Relevance attack on detectors

This paper focuses on high-transferable adversarial attacks detectors, which are hard to attack in a black-box manner, because of their multiple-output characteristics and the diversity across architectures. To pursue high transferability, one plausible way is find common property facilitates discovery weaknesses. We first suggest that relevance map from interpreters for detectors such property. Based it, we design Relevance Attack Detectors (RAD), achieves state-of-the-art exceeding existing results by above 20%. On MS COCO, detection mAPs all 8 architectures more than halved segmentation also significantly influenced. Given great transferability RAD, generate dataset object instance segmentation, i.e., Adversarial Objects COntext (AOCO), helps quickly evaluate improve robustness detectors.