Defending Against Saddle Point Attack in Byzantine-Robust Distributed Learning