DeepFESL: Deep Federated Echo State Learning-Based Proactive Content Caching in UAV-Assisted Networks

Unmanned aerial vehicles (UAVs) have proven to be useful in a variety of applications, including aerial base station relay. UAVs can be used to relay network access from the air to user equipment (UE) on the ground. To improve the quality of experience (QoE) provided by UAVs, proactive content frequently requested by UEs can be cached to the UAVs. However, it is difficult to predict the chaotic and nonstationary patterns of mobile UEs because of the QoE requirements, the frequent location changes, the delays in content delivery, and scalability issues. Thus, this paper proposes a novel Deep Federated Echo State Learning (DeepFESL) content caching scheme to predict popular content requests and mobility patterns, using the context information of UEs, such as device type, location, and other behaviors on the device. The adverse effects of training models with such sensitive data make it inevitable to consider the privacy issues related to using such data. As a result, the proposed DeepFESL scheme employs a decentralized technique, federated learning to collaboratively train shared models on UEs' devices, where shared weights are aggregated at the control server. Popular contents are cached at the UAVs, where the UAVs form associations with the UEs in clusters based on predictions of the mobility patterns of the UEs. Experiment results show that our proposed DeepFESL scheme outperforms existing schemes in predicting popular UE contents and locations, increasing cache hit rate and cache utilization, reducing content transmission delays, and improving UEs' QoE.