Phase-Synchronization-Based Federated Learning for Edge Computing Over Decentralized Networks

Federated learning (FL) enables collaborative model training over decentralized edge devices while preserving data privacy; however, its performance is often degraded by client drift caused by non-IID data distributions and partial client participation. Motivated by the need for a principled mechanism to suppress directional inconsistency among distributed updates, we propose FedKuramoto, a novel FL algorithm inspired by the phase-synchronization dynamics of the Kuramoto oscillator model. In FedKuramoto, each client update is interpreted as an oscillator whose phase is defined by its angular deviation from the global update direction. The proposed method enforces phase coherence via a phase-constrained projection and introduces a coherence-aware momentum that acts as a restoring force, without incurring additional communication overhead. Extensive experiments on various benchmarks under varying heterogeneity and participation ratios demonstrate that FedKuramoto consistently accelerates convergence and improves accuracy over conventional FL algorithms. These results indicate that importing synchronization mechanisms from nonlinear dynamical systems offers an effective and robust design paradigm for decentralized learning.