FairDP: Certified Fairness with Differential Privacy

2023-05-25

AI Summary

This paper addresses the challenge of simultaneously ensuring group fairness and differential privacy (DP) in machine learning. We propose FairDP, a training framework that achieves provably certified group fairness under ε-differential privacy via group-wise independent training, group-level gradient clipping, and Gaussian noise injection. Crucially, we innovatively leverage DP noise to quantify fairness statistics and derive theoretical fairness bounds—establishing, for the first time, confidence bounds on fairness grounded in DP mechanisms—while guaranteeing equitable contribution from all demographic groups. To mitigate utility degradation, FairDP incorporates progressive knowledge distillation. Extensive experiments on multiple benchmark datasets demonstrate that FairDP improves group fairness by over 65% on average while incurring less than 4% utility loss, significantly outperforming existing baselines that jointly optimize privacy and fairness.

Abstract

This paper introduces FairDP, a novel training mechanism designed to provide group fairness certification for the trained model's decisions, along with a differential privacy (DP) guarantee to protect training data. The key idea of FairDP is to train models for distinct individual groups independently, add noise to each group's gradient for data privacy protection, and progressively integrate knowledge from group models to formulate a comprehensive model that balances privacy, utility, and fairness in downstream tasks. By doing so, FairDP ensures equal contribution from each group while gaining control over the amount of DP-preserving noise added to each group's contribution. To provide fairness certification, FairDP leverages the DP-preserving noise to statistically quantify and bound fairness metrics. An extensive theoretical and empirical analysis using benchmark datasets validates the efficacy of FairDP and improved trade-offs between model utility, privacy, and fairness compared with existing methods. Our empirical results indicate that FairDP can improve fairness metrics by more than 65% on average while attaining marginal utility drop (less than 4% on average) under a rigorous DP-preservation across benchmark datasets compared with existing baselines.

Problem

• Ensures group fairness in model decisions
• Guarantees differential privacy for training data
• Balances privacy, utility, and fairness effectively

Innovation

• Group-specific model training
• DP-preserving noise integration
• Fairness certification with DP