Online Program Home
My Program

Abstract Details

Activity Number: 458 - Differential Privacy Research and Applications at the U.S. Census Bureau
Type: Topic Contributed
Date/Time: Wednesday, July 31, 2019 : 8:30 AM to 10:20 AM
Sponsor: Government Statistics Section
Abstract #304724 Presentation
Title: Estimating the Variance of Complex Differentially Private Algorithms
Author(s): Robert Ashmead*
Companies: Ohio Colleges of Medicine Government Resource Center
Keywords: Privacy; Disclosure Avoidance; Differential Privacy; Variance Estimation

Simple differentially private mechanisms such as the Laplace or geometric, and even more complicated methods such as the matrix mechanism have closed-formed expressions for the variance of the output. These expressions are not affected by changing the input dataset. However, closed-form expressions for more complex differentially private algorithms such as the 2020 Decennial Census Disclosure Avoidance System (DAS) that utilize post-processing techniques such as non-negativity and integer solutions are not available and depend on their input dataset. Spending additional privacy-loss budget, one could estimate the error of a particular realization of the mechanism or use Monte Carlo simulations repeatedly applying the mechanism to the input data to estimate the variance. However, either option is costly in terms of the privacy-loss budget. Instead, we propose a bootstrap-inspired method using the single production run of the algorithm to estimate the variance without spending additional privacy-loss budget. Through simulations utilizing the 1940 publicly available Census data, we illustrate the utility of this methodology.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2019 program