HandoffFL: Efficient Communication in Federated Learning Via Asynchronous Layer-Wise Handoff Training

Abstract

Federated Learning (FL) enables collaborative training without sharing raw data, yet conventional synchronous approaches suffer from poor scalability and high communication costs, imposing heavy burdens on resource-constrained clients in heterogeneous edge environments. While asynchronous FL reduces heterogeneous stragglers, it introduces severe staleness and convergence degradation. Recent layer-wise FL methods reduce resource consumption by decomposing training into sequential layer-level tasks; however, they still rely on rigid round-based synchronization and therefore lack adaptability to dynamic client availability and network fluctuations. We propose HandoffFL, an event-driven asynchronous layer-wise FL framework built on a real-time message-passing monitoring architecture. HandoffFL organizes training into layer-level stages, where clients train one layer at a time with frozen prefixes. Transitions between layers are triggered by adaptive convergence events, enabling clients to progressively hand off converged layers and advance at their own pace. To support convergence detection at intermediate layers, the framework introduces temporal classifiers and further applies layer-specific staleness-aware aggregation with exponential decay and truncation. Extensive experiments on MNIST, Fashion-MNIST, and CIFAR-10 with multiple model architectures demonstrate that HandoffFL substantially reduces wall-clock training time and communication overhead compared to synchronous layer-wise FL, while maintaining comparable accuracy and robustness under heterogeneous client conditions compared with asynchronous FL.