Color image sparse adversarial samples generation via color channel Volterra expansion

LingFeng Cheng, Heying Zhang, Jiashu Zhang

Abstract: The existing sparse adversarial attack methods for color images often overlook the interplay between RGB channels. In saliency analysis, only the individual impact of each input variable on the output is considered, while their interaction effects are disregarded. In order to overcome the lack of sparsity caused by inaccurate saliency analysis methods, we propose a sparse adversarial example generation method for color images based on Volterra expansion of color channels. Firstly, Volterra expansion is performed on different RGB channels, and a saliency analysis method is implemented to relate the interaction effects of different RGB channels. Then, the sparse adversarial perturbation is generated by a heuristic method with multiple iterations. Compared with the existing sparse adversarial attack methods for white-box color images, our proposed method can obtain sparser results with the same attack effect.