Adversarial Training of Anti-Distilled Neural Network with Semantic Regulation of Class Confidence