Improving the quality of Federated Learning processes via Software Defined Networking

Federated Learning (FL) is rapidly gaining popularity as an effective cooperative and distributed approach, widely used by edge devices, to train machine learning models. Several aspects shall be managed to ensure a FL process that can more precisely match the QoS requirements of the applications that use it. The heterogeneity in the dataset available to each participant in the process, the variability in computational/memory capabilities, and the different availability of communication resources to connect the clients to the server are among the most critical. In this paper we will focus on the latter issue, less investigated in the literature, with particular reference to the case where the FL is used to support time-sensitive applications. Specifically, we will focus on studying the potential of an approach that leverages the Software-Defined Networking paradigm (SDN) to maintain the distributed learning process at high levels of effectiveness and efficiency even in the presence of edge client devices that may be delayed in delivering the result of their training due to the overload conditions experienced in the communication paths to the server. It will be shown, via a proof-of-concept performance evaluation campaign, how the proposed SDN support to the FL can guarantee significant overall reductions in process time at the cost of limited signaling overhead due to traffic to and from the controller.