Paper accepted at “Applied Soft Computing” Journal

Title: GNS-GAN: A novel GAN model based on gradient noise suppression

Authors: Hongyou Chen, Lingfeng Qu, Baodan Tian, Yutong He, Yong Fan, Hadi Amirpour, Christian Timmerer and Yao Xin

Abstract: Generative adversarial networks (GANs) are widely applicable generative models. However, ensuring stability in adversarial learning remains a significant challenge in current GAN training. Gradient noise, among other factors, significantly impacts the stability of adversarial learning in GAN training. To improve the stability of adversarial learning, a gradient noise suppression generative adversarial network model (GNS-GAN) is proposed. This novel GAN addresses gradient noise by establishing stochastic differential equations (SDEs) for gradient noise in both the discriminator and the generator. The factors affecting the stability of adversarial learning are then analyzed using the assumed gradient noise distribution. Subsequently, an adversarial learning method is designed for the discriminator and generator to suppress gradient noise, thereby completing the adversarial training of GNS-GAN. To verify the performance of GNS-GAN, the experimental results are compared and analyzed using CELEBA, BEDROOM, and CIFAR10 datasets. The FID (Fréchet Inception Distance) values are 23.04 for CELEBA, 18.04 for BEDROOM, and 26.59 for CIFAR10. The GNS-GAN model has stable training performances in the tested datasets. These results demonstrate that the novel GAN model enhances the stability of adversarial learning and the quality of the generated images.