Connecting the clues: Detecting adversarial perturbations using contextual inconsistency

The researchers propose a method for detecting adversarial perturbations using contextual inconsistency, with detection ROC-AUC mostly exceeding 0.95 against multiple adversarial attacks on PASCAL VOC and MS COCO. The method learns contextual consistency rules among objects in a scene during training, training one autoencoder for each object category, and outputs the difference between input and output when a perturbation violates contextual rules. Compared with state-of-the-art context-independent methods, detection performance improves by more than 20%.