Research On Privacy Machine Learning Method Based On Multi-key Homomorphism

In recent years,more and more machine learning algorithms rely on cloud computing.When machine learning is trained or classified in a cloud environment,it is easy to trigger malicious acquisition and utilization of data by attackers.Therefore,this research is supported by the National Natural Science Foundation of China.”Car network privacy protection and Byzantine fault-tolerant synergy fusion mechanism under big data environment”,Chongqing Municipal Education Commission Science and Technology Research Key Project ”Secure Transmission and Formal Verification Method of Car Network Information in Multi-source Heterogeneous Data Environment”,Chongqing Basic Science With the support of the frontier basic research project ”Research on the Collaborative Fusion Method of Mass Data Privacy Protection in the Internet of Vehicles in the Mountain Environment” and other funds,related research has been done on the privacy machine learning method based on the multi-key homomorphic method.The main work of the paper is as follows:1.In order to solve the problem of over-fitting training caused by the redundancy of encrypted data in machine learning,a privacy storage and communication scheme that can prevent redundancy on the cloud is constructed.This scheme uses number theory research unit to construct a Two-party repetitive data cleaning protocol.Specifically,the client provides a deduplication tag of the data to be uploaded,and the cloud server homomorphically checks whether there is redundancy in the tag list.In the case of duplication,the corresponding tag and the file in the server and the corresponding decryption key will be returned,in order to give the client the ability to download and decrypt files.Analysis shows that the security of this scheme can be reduced to the closest vector problem on the lattice,and it fully realizes semantic security.At the same time,the solution does not require any trusted third-party assumptions.2.On the basis of the plaintext data training model,combining with the input data real number encoding,the re-linearization technology,decoding and polynomial approximation of the activation function,the neural network regression of the encryption method is realized.The existing model is transformed into a ciphertextdriven model using a multi-key homomorphic encryption algorithm,which solves the problem of regression privacy in the machine learning process.In this scheme,the data can be encrypted with multi-key homomorphic,and the data prediction can be completed under the premise of semantic security.The experimental results show that,The regression accuracy after data encryption is very close to the prediction accuracy before encryption,and in the whole process,the real information of the server and the data provider will not be exposed to each other.3.In order to further protect the privacy of the machine learning training process in the multi-data provider scenario,a neural network privacy training and regression method for multi-party data and a single model at the same time is proposed.Each data provider performs an encrypted interaction with the cloud so as to realize the training and regression of private machine learning.Combining the difficulty of conjugate search problem and discrete logarithm problem,the confidentiality of training data and system model can be attributed to fully researched security assumptions.In terms of efficiency,due to all the information are encoded as low-dimensional matrices,compared with the plaintext implementation,the expansion rate of storage and calculation overhead is linear,and there is no loss of precision.Because this scheme is based on a fully homomorphic encryption algorithm,it is easy to be transplanted to other machine learning model involving multiple parties.