Unlocking Trustworthy ML: Privacy and Fairness in Focus

Machine learning (ML) systems increasingly process sensitive data and inform critical decisions, raising serious concerns about privacy and fairness.

Driven by ethical considerations and regulations such as the AI Act and GDPR, there is an urgent need for high-utility ML systems that guarantee both.

Yet, the development of such systems is hindered by the current lack of theoretical knowledge about the interplay between privacy and fairness.

The proposed project addresses this knowledge gap by studying the fundamental relationships between privacy and fairness.

A key novelty in the project is the use of pointwise maximal leakage, a highly robust and flexible privacy measure introduced recently.The research has three specific aims: investigating how fairness can be achieved using privatized data, identifying synergies between privacy and fairness, and exploring the fundamental limits of utility, privacy, and fairness.

The project is primarily theoretical, situated at the intersection of computer science and information theory, and will be supplemented by empirical experiments using open-source datasets. It will be completed over 24 months at Inria and KTH.The significance of the project lies in its impacts.

In the short term, the project will enable the development of high-utility ML algorithms that synergistically address privacy and fairness.

Long term, the project will advance individuals’ digital rights by contributing to the development of human-centric ML systems.