, we investigate when robust explanations are necessary and what they cost. We prove that the robustness of explanations is determined by the robustness of the model to be explained. Therefore, we can have robust explanations for free for a robust model. To have robust explanations for a non-robust model, composing the original model with a kernel is proved as an effective way that returns strictly more robust explanations. Nevertheless, we argue that this also incurs a robustness-faithfulness trade-off, i.e., contrary to common expectations, an explanation method may also become less faithful when it becomes more robust. This argument holds for any model. We are the first to introduce this trade-off and theoretically prove its existence for SmoothGrad. Theoretical findings are verified by empirical evidence on six state-of-the-art explanation methods and four backbones.