GBMix: Enhancing Fairness by Group-Balanced Mixup

Mixup is a powerful data augmentation strategy that has been shown to improve the generalization and adversarial robustness of machine learning classifiers, particularly in computer vision applications. Despite its simplicity and effectiveness, the impact of Mixup on the fairness of a model has not been thoroughly investigated yet. In this paper, we demonstrate that Mixup can perpetuate or even exacerbate bias presented in the training set. We provide insight to understand the reasons behind this behavior and propose GBMix, a group-balanced Mixup strategy to train fair classifiers. It groups the dataset based on their attributes and balances the Mixup ratio between the groups. Through the reorganization and balance of Mixup among groups, GBMix effectively enhances both average and worst-case accuracy concurrently. We empirically show that GBMix effectively mitigates bias in the training set and reduces the performance gap between groups. This effect is observed across a range of datasets and networks, and GBMix outperforms all the state-of-the-art methods.