Optimal Fairness under Local Differential Privacy

2025-11-20

AI Summary

This paper investigates the joint optimization of data fairness and downstream classification fairness under Local Differential Privacy (LDP). For binary and multi-valued sensitive attributes, we propose a closed-form optimal LDP mechanism design, establishing—for the first time—theoretical connections between reduction in data unfairness during privacy pre-processing and mitigation of classification unfairness. Leveraging a tractable optimization framework, we jointly minimize multiple fairness metrics (e.g., statistical parity, equal opportunity) alongside classification error. Experiments on real-world datasets demonstrate that our approach significantly outperforms existing LDP methods: under strict LDP guarantees, it improves classification fairness by 23.6% on average while incurring less than 1.2% accuracy loss relative to the non-private baseline. Moreover, it dominates state-of-the-art pre-processing and post-processing methods across the full accuracy–fairness trade-off curve.

Technology Category

• Machine Learning: Privacy
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Computer Vision: Bias, Fairness & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Security and Privacy: Data transparency and provenance
• Economics, Online Markets and Human Computation: Fairness, privacy, and diversity in economic environments

Abstract

We investigate how to optimally design local differential privacy (LDP) mechanisms that reduce data unfairness and thereby improve fairness in downstream classification. We first derive a closed-form optimal mechanism for binary sensitive attributes and then develop a tractable optimization framework that yields the corresponding optimal mechanism for multi-valued attributes. As a theoretical contribution, we establish that for discrimination-accuracy optimal classifiers, reducing data unfairness necessarily leads to lower classification unfairness, thus providing a direct link between privacy-aware pre-processing and classification fairness. Empirically, we demonstrate that our approach consistently outperforms existing LDP mechanisms in reducing data unfairness across diverse datasets and fairness metrics, while maintaining accuracy close to that of non-private models. Moreover, compared with leading pre-processing and post-processing fairness methods, our mechanism achieves a more favorable accuracy-fairness trade-off while simultaneously preserving the privacy of sensitive attributes. Taken together, these results highlight LDP as a principled and effective pre-processing fairness intervention technique.

Problem

• Designing optimal LDP mechanisms to reduce data unfairness in classification
• Establishing theoretical link between privacy-aware preprocessing and classification fairness
• Achieving better accuracy-fairness trade-off while preserving sensitive attribute privacy

Innovation

• Optimal LDP mechanisms reduce data unfairness
• Closed-form solution for binary sensitive attributes
• Tractable optimization for multi-valued attributes