Proof of Concept for FedRelax on Kubernetes: An Implementation Guide

Loading...
Thumbnail Image

URL

Journal Title

Journal ISSN

Volume Title

School of Electrical Engineering | Master's thesis

Date

2024-11-27

Department

Major/Subject

Electronic and Digital Systems

Mcode

Degree programme

Master's Programme in Automation and Electrical Engineering

Language

en

Pages

107

Series

Abstract

In recent years, federated learning has been introduced and gained interest in different research areas. The evolution of federated learning has enabled an ability to train models on distributed datasets residing on multiple devices while preserving data privacy. FedRelax, a novel federated learning approach, is designed to train a heterogeneous network of models. Kubernetes is the most popular container orchestration tool to handle distributed environment and facilitate the application deployment, scaling and management. This thesis aims to study the integration of cloud technologies and federated learning system to investigate the behavior of leveraging these technologies together in real-world scenarios. Specifically, the study explores the feasibility of implementing the FedRelax algorithm, an novel federated learning method, on a Kubernetes cluster. The thesis study includes the design and implementation documentation for a Proof of Concept (POC). The performance, scalability and fault tolerance of the POC will be evaluated through different experiments and test scenarios. A baseline approach that runs FedRelax locally will be used as a benchmark to analyze the performance of this POC. When comparing the POC with baseline approach, a model generated by POC has less accuracy and higher runtime due to the complexity of distributed environment. However, the implementation enable a reliable scaling and fault tolerance mechanisms. Even though there are still potential areas to improve, the results shows that it is feasible to implement FedRelax algorithm on Kubernetes environment and the algorithm can perform training in a distributed settings.

Description

Supervisor

Jung, Alex

Thesis advisor

Sarcheshmehpour, Yasmin

Keywords

federated learning, cloud native technologies, FedRelax, docker, kubernetes, software development

Other note

Citation