Privacy in machine learning

Rich data is privacy sensitive, which may preclude uploading to the data center. New approaches have been proposed to decouple the training process from the access to the training data using decentralized optimization. In this case, data is allowed to be distributed locally, and individual processors are training under coordination of the master server. However, the server still has to aggregate update information from all processors, through which the privacy is going to be at risk as it may be revealed by the update information. My work is going to research to which extent is this privacy leakage and how to prevent it.