DISSEC-ML ˸ vers un apprentissage automatique distribué et sécurisé dans le cloud personnel
Julien Mirval · theses.fr (ABES) · 2024
Initiatives such as the Blue/Green Button, MesInfos, MiData and new regulations such as the GDPR at the European level aim to enable individuals to retrieve their personal data from the companies or organizations which collected it. At the same time, platforms called Personal Data Management Systems (PDMS), Personal Information Management Systems (PIMS) or Personal Clouds are developing rapidly and allow users to store all their digital assets: data directly generated by their devices (e.g., connected objects, home automation, photos) and data from their interactions (e.g., preferences, social data, health, banking). Users can then use their PDMS for personal applications or for the benefit of the community. Thus, the PDMS paradigm promises to pave the way for new and innovative usage developed around personal data, including distributed computations across a large number of PDMS (e.g., automatic classification, recommendations, participatory studies). Such examples often require the formation of an artificial intelligence (AI) model based on a large volume of user data. However, this approach also raises significant privacy protection and performance challenges. The organization of secure and efficient distributed computing across a large number of PDMS can be complex, especially in the presence of a potentially large number of corrupted nodes. This CIFRE thesis is realized with the Cozy Cloud company which provides a Personal Cloud solution named Cozy. The objective is to make an in-depth study of this new and crucial problem and to propose appropriate solutions to effectively train an AI model (e.g., a deep neural network) in a fully distributed system while providing strong security guarantees to the participating nodes. The results, in the form of protocols and distributed, secure and reliable execution algorithms, will be applied to practical cases provided by the Cozy Cloud company, which offers a PDMS-type solution.