MAAD Private: Multi-Attribute Adversarial Debiasing with Differential Privacy

Niloy Purkait, Emmanuel Keuleers, Henry Brighton

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: differential privacy, fair classification, adversarial learning

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

TL;DR: Adversarial framework for debiasing classifiers in senarios with multiple sensitive attributes, trained under differential privacy .

Abstract:

Balancing the trade-offs between algorithmic fairness, individual privacy, and
model utility, is pivotal for the advancement of ethical artificial intelligence. In
this work, we explore fair classification through the lens of differential privacy.
We present an enhancement to the adversarial debiasing approach, enabling it
to account for multiple sensitive attributes while upholding a privacy-conscious
learning paradigm. Empirical results from two tabular datasets and a natural language
dataset demonstrate our model’s ability to concurrently debias up to four
sensitive attributes and meet various fairness criteria, within the constraints of differential
privacy.

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: 7256