Distributed GANs (DGANs)

A distributed GAN (DGAN) trains a generative adversarial network when the real data is spread across several sites that can't share it — hospitals with patient scans, for example.

Common designs: each site keeps a local discriminator that sees its private data, while a central generator learns from all the discriminators' feedback (as in MD-GAN). Or the generator and discriminator are both trained federated-style, with parameters averaged across sites (FedGAN).

The result is a generator that captures the combined data distribution, which can then produce synthetic data for sharing, augmentation or analysis — without any site revealing its raw records.