Abstract

Efficient edge caching is essential for meeting the strict latency and reliability requirements of Ultra-Reliable Low Latency Communication (URLLC) in 5G and beyond networks. This paper presents a federated learning-based framework for predictive edge caching, addressing the challenges of data privacy, client heterogeneity and dynamic user behavior. First, a URLLC edge scenario was simulated using CloudSim Plus, generating realistic user request traces and network dynamics. The resulting dataset was used to train a federated neural network CacheHitPredictor model using the Flower framework with varying Dirichlet factors to study the effects of data heterogeneity. The trained model was deployed via a Flask REST API and integrated with the CloudSim simulation to enable online, modeldriven cache decisions. Experimental results demonstrate that the federated model outperforms the conventional Least Frequently Used (LFU) caching policy, achieving higher cache hit rates as the number of requests increases, with an average latency of 8 milliseconds. The study highlights the impact of non-IID data on federated learning model convergence and shows that increased training rounds can mitigate these effects. Overall, the proposed approach enhances cache efficiency, demonstrating its practical potential for URLLC edge caching in next-generation wireless networks.