Privacy-preserving semi-parallel logistic regression training with fully homomorphic encryption.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomic Data Analysis

Background:

• Privacy-preserving computations are crucial for secure medical and genomic data research.
• Secure cloud deployment of algorithms on encrypted data enables innovation without revealing individual genomes.
• Advancements in secure computation methods face challenges with large biomedical datasets.

Purpose of the Study:

• To address the challenge of feature selection in encrypted genomic databases for machine learning.
• To improve the quality of regression models using features from encrypted data.
• To develop practical solutions for privacy-preserving analysis of large-scale biomedical data.

Main Methods:

• Utilized the Chimera hybrid framework, integrating TFHE and HEAAN fully homomorphic encryption schemes.
• Focused on a practical machine learning scenario for feature selection in encrypted databases.
• Developed a bootstrapped approach for flexible parameter application without re-encryption.

Main Results:

• The solution was a finalist in the iDash 2018 competition (HE Track).
• The proposed method is the only bootstrapped approach applicable across different parameter sets.
• Demonstrated practicality for real-world applications by avoiding database re-encryption.

Conclusions:

• This work represents a significant step towards solving the general feature selection problem for large encrypted databases.
• The developed method enhances the security and practicality of machine learning on sensitive genomic data.
• Paves the way for more advanced privacy-preserving analyses in biomedical research.