This article says that the training set was small. And presumably the misclassified images were unlike images in the training set. If the training set includes

This article says that the training set was small. And presumably the misclassified images were unlike images in the training set. If the training set includes images intended to mislead, then wouldn't the classifier then be more tolerant?