Binary Convolutional Neural Network Based On Homomorphic Encryption

With the rapid development of computers and information technologies,more and more users store massive data on cloud platforms and enjoy the convenience of cloud computing.The data stored on the cloud platform is always involving users' privacy.once the data is stolen,it will bring huge losses to users.In response to this problem,scholars and experts have proposed neural network models based on homomorphic encryption to complete inference tasks on the cloud.The user uses the homomorphic encryption technology to encrypt the private data,and upload it to the cloud platform.The cloud platform processes the ciphertext without decryption,and then sends the result to the user.Due to the high complexity of homomorphic encryption technology,the calculation efficiency of ciphertext on the cloud platform is low.At the same time,the size of homomorphic ciphertext state is large,which increases the pressure of data transmission and cloud platform storage space.Besides,the existing homomorphic operations are limited.This results that the pooling layer and activation function of the neural network model calculate the ciphertext inefficiently.In order to solve these problems,this thesis has carried out in-depth research.The main research work is as follows:1)To solve the problem of low computational efficiency of neural network on cloud platform,this thesis uses the ring of the homomorphic encryption(Fully Homomorphic Encryption over the Torus,TFHE)to complete data encryption and proposes the binary convolutional neural network based on TFHE.In this thesis,the parameters of model's convolution layer and inputs data are binarized in binary convolutional neural network.When the ciphertext data and the plaintext parameters are performed convolutional operations in first layer,this thesis uses a homomorphic XNOR gate operation to replace the multiply operation.This improves the model's computational efficiency.In the pooling layer,this thesis uses the bitwise logic gates to improve the efficiency of the ciphertext comparison operation.In the activation function,the bitwise calculation of the AND gate and the NOT gate are used to implement the Re LU operation,which increases the nonlinear expression ability of the neural network model.Compared with the existing neural network models based on homomorphic encryption,the calculation speed of the model on the MNIST and Wisconsin Breast Cancer datasets has been increased by 1.4 and 3.6 times.2)To solve the problem of high computational complexity and large size of ciphertext in fully homomorphic encryption,this thesis proposed a binary convolutional neural network based on ciphertext conversion.In this model,this thesise uses TFHE leveled homomorphic scheme to encrypt user data.Compared with existing models,the size of the ciphertext has been reduced to 1/6 of its original size.In the convolutional layer,the ciphertext is calculated in a leveled homomorphic form to improve the calculation efficiency.Then,this thesis uses ciphertext conversion operation to convert the leveled homomorphic ciphertext into the fully homomorphic ciphertext,which can support bitwise computation.Compared with the previous neural network model,the test results show that the calculation speed of the model on the MNIST data set is significantly improved,and the recognition rate increases to 97.27%.In this thesis,the scheme is proposed to solve the problem of high complexity of homomorphic encryption,the non-linearity calculation and comparison operation of neural network.The experimental results show that the binary convolutional neural network based on TFHE homomorphic encryption not only guarantees the data security but also significantly improves the computational speed of the neural network on the ciphertext.After the binarization operation,the size of the ciphertext is reduced.This scheme surpasses many current neural network models based on homomorphic encryption in recognition rate and calculation efficiency,and has certain practical value.