Abstract
Convolutional neural networks (CNNs) have become indispensable to medical image diagnosis research, enabling the automated differentiation of diseased images from extensive medical image datasets. Due to their efficacy, these methods raise significant privacy concerns regarding patient images and diagnostic models. To address these issues, some researchers have explored privacy-preserving medical image diagnosis schemes using fully homomorphic encryption (FHE). However, these schemes often support and are suitable for only a limited number of non-linear layers, resulting in less effective diagnoses and potentially inaccurate results. To improve upon these limitations, we propose and design a robust privacy-preserving medical diagnosis scheme that maintains both diagnostic accuracy and effectiveness at the same time. First, we utilize FHE to encrypt both the image and the model to safeguard the confidentiality of medical data and the model itself. Then, we introduce batch normalization to facilitate the use of multiple non-linear layers in deep convolutional neural networks within a ciphertext context. Furthermore, we employ a 2-degree polynomial function to approximate the ReLU activation function effectively. Finally, we introduce two innovative network depth optimization techniques to solve the issue of CNN depth insufficiency. Both theoretical and empirical analyses confirm that our scheme not only protects the confidentiality of medical images and diagnostic models but also ensures practicality and efficiency.