[Paper Review] RepFair-GAN: Mitigating Representation Bias in GANs Using Gradient Clipping

Patrik Joslin Kenfack, Kamil Sabbagh+2 more|arXiv (Cornell University)|Jul 13, 2022

Adversarial Robustness in Machine Learning4 citations

TL;DR

This paper proposes RepFair-GAN, a method to mitigate representation bias in GANs by applying group-wise gradient norm clipping during discriminator training. By controlling the magnitude of gradients per sensitive attribute group, the generator learns to sample uniformly from all groups at test time, achieving fairer data generation without sacrificing sample quality, as validated on MNIST and SVHN with both balanced and imbalanced group distributions.

ABSTRACT

Fairness has become an essential problem in many domains of Machine Learning (ML), such as classification, natural language processing, and Generative Adversarial Networks (GANs). In this research effort, we study the unfairness of GANs. We formally define a new fairness notion for generative models in terms of the distribution of generated samples sharing the same protected attributes (gender, race, etc.). The defined fairness notion (representational fairness) requires the distribution of the sensitive attributes at the test time to be uniform, and, in particular for GAN model, we show that this fairness notion is violated even when the dataset contains equally represented groups, i.e., the generator favors generating one group of samples over the others at the test time. In this work, we shed light on the source of this representation bias in GANs along with a straightforward method to overcome this problem. We first show on two widely used datasets (MNIST, SVHN) that when the norm of the gradient of one group is more important than the other during the discriminator's training, the generator favours sampling data from one group more than the other at test time. We then show that controlling the groups' gradient norm by performing group-wise gradient norm clipping in the discriminator during the training leads to a more fair data generation in terms of representational fairness compared to existing models while preserving the quality of generated samples.

Motivation & Objective

• To formalize a new fairness notion—representational fairness—for generative models, requiring uniform distribution of protected attributes (e.g., gender, race) in generated samples at test time.
• To investigate the root cause of representation bias in GANs, particularly why generators favor certain groups even when training data is balanced.
• To develop a simple, effective, and generalizable method to enforce fair generation in GANs without modifying the generator or requiring additional data.
• To evaluate the method’s effectiveness across different levels of group imbalance and across multiple datasets (MNIST, SVHN).
• To demonstrate that controlling gradient norms per group during discriminator training leads to more uniform and fair sampling at inference time.

Proposed method

• The method introduces group-wise gradient norm clipping during discriminator training, where gradients for each sensitive attribute group are clipped independently if they exceed a predefined maximum norm.
• The clipping is applied per batch by splitting the batch into groups based on protected attributes (e.g., digit class in MNIST, digit type in SVHN), then clipping each group’s gradient norm separately.
• The approach maintains the standard GAN training objective but modifies the discriminator’s backpropagation step to ensure balanced gradient updates across groups.
• The maximum gradient norm (C) is a hyperparameter tuned via validation; experiments show optimal performance at C=2 for MNIST and SVHN.
• The method is compatible with various GAN variants as it only modifies the discriminator’s gradient update step and does not require changes to the generator or data preprocessing.
• The approach is lightweight and efficient, requiring only minor modifications to standard GAN training pipelines.

Experimental results

Research questions

• RQ1 Why do GANs generate biased samples even when training data has balanced group representation?
• RQ2 How does disparity in gradient norms across sensitive attribute groups influence the generator’s sampling behavior at test time?
• RQ3 Can controlling group-specific gradient norms during discriminator training lead to more uniform and fair generation across groups?
• RQ4 How does the method perform under varying degrees of group imbalance in the training data?
• RQ5 Does the proposed method preserve the quality of generated samples while achieving fairness?

Key findings

• Vanilla GANs exhibit strong representation bias at test time, favoring well-represented groups even when training data is balanced, due to disparities in gradient norms during discriminator training.
• When the magnitude of gradients for one group is significantly larger than others during discriminator training, the generator learns to favor sampling from that group at test time.
• Group-wise gradient norm clipping at C=2 achieves uniform group representation in generated samples on both MNIST and SVHN, even when training data is imbalanced (e.g., 20% vs 80%).
• For maximum gradient norm values below 2 (e.g., 0.1, 0.5, 1), training fails to converge due to vanishing gradients, resulting in random, low-quality samples.
• For maximum gradient norms above 20, representation bias re-emerges, indicating that overly large clipping thresholds fail to constrain gradient disparities.
• RepFair-GAN maintains sample quality comparable to vanilla GANs while achieving significantly fairer group representation, as shown in qualitative and quantitative evaluations across datasets.