CSU-JPG/VVA-Bench
收藏资源简介:
VVA-Bench是一个诊断性基准,用于评估图像到视频生成模型在隐式视觉提示攻击下的安全性。与仅依赖显式不安全文本提示不同,该基准研究在视觉输入中嵌入有害时间意图的情况,例如箭头、表情符号、姿势标记、草图、相机控制提示或前后帧对。发布的标注文件包含452个不安全攻击样本。每个样本将一个弱或未充分指定的文本提示与一个或多个图像输入以及描述视觉攻击格式和目标风险类别的元数据配对。数据集涵盖两种主要攻击形式:双帧时间攻击和单图像视觉提示攻击,其中视觉提示子集进一步分为五种代表性机制。数据集用于安全研究、模型评估、红队测试和防御开发,可能包含敏感或令人不安的概念,应仅在受控研究环境中使用。
VVA-Bench is a diagnostic benchmark for evaluating the safety of image-to-video (I2V) generation models under implicit visual prompt attacks. Instead of relying only on explicit unsafe text prompts, the benchmark studies cases where harmful temporal intent is embedded in visual inputs, such as arrows, emojis, pose marks, sketches, camera-control cues, or before/after frame pairs. The released annotation file contains 452 unsafe attack samples. Each sample pairs one weak or underspecified text prompt with one or more image inputs and metadata describing the visual attack format and target risk category. The dataset covers two major attack forms: two-frame temporal attacks and single-image visual-prompt attacks, with the visual-prompt subset further divided into five representative mechanisms. It is intended for safety research, model evaluation, red-teaming, and defense development, and may contain sensitive or disturbing concepts, so it should be used only in controlled research settings.




