Enhancing healthcare data privacy with homomorphic encryption

Abstract

Artificial intelligence in healthcare has birthed advances in fields such as medical diagnostics and personalised treatments. However, the use of patient information raises pertinent privacy issues, creating the challenge of protecting patient data privacy and confidentiality. This dissertation focused on creating a privacy-preserving artificial intelligence application for melanoma diagnosis by employing Fully Homomorphic Encryption (FHE) to bolster data confidentiality. The main objectives included examining the challenges of managing health data for AI, exploring FHE applications, developing an application that integrates FHE for confidential processing of health data, and verifying its accuracy, efficiency and privacy assurances. The methodology included conducting a literature review, requirement analysis, designing an FHE-enabled architecture using Concrete ML, system development, testing and validation. The application was developed using the Streamlit framework and TensorFlow’s ResNet50 model combined with clinically pre-trained weights calibrated for melanoma detection. A minimal custom model designed for FHE compatibility was then implemented through a knowledge distillation process from the ResNet50 teacher model. This was to address the inherent computational constraints of homomorphic encryption while preserving diagnostic capabilities. Thorough testing and validation was carried out to ensure the application functions as per the requirements and achieved the set goals. A verified security expert attempted to access unencrypted patient data during FHE operations and confirmed that the data remained securely encrypted throughout the processing pipeline. The application offered accurate melanoma detection while ensuring verifiable endto- end encryption of patient data throughout the process in compliance with healthcare privacy laws. This demonstrates FHE’s potential in facilitating the adoption of privacy-preserving AI in healthcare and other data-sensitive fields. KEY WORDS: Artificial Intelligence, Concrete ML, Confidentiality, Fully Homomorphic Encryption, Healthcare, Machine Learning, Melanoma, Privacy, ResNet50, Streamlit.