Reproduce “DP-EM: Differentially Private Expectation Maximization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mijung Park, Jimmy Foulds, Kamalika Chaudhuri, Max Welling
DP-EM privatizes expected sufficient statistics used in expectation-maximization updates, adding calibrated noise before mapping them to model parameters. Its privacy analysis composes repeated updates using the moments accountant or zero-concentrated differential privacy, for exponential-family models under boundedness assumptions. In Gowalla k-means, DP-EM with zCDP achieved NICV 0.0101, versus 0.0236 for DPLloyd with linear composition; this is a clustering result, not a mixture-model likelihood result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.