Development and experimental evaluation of distributed split learning schemes over low-cost microcomputers (Raspberry Pi)

Abstract

This thesis explores the feasibility and practical deployment of distributed machine learning schemes on resource-constrained devices, focusing on Split Learning (SL) and Split Federated Learning (SFL). By implementing both approaches on Raspberry Pis, it provides an in-depth performance evaluation under real-world conditions, demonstrating practical evidence on how distributed learning schemes behave beyond simulation studies or purely theoretical models. Through systematic experimentation and monitoring of key metrics such as training time, accuracy, resource usage and network efficiency, this work uncovers the benefits and limitations of each scheme under varying data heterogeneity scenarios. These findings offer substantial insights for the scalability, privacy and resource efficiency of distributed learning deployed on IoT devices, and propose concrete directions for overcoming current challenges in distributed machine learning applications across resource-constrained devices and networks.