Topics in Theory for Society: Differential Privacy
COMPSCI 2260
Subject & Catalog Number
Course Information
Description
Differential Privacy is a mathematically rigorous definition of privacy that has become the de facto standard for statistical analysis of large datasets. Differential privacy provides a concrete measure of privacy loss, and differentially private algorithms are equipped with a parameter for controlling this loss. A signal property of differential privacy is closure under composition, meaning that we can understand and control the cumulative privacy loss as the data are subjected to multiple analyses. In consequence, differential privacy is programmable: one can combine simple differentially private computational primitives in creative ways to obtain privacy-preserving algorithms for complex analytical tasks. The course will cover (1) the basics of differential privacy: the definition and its properties, computational primitives, and composition theorems; (2) selected advanced differentially private algorithms drawn from the literature and a wide range of application areas from industry to the US Census; and (3) applications of differential privacy to validity and replicability of data analyses.
Course Notes
This course was previously numbered CS 226R.
Available for Harvard Cross Registration
