Improving the Transferability of Adversarial Examples Through Spatial-Based Color Modification

Unrestricted adversarial attacks that modify the attributes or content of images have demonstrated significant potential. Among unrestricted adversarial attacks, color modification attacks that targeting image color have increasingly obtained attention from researchers. By perturbing the color information of images, color modification attacks can be more imperceptible to human vision compared with traditional adversarial perturbations constrained by$ L_{p}$-norm. Furthermore, the alteration of image attributes significantly disrupts the low-level features extracted by Deep Neural Networks (DNNs), effectively enhancing the transferability. However, current work focus solely on global image attributes, which limit the attack effectiveness. Therefore, this paper proposes Local-Global-Local (LGL), a novel color modification attack framework that incorporates spatial and color information of the image. The color distribution of the image is initialized at first, and then the color filter and spatial masks are used to adjust the global and local color distribution in more detail. Experimental results show that the proposed method achieves superior performance in terms of adversariality and transferability, while also demonstrating robustness against defense methods.