CADA
收藏资源简介:
CADA数据集由北京航空航天大学等机构创建,包含18,808个对抗性视觉-问答对,旨在评估和增强自动驾驶视觉语言模型(VLM)的鲁棒性。该数据集通过生成对抗性扰动,模拟自动驾驶场景中的高风险情境,帮助研究人员识别和解决模型在复杂驾驶环境中的潜在漏洞。数据集的应用领域主要集中在自动驾驶系统的安全性评估和对抗性攻击防御研究,旨在提升自动驾驶系统的可靠性和安全性。
The CADA dataset, constructed by Beihang University and other affiliated institutions, encompasses 18,808 adversarial visual question-answer pairs. It is primarily intended to evaluate and strengthen the robustness of visual language models (VLMs) deployed in autonomous driving systems. By generating adversarial perturbations to simulate high-risk scenarios within autonomous driving environments, this dataset enables researchers to detect and mitigate potential vulnerabilities of models under complex driving conditions. Its core application domains include safety assessment of autonomous driving systems and adversarial attack defense research, with the ultimate objective of enhancing the reliability and safety of autonomous driving platforms.

- 1Black-Box Adversarial Attack on Vision Language Models for Autonomous Driving北京航空航天大学, 新加坡国立大学, 中国航空工业发展研究中心, 南洋理工大学 · 2025年



