Measuring Overhead Costs of Federated Learning Systems by Eavesdropping.

This paper addresses the issue of communication overhead costs of federated learning including transmission bandwidth and synchronisation efforts. These costs typically consist of locally observable costs on executing components, but there are also hidden costs that can only be measured from a system-wide perspective. The goal is to provide insight into these hidden costs, measure them and identify strategies for reducing them. We propose an approach to tackle the hidden costs by establishing a methodology consisting of an eavesdropping concept and an evaluation strategy. This way we obtain a refined analysis of directly observable costs contrasting hidden costs, which is underpinned by experiments based on a 40-client-spanning federated learning system and the FEMNIST dataset.