Deep Learning with Differential Privacy

Machine learning techniques based on neural networks are achieving remarkable\nresults in a wide variety of domains. Often, the training of models requires\nlarge, representative datasets, which may be crowdsourced and contain sensitive\ninformation. The models should not expose private information in these\ndatasets. Addressing this goal, we develop new algorithmic techniques for\nlearning and a refined analysis of privacy costs within the framework of\ndifferential privacy. Our implementation and experiments demonstrate that we\ncan train deep neural networks with non-convex objectives, under a modest\nprivacy budget, and at a manageable cost in software complexity, training\nefficiency, and model quality.\n