Research And Development Of An Encrypted Federated Learning Platform For Electronic Medical Record Data

With the advancement and rise of medical informatization and big data,medical institutions obtain well-performing machine learning models by sharing electronic medical record data.However,electronic medical record data contains a large amount of private information of patients.With the promulgation of relevant privacy protection laws and the enhancement of people’s awareness of privacy protection,medical institutions cannot share relevant electronic medical record data sets,thus forming "data islands",leading to medical treatment and data cannot be used to the maximum.Although the federated learning proposed by Google allows the client to learn the shared model in a collaborative manner when the data is not available locally,it only protects the privacy of the training data to a certain extent and is a potential adversary through existing attack methods;therefore,the model optimization parameters of the client can be stolen by corrupting the server to restore the local training data.This paper mainly studies the problem of model training with privacy protection in the federated learning environment.Aiming at the "data islands" between medical institutions and the security issues of federated learning server-side model parameters,we propose a horizontal federated learning framework based on multi-key homomorphic encryption.First of all,the solution we designed mainly deals with the situation that the feature space of the medical data set is basically the same,and the sample space is different.Second,the federated learning framework in this paper uses a multi-key homomorphic encryption scheme to protect the security of model parameters in the server-side and then protect the privacy of the client’s training data,because multi-key homomorphic encryption has semantic security.Third,in the face of the multi-user scenario of federated learning,each client holds its own public and private key pair,which not only breaks the restriction that multiple users need to share the same public and private key pair in advance,but also facilitates the client to join the model update at any time,and increase.This improves the flexibility and scalability of the system,and improves the overall safety of the training process.Fourth,the core of the solution is to protect the security of the server-side model aggregation calculation,so we set the server side as two non-colluding dual servers,and only one of the servers can interact with the client to prevent collusion between the server and the client.to enhance the security of the program.Fifth,experimental analysis shows that under the premise of ensuring the safety of model parameters,the accuracy of the model is not much different from the accuracy of the model trained in the plaintext state.Finally,we designed and developed an encrypted federated learning platform for the privacy protection of electronic medical record data based on the above scheme.Through demand analysis and detailed design of the platform,the functional architecture and technical architecture of the platform are determined.The main part of the platform is developed in Java language,and core functions such as machine model training and model parameter encryption are implemented in Python language.Finally,we tested the functions of the platform and demonstrated the usability of the platform’s visual interface to users.The test results proved that the usability of the platform met the requirements.