Scalable and interoperable edge-based federated learning in IoT contexts

The analysis of data coming from massively deployed Internet of Things (IoT) devices pave the way to a myriad of intelligent applications in several vertical domains. Federated Learning (FL) has been recently proposed as a prominent solution to train Machine Learning (ML) models directly on top of devices (FL clients) generating data, instead of moving them to centralized servers in charge of training procedures. FL provides inherent benefits, mainly in terms of privacy preservation and reduction of network congestion for datasets exchange. Despite the huge recent research efforts, it still faces challenges for a practical implementation effectively targeting low communication footprint, robustness and interoperability. To fill this gap, in this work we propose a novel comprehensive framework built upon the Message Queue Telemetry Transport (MQTT) publish/subscribe messaging protocol and the Open Mobile Alliance (OMA) Lightweight Machine-to-Machine (LwM2M) semantics to facilitate FL operations and make them more suited to handle IoT devices acting as FL clients. The viability of the proposal as well as its communication efficiency compared to a literature solution are evaluated through a realistic Proof-of-Concept (PoC) under different link settings and for different datasets.

Automated summary

This paper presents a comprehensive framework based on MQTT and LwM2M semantics to address the challenges of Federated Learning in low communication footprint, robustness and interoperability for IoT devices acting as FL clients. The viability of the proposal and its communication efficiency compared to a literature solution are evaluated through a realistic Proof-of-Concept (PoC) under different link settings and for different datasets.