The authors analyze the product-distribution pairs used to bound the privacy of repeated, potentially adaptive mechanisms, including differentially private stochastic gradient descent.
Gaussian privacy bounds hold under milder conditions for subsampled algorithms
Feldman and Shenfeld bound composition using truncated privacy-loss means and variances, recovering the predicted scaling for subsampled Gaussian mechanisms at substantially lower noise levels.
Big Tech
Vitaly Feldman · Moshe Shenfeld
Apple · The Hebrew University of Jerusalem
Research Digest··2 min read
Vitaly Feldman and Moshe Shenfeld show that accumulated privacy loss can be approximated by a Gaussian when each step’s loss is small except on rare events.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.