Federated Machine Learning (FL) can be used effectively in distributed datasets, where data owners hesitate to share their raw data, as a reliable approach to train an ML algorithm. However, in the case of sensitive healthcare datasets, additional privacy measures before feeding into machine learning mechanisms are also necessary. Our approach uses the federated learning framework, which removes the necessity of sharing patients' sensitive data in a raw format outside the premise. First, the data owners agree on a list of features selected by the correlation; then, after training the local models, the obtained local models are transmitted to the central server for aggregation. The differential privacy (DP) approach is adopted to perturb the local models before transmission to add an extra privacy layer. As a result, our framework achieves improved utility as the feature selection reduces the data dimension. Finally, based on the patient's genomic data, the framework establishes a practical healthcare application to privacy-predict certain heart failure/cancer diseases.
