Enhancing Transformation-Based Defenses Against Adversarial Attacks with a Distribution Classifier