Amélioration de l’apprentissage automatique différentiellement confidentiel : Optimisations pour les scénarios de requêtes répétées

Deep neural networks and other machine learning models have experiencedunprecedented growth in recent years. Alongside this enthusiasm, there hasbeen an increasing and well-founded concern about the privacy of the vastamounts of data required to train these models. The combination of thesetwo factors has been a key driver of interest in privacy-preservingmachine learning techniques. Differential Privacy has emerged as the goldstandard for measuring privacy. This framework is now applied on awide range of data-driven tasks, such as machine learning and collaborativeanalysis, where multiple stakeholders wish to query shared data withoutexposing their own. The main challenge in this domain lies in balancingprivacy guarantees with the utility of the results. Indeed,privacy-preserving techniques often come at the cost of reduced utility.This thesis focuses on techniques to improve machine learning models andtools for analyzing them, while ensuring a satisfactory level ofprivacy for the underlying data. First, it introduces an innovativeapproach to privacy-preserving gradient descent methods by addressingthe bias introduced by existing methods. By leveraging properties ofgradient regularity rather than clipping the gradient, as it iscommonly done in popular methods, our approach effectively reducesbias and the noise added to the gradient. We propose a new algorithmthat surpasses the state of the art across various datasets.Second, the thesis explores techniques for computing privacy-preservingempirical cummulative distribution functions, even in cases where thedata is distributed across multiple entities. This study proposes anovel method compatible with different security protocols,offering provable privacy guarantees and an analysis of computational costs.range of applications are explored, and experimental results arepresented to validate the utility of these methods.By analyzing optimization mechanisms and distribution functions,this thesis contributes to the development of more practical andefficient privacy-preserving machine learning and data analysis techniques.