A Unified Approach to Interpreting and Boosting Adversarial\n Transferability

In this paper, we use the interaction inside adversarial perturbations to\nexplain and boost the adversarial transferability. We discover and prove the\nnegative correlation between the adversarial transferability and the\ninteraction inside adversarial perturbations. The negative correlation is\nfurther verified through different DNNs with various inputs. Moreover, this\nnegative correlation can be regarded as a unified perspective to understand\ncurrent transferability-boosting methods. To this end, we prove that some\nclassic methods of enhancing the transferability essentially decease\ninteractions inside adversarial perturbations. Based on this, we propose to\ndirectly penalize interactions during the attacking process, which\nsignificantly improves the adversarial transferability.\n