Predicting the evolution of a dynamic network—the addition of new edges and the removal of existing edges—is challenging. In part, this is because: (1) networks are often noisy; (2) various performance measures emphasize different aspects of prediction; and (3) it is not clear which network features are useful for prediction. To address these challenges, we develop a novel framework for robust dynamic network prediction using an adversarial formulation that leverages both edge-based and global network features to make predictions. We conduct experiments on five distinct dynamic network datasets to show the superiority of our approach compared to state-of-the-art methods.