Adaptive sampling and clipping for differentially private worst-group optimization

The researchers propose a differentially private worst-group optimization algorithm, ASC; experiments show its worst-group accuracy is significantly higher than prior methods, without sacrificing overall average accuracy. ASC adaptively adjusts the sampling rate and clipping threshold of each group's gradient contribution, reweights the training objective toward harder-to-learn groups, and keeps the noise required for privacy low enough to maintain model utility. Prior fairness-oriented worst-group optimization and differentially private SGD often conflict, and there is a lack of practical methods that address both.