Research Of Privacy Preservation In Logistic Regression Based On Homomorphic Encryption

For all those years,with the development of the machine learning techniques,it has provided new solutions for many problems.However,though with its convenience brought to people,machine learning also has privacy and security issues at the same time.Nowadays,people have stressed more on privacy.Therefore,it has become a research focus on how to assure effective machine techniques while protecting users' private data.Homomorphic encryption technology is the key to solve this problem.But it's a very challenging task for machine learning in reality.The purpose of this essay is to provide actual support for some machine learning techniques in ciphertext,such as linear regression and logistic regression.Our research was based on homomorphic encryption on integer vectors and machine learning algorithm.In order to apply the machine learning algorithm in ciphertext,machine learning algorithm was decomposed into addition,multiplication,activation function and other parts,and the fitting of the activation function were studied.With a series of methods,the two problems in combination of machine learning algorithm and ciphertext were solved.We used Python to realize an effective scheme for integer vector homomorphic encryption and carried out expansion and optimization for encryption scheme.With the improvement of homomorphic encryption scheme,it can make machine learning algorithm trained in ciphertext to protect users' privacy.And it can also make data trained in an unsecured environment.At last,we proposed a training scheme based on ciphertext logistic regression model of integer vector homomorphic encryption.This scheme can protect training models and users' data information from leaking.We used real datasets to access the scheme and to prove the probability in efficiency and computational cost.Shown from the lab result and analysis,the scheme mentioned in this essay can protect users' privacy safely and effectively.Meanwhile,we speak highly of our scheme with high efficiency and accuracy.Based on the two-class ciphertext logistic regression model,we also extends it to multi-class logistic regression,and the experiments prove the correctness of this scheme.