High performance logistic regression for privacy-preserving genome analysis.

Area of Science:

• Biomedical data security
• Machine learning cryptography

Background:

• Biomedical data is often siloed due to privacy regulations.
• Training machine learning models on distributed, private data presents a significant challenge.
• Secure multi-party computation (SMPC) offers a solution by combining machine learning and cryptography.

Purpose of the Study:

• To develop a practical and efficient secure machine learning protocol for distributed biomedical data.
• To address the computational overhead and runtime increase associated with SMPC.
• To enable collaborative training of models on sensitive datasets without violating privacy.

Main Methods:

• Utilized secure two-party computation protocols with a trusted initializer.
• Implemented a gradient descent algorithm for logistic regression with a clipped ReLu activation function.
• Developed a novel cryptographic protocol for secure activation function computation, avoiding secure comparisons and Yao's garbled circuits.

Main Results:

• Trained a model on a large gene expression dataset requiring over 7 billion secure multiplications.
• Achieved a training completion time of approximately 26.90 seconds on a local area network.
• The implementation is an optimized version that won first place in the iDASH 2019 secure genome analysis competition.

Conclusions:

• Presented a secure logistic regression training protocol with an efficient subprotocol for activation function computation.
• Demonstrated the fastest known SMPC implementation for training logistic regression models on high-dimensional, distributed genome data.
• The approach facilitates privacy-preserving machine learning in biomedical applications.