Special Topics in Differential Privacy, STAT598

In Fall 2022, Prof Jordan Awan will be teaching a special topics course STAT598 in Differential Privacy.

Course Title: STAT598: Differential Privacy

Informal prerequisites: STAT 517, STAT 519, familiarity with R

Course Summary:

Differential privacy (DP) has arisen as the state-of-the-art framework for formal privacy protection, and has been adopted by tech companies such as Apple, Google, Microsoft, as well as by the US Census. Differential privacy requires the introduction of randomness into a statistical procedure, to make it difficult to determine what the contribution of any particular individual was to the dataset. Several interesting questions in privacy are 1) how should we define privacy, and what properties should a privacy definition have, 2) how do we design algorithms for different tasks to satisfy a privacy guarantee, 3) once a randomized privacy-preserving statistic is produced, how do we incorporate this additional randomness to perform valid statistical inference. 

In this class, we will study differential privacy, and tackle each of the above questions. To start, we will consider what properties an algorithm must have to avoid being "blatantly non-private." Then we will propose differential privacy as a formal privacy-preserving framework. We will derive several fundamental properties of the DP framework, and develop several general-purpose DP mechanisms. We will also study extensions and modifications of differential privacy: local DP, approximate DP, concentrated DP, Renyi DP, and f-DP. Finally, we will consider techniques to perform valid statistical inference on privatized data.