RepFair-GAN: Mitigating Representation Bias in GANs Using Gradient Clipping

Kamil Sabbagh, Patrik Joslin Kenfack, Adín Ramírez Rivera, Adil Khan

Keywords: Fairness, Representation bias, GANs

TL;DR: Introducing a new GAN training technique using gradient clipping to reduce representation bias.

Abstract: This work introduces a new notion of fairness, \textit{representational fairness}, for generative models, which ensures uniform representation of demographic groups in the generated data. Vanilla GANs violate this notion even when groups are equally represented. The proposed solution is to use \textit{group-wise} gradient norm clipping to control gradient groups’ magnitude during discriminator training. Experiments show that this method improves \textit{representational fairness} while maintaining sample quality.