Privacy Amplification for Correlated-Noise Mechanisms via b-Min-Sep Subsampling
Speaker at Google Research & TechTalks
A rigorous examination of privacy amplification techniques for correlated-noise mechanisms, introducing b-min-sep subsampling as a novel batching scheme for DP-MF. Explores advanced topics in private model training.
A Google TechTalk, 2026-02-18, presented by Andy Dong ABSTRACT: DP-SGD remains the standard approach for private model training. A variety of techniques have been developed to improve its privacy–utility tradeoff, including privacy amplification through data subsampling, leveraging structured randomness in the training process, and correlated-noise mechanisms such as DP matrix factorization (DP-MF). While DP-MF can improve utility, its interaction with subsampling-based amplification is less explored than in the classical DP-SGD setting.