Decentralised federated machine learning at the edge on iot devices for healthcare
File(s)
Author(s)
Calo, James
Type
Thesis
Abstract
Deep learning has revolutionised healthcare, transforming the analysis of medical data and achieving performance on par with medical experts in a wide range of imaging tasks. This paradigm shift has established artificial neural networks as pivotal tools in healthcare applications, including disease diagnosis, treatment planning, and surgical interventions. As such, machine learning has become a cornerstone of modern medical innovation, driving progress toward improved clinical outcomes and personalized care.
However, training deep neural networks directly on the Internet of Things (IoT) devices at the edge poses significant challenges due to their constrained computational resources. Consequently, the vast potential of IoT devices in healthcare remains underutilized. These devices, widely available in medical environments, offer a unique opportunity to decentralise artificial intelligence (AI) by shifting computation from centralised cloud systems to the edge. Exploiting mist architectures already deployed in hospitals, this approach has the potential to revolutionise healthcare by enabling real-time processing, reducing latency, and enhancing data privacy and security.
Despite this promise, the adoption of machine learning in healthcare is hindered by critical challenges, including data privacy concerns, interoperability barriers, and the need for equitable access to technology. Innovative solutions are required to address these obstacles and fully realise the transformative potential of AI-driven healthcare.
This thesis addresses these challenges by designing a novel, decentralised federated learning framework tailored for healthcare applications. This framework facilitates privacy-preserving inter-hospital collaboration by leveraging existing mist architectures and the ubiquity of IoT devices. A key innovation of this research is the development of a bespoke blockchain consensus mechanism integrated with masked autoencoders (MAEs), providing robust privacy protection and enabling advanced federated learning capabilities. This work represents a significant step toward unlocking the full potential of decentralised machine learning in healthcare, paving the way for secure, collaborative, and scalable AI solutions in clinical practice.
However, training deep neural networks directly on the Internet of Things (IoT) devices at the edge poses significant challenges due to their constrained computational resources. Consequently, the vast potential of IoT devices in healthcare remains underutilized. These devices, widely available in medical environments, offer a unique opportunity to decentralise artificial intelligence (AI) by shifting computation from centralised cloud systems to the edge. Exploiting mist architectures already deployed in hospitals, this approach has the potential to revolutionise healthcare by enabling real-time processing, reducing latency, and enhancing data privacy and security.
Despite this promise, the adoption of machine learning in healthcare is hindered by critical challenges, including data privacy concerns, interoperability barriers, and the need for equitable access to technology. Innovative solutions are required to address these obstacles and fully realise the transformative potential of AI-driven healthcare.
This thesis addresses these challenges by designing a novel, decentralised federated learning framework tailored for healthcare applications. This framework facilitates privacy-preserving inter-hospital collaboration by leveraging existing mist architectures and the ubiquity of IoT devices. A key innovation of this research is the development of a bespoke blockchain consensus mechanism integrated with masked autoencoders (MAEs), providing robust privacy protection and enabling advanced federated learning capabilities. This work represents a significant step toward unlocking the full potential of decentralised machine learning in healthcare, paving the way for secure, collaborative, and scalable AI solutions in clinical practice.
Version
Open Access
Date Issued
2024-12-01
Date Awarded
01/04/2025
Advisor
Lo, Benny
Sponsor
Engineering and Physical Sciences Research Council
Grant Number
EP/R513052/1
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)