NEBULA - Decentralized Federated Learning for Heterogeneous Networks
Enrique Tomás Martínez Beltrán, Gérôme Bovet, Gregorio Martínez Pérez, Alberto Huertas Celdrán · 2025
Federated learning (FL) enables collaborative model training without sharing raw data, which is pivotal for maintaining privacy. However, existing FL frameworks often rely on a central coordinator, posing risks in heterogeneous networks. This work presents NEBULA, a decentralized FL platform that unifies centralized and peer-to-peer FL paradigms, integrating network awareness and autonomous adaptation for improved resilience and efficiency. Key contributions include: (1) a unified architecture supporting both server-coordinated and fully decentralized operation; (2) network-aware orchestration for dynamic communication and aggregation optimization; and (3) built-in mechanisms for robust operation. The demonstration will showcase real-time performance, defense against adversarial attacks, and adaptive client participation in challenging network scenarios.