Federated-Learning-Driven Radio Access Networks

In this work, we present federated learning (FL) driven radio access mechanisms, which can be employed at the future radio access networks (RANs) to achieve high accuracy rather than data rate that was the objective in RANs without FL application so far. We first present low feedback FL techniques based on local aggregation updates such as sparsification and quantization, which enable a communication efficient FL application in RANs. Next, the concept of FL-driven radio access mechanisms is presented, which is considered the joint design/optimization of FL with beamforming, retransmission, resource allocation and scheduling respectively. The aim of each joint optimization is to achieve the highest accuracy of the considered training model on data originated from the connected devices. Finally, a simulation setup is developed to obtain simulation results, which show the advantage of using the proposed radio access mechanisms comparing with FL without joint design with RAN in terms of accuracy, feedback overhead and communication rounds.