Generative Neural Network Mixture Models for Differential Privacy

The researchers propose a differential privacy technique that uses a mixture of k generative neural networks to publish generative models and high-dimensional datasets in a privacy-preserving manner. The data is first clustered into k clusters via differentially private kernel k-means, and each cluster is assigned to a separate restricted Boltzmann machine or variational autoencoder, trained with differentially private gradient descent. Evaluation on MNIST, call detail records, and public transportation datasets shows that the synthetic samples generated by this method are realistic and that it can accurately compute any number of counting queries.