Mutual Information Optimally Local Private Discrete Distribution Estimation
read the original abstract
Consider statistical learning (e.g. discrete distribution estimation) with local $\epsilon$-differential privacy, which preserves each data provider's privacy locally, we aim to optimize statistical data utility under the privacy constraints. Specifically, we study maximizing mutual information between a provider's data and its private view, and give the exact mutual information bound along with an attainable mechanism: $k$-subset mechanism as results. The mutual information optimal mechanism randomly outputs a size $k$ subset of the original data domain with delicate probability assignment, where $k$ varies with the privacy level $\epsilon$ and the data domain size $d$. After analysing the limitations of existing local private mechanisms from mutual information perspective, we propose an efficient implementation of the $k$-subset mechanism for discrete distribution estimation, and show its optimality guarantees over existing approaches.
This paper has not been read by Pith yet.
Forward citations
Cited by 3 Pith papers
-
Optimal quantum locally differentially private mechanisms in the high-privacy regime
Optimal QLDP mechanisms achieve the same asymptotic Q/C ratio as classical LDP for Holevo information and hypothesis-testing error exponents, with Q/C >= 3/2 when protecting n-ary data for n >= 3.
-
Beyond Epsilon: A Principled QIF Framework for Local Differential Privacy
QIF framework using Blackwell ordering classifies seven LDP frequency estimation protocols and shows some previously optimal ones are incomparable or strictly dominated.
-
Multi-tier Differential Private Query Release
A framework for multi-tier differential private query release that bounds cumulative privacy loss to the maximum budget while achieving optimal utility comparable to single-tier mechanisms.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.