Tram-FL: Routing-based Model Training for Decentralized Federated Learning
Kota Maejima, Takayuki Nishio, A. Yamazaki, Yuko Hara–Azumi · 2024
In decentralized federated learning (DFL), the dual challenges of extensive inter-node communication and non-independent, identically distributed (non-IID) data impede the attainment of high-accuracy models while maintaining minimal communication traffic. We propose Tram-FL, which progressively refines a global model by transferring it sequentially amongst nodes. We also introduce a dynamic model routing algorithm for optimal route selection, aimed at enhancing model precision with minimal forwarding. Our experiments demonstrate that Tram-FL with the proposed routing delivers high model accuracy, outperforming baselines while reducing communication costs.