Practical Privacy-Preserving Convolutional Neural Network Inference Framework With Edge Computing for Health Monitoring

Abstract

Using Convolutional Neural Network (CNN) model to analyze monitoring data in Body Area Network (BAN) has become an important way to solve health related issues in the current large sub-health population and aging population. However, the inference and analysis process of BAN data needs to ensure efficiency and security. At present, ensuring a balance of efficiency and security in the inference of CCN models is challenging. Therefore, an efficient and secure CNN inference scheme is proposed based on two Edge-Cloud-Servers (CS $_{0}$ and CS $_{1}$ ). By analyzing the CNN model and combining two secret sharing semantics, we optimize the communication overhead of inference. Specifically, a new non-interactive secure convolutional layer computation protocol is designed to significantly reduce the number of interactions between CS $_{0}$ and CS $_{1}$ . For non-linear layers, we propose a simpler secure comparison computation functionality to reduce the communication overhead. Moreover, we also design some lightweight secure building blocks based on secret sharing to improve computing efficiency. We implement our proposed scheme on two standard datasets. Through the theoretical analysis and experimental comparison, our scheme improves the computational efficiency.