The goal of differentially private data analysis is to design algorithms for analyzing datasets while ensuring that sensitive information about individuals is not revealed. In this talk I will present both new lower bounds and new algorithms for differentially private data analysis. On the negative side, I will present some new, nearly-optimal lower bounds on the amount of data required to release differentially private statistics on high-dimensional datasets. These results show that there is a significant "price of differential privacy" in high-dimensional datasets. We prove these lower bounds using a cryptographic primitive called a fingerprinting code that we show is closely connected to differentially private data analysis. On the positive side, I will present efficient algorithms for computing differentially private contingency tables, using techniques from computational learning theory.
Jon Ullman is a postdoctoral fellow at the Center for Research on Computation and Society at Harvard University. He recently completed his Ph.D., also at Harvard, where he was advised by Salil Vadhan. He is interested in the foundations of data privacy and its connections to areas like cryptography, learning theory, and game theory.