Research On Outsourced Privacy-preserving Multi-party Computation

With the generation of massive data and the rapid development of the computer industry,cloud services have entered people's lives.It utilizes its powerful storage and computing ability to collect,manage and compute distributed digitized data.These data can be collected through the mobile Internet of things,the Internet,sensors,smart processing devices and so on.The emergence of cloud servers connects computing nodes in different locations and turns them into a service platform with strong computing power.Users can outsource their own data to the cloud server for entrusted computing.This not only frees users from self-burdensome computing,but also makes the results of big data computing more accurate and effective.As a commercial service model,cloud servers can indeed provide users with a variety of personalized services.However,while the cloud server provides us with convenient services,it also brings serious privacy leakage problems.Because the essence of cloud computing services is that users entrust cloud service providers to perform tasks that they cannot solve or are unwilling to solve.After being authorized by users,cloud servers can directly touch users' data anytime and anywhere,which will result in the disclosure of users' private data and sensitive information.At present,the mainly problems in the multi-party outsourced calculation are as follows:(1)Users' privacy cannot be protected.During some computing services,users' uploaded data often contain the privacy.Although sometimes cloud servers may follow the protocol to complete calculation,but cloud servers may through conspiracy attack or other malicious means to speculate about users' privacy or results from intermediate results or the final result,this poses a threat to the user's privacy.(2)The efficiency of the scheme is too low.At present,for the multi-party secure outsourced computing privacy-preserving schemes,fully homomorphic encryption is often used.Because by this way,the cloud server calculates on the encrypted data of the user and then decrypts the result,which is equal to the result calculated on the plaintext data of the users,but this leads a large amount of time consumption.(3)The correctness of cloud server's calculation results cannot be verified.For clients with weak computing power,it is difficult to verify the correctness of the results computed by the cloud server.(4)The secure outsourced privacy preserved scheme can only be applied to the scenario where a single user requests the cloud server for outsourced computing,but not to the scenario where multiple users' distributed data is sent to the cloud server for computing.However,at present,users' data is basically distributed,and cloud server's job is to deal with and compute these distributed data.Thus,for secure outsourced computing privacy-preserving schemes,there is an urgent need to solve the problem is to construct the corresponding data encryption method and validation algorithm,so it can ensure the safety of user privacy data and sensitive information,the correctness of the results calculated by the cloud servers can be verified by the authentication server,so as to realize the security of users' privacy of the data and the verifiability of the results.With the existing secure outsourced schemes of various important mathematical computation in the machine learning,this paper makes in-depth analyses of the problems and shortcomings in the existing schemes.Taking the gradient descent methods,matrix multiplication and non-negative factorization of matrix as the focus,we focus on the privacypreserving multi-party outsourced computation from the following aspects.(1)The design and analysis of the outsourced privacy-preserving multi-party gradient descent scheme(OPPGD).In the current research on secure outsourcing gradient descent privacy protection,the prediction model is owned by the cloud server when the user training samples with the help of the cloud server.But in real life,predictive models are usually kept by the model owner rather than the cloud servers.At the same time,the existing schemes are not applicable to all gradient descent methods.Aiming at the above defects,this paper uses El Gamal encryption algorithm and Gentry's Evaluation circuit to improve the existing secure outsourcing gradient descent privacy protection scheme,and proposes a new multi-party outsourcing gradient descent privacy protection scheme.The scheme realizes that the training model is owned by the model owner instead of the cloud server,and the training dataset is a collection of samples from multiple data owners.The dataset can be partitioned either horizontally or vertically.At the same time,the model owner can use any gradient descent method he wants to optimize the training model,which is impossible in all the previous schemes.According to the correctness and privacy-preserving analyses,it can be concluded that the multi-party outsourcing gradient descent privacy-preserving scheme constructed in this paper can not only protect the data's privacy of the data owner and the training model's privacy of the model owner in the whole training process.At the same time,the scheme can also be applied to all the outsourced gradient descent methods.This greatly reduces the particularity of the multi-party outsourced gradient descent privacy-preserving scheme and improves its applicability and universality.(2)The design and analysis of the outsourced privacy-preserving multi-party matrix multiplication scheme(OPPMM).In the current research on privacy protection of security outsourcing matrix multiplication,a single user preprocesses the two matrices they want to multiply and sends them to the cloud server.The cloud server then multiplies the two matrices,and finally returns the encrypted results back to the user.The user can get the desired final results after decrypting the encrypted results.But today,when cloud servers perform machine learning for users,they often compute the product of multiple matrices which are from different users.However,the existing secure outsourcing privacy-preserving schemes can neither solve the problem of multiple matrix multiplication nor solve the problem of multiple users.Therefore,aiming at the existing problems,this paper constructs a new privacy-preserving protocol for multi-party outsourcing matrix multiplication by using Kronecker function.The protocol improves the existing schemes for secure outsourcing matrix multiplication.The scheme proposed in this paper not only realizes that the cloud server can calculate the multiplication of matrix from different users,but also that the correctness of the calculation results of the cloud server can be verified by the verification server.This paper makes a detailed theoretical proof,and concludes that if the result of cloud server is wrong or maliciously fails to execute the protocol,the result of cloud server will always fail to pass the verification of the verification server with a probability that cannot be ignored.Therefore,the outsourced privacy-preserving multi-party matrix multiplication scheme proposed in this paper not only enhances the applicability of the scheme but also guarantees the verifiability of the cloud server results.(3)The design and analysis of the outsourced privacy-preserving multi-party matrix factorization scheme(OPPMF).In the current research of secure outsourcing matrix factorization privacy-preserving schemes,most of the solutions are focus on the same scenario: a single user preprocesses the matrix he wants to decompose and sends it to the cloud server for delegate computing.The cloud server returns the encrypted results to the user after making calculations,and then the user can get the desired results after decrypting them.However,in machine learning,especially in recommendation systems,the user-rating matrix decomposed by the cloud server is usually a matrix with a large dimension,and the matrix is composed of samples of different users.The matrix can be divided either horizontally or vertically.At present,the only solution is using the fully homomorphic encryption,but this leads to high complexity,and the cloud server results are not available to the valid validation.Aiming at this scenario and the problem in the existing scheme,this paper proposes a new privacy preserving scheme of multi-party outsourcing matrix factorization by using Kronecker function.The scheme can not only adapt to the scenario of multi-party outsourcing matrix factorization but also avoid using fully homomorphic encryption technology.The scheme can validate the correctness of the results calculated by the cloud server.According to the correctness proof and the privacy preserving proof of the scheme,and the experimental evaluation results,we can conclude that the scheme proposed in this paper not only guarantees the correctness and the privacy of users,but also owns high efficiency.