Asynchronous Attention-Weighted Federated Learning for Edge Cache Optimization in Mobility-Aware Vehicles

Vehicular Edge Computing (VEC) utilizes RSUs for low-latency content caching in dynamic vehicular environments, yet faces challenges from vehicles’ high mobility causing frequent coverage exits, resulting in outdated model updates and compromised freshness. To address this, we propose the Asynchronous Weighted-Attention Federated Learning (AWAFL) model with three core innovations: (1) A dual-path attention mechanism dynamically assesses vehicle contributions by integrating positional relevance and channel quality; (2) A delayed gradient compensation strategy mitigates outdated updates by incorporating decay-weighted historical models from transient vehicles; (3) A phased aggregation process prioritizes high-confidence contributions from stable participants during global model updates. Experiments demonstrate AWAFL’s superiority over FedAvg, Thompson Sampling, and Random methods, achieving cache hit rate improvements of approximately 2%, 11%, and 35% respectively at 400 cache capacity, validating its effectiveness in highly dynamic networks.