Towards a Unified Game-Theoretic View of Adversarial Perturbations and Robustness

How to explain adversarial robustness from the perspective of feature representation? • We discover that adversarial attacks mainly affect high-order interactions between input variables. • Adversarial training boosts the robustness of DNNs by learning more discriminative low-order interactions. • We proposed a unified explanation for several adversarial defense methods.