Mitigating Disparate Impact of Differentially Private Learning through Bounded Adaptive Clipping

2025-06-02

AI Summary

In differentially private (DP) learning, adaptive gradient clipping often excessively suppresses gradients from minority-group samples, exacerbating performance disparities across groups. To address this, we propose Bounded Adaptive Clipping (BAC), the first adaptive clipping mechanism incorporating a tunable lower bound on the clipping threshold. This constraint prevents threshold collapse during training, thereby preserving gradient information from underrepresented groups while maintaining ε-differential privacy. By explicitly integrating group fairness into the DP optimization objective, BAC mitigates the degradation of worst-group accuracy. Experiments on skewed MNIST and Fashion MNIST demonstrate that BAC improves worst-class accuracy by over 10 percentage points compared to unbounded adaptive clipping, and by over 5 percentage points relative to fixed clipping. These results show substantial gains in group fairness for DP models without compromising privacy guarantees.

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: Privacy-enhancing technologies
• Economics, Online Markets and Human Computation: Fairness, privacy, and diversity in economic environments

Abstract

Differential privacy (DP) has become an essential framework for privacy-preserving machine learning. Existing DP learning methods, however, often have disparate impacts on model predictions, e.g., for minority groups. Gradient clipping, which is often used in DP learning, can suppress larger gradients from challenging samples. We show that this problem is amplified by adaptive clipping, which will often shrink the clipping bound to tiny values to match a well-fitting majority, while significantly reducing the accuracy for others. We propose bounded adaptive clipping, which introduces a tunable lower bound to prevent excessive gradient suppression. Our method improves the accuracy of the worst-performing class on average over 10 percentage points on skewed MNIST and Fashion MNIST compared to the unbounded adaptive clipping, and over 5 percentage points over constant clipping.

Problem

• Mitigating disparate impact in differentially private learning
• Addressing accuracy loss for minority groups in DP models
• Preventing excessive gradient suppression via bounded clipping

Innovation

• Bounded adaptive clipping prevents excessive gradient suppression
• Tunable lower bound enhances accuracy for minority groups
• Improves worst-performing class accuracy significantly