FairPATE: Exposing the Pareto Frontier of Fairness, Privacy, Accuracy, and Coverage

Mohammad Yaghini, Patty Liu, Franziska Boenisch, Nicolas Papernot

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, privacy, pate, pareto frontier

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

Abstract: Deploying machine learning (ML) models often requires both fairness and privacy guarantees. In this work, we study the challenges of integrating group fairness interventions into the Private Aggregation of Teacher Ensemble (PATE) framework. We show that in the joint fairness-privacy setting, the placement of the fairness intervention before, or after PATE’s noisy aggregation mechanism (which ensures its differential privacy guarantees) leads to excessive fairness violations, or inefficient privacy budgeting, respectively. With this in mind, we present FairPATE which adds a rejection mechanism due to fairness violations. Through careful adjustment of PATE’s privacy accounting, we match the DP-SGD-based state-of-the-art privacy-fairness-accuracy trade-offs (Lowy et al., 2023) in demographic parity, and improve on them for equality of odds with 2% lower disparity at similar accuracy levels and privacy budgets. We also evaluate FairPATE in the setting where exact fairness guarantees need to be enforced by refusing to provide algorithmic decisions at inference-time (for instance, in a human-in-the-loop setting) thus trading off fairness with coverage. Based on our FairPATE, we provide, for the first time, empirical Pareto frontiers for fairness, privacy, accuracy, and coverage on a range of privacy and fairness benchmark datasets.

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: 9455