MARVEL D3.4 - MARVEL's federated learning realization

This document reports on MARVEL’s federated learning realisation consisting of the component FedL and several novel federated learning methodologies that have been developed within the course of the project. First, FedL is described in general terms, and subsequently in MARVEL R1 realisation for visual crowd counting tasks, on the three MARVEL pilots: 1) GRN in Malta, for the use case GRN4 Junction traffic tra- jectory collection; 2) MT in Trento, for the use case MT1 Monitoring of crowded areas; and 3) the experimental pilot UNS in Novi Sad, for the use case UNS1 Drone experiment, for monitoring large space public events. Further, novel federated methodologies for model personalisation and clustering, decentralised and unsupervised anomaly detection, and, distributed inference and social learning are described, with particular implications and relevance for MARVEL. Finally, a novel strategy for the design of federated learning protocols based on the metric of large deviations is presented.