Privacy-Preserving Convolution Neural Network Inference with Edge-assistance

Abstract

Nowadays the wide application of Convolution Neural Network(CNN) has promoted the rapid development of new intelligence fields. However, the inference of complex CNN usually includes a large number of computational operations. It is hard for the resource-limited user to complete the complicated CNN inference by his own. Privacy-preserving convolution neural network inference schemes allow the user to leverage the edge/cloud server to complete CNN inference under the condition that the privacy is protected. Nevertheless, existing schemes either require time-consuming cryptographic techniques, such as secure multi-party computing and homomorphic encryption, or fail to protect the privacy of the trained CNN model. To address this problem, we propose a novel privacy-preserving CNN inference scheme with edge-assistance. In this scheme, the privacy of the input data and the trained model is well protected and no complex cryptographic technique is involved. We blind the data to be predicted and the trained model to the edge servers, and then the edge servers calculate the most time-consuming layers. The user only deals with computationally efficient layers, fast encryption and recovery. Analysis and experimental results demonstrate that the proposed scheme is secure and efficient, which obviously saves the computational overhead of the user.