|
Abstract:
|
In 2018-2019, the National Institute for Standards and Technology (NIST) organized the differentially-private synthetic data challenge. Differential privacy is seeing increased adoption among companies and government organizations due to the rigorous privacy guarantees it provides to individuals. Synthetic data is appealing because it looks just like the true data, so statistical analysis designed for the true data can also be used on the synthetic data. For these reasons, differentially-private synthetic data was identified as an important open problem for the privacy community. The competition encouraged competitors to think hard about the problem and design novel and practical solutions. In this presentation, I will talk about the challenges of designing high-utility privacy mechanisms for synthetic data, and the key ideas behind the winning solution of the competition. I will also talk about the inherent limitations of differentially-private synthetic data, and how these limitations may impact statistical workflows. I will conclude with open problems for the research community.
|
