Edge-to-Cloud Federated Learning with Resource-Aware Model Aggregation in MEC

The rapid increase in the number of connected devices paired with the adoption of Machine Learning (ML) applications dramatically augments the computation and communication requirements imposed on today's telecommunication networks. New ML techniques and networking paradigms such as Federated Learning (FL), Multi-access Edge Computing (MEC), and Software-Defined Wide Area Networks (SD-WANs) are needed to cope with these requirements. However, to run FL in MEC SD-WANs, intelligent resource management strategies and an evaluation of the impact of FL on the network resources are necessary. In this work, we discuss online resource management strategies for FL model aggregation enhanced by intermediate aggregation at edge nodes. Our analysis shows that a layer of intermediate aggregators (edge aggregators) alleviates the traffic on network links and allows us to take advantage of edge computing nodes, but the risk of congestion in the back-haul network is still high. We thus propose a new aggregation scenario deploying an aggregator overlay network and present an algorithm optimizing the routing of edge aggregators. Our proposed solution can adapt better to resource utilization in the network, achieving a decrease of the failure rate of FL training rounds by up to 15 percent while reducing cloud link congestion.