Adversarial Latent Feature Augmentation for Fairness

Hoin Jung, Junyi Chai, Xiaoqian Wang

Primary Area: societal considerations including fairness, safety, privacy

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Fairness, Data Augmentation, Adversarial Attack

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: As fairness in machine learning has been increasingly important to mitigate bias in models, various methods to enhance fairness have been proposed. Among them, the data augmentation approach has shown promising results in improving fairness. However, existing data augmentation methods on either input or latent features provide limited evidence of how they discover bias and rectify it. In this paper, we propose the Adversarial Latent Feature Augmentation (ALFA) for fairness, which effectively merges adversarial attacks against fairness and data augmentation in the latent space to promote fairness. Though the adversarial perturbation against fairness has been discussed in existing literature, the effect of such adversarial perturbations has been inadequately studied only as a means to depreciate fairness. In contrast, in this paper, we point out that such perturbation can in fact be used to augment fairness. Drawing from a covariance-based fairness constraint, our method unveils a counter-intuitive relationship between adversarial attacks against fairness and enhanced model fairness upon training with the resultant perturbed latent features by hyperplane rotation. We theoretically prove that our adversarial fairness objective assuredly generates biased feature perturbation, and we validate with extensive experiments that training with adversarial features significantly improve fairness.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7792