In this paper, a new average gradient-based adversarial attack is proposed. In our proposed method, via utilizing the gradient of each iteration in the past, a dynamic set of adversarial examples is constructed first in each iteration. Then, according to the gradient of the loss function with...