In binary classification, this is basically a precision/recall trade-off. More precision (ie, fewer false positives) necessarily means less recall (ie, more fal

In binary classification, this is basically a precision/recall trade-off. More precision (ie, fewer false positives) necessarily means less recall (ie, more false negatives). Classifier statistics like AUC basically tell you how good the classifier is in terms of how bad the trade-off is. Adversarial situations - any time humans are trying to trick your classifier, in other words - will always make the trade-off worse.