Towards Certifying the Asymmetric Robustness for Neural Networks: Quantification and Applications