Web-Based Privacy-Preserving Multicenter Medical Data Analysis Tools Via Threshold Homomorphic Encryption: Design and...

Area of Science:

• Medical Informatics
• Cryptography
• Machine Learning

Background:

• Multicenter medical research benefits from data sharing for generalizability and discovery.
• Data sharing is hindered by privacy concerns and potential sensitive information leaks.
• Existing secure machine learning frameworks often lack practicality for multicenter applications due to single-key limitations.

Purpose of the Study:

• To develop a privacy-preserving machine learning protocol for multiple data providers and researchers.
• To enable secure training and evaluation of models (e.g., logistic regression) on pooled medical data.
• To ensure robust privacy protection for both sensitive data and the resulting learned models.

Main Methods:

• Adaptation of a novel threshold homomorphic encryption scheme for privacy guarantees.
• Development of new relinearization key generation techniques for enhanced scalability and multiplicative depth.
• Implementation of novel model training strategies, including x-fold cross-validation for simultaneous multiple model training.

Main Results:

• Evaluation using a client-server architecture demonstrated protocol performance.
• Privacy-preserving logistic regression model training and evaluation were achieved over 10 attributes in a large dataset (49,152 samples).
• The protocol completed training in approximately 7 minutes and evaluation in approximately 20 minutes with 10-fold cross-validation.

Conclusions:

• Presentation of the first privacy-preserving multiparty logistic regression protocol using threshold homomorphic encryption.
• The developed protocol is practical for real-world applications.
• This advancement has the potential to significantly promote multicenter medical research.