CSA: Crafting adversarial examples via content and style attacks

Abstract

Most existing black-box attacks fall into two categories: gradient-based attacks and unrestricted attacks. The former injects adversarial perturbations into the original clean examples under the $L_p$-norm constraint, while the latter tends to attack by changing the shape, color, and texture of the original image. However, the adversarial examples generated by the gradient-based attacks are vulnerable to defense methods and unnatural to the human eye. Meanwhile, unrestricted attacks have poor transferability of adversarial examples compared to gradient-based attacks. Therefore, we propose a novel attack that combines gradient-based and unrestricted attacks, i.e., Content and Style Attack (CSA). Specifically, we utilize an encoder to extract the content features of the original image and train a reconstructor to generate an image consistent with these features. A gradient-based method is then employed to inject perturbations, followed by using the encoder to extract the content features of the altered image. We implement a momentum-based approach to search for malicious style information, which is then fused with the adversarial content features to create the final attack features. Extensive experiments on the ImageNet standard dataset demonstrate that our method is capable of generating adversarial examples that are both natural-looking and possess high transferability.