Secure distributed machine learning in healthcare: a study on FAIR, compliance and cybersecurity for federated learning

Abstract

Artificial intelligence has reshaped the field of health care: from algorithms that can instantaneously diagnose radiographs to analyses of repositories spanning millions of viral genomic sequences. Despite these advances, we are facing the reality of the challenges and risks involved in storing and processing enormous amounts of sensitive data.

This dissertation explores how those challenges can be addressed with methods of computation that prevent the need to transfer data, instead learning from local data and sharing those learnings across distributed models. The dissertation proposes combining the concept of federated learning with methods from cryptographic computation to introduce a methodology that provides both security and privacy by design.

This research covers three broad themes around federated learning: FAIR, compliance and cybersecurity across four research articles. The articles cover that FAIR enables cross-country collaboration and machine interpretability, a pre-requisite for federated learning in an international setting. The research shows that a fully federated data architecture utilising FAIR principles allows for GDPR-compliance by design.

The dissertation concludes by introducing secure distributed machine learning, a cryptographically secure variant of federated learning. It is demonstrated that this methodology incurs minor computational overhead, provides cryptographic security and has efficacy equivalence with federated learning.