Logistic regression model training based on the approximate homomorphic encryption.

Area of Science:

• Cryptography
• Machine Learning
• Data Security

Background:

• Big data analysis raises security concerns due to sensitive individual information in training datasets.
• Secure computation is explored as a privacy protection solution by the cryptography community.
• Research focuses on the efficiency of cryptographic primitives for practical privacy-preserving applications.

Purpose of the Study:

• To develop a method for training logistic regression models without information leakage.
• To enhance the efficiency and practicality of privacy-preserving machine learning algorithms.
• To address security concerns in big data analysis involving sensitive personal data.

Main Methods:

• Utilized the homomorphic encryption scheme by Cheon et al. (ASIACRYPT 2017) for efficient real number arithmetic.
• Devised a novel encoding method to minimize the storage requirements of encrypted databases.
• Adapted Nesterov's accelerated gradient method to reduce computational cost and iterations while preserving model quality.

Main Results:

• Achieved state-of-the-art performance for homomorphic encryption in a real-world application.
• The developed method was recognized as the best solution in Track 3 at the iDASH privacy and security competition 2017.
• Generated a logistic regression model from a dataset (1579 samples, 18 features) in approximately six minutes.

Conclusions:

• Presented a practical solution for securely outsourcing logistic regression analysis.
• Ensured data confidentiality is maintained during outsourced analytical processes.
• Demonstrated the feasibility of privacy-preserving machine learning for sensitive datasets.