PubSub-ML: A Model Streaming Alternative to Federated Learning

Federated learning is a decentralized framework where participating sites are engaged in tight collaboration, forcing them into symmetric sharing and the agreement terms of data samples, feature spaces, model types architectures, privacy settings, training processes. We propose PubSub-ML, Publish-Subscribe for Machine Learning, as solution loose collaboration setting each site maintains local autonomy on these decisions. In either publisher or subscriber both. The publishers publish differentially private machine models subscribers subscribe to published order construct customized use, essentially benefiting from other sites' by distilling knowledge publishers' while respecting privacy. term “model streaming” comes extension PubSub-ML streams with concept drift. Our extensive empirical evaluation shows that outperforms federated methods significant margin.