Some databases may contain confidential information. To respect the privacy of its elements, data must be handled carefully. For example, it is not sufficient to remove personal information from a database to guarantee privacy when releasing it. These considerations also apply to artificial intelligence. As artificial intelligence advances, the need for data continues to grow. Machine learning algorithms depend on the data that was used during their training and it has been demonstrated that a machine learning algorithm can leak private information if used without privacy preserving mechanisms. Synthetic data comes as a simple and flexible solution to this problem. However, synthetic data can also be the source of privacy breaches. Differential privacy has emerged as the mathematical framework of reference for the protection of randomized algorithms. We will focus on differential privacy in the context of synthetic data. We will look for efficient synthetic data generation algorithms and we will specifically focus on the protection of correlated data. We will also study the theoretical and practical limits of privacy that can be achieved through private data generation.
- M. Marc TOMMASI Professeur des universités Université de Lille Directeur de thèse, - M. Aurélien BELLET Directeur de recherche Antenne Inria de l'Université de Montpellier Co-directeur de thèse, - M. Olivier CAPPé Directeur de recherche CNRS - Département d'informatique de l'ENS Rapporteur, - Mme Catuscia PALAMIDESSI Directrice de recherche Inria Saclay - LIX, École polytechnique Rapporteure, - M. Liva RALAIVOLA Professeur des universités Université d'Aix-Marseille et Criteo AI Lab Examinateur, - M. Bogdan KULYNYCH Docteur Lausanne University Hospital (CHUV) Examinateur, - M. Matthieu BOUSSARD Craft AI Invité, - M. Peter KAIROUZ Google Invité.
Thesis of the team MAGNET defended on 25/11/2026