| Designing medical systems that can automatically diagnose patient's conditions from test data can greatly improve healthcare systems. With the help of machine learning tools and differential privacy consideration, this system can be made more efficient and powerful. Empirical risk minimization is a common and useful technique with which we can obtain a good approximation of globally optimal classifier and thus give good statistical classification result. Firstly we introduce three models for medical data learning and two methods for distributed model. Then we compare a novel distributed classification method which we called the "feature method" with traditional averaging method on different real world data sets to gain an insight into their performance and properties. Next we give analysis on the performance of the feature method under non-private and differentially private conditions and conduct some experiments to draw several important conclusions from them. Finally we conclude that the distributed learning system we recommend achieve the best result among the three models. |
