Generating Imperceptible Adversarial Patch Based on Vulnerable Targeted Attack

The vulnerability and sensitivity of deep neural networks to adversarial examples have attracted researchers' attention towards the security application of deep model. Current methods mainly focus on the aggressiveness and size of the patches, while ignoring the imperceptibility and concealment of the patches. In this paper, we propose a two-step attack algorithm to generate imperceptible adversarial patches. Specifically, the first step of our algorithm is to obtain the vulnerable target category that is most likely to be misclassified; and the second step is to perform targeted attack using the vulnerable target coming from the first step. The proposed algorithm is very robust and can be generalized to other untarget attack approaches, such as Shadow attack and IFGSM. Extensive experiments on widely used classifiers demonstrate the availability and efficiency of the proposed algorithm.