Development of Privacy-preserving Deep Learning Model with Homomorphic Encryption: A Technical Feasibility Study in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cryptography

Background:

• Deep learning models are effective for analyzing CT images of renal masses.
• However, privacy concerns limit the use of sensitive patient data in these models.
• Homomorphic encryption offers a potential solution for privacy-preserving analysis.

Purpose of the Study:

• To assess the technical feasibility of integrating homomorphic encryption into deep learning for privacy-preserving CT renal mass analysis.
• To evaluate the impact of encryption on diagnostic performance and computational demands.

Main Methods:

• Developed a deep learning system with three phases: reference CNN (Ref-CNN), modified for encryption (Approx-CNN), and fully homomorphic encrypted (HE-CNN) using the CKKS scheme.
• Trained and evaluated models on 12,446 CT images of renal masses.
• Assessed performance using Area Under the Receiver Operating Characteristic Curve (AUC) and Area Under the Precision-Recall Curve (AUPRC).

Main Results:

• All models achieved high diagnostic accuracy (AUC: 0.89-0.99, AUPRC: 0.67-0.99).
• Minimal performance trade-off observed between Ref-CNN and Approx-CNN.
• Homomorphic encryption significantly increased storage (65KB to 32MB) and computation time (50 min CPU vs. 90 sec GPU).

Conclusions:

• Privacy-preserving deep learning inference using homomorphic encryption is technically feasible for renal mass classification on CT images.
• Comparable diagnostic performance is achievable with end-to-end data encryption.
• Significant increases in computational and storage requirements necessitate optimized hardware acceleration.