Research Of Privacy Protection In The Data Usage Phase Of Intelligent Medical Cloud Platform

Intelligent medical is unprecedentedly hot,and various medical cloud platforms rise rapidly.The data use stage makes the medical cloud platform give full play to the value of medical data.However,medical data contains a large amount of personal privacy information,and improper data processing will lead to the disclosure of personal privacy.Meanwhile,due to the value of its data,the methods and number of attacks against medical data increase significantly.Data using the primary means of data release,intelligent diagnosis model applications and statistical histogram and so on,based on the above three kinds of using phase of data security and data security and availability cannot be effectively balance problems,according to the "data standardization-data privacy protection-privacy data safety training-classification samples after security histogram" research train of thought,from the perspective of the nature of the data to the data used at various stages,total contribution and innovation make the following aspects:(1)In view of the general deviation of medical data quality and the problems of traditional normalization methods,an improved normalization processing method is proposed,which improves the problem of min-max non-zero center and the relatively complex z-score calculation in traditional normalization methods.Quality of medical data.And data normalization and DP(Differential Privacy)are associated,thereby simplifying the application of differential privacy in data privacy protection.At the same time,the sensitivity in differential privacy protection is fixed through the normalization process,and then the test accuracy of the data after differential privacy processing is used to mediate the scope of the specification in this normalization,which reduces the negative impact of differential privacy on the data and increases the flexibility of normalization.(2)Aiming at the problem that the traditional differential privacy will protect the data after the release of data protection,this paper will combine the differential privacy protection and the decision tree model,and propose the DPDT(Differential Privacy and Decision Tree)algorithm.The DPDT algorithm establishes an attribute weight calculation system based on CART(Classification and Regression Tree),which uses attribute weights to change the traditional differential privacy noise addition method,thereby further reducing the negative impact on the availability of medical data during the privacy protection process.Medical data release provides strong privacy guarantee.(3)Aiming at the problems of gradient bounce and accuracy drop after differential privacy on the intelligent diagnosis model security protection,this paper combines differential privacy protection with minimum batch gradient descent and proposes DPMB(Differentially Private Mini-Batch Gradient Descent Algorithm)algorithm.The DPMB algorithm provides the privacy protection of the underlying training data for the intelligent diagnosis model,which effectively prevents the theft of training data by attack methods such as member attacks and gradient inversion attacks.The dynamic learning rate is added to the DPMB algorithm,which solves the problem of too large gradient jump caused by noise addition when the model is near-fitting,speeds up the model’s fitting speed,and improves the fitting quality.At the same time,the moment accountant is used to limit the privacy loss of the DPMB algorithm and the excessive addition of noise in differential privacy protection,which enhances the security and accuracy of the model.(4)For the application of statistical histogram,based on the big data of the "Rainbow Cloud Health Service System" platform,the user data is digitally processed and some of the data information is hidden.Afterwards,DPDT and DPMB algorithms are used for privacy processing and classification of the data.The statistical analysis results of the original data are compared with the statistical analysis results of DPDT processing and DPMB classification,which proves that DPDT and DPMB ensure the safety of data and models,and do not affect the overall trend of statistical analysis.