An Adversarial Attack Based On Multi-Objective Optimization In The Black-Box Scenario: Moea-Apga Ii

Various approaches have been proposed to exploit the vulnerability to challenge the robustness of victim models, in the black-box scenario, it is difficult to generate barely noticeable adversarial examples while guaranteeing the attack success rate. Although some methods could solve this problem to some extent, the imperceptibility of the generated perturbations is still far from that of the most advanced attack, worse still, it is infeasible to attack the color image datasets due to its inefficiency. In MOEA-APGA II, We propose the new objective function and the novel population evolution strategies to reduce the average distortion without sacrificing the attack success rate, and compared to the state-of-the-art black-box attack (ZOO), our method achieves a better attack success rate under fewer queries on the benchmark datasets.