Combination of fairness and privacy in training data

• Auteur / Autrice : Karima Makhlouf
• Direction : Catuscia Palamidessi
• Type : Projet de thèse
• Discipline(s) : Informatique, données, IA
• Date : Inscription en doctorat le 01/09/2021
• Etablissement(s) : Institut polytechnique de Paris
• Ecole(s) doctorale(s) : École doctorale de l'Institut polytechnique de Paris
• Partenaire(s) de recherche : Laboratoire : LIX - Laboratoire d'informatique

Mots clés

Mots clés libres

Utilité

Équité

Confidentialité

Causalité

Accuracy

Fairness

Privacy

Causality

Résumé

The objective of this thesis project is to investigate the combination of privacy, fairness, and accuracy in machine learning. We aim to develop (a) mechanisms for differential privacy that provide a good trade-off with accuracy, and (b) approximate notions of fairness that line up with the nature of differential privacy, and respect the correlation between decisions and legitimate features. We will then study how to process the training data so to guarantee that the resulting models provide fair decisions and offer a good compromise between prediction accuracy and robustness against leakage of sensitive data (pre-processing). We will also compare our approach with the other typical methods to ensure that models resulting from machine learning satisfy certain properties. Namely, in-processing (modifications of the learning algorithm), and post-processing methods.