Privately Evaluating Untrusted Black-Box Functions

Explore a plenary talk on privately evaluating untrusted black-box functions in the context of differential privacy. Delve into tools for sharing sensitive data when data curators face unknown queries from untrusted analysts. Learn about privacy wrappers, algorithms that provide differentially private approximations of black-box functions on sensitive datasets. Examine two settings: automated sensitivity detection and provided sensitivity bounds. Discover the first privacy wrapper for automated sensitivity detection and improvements in accuracy and query complexity for the provided sensitivity bound setting. Investigate lower bounds for both scenarios and understand how these mechanisms contribute to the feasibility of differentially private release of general function classes.