AutoSplice
收藏资源简介:
AutoSplice数据集由纽约州立大学布法罗分校的研究团队创建,包含5894张图像,其中3621张为通过DALL-E2模型基于文本提示生成的伪造图像,其余为真实图像。数据集利用视觉新闻数据集中的图像和标题,通过自动媒体分析工具和人工标注提取可操作的对象区域,并替换相应的文本提示,生成高质量的伪造图像。该数据集主要用于媒体取证领域的伪造检测和定位,旨在解决当前媒体取证技术在检测语言-图像模型生成图像方面的挑战。
The AutoSplice dataset was created by a research team from the University at Buffalo, State University of New York. It contains 5,894 images, among which 3,621 are forged images generated by the DALL-E 2 model based on text prompts, while the remaining images are authentic. This dataset leverages images and captions from visual news datasets, extracts actionable object regions via automated media analysis tools and manual annotations, and replaces the corresponding text prompts to generate high-quality forged images. It is primarily used for forgery detection and localization in the field of media forensics, aiming to address the current challenges faced by media forensics technologies when detecting images generated by text-to-image models.

- 1AutoSplice: A Text-prompt Manipulated Image Dataset for Media Forensics纽约州立大学布法罗分校 · 2023年



