A training technique that flips gradient signs to force a model to learn features that fool an adversarial classifier.