Fairness-Aware Cooperative Caching in Vehicular Networks: An Asynchronous Federated Learning and Attention-Enhanced Multiagent DRL Approach

Cooperative caching is regarded as one of the most promising technologies for vehicular networks, as it significantly reduces content delivery latency by prestoring popular content at roadside units (RSUs) closer to vehicle users (VUs). However, accurately predicting popular content and subsequently making caching decisions to enhance the Quality of Experience (QoE) remains a complex challenge, with varied content preferences and performance levels across VUs. To address this issue, we propose a fairness-aware collaborative edge caching system that formulates the factors influencing QoE as a multiagent Markov decision process, with the aim of maximizing the long-term system revenue. Furthermore, we design an activity-based VU clustering and weighted prediction model and propose a training framework based on asynchronous federated learning for global model updates. The trained model extracts key features and accurately predicts popular content from historical data. Additionally, we present an attention-enhanced multiagent discrete soft actor-critic algorithm to tackle the complex caching decision problem and promote collaboration among RSUs. Extensive simulation experiments validate the superiority of our solution. Compared with other benchmarks, the proposed method improves service fairness by 50% and increases average system revenue by 16.7%.