Secure Logistic Regression Based on Homomorphic Encryption: Design and Evaluation.

Area of Science:

• Cryptography
• Machine Learning
• Data Security

Background:

• Secure data outsourcing and privacy-preserving machine learning are critical research areas.
• Homomorphic encryption (HE) offers potential for secure computation on encrypted data.
• Current HE frameworks are limited to simplified models, hindering real-world applications.

Purpose of the Study:

• To develop a practical homomorphic encryption solution for mainstream machine learning models, specifically logistic regression.
• To enable secure outsourcing of sensitive data for analysis without decryption.

Main Methods:

• Adapted a novel homomorphic encryption scheme optimized for real number computations.
• Developed least squares approximation for the logistic function to enhance accuracy and efficiency.
• Implemented new packing and parallelization techniques for improved performance.

Main Results:

• Demonstrated the feasibility of homomorphically encrypted logistic regression using real-world datasets.
• Achieved practical performance in terms of speed and memory consumption (e.g., 116 minutes for training on the Edinburgh dataset).
• Obtained accurate predictions on the testing dataset, validating the model's effectiveness.

Conclusions:

• Presented the first homomorphically encrypted logistic regression outsourcing model.
• Leveraged the observation that precision loss in classification models does not significantly alter the decision boundary.
• Established a foundation for secure and private machine learning on sensitive data.