Discriminator-Free Generative Adversarial Attack Shaohao Lu Yuqiao Xian Ke Yan School of Computer Science and School of Computer Science and Tencent Youtu Lab Engineering Engineering Shanghai, China Guangzhou, China Guangzhou, China
[email protected] [email protected] [email protected] Yi Hu Xing Sun Xiaowei Guo Tencent Youtu Lab Tencent Youtu Lab Tencent Youtu Lab Shanghai, China Shanghai, China Shanghai, China
[email protected] [email protected] [email protected] Feiyue Huang Wei-Shi Zheng Tencent Youtu Lab School of Computer Science and Shanghai, China Engineering
[email protected] Guangzhou, China
[email protected] (a) AdvGAN (A method w/ discriminator.) (b) Ours (A method w/o discriminator.) Figure 1: Both the images above can make ResNet18 [11] fail without changing in appearance significantly. The one generated by our method gets better visual quality without discriminator, while the other one generated by AdvGAN [35] has conspicuous circular noises. arXiv:2107.09225v1 [cs.CV] 20 Jul 2021 ABSTRACT the existing works for adversarial attack are gradient-based and suf- The Deep Neural Networks are vulnerable to adversarial exam- fer from the latency efficiencies and the load on GPU memory. The ples (Figure 1), making the DNNs-based systems collapsed by generative-based adversarial attacks can get rid of this limitation, adding the inconspicuous perturbations to the images. Most of and some relative works propose the approaches based on GAN. However, suffering from the difficulty of the convergence oftrain- Permission to make digital or hard copies of all or part of this work for personal or ing a GAN, the adversarial examples have either bad attack ability classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation or bad visual quality.