Adaptive Sampling for Private Worst-Case Group Optimization

2026-02-11

AI Summary

This work addresses the limitations of existing worst-group optimization methods under differential privacy constraints, which suffer from inadequate privacy guarantees and poor performance for minority groups due to fixed weighting schemes. To overcome this, we propose the Adaptive Sampling and Clipping (ASC) algorithm, which jointly and adaptively tunes group-specific sampling rates and gradient clipping thresholds. ASC ensures a uniform differential privacy bound across all groups while significantly improving worst-group accuracy. By effectively reducing gradient variance, ASC yields tighter privacy bounds and achieves a favorable trade-off among fairness, privacy uniformity, and model utility—without compromising overall average performance.

Technology Category

• Machine Learning: Privacy
• Search and Optimization: Sampling/Simulation-based Search
• Computer Vision: Bias, Fairness & Privacy

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Responsible Web: Algorithmic accountability and transparency on the web
• Security and Privacy: Privacy-enhancing technologies

Abstract

Models trained by minimizing the average loss often fail to be accurate on small or hard-to-learn groups of the data. Various methods address this issue by optimizing a weighted objective that focuses on the worst-performing groups. However, this approach becomes problematic when learning with differential privacy, as unequal data weighting can result in inhomogeneous privacy guarantees, in particular weaker privacy for minority groups. In this work, we introduce a new algorithm for differentially private worst-case group optimization called ASC (Adaptively Sampled and Clipped Worst-case Group Optimization). It adaptively controls both the sampling rate and the clipping threshold of each group. Thereby, it allows for harder-to-learn groups to be sampled more often while ensuring consistent privacy guarantees across all groups. Comparing ASC to prior work, we show that it results in lower-variance gradients, tighter privacy guarantees, and substantially higher worst-case group accuracy without sacrificing overall average accuracy.

Problem

• differential privacy
• worst-case group optimization
• privacy guarantees
• group fairness
• adaptive sampling

Innovation

• adaptive sampling
• differential privacy
• worst-case group optimization
• gradient clipping
• fairness