Federated Learning for Scalable Edge Caching in Dynamic D2D Communication Networks

Device-to-device (D2D) communication networks face significant challenges due to their dynamic nature, including limited bandwidth, high latency, and frequent disruptions. These factors complicate content caching and timely user request fulfillment. To address these issues, we propose a federated learning (FL)-based edge caching framework that utilizes density-based spatial clustering of applications with noise (DBSCAN) to cluster devices based on content similarity and Euclidean distance. Within each cluster, a master user equipment is selected based on the signal-to-noise ratio, willingness, and battery level, while slave user equipment devices share content based on popularity. A graph neural network (GNN) is employed for local caching prediction, effectively capturing complex spatiotemporal dependencies. FL enables global model training without centralizing raw data, improving scalability and privacy. Additionally, deep Q-learning (DQL) optimizes caching decisions for each device. Simulation results demonstrate the effectiveness of the proposed framework, achieving a high cache hit ratio of 94.6%, an average delay of 33ms for 100 clusters, and energy consumption between 1.727 to 1.746 bit/J. These results highlight the framework’s efficiency in reducing delay and power consumption, outperforming existing techniques in terms of these performance metrics.