Unsupervised Domain Adaptation with Differentially Private Gradient Projection

Domain adaptation is a viable solution for deep learning with small data. However, domain models trained on data sensitive information may be violation of personal privacy. In this article, we proposed unsupervised adaptation, called DP-CUDA, which based differentially private gradient projection and contradistinguisher. Compared the traditional process, DP-CUDA involves searching domain-invariant features between source target first then transferring knowledge. Specifically, model in by supervised from labeled During training model, feature used to solve classification task an end-to-end manner using unlabeled directly, noise injected into gradient. We conducted extensive experiments variety benchmark datasets, including MNIST, USPS, SVHN, VisDA-2017, Office-31, Amazon Review, demonstrate our method’s utility privacy-preserving properties.