A Highly Transferable Camouflage Attack Against Object Detectors in the Physical World

Abstract

To assess the vulnerability of deep neural networks in the physical world, many studies have introduced adversarial examples and applied them to computer vision tasks such as object detection in recent years. Compared to patch-based adversarial attacks, camouflage-based attacks have received more and more attention due to their ability to attack detectors from multiple viewpoints. However, existing adversarial examples often rely on glass-box models and exhibit limited transferability to closed-box models, which remains a significant challenge. To address this issue, we propose the highly transferable camouflage attack, a novel physical adversarial attack framework designed to generate robust and efficient adversarial camouflage that can mislead object detectors in diverse scenarios. Specifically, we introduce a distraction method to distribute the features of the attention map between models, and propose enhanced transfer strategies to improve adversarial transferability through augmenting the input data and the attacked models. Extensive experiments demonstrate that our highly transferable camouflage attack can effectively mislead object detectors in both digital and physical worlds, enhancing the transferability of adversarial camouflage on multiple mainstream detectors.