Beating Attackers At Their Own Games: Adversarial Example Detection Using Adversarial Gradient Directions

Adversarial examples are input that specifically crafted to deceive machine learning classifiers. State-of-the-art adversarial example detection methods characterize an as either by quantifying the magnitude of feature variations under multiple perturbations or measuring its distance from estimated benign distribution. Instead using such metrics, proposed method is based on observation directions gradients when crafting (new) play a key role in characterizing space. Compared use perturbations, efficient it only applies single random perturbation example. Experiments conducted two different databases, CIFAR-10 and ImageNet, show achieves, respectively, 97.9% 98.6% AUC-ROC (on average) five attacks, outperforms state-of-the-art methods. Results demonstrate effectiveness gradient for detection.