Enhancing the Transferability of Targeted Attacks with Adversarial Perturbation Transform

Abstract

The transferability of adversarial examples has been proven to be a potent tool for successful attacks on target models, even in challenging black-box environments. However, the majority of current research focuses on non-targeted attacks, making it arduous to enhance the transferability of targeted attacks using traditional methods. This paper identifies a crucial issue in existing gradient iteration algorithms that generate adversarial perturbations in a fixed manner. These perturbations have a detrimental impact on subsequent gradient computations, resulting in instability of the update direction after momentum accumulation. Consequently, the transferability of adversarial examples is negatively affected. To overcome this issue, we propose an approach called Adversarial Perturbation Transform (APT) that introduces a transformation to the perturbations at each iteration. APT randomly samples clean patches from the original image and replaces the corresponding patches in the iterative output image. This transformed image is then used to compute the next momentum. In addition, APT could seamlessly integrate with other iterative gradient-based algorithms, incurring minimal additional computational overhead. Experimental results demonstrate that APT significantly enhances the transferability of targeted attacks when combined with traditional methods. Our approach achieves this improvement while maintaining computational efficiency.