Adversarial patch attacks multispectral aerial object detectors system

Multispectral imaging systems are widely used in many security-sensitive areas, so it is crucial to study multispectral adversarial attack methods to assess the robustness of their detection systems. However, current research has focused on single-spectrum pedestrian detectors, and multispectral pedestrian detection adversarial attacks for aerial photography views have not been fully investigated. In this paper, we propose a new multispectral aerial photography adversarial attack. We establish an adversarial patch generation framework for attacking the aerial photography pedestrian detection system, design an adversarial patch with a barcode-like structure, and use an SPA-oriented optimization algorithm to study the effects of the shape, position and angle of the adversarial patch on the attack effect. Moreover, our designed adversarial patches have the same structure under visible light and infrared conditions, which is convenient for the subsequent physical transferability study. The experimental results show that the generated adversary patches produce good attack effects on two aerial multispectral detectors, and the effectiveness of this method in visible and infrared attacks is 72.46% and 44.22%, respectively.