A Multifunctional Secure Outsourcing Computing Tool Based On Homomorphic Encryption And Its Application On Machine Learning

The enormous strides of big data technologies are expected to revolutionize the daily life of human beings,with massive data generated by digital devices every day.As a computing task requiring a lot of computing power and resources,machine learning can help users to mine potentially valuable information in data and provide suggestions for users to make decisions.However,many users are limited by their own limited local resources so that they are unable to carry out mass data storage and computing tasks.Cloud computing,as a rapidly emerging technology,can provide massive storage resources and sufficient computing capacity.Nevertheless,cloud servers are not always trusted,the security of outsourced computing has become an important issue that people are increasingly concerned about.Therefore,how to design effective secure outsourcing computing tools and realize the outsourcing computing of machine learning algorithms such as clustering and neural network under the premise of protecting privacy is a big challenge for researchers.In addition,current secure outsourcing clustering schemes only provide clustering results,but do not provide the evaluation of clustering results,so that users with limited resources cannot obtain the quality of clustering results.On the other hand,the emergence of model reverse attacks makes it necessary to protect the privacy of both data and models when training neural network models in the cloud.In response to the above problems and challenges,this paper proposes a multifunctional secure outsourcing computing tool supporting real number operations based on homomorphic encryption,and implements the secure outsourcing computing of machine learning algorithms including clustering and neural network.The main contributions of this paper are summarized as follows:1.This paper implements a multi-functional secure outsourcing computing tool that can support real number operations.Based on the Paillier Cryptosystem With Partial Decryption and dual-server architecture,this paper realizes the storage of real number ciphertexts and the secure outsourcing computing protocols such as addition,subtraction,multiplication and division.Furthermore,this paper provides mutual transformation protocols between different types of ciphertext,so as to balance efficiency and precision in complex computing tasks.2.This paper implements an estimable clustering algorithm on ciphertexts.In this paper,homomorphic K-means clustering algorithm and homomorphic hierarchical clustering algorithms are realized.Furthermore,homomorphic silhouette coefficient is provided which can evaluate clustering results on ciphertexts,so that the user can obtain the evaluation of the clustering results while obtaining the clustering results.3.This paper implements homomorphic neural networks supporting inference and training.This paper constructs secure forward propagation and back propagation algorithms on ciphertexts for common neural network modules such as dense layer and convolution layer,so that the cloud server can perform model training or inference tasks requested by user on the premise of protecting the privacy of data and neural network model.In this paper,comprehensive experiments are performed on the proposed multifunctional secure outsourcing computing tool,and its security is proved.From the perspective of computing time and communication cost,this paper compare and analyze the secure outsourcing computing protocols of different ciphertext types.Moreover,the effects of different parameters on the computing efficiency of homomorphic clustering algorithms and homomorphic neural networks are studied.