Evaluating Adversarial Robustness on Document Image Classification