CS 798L: Differential Privacy in Machine Learning
Who can take this course
It is desirable but not mandatory that a student interested in this course have done CS203 (or, equivalent) or CS771 (or. equivalent). The instructor's consent will be needed for enrolling in this course.
Objectives
How can we extract insights from a dataset containing sensitive information while ensuring the privacy of the individuals it includes? This course addresses this question by examining the limitations of simple approaches and advancing to solutions involving differential privacy. The class will cover fundamental principles of differential privacy, delve into algorithms for attaining privacy, and explore applications in statistics and machine learning.
Course Content
The following topics will be covered:
- Attacks on statistical data privacy
- Introduction to differential privacy
- Definition: Pure, approximate, concentrated and Renyi differential privacy
- Properties of Differential Privacy: Post-processing, basic and advanced composition, privacy conversions
- Algorithms for achieving Differential Privacy: Laplace mechanism, Exponential mechanism, Gaussian mechanism. Sparse vector technique, Binary tree mechanism
- Differential Privacy in Statistics: Private mean estimation, Adaptive data analysis
- Differential Privacy in Machine Learning: Private empirical risk minimization, Private gradient descend and moment accountants, Private FTRL
- Other topics: Privacy in game theory and mechanism Design, Privacy in online learning and multi-armed bandits, Private Synthetic data generation
