CrossVLAD
收藏资源简介:
CrossVLAD是一个专门用于评估统一视觉语言模型(VLMs)跨任务对抗攻击的基准数据集。它基于MSCOCO数据集构建,并使用GPT-4辅助进行注释。数据集包含了3000张经过精心挑选的图像,涉及10个语义类别的79个变换对。CrossVLAD的构建过程采用了严格的筛选标准,确保了对象的大小限制、类别独特性、标题验证等。数据集的设计旨在模拟跨任务对抗攻击的场景,要求对抗性样本能够一致地改变目标对象在四个下游任务中的分类。CrossVLAD的引入填补了统一VLMs安全评估中的一个关键研究空白,为评估跨任务对抗攻击的成功率提供了一个全面的评估框架。
CrossVLAD is a benchmark dataset specifically designed for evaluating cross-task adversarial attacks against unified Vision-Language Models (VLMs). It is built upon the MSCOCO dataset, with annotations assisted by GPT-4. The dataset contains 3000 carefully curated images, along with 79 transformation pairs covering 10 semantic categories. The construction of CrossVLAD adopts stringent screening criteria to enforce object size constraints, category uniqueness, caption validation, and other relevant requirements. This dataset is engineered to replicate cross-task adversarial attack scenarios, wherein adversarial examples must consistently alter the classification of target objects across four downstream tasks. The introduction of CrossVLAD addresses a critical research gap in the safety assessment of unified VLMs, offering a comprehensive evaluation framework for quantifying the success rate of cross-task adversarial attacks.
CRAFT数据集概述
数据集基本信息
- 数据集名称:CRAFT
- 对应论文:One Object, Multiple Lies: A Benchmark for Cross-task Adversarial Attack on Unified Vision-Language Models
数据集内容
- 数据集类型:代码实现(对应论文的代码实现)
- 功能定位:跨任务对抗攻击基准(针对统一视觉-语言模型)
当前状态
- 开发状态:代码整理中(尚未完全开源)
- 开源计划:计划开放源代码和模型
注意事项
- 当前版本尚未完全开放
- 需关注项目更新以获取完整资源




