SCA:Segmentation based adversarial Color Attack

Deep Neural Networks (DNNs) are widely used nowadays. Some researchers have found an effective way to do harm to the DNNs, which is called adversarial examples. Via the way of generating the adversarial examples can divide the attack into two directions. Part of the researchers is focusing on generating imperceptible images whose noise is restricted by Lp- Norm. However, these adversarial examples have poor transferability. Therefore, some researchers are focusing on generating images with large unrestricted corruption. However, existing unrestricted images are noticeable, arousing human suspicion. Among them, the attack that modifies the color of images may generate images with better image quality. In this work, we proposed an unrestricted adversarial attack that modifies the color of original images with different strengths in each region. The regions are segmented by a segmentation network. The experiment shows that our method has strong attack strength with a quite high image quality compared with existing unrestricted methods.