N-CRiPT Public Seminar: Trustworthy Federated Learning

Speaker: NUS Presidential Young Professor Reza Shokri

Date & Time: Thursday 26 November, 3.00pm – 4.00pm

Federated learning enables multiple parties to train a global model on all their data, without sharing their local data with each other. This keeps all data private but would not make federated learning a privacy-preserving scheme. The information exchange between parties in federated learning indirectly leaks a significant amount of information about the parties’ sensitive data. In this talk, we provide an overview of what federated learning is, show how it leaks private information, and discuss the way forward for building algorithms that can be trusted as privacy-preserving. We will also analyze the robustness of federated learning algorithms with respect to poisoning attacks where a subset of parties try to manipulate the global model. We then discuss ideas on how to build robust federated learning algorithms.