Fictitious Facial Identity VQA Dataset
收藏资源简介:
Fictitious Facial Identity VQA Dataset是由威斯康星大学麦迪逊分校等机构创建的一个用于视觉语言模型(VLM)遗忘评估的数据集。该数据集包含400个合成面部图像,每个图像关联20个关于个人背景、健康记录和犯罪历史的问答对,总计8000条数据。数据集的创建过程包括从SFHQ数据集中筛选面部图像,并使用GPT-4生成问答对。该数据集主要用于评估在“被遗忘权”背景下,VLM能否有效遗忘隐私信息,旨在解决视觉语言模型中的隐私保护问题。
The Fictitious Facial Identity VQA Dataset is a dataset developed by institutions including the University of Wisconsin-Madison for evaluating the forgetting performance of Vision-Language Models (VLMs). It contains 400 synthetic facial images, each paired with 20 question-answer pairs covering personal background, health records and criminal history, totaling 8,000 data entries. The dataset construction process includes filtering facial images from the SFHQ dataset and generating question-answer pairs using GPT-4. Its main purpose is to assess whether VLMs can effectively erase private information under the context of the "right to be forgotten", aiming to address privacy protection issues in vision-language models.
FIUBench 数据集概述
许可证
- 许可证类型: Apache 2.0




