Differential Privacy and Statistical Inference under Differential Privacy

Course Content
Access to data is key for scientific progress. Yet, data is often sensitive in nature and must, therefore, fulfil certain requirements to ensure privacy. For instance, publicly available data should be arranged such that nobody can trace back the individuals in the dataset. This is exactly what Differential Privacy is concerned with.

Differential privacy comes with two main challenges: an algorithmic one and a statistical one. The algorithmic challenge boils down to the following question: how to ‘inject noise’ into the data in such a way that identification of individuals can be avoided? As the injection of noise perturbs estimates, inferences, and predictions from the data, this leads to the statistical challenge: how to properly account for this noise using statistical methods?

A well-known example for differential privacy is the U.S. Census 2020, for which the U.S. Census Bureau applied differential privacy to mask information about individuals. The scientific importance of this topic is illustrated by recent articles published in journals such as Science Advances.

Teaching Methods
Lectures (4 hours a week) and tutorials (2 hours a week).

Type of Assessment
Written exam and assignments.