Latency Analysis of Multi-Exit Split Federated Learning via Testbed Implementation

Federated Learning (FL) offers a decentralized approach to training machine learning models across distributed devices, addressing data privacy concerns and reducing communication overhead inherent in centralized learning methods. However, as deep neural networks (DNNs) increase in complexity, the memory limitations of client devices pose significant challenges, particularly in real-time applications. To overcome these challenges, Split Learning (SL) has been proposed, which partitions DNN models between client devices and a central server, thereby reducing the burden on clients while maintaining data privacy. This paper extends SL by integrating a multi-exit strategy, allowing for early inference on less complex data samples to further enhance inference speed and efficiency. We validate our approach using a real-world testbed that replicates actual communication environments, contrasting it with existing methods under network conditions. Our results demonstrate that the Multi-Exit Federated Split Learning (ME-FedSL) algorithm not only reduces training and inference times but also improves the feasibility of deploying complex DNN models in federated learning environments, particularly in scenarios with bandwidth constraints.