Towards Effective and Robust Neural Trojan Defenses via Input Filtering