Customized attacks and defense strategies for robust and privacy-preserving federated learning

In this talk, we will review potential attacks against the robustness and privacy of federated learning with a particular interest in those customized to the actual setting or the underlying machine learning approach. We will first consider the existence of stragglers (slow, late-arriving clients) and their impact on the performance of the FL framework and study potential defense strategies. We will then focus on federated graph learning and explore dedicated attacks against and defenses for the privacy of the graph.