A differentially private "Draw and Discard" machine learning framework

The authors propose "Draw and Discard," a differential privacy framework for client-side distributed machine learning that provides differential privacy guarantees under the local privacy model and uses asynchronous client-server communication. The framework achieves load distribution and scalability through random model sampling, while also bringing server-side privacy protection and improved model quality through averaging. The authors apply it to generalized linear model learning, analyze privacy guarantees against multiple adversaries, and experiments show the framework is feasible in practical deployment.