ICLR 2025Poster

Privately Counting Partially Ordered Data

作者: Matthew Joseph, Mónica Ribero, Alexander Yu

摘要

We consider differentially private counting when each data point consists of $d$ bits satisfying a partial order. Our main technical contribution is a problem-specific $K$-norm mechanism that runs in time $O(d^2)$. Experiments show that, depending on the partial order in question, our solution dominates existing pure differentially private mechanisms and can reduce their error by an order of magnitude or more.

PDFOpenReview