Privacy-Preserving Artificial Intelligence: Application to Precision Medicine.

Area of Science:

• Medical Informatics
• Cryptography
• Machine Learning

Background:

• Machine Learning (ML) shows promise for personalized medicine.
• Patient data privacy regulations hinder clinical adoption of Deep Learning (DL).

Purpose of the Study:

• To propose a privacy-preserving solution for DL in healthcare using Fully Homomorphic Encryption (FHE).
• To evaluate the feasibility and performance of FHE-based DL on sensitive health data.

Main Methods:

• Utilized the MORE scheme of FHE for computations on encrypted floating-point health data.
• Demonstrated feasibility on MNIST digit recognition and a Whole-Body Circulation (WBC) model estimation task.
• Analyzed the security aspects of the chosen FHE scheme.

Main Results:

• Achieved comparable accuracy to unencrypted DL models on both benchmark and complex tasks.
• Showcased that DL computations can be performed on encrypted data with reasonable overhead.
• Identified practical applications where the FHE scheme's balance of performance and security is suitable.

Conclusions:

• FHE offers a viable solution for privacy-preserving DL in healthcare.
• The proposed method enables personalized medicine applications without compromising patient data.
• The MORE FHE scheme provides a practical approach for sensitive health data analysis.