Imperceptible adversarial attack with entropy feature and segmentation-based constraint

Rongdong Li, Qinliang Lin, Yinglong Fu, Weicheng Xie, Linlin Shen

Abstract: Methods of adversarial attack and defense have attracting increasing attention in the fields of security and protection related applications. However, current algorithms carry out perturbations on entire images and mostly consider their imperceptibility to machines, while does not take their human imperceptibility into account. In this work, we propose a constrained adversarial attack algorithm with both machine and human imperceptibility based on image entropy feature and accurate segmentation. The proposed algorithm has three merits. First, image entropy-based feature for quantifying the imperceptibility of a semantic region is introduced, which is simple yet efficient to implement. Second, in terms of the imperceptibility metric, accurate target regions for adversarial perturbation are obtained based on scene-aware segmentation and merging. Third, a general adversarial attack based on segmentation region constraint is proposed to induce both machine and visual imperceptibility. Experimental results in terms of qualitative and quantitative analysis reflect the effectiveness of the proposed algorithm compared with the state of the arts.