Boosting Adversarial Attacks on Neural Networks with Better Optimizer

Algorithm 1

Adam iterative fast gradient method.

	Input: A convolutional neural network and the corresponding cross-entropy loss function ; an original image and the corresponding ground-truth label ; the number of iterations ; the iteration time step ; the dimension of the input image ; the size of the perturbation ; Adam decay factors and ; and a denominator stability factor .
	Output: An adversarial example , s.t. .
(1) , , , and ;
(2)
(3) while do:
(4)
(5)  ;
(6)  ;
(7)  ;
(8)  ;
(9)  ;
(10)  ;
(11)  ;
(12) end while
(13)	return .