HiMAC: Hierarchical Federated Learning with Mobility-Aware Collaboration

Abstract

Mobility-aware hierarchical federated learning (MobiHFL) is an emerging paradigm to support urban applications where clients continuously move across edge-server coverage areas. Existing methods mainly treat mobility as a scheduling challenge and often train a single global model across all edge-servers. This overlooks the spatial non-IID characteristics of urban environments, where regions exhibit distinct yet correlated data distributions. Region-wise personalization is a natural remedy, but mobility makes it difficult. Each edge model is trained on its current client population, while future arriving and departing clients may follow different distributions, causing cold-start degradation and feature-classifier mismatch after handover. We propose \sysname, a personalized MobiHFL framework that turns predicted client movement into proactive inter-edge collaboration. \sysname uses per-client LSTM trajectory predictors to estimate future client flows and builds a round-level transition matrix to guide collaboration. With a decoupled backbone-classifier architecture, \sysname prepares each edge for expected incoming clients through incoming-aware classifier initialization and improves compatibility with likely destination edges through outgoing-regularized backbone training. Extensive experiments on image classification and object detection benchmarks show improvements of up to 12.18 percentage points in image-classification accuracy and 6.50 points in mAP@50 over state-of-the-art baselines across varying mobility intensities.