Abstract:

Deep neural networks have high accuracy in image classification. However, when small adversarial perturbation is added to the original image, the accuracy of classification will decrease significantly. Studies show that there is an universal adversarial perturbation for a classifier and a data set, which can attack most of the original images. This paper designs a method for making universal adversarial perturbation with generative adversarial network. Through the training of the generative adversarial network, the generator can make an universal adversarial perturbation which added to the original image to make the adversarial sample, so as to achieve the purpose of attack. This paper conducts no target attack, target attack and transfer attack experiments on the CIFAR-10 dataset. Experiments show that the universal adversarial perturbation generated by the generative adversarial network can reach an attack success rate of 89% under lower norm constraints, and the trained generator can produce a large number of adversarial samples in a short time, which is conducive to the robustness research of deep neural network.

Key words: deep neural network, universal adversarial perturbation, generative adversarial network