Design and Analysis of Federated Learning Systems

Federated Learning (FL) is a privacy-preserving machine learning strategy that enables network-edge clients to train a shared model under the orchestration of a central server. This thesis deals with the design and analysis of FL systems. Specifically, the focus is on three primary design choices, which have a predominant effect on the communication and operational costs of an FL system: (i) the communication protocol used for the interactions among the server and the clients; (ii) the database employed to store the shared model parameters; (iii) the internal architecture of the system to handle synchronization and model update processes. Identifying the most efficient and low-cost configurations requires the availability of ad-hoc simulations and profiling tools able to emulate different implementation solutions and assess their performance under variable load profiles. This thesis addresses this need by proposing an emulation environment that takes an FL system configuration as input and returns its performance metrics, i.e., the communication costs and the processing time for model update, and the resource utilization, i.e., RAM and CPU usage. After testing multiple FL system implementations with the proposed tool, the collected results underscored that, depending on the selected configuration, it is possible to reduce communication by 87% and processing time by 86% compared to the least-performing system. These findings highlight the benefits of design space exploration, as provided by the proposed tool, in supporting the implementation of efficient and low-cost FL systems.