Prediction Of Private Data By Convolutional Neural Network Based On Homomorphic Encryption

With the progress of the digital age,and machine learning to predict based on cloud computing is also becoming more and more widely,but the traditional machine learning algorithms need to access the raw data,it brings a potential security risk to the user data,especially some private data,such as medical records,financial data and other sources of information into a third party.In order to protect private data from disclosure and realize secure machine learning prediction,the original data can be encrypted before the prediction.Homomorphic encryption technology can calculate the encrypted private data,and this privacy protection method makes the security prediction possible.However,because of the huge overhead of the encryption scheme and the large number of parameters of the neural network,it brings a huge amount of computation and requires a lot of computation time.This paper presents a faster neural network prediction scheme for ciphertext data.In order to predict the image data,this paper starts from the convolutional neural network,and improves each level of the network to make it more friendly to use with the homomorphic encryption technology.In order to make the prediction of neural network under homomorphic encryption possible,for the activation function layer in the prediction model,this paper selects the square function with multiplication operation only,and for the pooling layer in the prediction model,this paper selects the average pool with addition and multiplication operation only.In order to speed up the neural network prediction under homomorphic encryption scheme,for homomorphic encryption cipher and the largest amount of calculation in the operation between ciphertext multiplication,in this paper,the activation function is optimized,by introducing more convolution kernel layer to reduce the activation function of the input data,thereby reducing activation function layer ciphertext in cipher multiplication calculation;For the classification layer with the largest number of parameters in the prediction network model,a more homomorphic classification layer,namely the global average pooling layer,is adopted.The whole network model becomes a homomorphic encryption-friendly neural network structure.And the appropriate homomorphic encryption scheme is selected to encrypt the user's data,and the secure neural network prediction is realized.Finally,the experiment proves that,on the premise of basically maintaining the accuracy of the prediction model,it not only ensures the data security of users and cloud server,but also realizes the efficient prediction of ciphertext image data.