Research On Key Technologies Of Privacy-preserving Machine Learning Based On Homomorphic Encryption

At present,machine learning is widely used in many application scenarios,such as image recognition,speech recognition,text processing,etc.These machine learning models tend to be more accurate when training and learning massive amounts of data collected from different sources.However,the collection of massive amounts of data often leads to concerns about data privacy breaches.Especially with the popularization of cloud computing,data owners tend to outsource their data and machine learning models to clouds with powerful resources.However,direct outsourcing can compromise data privacy because the cloud is not entirely trustworthy.To avoid privacy leaks,a good strategy is to encrypt the data before uploading it to the cloud.However,it also poses a new challenge to implement machine learning algorithms in ciphertext domain.In particular,non-interactive measures of similarity between data sets are critical to ensuring the functionality and efficiency of machine learning.This thesis presents three typical machine learning algorithms based on Vector Homomorphic Encryption(VHE),including classification,clustering and regression.In this thesis,three typical machine learning algorithms based on vector homomorphic encryption are studied.The main research content and results of this paper mainly include the following aspects:1.Study the homomorphic encryption scheme suitable for machine learning.Since the basic elements of many machine learning operations are vectors,we study efficient homomorphic encryption scheme based on vector.The homomorphic encryption scheme mainly includes homomorphic addition,homomorphic multiplication,homomorphic linear transformation,homomorphic weighted inner product and so on of support vectors.2.Study the classification method of k NN based on vector homomorphic encryption.The efficient k NN classification method in outsourcing environment is studied.On this basis,we improve the scheme so that our scheme can be parallelized.At the same time,we study the encryption scheme of the function to ensure the privacy of the operation function.3.Study the clustering method of k-means based on vector homomorphic encryption.This paper studies the clustering method of k-means with completely outsourced security.Therefore,we need to study the similarity measurement scheme of vectors in ciphertext domain.On the basis of ensuring accuracy,this paper studies how to avoid non-integer operation in ciphertext domain.4.Study the linear regression method based on vector homomorphic encryption.The linear regression method of complete outsourcing security is studied.Therefore,we need to study the loss function of privacy protection and the least square method.How to use gradient descent method to optimize the loss function under the premise of data privacy is studied.