Cloud-based homomorphic encryption for privacy-preserving machine learning in clinical decision support

  • Jim Basilakis

Western Sydney University thesis: Doctoral thesis

Abstract

While privacy and security concerns dominate public cloud services, Homomorphic Encryption (HE) is seen as an emerging solution that ensures secure processing of sensitive data via untrusted networks in the public cloud or by third-party cloud vendors. It relies on the fact that some encryption algorithms display the property of homomorphism, which allows them to manipulate data meaningfully while still in encrypted form; although there are major stumbling blocks to overcome before the technology is considered mature for production cloud environments. Such a framework would find particular relevance in Clinical Decision Support (CDS) applications deployed in the public cloud. CDS applications have an important computational and analytical role over confidential healthcare information with the aim of supporting decision-making in clinical practice. Machine Learning (ML) is employed in CDS applications that typically learn and can personalise actions based on individual behaviour. A relatively simple-to-implement, common and consistent framework is sought that can overcome most limitations of Fully Homomorphic Encryption (FHE) in order to offer an expanded and flexible set of HE capabilities. In the absence of a significant breakthrough in FHE efficiency and practical use, it would appear that a solution relying on client interactions is the best known entity for meeting the requirements of private CDS-based computation, so long as security is not significantly compromised. A hybrid solution is introduced, that intersperses limited two-party interactions amongst the main homomorphic computations, allowing exchange of both numerical and logical cryptographic contexts in addition to resolving other major FHE limitations. Interactions involve the use of client-based ciphertext decryptions blinded by data obfuscation techniques, to maintain privacy. This thesis explores the middle ground whereby HE schemes can provide improved and efficient arbitrary computational functionality over a significantly reduced two-party network interaction model involving data obfuscation techniques. This compromise allows for the powerful capabilities of HE to be leveraged, providing a more uniform, flexible and general approach to privacy-preserving system integration, which is suitable for cloud deployment. The proposed platform is uniquely designed to make HE more practical for mainstream clinical application use, equipped with a rich set of capabilities and potentially very complex depth of HE operations. Such a solution would be suitable for the long-term privacy preserving-processing requirements of a cloud-based CDS system, which would typically require complex combinatorial logic, workflow and ML capabilities.
Date of Award2020
Original languageEnglish

Keywords

  • cloud computing
  • security measures
  • computer security
  • data encryption (computer science)
  • data protection

Cite this

'