Expanding the cloud-to-edge continuum to the IoT in serverless federated learning

Serverless computing enables greater flexibility and efficiency in the cloud-to-edge continuum. Artificial Intelligence and Machine Learning (AI/ML) applications benefit greatly from this paradigm, as they need to gather, preprocess, aggregate and analyze data at various scales. In such contexts, the increasing hardware/software resource availability of Internet of Things (IoT) devices provides the opportunity to exploit them not only as data sources in AI/ML infrastructures, but also as computational nodes for model training and inference; nevertheless, comprehensive frameworks are still mostly missing. This work introduces an innovative serverless computing architecture which expands the cloud-to-edge continuum toward IoT devices. The same functions can run on IoT, edge and cloud nodes with minimal to no code modification and they can be invoked through a uniform interface. A federated learning framework is defined based on the proposed architecture, exploiting an existing IoT-oriented ML algorithm in a novel way. Notably, IoT nodes are used for both federated training and local inference tasks. A full prototype implementation has been built with off-the-shelf technologies and devices. A case study on federated machine learning for activity recognition and experiments have been conducted to validate key elements of the proposal.