SlaClip: Gradient Norm Slacks can be Indicator for Adaptive Clipping in DP-SGD (ICML 2026 Spotlight)

This work introduces SlaClip, a novel method that helps differentially private machine learning models choose a better gradient clipping threshold automatically during training. It reuses “slack” information already produced by standard Differentially Private SGD process, so it requires no extra private queries and keeps the same privacy cost as vanilla DP-SGD. Experiments show that SlaClip improves performance over leading DP-SGD methods, including strong baselines from Google and Amazon.