Physical Attack Naturalness (PAN) dataset

PAN数据集是由北京航空航天大学SKLSDE实验室创建的，专门用于研究物理世界攻击的自然性。该数据集包含2688张图像，这些图像是在自动驾驶场景中生成的，包括5种常用的攻击方式和2种良性模式。数据集通过126名参与者的主观评级和凝视信号来评估图像的自然性。PAN数据集首次揭示了环境特征（如光照、角度等）和行为特征（如人类凝视）对攻击自然性的影响，并提出了Dual Prior Alignment (DPA)网络来自动评估攻击的自然性，该网络通过模仿人类的评级和凝视行为来提高评估的准确性。PAN数据集的应用领域主要集中在提高物理世界攻击的自然性和隐蔽性，以及开发更有效的自动评估方法。

The PAN Dataset was created by the SKLSDE Laboratory at Beihang University, specifically for researching the naturalness of physical-world adversarial attacks. This dataset comprises 2688 images generated in autonomous driving scenarios, covering 5 commonly used adversarial attack methods and 2 benign patterns. The naturalness of these images is evaluated through subjective ratings and gaze signals collected from 126 participants. The PAN Dataset is the first to reveal the impacts of environmental characteristics (e.g., illumination, viewing angle, etc.) and behavioral characteristics (e.g., human gaze) on the naturalness of adversarial attacks. It also proposes the Dual Prior Alignment (DPA) network for automatic naturalness assessment of adversarial attacks, which enhances evaluation accuracy by mimicking human rating and gaze behaviors. The primary application areas of the PAN Dataset focus on improving the naturalness and stealth of physical-world adversarial attacks, as well as developing more effective automatic evaluation methods.