Hardening Inline DGA Classifiers Against Adversarial Attacks