Adversarial example attack of SAR image based on Grad-CAM mask extraction

The adversarial example attack on SAR images has been widely used in military confrontation in recent years. However, many existing adversarial example generation methods still exist problems such as excessive generation perturbation, unstable training, and susceptibility to detection. In response to the above issues, a SAR adversarial example generation method based on Grad-CAM mask extraction has been proposed, which uses Grad-CAM algorithm to extract masks from SAR images, and optimizes the pixel points in the mask for adversarial example search. Finally, the adversarial examples generated by proposed method have a structural similarity of over 0.99 between the adversarial examples and the original SAR images, while ensuring an adversarial success rate of over 90%. The number of perturbed pixels is also less than 1/5 of the total pixels in the SAR images. The proposed method effectively improves structural similarity and reduces the number of perturbed pixels in adversarial examples.