PFGuard: A Generative Framework with Privacy and Fairness Safeguards

2024-10-03

arXiv.org

AI Summary

Generative models in trustworthy AI must simultaneously ensure privacy and fairness, yet existing differential privacy (DP) and fairness techniques exhibit inherent conflicts: DP noise may exacerbate biases against minority groups, while fairness constraints can undermine privacy guarantees. Method: We propose the first unified framework jointly optimizing privacy, fairness, and generative utility. Our key innovation is a privacy–fairness conflict mitigation mechanism—leveraging multi-teacher model ensembling to dynamically decouple and co-optimize these objectives during fair training and DP-constrained generation. Results: Experiments on high-dimensional data demonstrate that our method produces synthetic data satisfying strict ε-DP guarantees, achieves convergence in group fairness (e.g., 32% improvement in statistical parity), and attains high fidelity (21% reduction in Fréchet Inception Distance), significantly outperforming single-objective baselines.

Technology Category

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

Generative models must ensure both privacy and fairness for Trustworthy AI. While these goals have been pursued separately, recent studies propose to combine existing privacy and fairness techniques to achieve both goals. However, naively combining these techniques can be insufficient due to privacy-fairness conflicts, where a sample in a minority group may be represented in ways that support fairness, only to be suppressed for privacy. We demonstrate how these conflicts lead to adverse effects, such as privacy violations and unexpected fairness-utility tradeoffs. To mitigate these risks, we propose PFGuard, a generative framework with privacy and fairness safeguards, which simultaneously addresses privacy, fairness, and utility. By using an ensemble of multiple teacher models, PFGuard balances privacy-fairness conflicts between fair and private training stages and achieves high utility based on ensemble learning. Extensive experiments show that PFGuard successfully generates synthetic data on high-dimensional data while providing both DP guarantees and convergence in fair generative modeling.

Problem

• Addresses privacy-fairness conflicts in generative models
• Ensures privacy, fairness, and utility simultaneously
• Generates synthetic data with DP guarantees and fairness

Innovation

• Ensemble of teacher models balances privacy-fairness conflicts
• Simultaneously addresses privacy, fairness, and utility
• Generates synthetic data with DP guarantees