Towards Assessing and Characterizing the Semantic Robustness of Face Recognition