Towards Automating Model Explanations with Certified Robustness Guarantees

Providing model explanations has gained significant popularity recently. In contrast with the traditional feature-level explanations, concept-based can provide in form of high-level human concepts. However, existing explanation methods implicitly follow a two-step procedure that involves intervention. Specifically, they first need to be involved define (or extract) concepts, and then manually compute importance scores these identified concepts post-hoc way. This laborious process requires effort resource expenditure due manual work, which hinders their large-scale deployability. practice, it is challenging automatically generate without intervention subjectivity defining units interpretability. addition, its data-driven nature, interpretability itself also potentially susceptible malicious manipulations. Hence, our goal this paper free from tedious process, while ensuring generated are provably robust adversarial perturbations. We propose novel interpretation method, not only prototype-based concept but certified robustness guarantees for explanations. conduct extensive experiments on real-world datasets verify desirable properties proposed method.