Thesis of Rémy Raes

On the constraints of mobile computing in cloud avoidance

Mobile devices are producing a deluge of data by leveraging a wide variety of embedded or connected sensors that capture the surrounding environment of end-users and their routines. However, this continuous data stream inevitably includes sensitive information that may jeopardize the privacy of end-users if processed by malicious stakeholders. While machine learning algorithms are nowadays widely adopted as a convenient keystone to process large datasets and infer actionable insights, they often require grouping the raw input data into a single place, thus imposing a privacy threat for end-users sharing their data. To address this ethical challenge, privacy-preserving machine learning and decentralized machine learning (DML) are revisiting state-of-the-art machine learning algorithms to enforce user privacy, among other properties. While most of the research contributions that have been achieved in this area remain at the theoretical level, the implementation and the deployment of these privacy-preserving algorithms in the field remain a challenging issue for most modern applications. Among difficulties, the limited resources of mobile devices and their partial device-to-device (D2D) connectivity makes it challenging to adopt DML algorithms for the masses. Nonetheless, we claim that leveraging fleets of mobile devices may provide promising opportunities for DML algorithms in the context of mobile crowdsourcing software systems. In particular, we believe that unsupervised machine learning algorithms (such as clustering), and federated learning algorithms, can benefit from nearby devices to reason upon a reduced set of relevant samples, captured as models by nearby devices sharing similar concerns and objectives. As promoted by DML approaches, the aggregation of such in situ models can effectively contribute to delivering personalized results (e.g. recommendations) to end-users without exposing their privacy.

Jury

M. Romain ROUVOY Professeur des universités Université de Lille Directeur de thèse, Mme Hélène COULLON Maîtresse de conférences IMT Atlantique Rapporteure, M. Antoine BOUTET Maître de conférences INSA Lyon Rapporteur, Mme Sonia BEN MOKHTAR Directrice de recherche CNRS Examinatrice, M. Stéphane DELBRUEL Maître de conférences Université de Bordeaux Examinateur, M. François TAÏANI Professeur des universités IRISA/INRIA Co-directeur de thèse, M. Davide FREY IRISA / Inria Invité, M. Adrien LUXEY-BITRI Université de Lille Invité.

Thesis of the team Spirals defended on 06/07/2026