Local Black-box Adversarial Attack based on Random Segmentation Channel

With the wide application of deep neural networks, the security problem of the model is becoming more prominent, adversarial attack is an important tool for evaluating the robustness and security of the model, adversarial attack can be categorized into white-box and black-box attacks. Aiming at the problem of huge perturbation and the low success rate of the adversarial example created in the transfer attack in the black box, a local region approach for randomly segmented channels is given. By randomly segmenting individual dimensions in order to improve the transferability of the adversarial example, the ScoreCAM method is introduced to extract localized focus regions of the image to generate the adversarial example. Experiments show that the performance of the method is better than the baseline algorithm; the fooling rate improved by up to 17%; and the average 2-norm module length decreased by 54.2%.