25 Sep 2019 (modified: 24 Dec 2019)ICLR 2020 Conference Blind SubmissionReaders: Everyone
  • Original Pdf: pdf
  • Keywords: Generative Adversarial Network, Robustness, Deep Learning
  • Abstract: Generative adversarial networks (GANs) are powerful generative models, but usually suffer from instability which may lead to poor generations. Most existing works try to alleviate this problem by focusing on stabilizing the training of the discriminator, which unfortunately ignores the robustness of generator and discriminator. In this work, we consider the robustness of GANs and propose a novel robust method called robust generative adversarial network (RGAN). Particularly, we design a robust optimization framework where the generator and discriminator compete with each other in a worst-case setting within a small Wasserstein ball. The generator tries to map the worst input distribution (rather than a specific input distribution, typically a Gaussian distribution used in most GANs) to the real data distribution, while the discriminator attempts to distinguish the real and fake distribution with the worst perturbation. We have provided theories showing that the generalization of the new robust framework can be guaranteed. A series of experiments on CIFAR-10, STL-10 and CelebA datasets indicate that our proposed robust framework can improve consistently on four baseline GAN models. We also provide ablation analysis and visualization showing the efficacy of our method on both generator and discriminator quantitatively and qualitatively.
4 Replies