SDN-Assisted Client Selection to Enhance the Quality of Federated Learning Processes

An emerging modality, increasingly used by edge devices, to train machine learning models in a distributed and cooperative way is Federated Learning (FL). It combines an increase in the quality of the learning process with data privacy needs. Alongside the advantages of this emerging paradigm, however, there is a critical factor that risks seriously affecting its effectiveness in future 5G and 6G application scenarios: the possible delays deriving from the scarcity of communication resources to connect the clients to the server, which risks slowing down the process excessively and making it less effective in the presence of new types of real-time applications typical of 5G/6G scenarios. To face this issue, the paper proposes a new approach to client selection that, unlike the various approaches to streamlining FL communications proposed so far, starts from a typically networking research point of view and makes use of the potential of the Software-Defined Networking (SDN) paradigm for the choice and continuous dynamic update of the clients participating in the FL process. This allows to keep the distributed learning process at high levels of effectiveness and efficiency, i.e., guaranteeing an overall time reduction of the FL process under different network traffic load conditions, as demonstrated by the performance evaluation campaign conducted through the implementation of a testbed platform.