|
Activity Number:
|
483
- Privacy and Work Force
|
|
Type:
|
Contributed
|
|
Date/Time:
|
Wednesday, August 10, 2022 : 2:00 PM to 3:50 PM
|
|
Sponsor:
|
Government Statistics Section
|
|
Abstract #320909
|
|
|
Title:
|
A New Bound for Privacy Loss of Bayesian Posterior Sampling
|
|
Author(s):
|
Xingyuan Zhao* and Fang Liu
|
|
Companies:
|
University of Notre Dame and Univerisity of Notre Dame
|
|
Keywords:
|
Bayesian;
data synthesis;
inherent privacy guarantees;
posterior sampling;
privacy for free;
privacy loss
|
|
Abstract:
|
Differential privacy (DP) is a state-of-the-art concept that formalizes privacy guarantees. We investigate the inherent privacy loss in releasing Bayesian posterior samples and derive a new bound for the privacy loss in the setting of DP. We show it is tighter than the existing bounds. We formulate the closed-form bounds in some common Bayesian models. Besides, the new bound is consistent with the likelihood principle. We also apply the derived inherent privacy guarantees to releasing synthetic data from Bayesian models. We run several experiments to demonstrate the improved utility of synthetic data obtained through the inherently private posterior samples compared to synthetic data generated by privatizing posterior distributions via explicitly designed DP mechanisms.
|
Authors who are presenting talks have a * after their name.
