The Role of SDN to Improve Client Selection in Federated Learning

In an ever-increasing number of contexts, it has now become common to use federated learning (FL) techniques, through which several heterogeneous devices cooperate in a distributed manner to increase the effectiveness in training machine learning (ML) models while maintaining confidentiality of the respective data. The federated learning process shows performance levels that are highly dependent, not only on the data available to each participant in the process, but also on the choice of devices that act as clients from time to time and that collaborate with each other. So far, much of the literature has focused on a client selection that considers the device computational/memory capabilities and the end-to-end delays of the process. However, no one so far has fully exploited the intrinsic capabilities of current and future programmable networks. This article introduces, for the first time, an approach to client selection that is augmented and made more effective by a joint orchestration carried out by the FL server and the controller of a software-defined networking (SDN) network. It will be demonstrated through a performance evaluation campaign that the use of typical SDN principles also in the client selection phase leads to significant advantages in terms of effectiveness and efficiency.