Seminar on fairness and privacy
By Gayane Taturyan, Achraf Azize

Gayane Taturyan (S2A): A Comprehensive View of Fairness through Distributional Stability

We view fairness as a property of distributional stability. Rather than assessing a predictor under a fixed data distribution, we study how its predictions change under perturbations that modify the composition of protected groups. A predictor is fair if it remains stable under such shifts. Under this perspective, several classical notions of fairness arise as stability with respect to specific perturbations, with the associated unfairness gap given by a Lipschitz constant of a prediction-rate functional. This formulation also yields guarantees that hold uniformly over a range of demographic compositions at test time, without requiring knowledge of the deployment distribution. It leads to a learning procedure based on convex combinations of reweighted predictors, formulated as a second-order cone program, for which we establish generalization bounds. Experiments on standard benchmarks illustrate the approach. Joint work with Charlotte Laclau and Stephan Clémençon.

Achraf Azize (ENSAE / Inria) : Differential Privacy for Statistics and Machine Learning: Guarantees, Trade-offs, and Limits

This talk will give an introduction to Differential Privacy, a rigorous framework for learning from sensitive data while limiting the influence of any single individual on the released output. I will first explain why classical anonymisation can fail, introduce the definition of Differential Privacy, and discuss its main interpretations and consequences. The second part of the talk will focus on the central question of private statistics and machine learning: how can we design algorithms that remain accurate while satisfying a strong privacy guarantee? I will illustrate this privacy–utility trade-off through several tasks, ranging from the release of simple statistics such as empirical means to private regression, differentially private stochastic gradient descent, and online learning problems such as bandits. Finally, I will discuss the complementary role of privacy attacks and auditing, as well as current limitations of Differential Privacy, including the choice of privacy budget, the cost of added noise, and the gap between formal guarantees and practical deployment.2022, an honorary distinction recognizing outstanding PhD research in NLP. Outside academia, she enjoys skiing, hiking, and photography.