Cynthia Dwork, Harvard University

Abstract

Modern cryptography has gifted us with an approach to creating digital manifestations of social phenomena in the physical world. As algorithms increasingly supplant human decision making and shape our view of the world, and as computers are increasingly leveraged to attack data analytic processes, the need for cryptographic thinking grows apace. In short, we need crypto on steroids.

This talk begins with differential privacy, a definition of privacy tailored to privacy-preserving analysis of large datasets together with a collection of techniques for achieving this goal and complementary lower bounds. We will see that the concept also protects against "p-hacking" (adaptive data analysis), allowing the construction of a reusable holdout set. We explore the cryptographic approach to algorithmic fairness, and see two roles for differential privacy in this endeavor. Finally, and time permitting, we touch on a tight connection to a problem in quantum computing.

Slides   Video