Robust GAN Based on Attention Mechanism

Abstract

AbstractDeep neural networks (DNNs) have been found to be easily mislead by adversarial examples that add small perturbations to inputs to produce false results. Different attack and defense strategies have been proposed to better study the security of deep neural networks. But these works only focus on an aspect such as attack or defense. In this work, we propose a robust GAN based on the attention mechanism, which uses the deep latent features of the original image as prior knowledge to generate adversarial examples, and it can jointly optimize the generator and discriminator in the case of adversarial attacks. The generator generates fake images based on the attention mechanism to deceive discriminator, the adversarial attacker perturbs the real images to deceive discriminator, and the discriminator wants to minimize the loss between fake images and adversarial images. Through this training, we can not only improve the quality of adversarial images generated by GAN, but also enhance the robustness of the discriminator under strong adversarial attacks. Experimental results show that our classifier is more robust than Rob-GAN [14], and the generator outperforms Rob-GAN on CIFAR-10.KeywordsRobustGANAdversarialAttention mechanism