GANs can generate useful synthetic data, and differential privacy adds a mathematical guarantee that no individual leaks from the training set. Together they enable collaborative, privacy-preserving deep learning - here is how.