A collaborative privacy-preserving deep learning system in distributed mobile environment
In the last couple years, deep learning gained great popularity in health and medical science. For analyzing personal health data, privacy of patients and their data is one of the biggest concerns. Traditional methods have the possibility of leaking data because of transferring raw data and storing all data in centralized houseware. Therefore, we proposed a collaborative privacy-preserving learning system based on deep neural network, which does not share local raw data. The system is implemented on an XMPP server and several mobile devices. In the experiments, reconstructed rate is proposed to evaluate the performance of distributed system compared with centralized training. The rate is over 90% in different scenarios. Furthermore, the network traffic while collaborative learning is also measured.
