Research On Privacy-preserving Machine Learning In Medical Environment

The telemedicine system provides users with more convenient medical diagnosis.Among them,in terms of telemedicine for chronic diseases(e.g.,heart failure,diabetes),medical users upload personal data to medical service providers,and real-time diagnosis results are returned to users by the medical service providers.Nevertheless,medical data usually contain sensitive information.The medical service providers may disclose medical data without authorization.Therefore,it is significant to construct a privacy-protected telemedicine system.To solve the above problems,the research on the algorithm design of machine learning to protect privacy in the medical environment,and the following research results are achieved:(1)In response to the problem of user data privacy in remote heart failure diagnosis being leaked by medical service providers,a lightweight remote heart failure pre-diagnosis system that protects the privacy of medical users’ data is designed.First,the system avoids computationally intensive and complex calculations and reduces the computational overhead of the system.Secondly,it is optimized for the efficiency and accuracy of the system.When the privacy-protecting Sigmoid and tanh activation functions are designed,this paper uses the least square method to construct safe Sigmoid and tanh functions,which avoids the problem of low fitting accuracy caused by the use of the Taylor series.Finally,experiments show that compared with the existing secret-based sharing schemes,the system has significantly improved computational overhead and accuracy.(2)To solve the problem that the patient’s private data in the diabetes remote-assisted diagnosis and treatment decision-making system may be leaked without authorization,an efficient diabetes drug dosage decision-making system that protects the patient’s privacy data is designed.First,the system allows medical institutions to outsource diabetes data to edge servers to assist doctors in deciding on drug dosages,but the patient’s private data will not be leaked.Secondly,a series of new protocols based on secret sharing to realize the sub-operations of the system are designed.Compared with the existing agreement,the new agreement significantly improves efficiency.Finally,numerous experiments have confirmed the efficiency of the system.