Robust generative adversarial network

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