Differentially Private Synthetic Data Can Be Accessible and Equitable

Abstract: 

Differentially Private (DP) mechanisms for synthetic data have been deployed in a variety of high-impact social settings (perhaps most notably in data release for the 2020 U.S. Census). However, assessing the impact of DP mechanisms on downstream data utility remains a major challenge, especially when it is known that these mechanisms have disparate impacts on protected sub-populations in the data. This talk will focus on these challenges. I will outline an approach to practitioner aligned DP synthetic data benchmarking, and discuss insights and simple but effective methods for mitigating adverse impacts of DP mechanisms on protected sub-populations.