TrueFake
收藏资源简介:
AI生成的合成媒体在现实世界场景中越来越多地被使用,通常用于通过社交媒体平台传播错误信息和宣传,其中压缩和其他处理可能会降低伪造检测线索。目前,许多取证工具未能考虑到这些现实挑战。在这项工作中,我们引入了TrueFake,一个包含60万张图像的大规模基准数据集,包括顶级生成技术和通过三个不同社交网络的分享。该数据集允许在非常现实和具有挑战性的条件下对最先进的伪造图像检测器进行严格评估。通过广泛的实验,我们分析了社交媒体分享如何影响检测性能,并确定了当前最有效的检测和训练策略。我们的发现强调了在反映现实世界使用条件下评估取证模型的必要性。
AI-generated synthetic media is increasingly being used in real-world scenarios, often for spreading misinformation and propaganda via social media platforms, where compression and other processing may degrade forgery detection cues. Currently, many forensic tools fail to account for these real-world challenges. In this work, we introduce TrueFake, a large-scale benchmark dataset containing 600,000 images, covering top-tier generative technologies and shares across three distinct social networks. This dataset enables rigorous evaluation of state-of-the-art forged image detectors under highly realistic and challenging conditions. Through extensive experiments, we analyze how social media sharing impacts detection performance, and identify the most effective detection and training strategies currently available. Our findings underscore the necessity of evaluating forensic models under conditions that reflect real-world usage scenarios.
TrueFake数据集概述
数据集基本信息
- 名称: TrueFake: A Real World Case Dataset of Last Generation Fake Images also Shared on Social Networks
- 作者: Stefano DellAnna, Andrea Montibeller, Giulia Boato
- 相关论文: TrueFake: A Real World Case Dataset of Last Generation Fake Images also Shared on Social Networks
- 数据集状态: 即将发布
数据集内容
- 规模: 600,000张图像
- 特点:
- 包含顶级生成技术生成的图像
- 包含通过三种不同社交网络分享的图像
- 模拟真实世界场景中的挑战性条件
数据集用途
- 用于在高度现实和挑战性条件下评估最先进的假图像检测器
- 分析社交媒体分享对检测性能的影响
- 识别当前最有效的检测和训练策略
提示信息
- 提示文件位置:
./prompts/prompts_faces.txtprompts_animals.txtprompts_landscapes.txtprompts_general.txt(来自sezenkarakus/image-description-dataset-v2)
- 生成脚本:
./prompts/src/
检测器信息
- 预训练网络: R50-E2P
- 权重下载: pretrained weights
- 环境设置: 使用
environment.yml创建虚拟环境 - 运行代码: 位于
./detector/目录下
引用信息
bibtex @misc{dellanna2025truefake, title={TrueFake: A Real World Case Dataset of Last Generation Fake Images also Shared on Social Networks}, author={Stefano DellAnna and Andrea Montibeller and Giulia Boato}, year={2025}, eprint={2504.20658}, archivePrefix={arXiv}, primaryClass={cs.MM}, url={https://arxiv.org/abs/2504.20658}, }




