ALittleRobustnessGoesaLongWay: Leveraging RobustFeaturesforTargetedTransferAttacks

However, we believe that targeted attacks are especially important for understanding neural-network classifiers, as they provide a tool to compare the features of two models. When a targeted attack transfers from one network to another, it suggests that the two networks rely on similar information for classification, and that they use the information in the same way.