Empirical Analysis of Federated Learning Algorithms: A Federated Research Infrastructure Use Case

Research Infrastructures provide resources and services for communities of researchers at large to conduct their experiments and foster innovation. Moreover, these can be used beyond research, e.g., for education or public service. The SLICES consortium is chartered to provide a fully programmable, distributed, virtualized, remotely accessible, European-wide, federated research infrastructure, providing advanced computing, storage, and networking capabilities, including interconnection by dedicated high-speed links. It will support large-scale, experimental research across various scientific domains. Data processing, in general, and especially Machine Learning, are of great interest to the potential audience of SLICES. According to these premises, this work aims to exploit such a peculiar Research Infrastructure and its Cloud-oriented development and deployment facilities to investigate Federated Learning (FL) approaches; in particular, here we evaluate the performance of two FL aggregation algorithms, i.e., FedAvg and FedProx, in settings, characterized by system heterogeneity, and statistical heterogeneity, that represent plausible, and possibly common, scenarios in forthcoming facilities, such as those mentioned above, community-oriented, shared Research Infrastructures. We have observed that the FedProx algorithm outperforms the FedAvg algorithm in such settings.