Adversarial Patches Exploiting Contextual Reasoning in Object Detection

Abstract

The utilization of spatial context to improve accuracy in most fast object detection algorithms is well known. The detectors increase inference speed by doing a single forward pass per image which means they implicitly use contextual reasoning for their predictions. We show that an adversary can exploit such contextual reasoning to fool standard detectors. We develop adversarial patches that make an object detector blind to a particular category even though the patch does not overlap with the missed detections. We also study methods to fix this vulnerability and show that limiting the use of contextual reasoning during object detector training acts as a form of defense that makes the detector robust. We believe defending against context based adversarial attack algorithms is not easy. We take a step towards that direction and urge the research community to give attention to this vulnerability.