An Introduction to Differential Privacy

In this talk, we will give an accessible introduction to the ideas and techniques underlying "Differential Privacy". Differential privacy has emerged in the last decade as a strong, semantically meaningful, mathematically formal definition of privacy in the context of data analysis. Despite its strong guarantees, it turns out that it is possible to carry out a vast collection of tasks to a high degree of accuracy, ranging from machine learning to combinatorial optimization, all while satisfying this constraint. We will explain why differential privacy is semantically meaningful, the basic building blocks that go into private algorithm design, and the compositional properties that allow the design of sophisticated algorithms from these basic pieces. No background will be assumed, besides basic familiarity with mathematical notation.