Research On Prediction Scheme Of Lightweight Convolutional Neural Network With Privacy Protection

In recent years,machine learning and deep learning have developed rapidly,and they have a wide range of applications in many fields,such as finance,medical treatment,face recognition,and autonomous driving.Because the calculation of prediction is very large,with the development of cloud computing,enterprises and individuals can obtain many benefits.Although powerful computing and storage capabilities of cloud servers can be used to implement outsourced computing,outsourced services often have privacy issues.We consider how to continue to use neural networks for prediction services while protecting user privacy.Consider a scenario in the medical field,where a user wants to use a cloud server to predict whether he has a certain disease.If a user uploads his personal health information directly to a cloud service,his privacy will inevitably be leaked.In view of the problems,this paper takes advantage of the homomorphic encryption to realize the privacy preserving convolutional neural network prediction task for the scenario in the medical field.The main contributions of this paper are as follows:· Lightweight convolutional neural network prediction algorithm based on FV encryp-tion: This thesis analyzes the shortcomings of the Crypto Nets scheme proposed by Mi-crosoft.In deep learning,feature complexity is usually positively related to the number of neural network layers.The more layers of the network,the more complex the features that can be extracted,and the higher the accuracy of the model.In order to reduce the multipli-cation operation in homomorphic encryption,Crypto Nets uses a shallow neural network,so it is not suitable for complex scenarios.This thesis proposes to apply the FV scheme to the lightweight convolutional neural network,which deepens the neural network level,trims the convolutional neural network without affecting the prediction accuracy,and implements a deep convolutional neural network privacy protection prediction scheme.It makes up for the shortcomings of shallow neural network in the Crypto Nets,and is suitable for more com-plex neural network prediction tasks.The thesis explained the selection of parameters in the FV scheme,selected different parameters,and made comparative experiments on their efficiency.· Lightweight convolutional neural network prediction algorithm based on Paillier encryption:This thesis improves on the previous scheme,and proposes to use the Paillier scheme on a lightweight convolutional neural network to achieve a privacy preserving prediction scheme.Paillier scheme is additive homomorphic encryption scheme,which is more efficient than FV.But if we want to use Paillier encryption to achieve ciphertext prediction,we can't do it with only one server.We propose a ciphertext multiplication protocol and a ciphertext division protocol.Based on these two protocols,we propose two scenarios,using two different convolutional neural networks to achieve security prediction.The same as the previous scheme,the convolutional neural network needs to be trimmed to realize the prediction scheme of the lightweight convolutional neural network with privacy protection.