Paper gives federated DP optimization linear communication cost
Original title: Federated Differentially Private Optimization via Low-Rank Preconditioning
The method uses DP rank selection, a DP power method, and a noisy low-rank covariance release under secure aggregation.
A cs.CR paper proposes federated differentially private PCA so a covariance estimate can precondition DP gradient descent for generalized linear models. It claims linear communication and central DP under secure aggregation, unlike quadratic central methods or local-DP baselines that worsen as clients grow. The authors give error bounds and say empirical results match a centralized quadratic-communication baseline at lower communication.
- The estimate is a DP feature-covariance matrix used as a Hessian approximation.
- Steps are DP rank selection, a DP power method, and a noisy subspace release.
- Authors report error bounds for the covariance estimate and the resulting gradient descent.
From the original article
1 min read at arxiv.orgWe consider differentially private (DP) optimization in the federated setting, where a DP estimate of the feature covariance matrix serves as a Hessian approximation for generalized linear models.
By Antti Koskela · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.