Understanding Black-Box Attacks Against Object Detectors from a User's Perspective

Abstract

Due to recent developments in object detection systems, and the realistic threat of black-box adversarial attacks on object detector models, we argue the need for a contextual understanding of the attacks from the users’ perspective. Existing literature reviews either do not provide complete and up-to-date summaries of such attacks or focus on the knowledge from the researchers’ perspective. In this research, we conducted a systematic literature review to identify state-of-the-art black-box attacks and extract the information to help users evaluate and mitigate the risks. The literature review resulted in 29 black-box attack methods. We analyzed each attack from the following main aspects: attackers’ knowledge needed to perform the attack, attack consequences, attack generalizability, and strategies to mitigate the attacks. Our results demonstrate an emerging increase in highly generalizable attacks, which now make up more than 50% of the landscape. We also reveal that more than 50% of recent attacks remain untested against mitigation strategies.