Research And Implementation Of Secure Multinomial Classification Logistic Regression Model Based On Homomorphic Encryption

With the widespread use of cloud computing,users are increasingly inclined to outsource complex machine learning tasks to cloud servers.However,how to protect the privacy of data on cloud servers has become an important issue.Homomorphic encryption technology can operate ciphertext without revealing plaintext information and get the same results as corresponding operations on plaintext after decryption.Therefore,homomorphic encryption is a feasible and potential solution for security outsourcing computing.However,the limitation of homomorphic operation and slow efficiency in homomorphic encryption scheme make it difficult to be widely used in business and daily life.At present,the research of privacy protection machine learning based on homomorphic encryption is constantly developing,but there is no secure multinomial classification model based on homomorphic encryption.The main work in this paper is as follows:(1)Analyze the process of encryption,decryption,homomorphic evaluation and rotation of HEAAN scheme,give the main functions of open source library and the operation process of homomorphic computation using this software library.(2)Analyze the logistic regression model and the approximate calculation method of sigmoid function through polynomial approximation,pack data matrix into a ciphertext to improve efficiency by using effective coding methods.(3)Use the OvR strategy to extend the two-class model to the multi-class model.And design an effective multinomial classification logistic regression model based on HEAAN scheme to protect data privacy,implement the security outsourcing computing.Analyze how to complete homomorphic evaluation of gradient descent method through basic operation of the HEAAN scheme,including the inner product calculation,the polynomial approximate calculation of the sigmoid function and the weight update process of ciphertext.(4)According to the idea of software engineering,implement the evaluation of security multinomial classification logistic regression model based on homomorphic encryption,and analyze each class and important function in detail.(5)Analyze the application scene of privacy protection multinomial classification problem,use the representative dataset to make a experiment on the model and record the test results,and make a comparative analysis with the multinomial classification model in plaintext.The results show that the designed secure multinomial classification logistic regression model can not only achieve privacy protection,but also the prediction accuracy is close to the calculation results on plaintext data.Through the above-mentioned work,a secure multinomial classification logistic regression model based on homomorphic encryption is designed.Homomorphic encryption is applied to machine learning to solve the multinomial classification problem of sensitive data,and the accuracy and feasibility of the model are verified through experiments.