Enabling Edge-based Federated Learning through MQTT and OMA Lightweight-M2M

The increasing processing capabilities of massively deployed Internet of Things (IoT) and edge devices coupled with the rising demands for fast and autonomous decision-making pave the way for accurate Machine Learning (ML) model training in proximity to where data is collected, as proposed by the Federated Learning (FL) paradigm. The devices contributing to the local training (clients) are typically large in number and differ in terms of computing and battery capabilities, experienced link quality conditions and owned datasets. If clients are not properly selected by the entity in charge of aggregating the individual models (FL server) both the convergence speed and the accuracy of the resulting model can be affected. The interest in the literature for the optimization of the client selection policies is manifest, but less attention has been devoted to the design of the practical implementation of such policies with focus on the client discovery procedures. In order to fill this gap, an edge-based framework is proposed to improve the FL client discovery procedures by leveraging (i) the Message Queue Telemetry Transport (MQTT) protocol for FL clients-server interactions aimed to retrieve the capabilities of the potential clients and (ii) the Open Mobile Alliance (OMA) Lightweight-Machine-to-Machine (LwM2M) for the semantic description of such capabilities. A proof-of-concept is provided to assess the communication footprint incurred by the proposal.