| Medical data is a key factor to promote the development of modern artificial intelligence medical care,and it is also an important basic strategic resource for the country.If the medical data of patients between various medical institutions can be effectively integrated,the disease prediction model can be trained and used rationally with the help of artificial intelligence,it can further promote the development of the medical industry and provide convenience for disease prediction,auxiliary diagnosis and treatment,telemedicine,scientific research,etc.However,due to the special sensitivities and importance of the medical field,the use and sharing of medical data has become a challenge for modern medicine.On the one hand,medical data involves the basic information of patients,physical condition,medical record information,medical insurance information,etc.,and some even involve medical experiments and scientific research data,which are not only related to the privacy and security of patients,but also related to industry development,even national security.On the other hand,due to the decentralized distribution characteristics of most medical institutions,coupled with increasingly stringent data security regulations,data cannot be communicated between medical institutions,and the amount of data in any hospital or data center is limited.Medical data forms many "data silos" among various institutions,so that a single institution lacks enough training samples for model training.Based on the thinking of the above problems,the main research contents of this thesis are as follows:(1)A multi-index federated aggregation algorithm based on trusted verification(TVFedmul)is proposed.This algorithm comprehensively considers the data quantity and data quality held by each participant to measure the weight of the participants in model aggregation.During aggregation,in order to truly reflect the training effect of each participant’s model,a public dataset is used for testing,and weights are calculated based on the test results.So the weights are dynamically adjusted according to the actual situation.At the same time,a hyperparameter is introduced to control the importance of the two factors,data quantity and data quality,so as to realize customized federated learning.The proposed algorithm has laid the algorithm foundation for the fourth chapter of the thesis.(2)A multi-hospital collaboration model based on blockchain and federated learning is proposed and the security verification mechanism is proposed.Based on the TVFedmul algorithm,the model combines the characteristics of blockchain that decentralization,non-tampering,traceability,and joint maintenance by multiple parties with the characteristics of federated learning that "data is available and invisible" and joint modeling by all parties to establish a set of distribution among multiple medical institutions.A secure and reliable collaboration mechanism.In addition,when building the system,a multi-party verification mechanism,consistent consensus and multi-party aggregated signatures are proposed.These three mechanisms not only improve the security and fairness of the entire system,but also reduce the operating cost. |
