| Today,the financial services industry generally uses machine learning techniques to build various models to predict situations ranging from financial transaction fraud to investment and targeted marketing activities.The use of logistic regression algorithms for supervised learning is a commonly used machine learning technique for this model.Before the actual learning stage,it is usually necessary to share and prepare a large amount of data with other data providers.Due to the requirements of privacy laws and confidentiality regulations,the data must be stored in the system and cannot be directly outsourced.Therefore,the privacy data needs to be calculated confidentially.Commonly used secret computing methods are differential privacy,federal learning,homomorphic encryption,and multi-party secure computing.The research content of this article is a multi-party joint learning method for privacy protection.The biggest contribution of this paper is to propose a machine learning solution based on homomorphic encryption,design a homomorphic logistic regression algorithm in combination with the logistic regression algorithm commonly used in the industry,and implement a multi-party based on Alibaba Cloud's PAI(Platform of Artificial Intelligence)platform Parallel homomorphic machine learning system,the system integrates MPI(Message Passing Interface)framework to provide parallel computing.We analyzed the application scena rios of homomorphic encryption and multi-party secure computing,introduced the background knowledge of homomorphic encryption in detail,and selected the fully homomorphic encrypted CKKS scheme to encrypt private data.We use the least squares fitting polynomial method to approximate the activation function,and use fusion coding technology to reduce the multiplication depth of the training process and reduce the algorithm complexity.We extended from two-party data communication protocol to multi-party data communication protocol,and designed a homomorphic machine learning model from two parties to multiple parties.Combined with Microsoft's SEAL library,a homomorphic machine learning system is implemented.The whole system is implemented in C++ coding,which runs fast and has strong system compatibility.It can be expanded into a homomorphic encrypted machine learning library.The resulting homomorphic machine learning system provides a drag-and-drop interface that allows users to easily build encrypted training and prediction processes without having to have expertise in cryptography.All they have to do is find other users to collaborate with,drag their data tables together,drag the machine learning module into the project,and wait for the results.T he experimental results of this paper on the public data set show that using 5 machines,a logistic regression model can be trained on the 4096?576 data set in 129 seconds.Under the experimental parameters,the system provides 128 bits of security,and the deviation from the correctness of the plaintext logistic regression algorithm is about 1%,which verifies the feasibility and accuracy of the algorithm. |
