Federated Learning Framework to Decentralize Mobility Forecasting in Smart Cities

The Federated Learning (FL) paradigm aims to provide performance advantages over centralized models, such as lower latency and communication overhead when doing most of the processing on the edge devices, better privacy as data does not travel over the network, easier handling in heterogeneous data sources and better scalability. However, the development of FL-based solutions is done through tools aimed for specialists as it always requires some programming. To cover this gap, we present an architecture for a lightweight container-based solution that offers a range of machine learning (ML) algorithms to build prediction engines for edge devices, which also includes the main options in algorithms/models for aggregation and refinement of models in the central server. The proposed framework allows the rapid build of containerized testbeds for the evaluation of ML and aggregation algorithms in the initial evaluation phase, and also later in the installation in real production infrastructures. We demonstrate the efficiency of our approach in estimating vehicle mobility in and out of the city of Aveiro, using real data collected by the communications and sensing infrastructure.