Mitigating Disparate Impact of Differentially Private Learning through Bounded Adaptive Clipping
Researchers have developed a new method called bounded adaptive clipping to address disparate impacts in differentially private machine learning. Standard adaptive clipping can disproportionately suppress gradients from minority groups, leading to reduced accuracy for these populations. The proposed technique introduces a tunable lower bound to prevent excessive gradient suppression, improving worst-class accuracy by up to 10 percentage points on benchmark datasets. AI
IMPACT Addresses fairness concerns in private ML, potentially enabling wider adoption of DP techniques.