Swarm Split Learning: A Fully Distributed Machine Learning System for Energy-Constrained IoT Systems

Abstract

The widespread implementation of deep learning across diverse domains is widely acknowledged. However, with the increasing complexity of models comes a greater demand for computational resources and extensive datasets. This presents significant challenges, particularly in resource-constrained environments where data acquisition is difficult. The healthcare sector is a prime example, facing strict privacy regulations governing data usage. To address the balance between data access and privacy protection, novel distributed learning methods have emerged. These approaches enable model training directly at the data source, eliminating the need to share raw data. This paper introduces a novel technique called Swarm Split Learning, which integrates distributed, privacy-conscious training with energy efficiency principles and dynamic server selection in the network. We evaluated our system's performance through comprehensive experiments and comparisons with alternative distributed training frameworks. Our findings demonstrate that our decentralized approach is more robust and energy-efficient, allowing devices with limited energy resources to remain active for significantly more training rounds. Consequently, this contributes to enhanced classification performance of the global model and improves its generalizability.

Authors 3

1. Ahmad Ayad Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Chair of Information Theory and Data Analytics,Aachen,Germany,D-52074

2. Mahdi Barhoush Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Chair of Information Theory and Data Analytics,Aachen,Germany,D-52074

3. Anke Schmeink Aachen Chair of Information Theory and Data Analytics

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University,Chair of Information Theory and Data Analytics,Aachen,Germany,D-52074