Privacy-preserving logistic regression with secret sharing

Area of Science:

• Computational statistics
• Data privacy
• Machine learning

Background:

• Logistic regression (LR) is crucial for binary outcome classification in medical research.
• Combining diverse datasets enhances statistical power but raises significant privacy concerns.
• Addressing privacy is essential for secure collaborative data analysis.

Purpose of the Study:

• To propose privacy-preserving protocols for logistic regression parameter estimation.
• To enable secure distributed training of logistic regression models.
• To address privacy challenges in multi-jurisdictional data aggregation.

Main Methods:

• Developed two protocols based on secure Multi-Party Computation (MPC).
• Utilized the Newton-Raphson method for parameter estimation.
• Protocols are designed for both honest and dishonest majority security settings.

Main Results:

• Protocols demonstrated high efficiency and accuracy on synthetic and real-world datasets.
• Experimental results show comparable performance to ordinary logistic regression.
• Algorithms effectively handle large, distributed datasets.

Conclusions:

• Introduced two iterative algorithms for privacy-preserving distributed logistic regression training.
• Implementation confirms the capability to manage large datasets from multiple sources.
• The work facilitates secure and powerful analysis of combined datasets.