Federated Learning (FL) is a promising approach for parameter normalisation in Machine Learning (ML) models, especially when data privacy and computing distribution are crucial. However, there are significant constraints in FL solutions, particularly concerning the handling of the mobility of participating nodes in the parameter aggregation processes, with a substantial impact on Vehicle to Everything (V2X) scenarios within the scope of smart cities. To address this challenge, we propose Mobile Federated Learning System (MobFedLS), a lightweight microservices-based framework capable of operating on various types of devices (mobile and non-mobile). MobFedLS features an interface to integrate ML models to cooperate in the FL process without intrusion between the parties. MobFedLS manages the entire federation process, from instantiating services on mobile nodes to the final parameter updates in the involved ML models and the release of resources used in all participating nodes. Additionally, MobFedLS handles node mobility and ensures the proper execution of federated processes, even with nodes entering and leaving at any stage of the aggregation process. To demonstrate the capabilities of MobFedLS, we use data collected through the city-scale infrastructure of Aveiro Tech City Living Lab (ATCLL), specifically the position of vehicles during their movement through the city. In the tests, we evaluate all phases of the aggregation process for mobile nodes. The results show that, even with intermittent connectivity to the city-infrastructure ATCLL, the MobFedLS system manages the node mobility and effectively handles node availability during the aggregation of ML model parameters.
Read full abstract