Do We Really Achieve Fairness with Explicit Sensitive Attributes?

Xiaotian Han, Zhimeng Jiang, Ninghao Liu, Na Zou, Qifan Wang, Xia Hu

Keywords: fairness, debias, demographic parity

Abstract: Recently the wide usage of machine learning models for high-stake decision-making raises the concerns about the fairness and discrimination issue. Existing works found that sensitive information of a sample could be leaked completely by sensitive attributes or partially by non-sensitive attributes, thus removing the sensitive attributes directly from the original features can not achieve fairness. The current fairness practice is to leverage the explicit sensitive attributes (i.e., as regularization) to debias the prediction, based on a strong assumption that non-sensitive attributes of all samples leak the sensitive information totally. However, we investigate the distribution of leaked sensitive information from non-sensitive attributes and make interesting findings that 1) the sensitive information distinctly varies across different samples. 2) the violation of demographic parity for samples prone to leak sensitive information (high-sensitive) are worse than that for low-sensitive samples, indicating the failure of current demographic parity measurements. To this end, we propose a new group fairness ($\alpha$-Demographic Parity) to measure the demographic parity for samples with different levels of sensitive information leakage. Furthermore, we move one step forward and propose to achieve $\alpha$-demographic parity by encouraging the independence of the distribution of the sensitive information in non-sensitive attributes and that of downstream task prediction, which is formulated as a cross-task knowledge distillation framework. Specifically, the sensitive teacher models the distribution of the sensitive information and the fair student models the distribution of the downstream task prediction. Then we encourage the independence between them by minimizing the Hilbert-Schmidt Independence Criterion. Our model can naturally tackle the limited sensitive attribution scenario since the teacher models can be trained with partial samples with sensitive attributes. Extensive experiments show the superior performance of our proposed method on the $\alpha$-demographic parity and performs well on limited sensitive attribute scenarios.

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.

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

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)

TL;DR: We found that different sample leak different amount of sensitive information and has different-level violation of demographic parity, thus we propose a new metric and method to address this problem.