Poisoning the Competition: Fake Gradient Attacks on Distributed Generative Adversarial Networks

Generative Adversarial Networks (GANs) have gained significant attentions due to their powerful generative abilities. However, the requirements for data privacy protection and data decentralization have led to the emergence of Distributed GANs (DGANs), where multiple discriminators collaboratively train the central generator model. However, the mode of distributed training poses the risk of fake gradient attacks where the discriminator forges the gradient to destroy or induce the training process of the generator. In this paper, we first explore the impact of gradient attacks on generator models in DGANs, with two types of fake gradient attacks proposed. The qualitative experiments on two simulated datasets demonstrate that the distribution characteristics of the generated data can be seriously affected by the two attacks. The attack reduces the similarity between the generated and real data and yields higher FID values than the most commonly used models. Especially, in the context of DGANs utilizing the Wasserstein GAN, our data-free attack increases the FID value of the generated images from 18.15% to 54.08%, and from 15.50% to 37.33% on the MNIST and CIFAR 10 datasets, respectively.