MGMD-GAN: Generalization Improvement of Generative Adversarial Networks with Multiple Generator Multiple Discriminator Framework Against Membership Inference Attacks
Nirob Arefin
TL;DR
MGMD-GAN addresses privacy risks in GANs by targeting the generalization gap that enables Membership Inference Attacks. It introduces a Multi-Generator Multi-Discriminator framework that partitions the training data into $K$ disjoint subsets and trains $K$ G–D pairs to learn a mixture distribution across partitions without a built-in privacy adversary. The authors provide a mathematical formulation with per-partition losses and show that the framework reduces overfitting and limits attack success in white-box MIA scenarios. Empirical results on MNIST demonstrate reduced generalization gap and competitive resistance to MIA compared with PAR-GAN, with performance depending on $K$ and the chosen objective $\phi$. This approach offers a privacy-enhancing option for synthetic data generation with potential applicability to larger, real-world datasets.
Abstract
Generative Adversarial Networks (GAN) are among the widely used Generative models in various applications. However, the original GAN architecture may memorize the distribution of the training data and, therefore, poses a threat to Membership Inference Attacks. In this work, we propose a new GAN framework that consists of Multiple Generators and Multiple Discriminators (MGMD-GAN). Disjoint partitions of the training data are used to train this model and it learns the mixture distribution of all the training data partitions. In this way, our proposed model reduces the generalization gap which makes our MGMD-GAN less vulnerable to Membership Inference Attacks. We provide an experimental analysis of our model and also a comparison with other GAN frameworks.