Descriptif

Extracting value from the data without violating privacy guarantees of individuals is a challenging task, but it is crucial to do it right when dealing with highly sensitive data, such as, healthcare or financial records. We begin by defining what privacy is and looking at several traditional privacy protection techniques, such as anonymization. We will see that they fail to provide robust privacy guarantees against attackers who have access to extra knowledge, and then we will move on to the notion _differential privacy_ (DP) that addresses those shortcomings and now is a de-facto standard when it comes to releasing sensitive information. We will discuss theoretical underpinnings of DP, present a range of differentially-private mechanisms tailored at specific data analytics queries, their composition and the fundamental trade-off between accuracy and utility. During the lab sessions, we are going to explore several open-source frameworks (e.g., Google DP library and OpenDP) and learn how DP can be applied in practical scenarios while paying attention to the common pitfalls and the ways to avoid them. We will also discuss existing variants of DP, such as local DP, the question of privacy budget management, and touch upon differentially-private machine learning algorithms.