Multi-attribute Semantic Adversarial Attack Based on Cross-layer Interpolation for Face Recognition

With the extensive research and application on Face Recognition (FR) model in daily life, the security of FR has attracted much attention as it is easily attacked by adversarial examples. Specifically, adversarial attacks can cause the model to make completely erroneous judgments by making very subtle changes to the source image. Therefore, it is of great significance for studying adversarial attacks that can improve the robustness and security of FR models. However, most of the existing attacks have low transferability of attack and high vulnerability to denoising defense models. To solve the above problems, a multi-attribute semantic adversarial attack based on cross-layer interpolation(C&A Adv) is proposed, which can generate imperceptible adversarial images whith high success rate and robustness to denoising defense methods. Particularly, C&A Adv semantically edit images by cross-layer feature space interpolation, which not only generates high quality adversarial images, but also has the robustness to partial denoising defense methods. In addition, to improve the success rate of the attack, several attributes are selected to edit instead of just one. According to the marginal gain of each attribute calculated in different face images, several attributes with the greatest marginal gain are selected to edit. Comparison and verification on CelebA dataset show that the C&A Adv achieves good experimental result.