Robust generative adversarial network

Generative Adversarial Networks (GANs) are one of the most popular and powerful models to learn the complex high dimensional distributions. However, they usually suffer from instability and generalization issues which may lead to poor generations. Most existing works focus on stabilizing the training for the discriminators of GANs while ignoring their generalization issue. In this work, we aim to improve the generalization capability of GANs by promoting the local robustness within the small neighborhood of the training samples. We prove that the robustness in the small neighborhood of the training sets can lead to better generalization. Particularly, we design a new robust method called Robust Generative Adversarial Network (RGAN) in which 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 Gaussian distribution used in most GANs) to the real data distribution, while the discriminator attempts to distinguish the real and fake distributions with the worst perturbations. Intuitively, the proposed RGAN can learn a good generator and discriminator that can even perform well on the worst-case input points. Strictly, we have proved that RGAN can obtain a tighter generalization upper bound than the traditional GANs under mild assumptions, ensuring a theoretical superiority of RGAN over GANs. We conduct our proposed method on five different baselines (five popular GAN models). And a series of experiments on CIFAR-10, STL-10 and CelebA datasets indicate that our proposed robust frameworks outperform five baseline models substantially and consistently.