Privacy-Preserving Feature Extraction for Medical Images Based on Fully Homomorphic Encryption

Abstract

Medical images contain sensitive patient information requiring privacy protection during cloud-based processing. This paper presents a novel privacy-preserving framework for medical image feature extraction based on fully homomorphic encryption (FHE). We develop an efficient encoding scheme that converts medical image data into polynomial representations suitable for homomorphic operations while preserving diagnostic accuracy. The framework implements specialized homomorphic algorithms for key point detection, feature description, and matching that operate entirely in the encrypted domain. Our approach incorporates SIMD (Single Instruction Multiple Data) optimization techniques to process multiple pixels simultaneously, reducing computational overhead and memory requirements. We introduce innovative methods for homomorphic comparison, division, and derivative operations essential for accurate feature extraction. Experimental evaluation on four medical imaging datasets demonstrates that our method achieves 93.6% feature extraction accuracy compared to plaintext processing, outperforming existing privacy-preserving approaches. Security analysis confirms 128-bit security with acceptable computational efficiency (75× slowdown versus plaintext) and minimal communication overhead. The proposed system enables secure outsourcing of medical image analysis to untrusted cloud environments without revealing sensitive patient data, facilitating privacy-compliant diagnostic assistance while maintaining clinical accuracy requirements.