Private pathological assessment via machine learning and homomorphic encryption.

Area of Science:

• Bioinformatics
• Computational Biology
• Medical Informatics

Background:

• Machine learning models like support vector machines (SVM) are crucial for medical data analysis.
• Protecting patient confidentiality during analysis is a significant challenge.
• Fully homomorphic encryption (FHE) offers a potential solution for privacy-preserving computations.

Purpose of the Study:

• To investigate the application of machine learning and FHE for private pathological assessment.
• To focus on the inference phase of SVM for classifying confidential medical data.
• To evaluate the feasibility of secure and efficient analysis of sensitive health information.

Main Methods:

• A novel framework employing the Cheon-Kim-Kim-Song (CKKS) FHE scheme for SVM inference on encrypted datasets.
• Implementation of a secure process that bypasses data decryption during analysis.
• Development of an efficient feature extraction technique for converting medical images into vector formats.

Main Results:

• The proposed system demonstrates practical applicability and efficacy across diverse datasets.
• Achieved classification accuracy and performance comparable to non-encrypted SVM inference.
• Maintained a 128-bit security level against cryptographic attacks on the CKKS scheme, with inference completed in seconds.

Conclusions:

• Fully homomorphic encryption (FHE) is viable for enhancing security and efficiency in bioinformatics.
• This approach has significant implications for privacy-preserving machine learning in healthcare.
• Potential benefits span cardiology, oncology, medical imagery, diagnostics, personalized medicine, and clinical research.