Sensitivity of Adversarial Perturbation in Fast Gradient Sign Method
Yujie Liu, Shuai Mao, Xiang Mei, Tao Yang, Xuran Zhao · 2019
Fast Gradient Sign Method is a well-known method for adversarial sample attack. New adversarial samples could be generated by adding a small perturbation to input images, however the small perturbation usually need users to select by themselves. This paper focuses on FGSM attack in face recognition scenario and empirically evaluate multiple factors for adversarial perturbation in terms of recognition performance. The results demonstrate adversarial perturbation is sensitive to many factors, such as size of perturbation, number of iterations, and granularity of the perturbation.