Training Medical-Diagnosis Neural Networks on the Cloud with Privacy-Sensitive Patient Data from Multiple Clients.

Abstract

Artificial neural networks (ANNs) are changing the paradigm in medical diagnosis. However, it remains an open problem how to outsource the model training operations to the cloud while protecting the privacy of distributed patient data. Homomorphic encryption suffers from high overhead over data independently encrypted from numerous sources, differential privacy introduces a high level of noise which drastically increases the number of patient records needed to train a model, while federated learning requires all participants to perform synchronized local training that counters our goal of outsourcing all training operations to the cloud. This paper proposes to use matrix masking for outsourcing all model training operations to the cloud with privacy protection. After outsourcing their masked data to the cloud, the clients do not need to coordinate and perform any local training operations. The accuracy of the models trained by the cloud from the masked data is comparable to the accuracy of the optimal benchmark models that are trained directly from the original raw data. Our results are confirmed by experimental studies on privacy-preserving cloud training of medical-diagnosis neural network models based on real-world Alzheimer's disease data and Parkinson's disease data.