Neural Network Privacy-preserving Computing Approach Based On Homomorphic Encryption

Machine learning algorithms represented by deep neural network play an important role in many fields with their excellent feature learning capabilities.However,training and using a neural network model requires more complex expertise in understanding the way neural networks operate and compute results.Meanwhile,users are faced with the problem of insufficient computing power.In order to solve these problems,they upload a large amount of data to cloud service providers(outsourcing processing).This service is called Machine Learning as a Service(MLaa S).However,the application of this service is restricted because these data contain sensitive information.It is risk to leak the data privacy.The emergence of homomorphic encryption provides a solution for privatepreserving computing in cloud services,but encrypted ciphertexts are not suitable to compute complex nonlinear operations in the application of homomorphic encryption technology.There are two methods to deal with the limitations of homomorphic encryption,one constructs a security protocol suitable for homomorphic encryption algorithms,the other replaces complex nonlinear operations with approximate operations.In order to apply homomorphic encryption in neural networks,the research is divided into two parts.And two different schemes are designed to complete encrypted computing,based on the above two methods:(1)Design a non-collusion dual-cloud system based on the method of constructing security protocol and partially homomorphic encryption algorithm.The system consists of the client and two cloud servers,which solves the limitations of homomorphic encryption operations in the neural network model,and has a secure collaborative processing mechanism: protects the user's data privacy through a homomorphic encryption algorithm,avoids the nonlinear operations in the homomorphic operation through a three-party protocol,and protects the private knowledge of the model by introducing fake neurons.In a word,this scheme strengthens the privacy protection mechanism,and reduces the prediction gap between ciphertext and plaintext.Further,it also improves efficiency compared with the existing schemes.And the usability of the system is verified through experiments.(2)Construct an encrypted neural network computing scheme based on the method of replacing complex operations and fully homomorphic encryption algorithm.The scheme consists of several modules: model training,preprocessing,data encryption,ciphertext calculation,and approximate activation.Data privacy is protected by encrypting data using a variant of the BFV encryption algorithm.The new encryption scheme allows the neural network model to increase the noise more slowly when computing the encrypted data.By replacing the nonlinear activation functions with polynomial functions,homomorphic encryption is well compatible with complex operations.This scheme provides an analysis of the polynomial function fitting method,and obtains the performance of different polynomial functions in the model.At end,the experiment accomplishes a higher prediction accuracy.Finally,the differences between the two schemes are compared and analyzed from multiple dimensions.Besides,thoughts and future research directions are given in the neural network privacy-preserving computing.