Fig. 2
From: Secure machine learning against adversarial samples at test time
DeepFool. The DeepFool attack approximates the decision boundary close to a targeted instance (triangle) by a linear line and searches for the shortest distance to cross the boundary and achieve targeted misclassification