Fairness without Sensitive attributes via Noise and Uncertain Predictions

Hongliang Ni, LEI HAN, Tong Chen, Shazia Wasim Sadiq, Gianluca Demartini

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, Fairness without Sensitive Attributes, Fairness without Demographics

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

Abstract: While model fairness improvement has been explored previously, existing methods invariably rely on adjusting explicit sensitive attribute values in order to improve model fairness in downstream tasks. However, we observe the trend of sensitive demographic information being inaccessible as public concerns around data privacy grow. In this paper, we propose a confidence-based hierarchical structure of variational autoencoder (VAE) architectures called ``Reckoner" for reliable fairness learning under the assumption of missing sensitive attributes. First, we present the results of exploratory data analyses conducted on the widely-used COMPAS dataset. We observed significant disparities in model fairness across different levels of confidence. Inspired by these findings, we devised a dual-model system in which the model initialised with a high-confidence data subset learns from the model initialised with a low-confidence data subset, enabling it to avoid biased predictions. To maintain predictiveness, we also introduced learnable noise into the dataset, forcing the data to retain only the most essential information for predictions. Our experimental results show that Reckoner consistently outperforms state-of-the-art baselines on both the COMPAS and the New Adult datasets in terms of both accuracy and fairness metrics.

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: 4806