Fairness and privacy in machine learning

Objectif

With the ubiquitous deployment of machine learning algorithms in nearly every area of our lives, the problem of unethical or discriminatory algorithm-based decisions becomes more and more prevalent. To partially address these concerns, new sub-fields of machine learning have emerged: fairness and privacy. The goal of the course is to introduce the audience to recent developments of fairness and differential privacy aware algorithms. The emphasise will be made on those methods which are supported by statistical guarantees and that can be implemented in practice. In the first part, we will study classification and regression problems under the so called demographic parity constraint—a popular way to define fairness of an algorithm. In the second part we will show classification and probability density estimation under differential privacy constraints.

Plan

Fairness: a) Introduction to concepts, definitions, discussion; b) Binary classification under demographic parity; c) Regression under independence constraints

Differential Privacy: a) Introduction, definitions; b) Estimation rates under global and local differential privacy; c) Binary classification and Density estimation under local differential privacy

Références

E. Chzhen, et al. (2019) Leveraging labeled and unlabeled data for consistent fair binary classification. NeurIPS 32

E. Chzhen and N. Schreuder (2022) A minimax framework for quantifying risk-fairness trade-off in regression. Ann. Statist. 

T.B. Berrett and C. Butucea (2020) Locally private non-asymptotic testing of discrete distributions is faster using interactive mechanisms. NeurIPS 34

C. Butucea, A. Dubois, M. Kroll and A. Saumard (2020) Local differential privacy: elbow effect in optimal density estimation and adaptation over Besov ellipsoids. Bernoulli, vol. 26, No 3, 1727-1764