While artificial intelligence largely relies on centralized infrastructures, growing concerns about data privacy, digital sovereignty, resilience, and scalability call for rethinking collaboration among distributed entities. Peer-to-peer architectures offer attractive robustness properties and avoid single points of failure, yet harnessing them efficiently for federated learning remains a major challenge.
This thesis addresses this problem through a layered architecture that bridges networking and decentralized learning. Its contributions unfold into four protocols. The peer-sampling protocols Elevator and Lift let a structured multi-hub topology emerge in a fully decentralized manner — a self-organized hub election for Elevator, made robust against Byzantine behaviour by Lift. HEAL then exploits this hierarchical structure to deliver decentralized learning whose convergence speed rivals federated learning while remaining resilient to crashes and churn. Finally, FLAIR adapts these principles to wireless sensor networks through a resource-aware election of cluster heads.
This work combines experimentation and theoretical analysis. The proposed protocols are evaluated at large scale through simulation (PeerSim, ns-3) and, for Elevator, through a real implementation over TCP/IP; in addition, a formal analysis establishes the convergence of Elevator toward its multi-hub topology.