Tight Group-Level DP Guarantees for DP-SGD with Sampling via Mixture of Gaussians Mechanisms
arXiv:2401.10294v23 citationsh-index: 1
Originality Synthesis-oriented
AI Analysis
This work provides incremental improvements in privacy analysis for machine learning practitioners by offering precise guarantees for DP-SGD.
The paper tackles the problem of computing tight group-level differential privacy guarantees for DP-SGD under various sampling methods, achieving results that are tight up to discretization errors when releasing all intermediate iterates.
We give a procedure for computing group-level $(ε, δ)$-DP guarantees for DP-SGD, when using Poisson sampling or fixed batch size sampling. Up to discretization errors in the implementation, the DP guarantees computed by this procedure are tight (assuming we release every intermediate iterate).