Abstract
Ultraviolet imaging offers unique advantages in low-light conditions, complex backgrounds, and specialized reconnaissance missions, and is widely used in fields such as covert target identification and security monitoring. However, existing adversarial sample generation methods are primarily focused on the visible light spectrum and are difficult to directly transfer to the ultraviolet spectrum. To address this, this paper proposes a method for generating adversarial samples in the ultraviolet spectrum based on 3D neural rendering, aiming to enhance the concealment and robustness of adversarial samples under ultraviolet imaging conditions. This method incorporates physical imaging constraints of the ultraviolet spectrum into a 3D neural rendering framework, modeling the characteristics of ultraviolet illumination and material responses to achieve refined modeling and optimization of target surface textures, thereby generating highly realistic ultraviolet adversarial samples. Additionally, an ultraviolet adversarial sample dataset covering various target categories, including pedestrians and vehicles, was constructed to support the evaluation of mainstream object detection models such as YOLO. Experimental results demonstrate that the generated UV adversarial samples can effectively deceive object detection models across various viewpoints and scenarios, validating the method’s effectiveness and practicality in the UV spectrum. This study provides a new technical approach for enhancing the robustness of UV object detection systems, offering significant theoretical implications and practical value.