Attacks Meet Interpretability: Attribute-steered Detection of Adversarial Samples

Adversarial sample attacks perturb benign inputs to induce DNN misbehaviors. Recent research has demonstrated the widespread presence and the devastating consequences of such attacks.