Personalized Hierarchical Federated Learning Framework for the Internet of Vehicles Based on Split Meta-Learning

The rapid popularization of the Internet of Vehicles demands efficient, privacy-preserving distributed learning. However, deploying Federated Learning in dynamic IoV environments faces a ”trilemma” of model adaptability, communication efficiency, and privacy, exacerbated by severe spatiotemporal data heterogeneity. To address this issue, we propose a personalized hierarchical framework that integrates split meta-learning within a ”vehicle-edge-cloud” architecture, referred to as pHFSML. At the vehicle-edge layer, semi-asynchronous split meta-learning protocols significantly reduce vehicular computational burdens, enabling rapid local adaptation. At the edge-cloud layer, gradient-sensitive momentum aggregation and loss-adaptive personalization ensure global stability while retaining local precision. Extensive experiments on non-IID benchmarks verify pHFSML’s superiority. Specifically, on the domain-relevant GTSRB dataset, pHFSML achieves 93.03% accuracy, outperforming state-of-the-art baselines by 0.80%, with a convergence speed 38.8% faster than FedAvg. Ablation studies further validate the necessity and synergistic effects of the proposed components.