Transferable Adversarial Attacks via High-Frequency Deformation Transformation

Ensuring the transferability of adversarial samples, particularly those crafted by surrogate models, to uncertain target models stands as a critical challenge. The transferabil-ity of adversarial samples can be improved through gradient- based or input transformation-based attacks. Previous input transformation-based methods enhance the transferability of gen-erated adversarial samples through approaches such as random mixing and transformation via resizing, as well as through elastic deformation. These methods can identify regions with increased attack potential. However, the perturbations in these regions are still not thorough. In this paper, we propose a novel approach termed High-frequency Deformation Transformation (HDT) for further improving transferability. HDT involves extracting high-frequency warped region first after performing a vanilla warping transformation on the image. Furthermore, HDT introduces additional perturbations by targeting high-frequency regions acquired during the first stage. The experimental results indicate that our method significantly enhances the transferability of the adversarial samples generated.