Explainable Adversarial Attacks on Coarse-to-Fine Classifiers