Self-Universal Cognitive Patch Attack for Targeted Transfer

Deep neural networks (DNNs) achieve remarkable results in image recognition, speech recognition, and other fields. However, their vulnerability to adversarial examples raises significant security concerns. Adversarial examples cause model misjudgment by introducing minimal perturbations to original samples. In the realm of current adversarial attack research, although traditional iterative attacks are used to generate adversarial examples, they still falter in target migration tasks. To address this, we propose a Self-Universal Cognitive Patch Attack (SUCPA) to boost the success rate of transfer-based targeted attacks. Our approach is motivated by the finding that generalized adversarial perturbations can enhance target migration attack performance. SUCPA innovatively avoids extra data usage, optimizing adversarial perturbations by maximizing feature similarity between adversarial global images and randomly cropped local regions. Specifically, we introduce a domainagnostic attention module and a patch attack module with weight decay to leverage robust intermediate feature representations, aggregate gradient information, and improve perturbation generality. Experiments on the ImageNet dataset demonstrate that SUCPA outperforms existing gradient-based attack methods on both standard and robustly trained models.