A Secure Convolutional Neural Network Inference Model Based on Homomorphic Encryption

The utilization of convolutional neural network (CNN) inference models has gained widespread adoption across various domains; however, it raises concerns regarding user privacy and security due to the requirement of plaintext information during the inference process. Homomorphic encryption has been proposed as a viable solution to address CNN's security issues. Nevertheless, this encryption algorithm encounters certain limitations such as significant ciphertext expansion, restricted nonlinear computing power, and limited number of multiplications, which hinder its application in CNN inference models. In this paper, we present an inference model framework that leverages the single instruction multiple data and ciphertext rotation functions of homomorphic encryption along with the additive secret sharing (ASS) concept to design and transform each network layer of CNN. Our approach effectively mitigates ciphertext expansion rate while enhancing computational efficiency and reducing communication volume without compromising on inference accuracy.