DFBench
收藏资源简介:
DFBench是一个大规模的深度伪造图像检测数据集,包含54万张真实、AI编辑和AI生成的图像,用于评估深度伪造检测模型的能力。数据集由8个来源收集了45k真实图像和15k AI编辑图像,并使用12个最先进的生成模型基于40k来自Flickr8k的提示生成了480k假图像。DFBench允许对检测模型和生成模型的检测准确性和规避能力进行双向评估。
DFBench is a large-scale deepfake image detection dataset containing 540,000 real, AI-edited and AI-generated images, which is used to evaluate the performance of deepfake detection models. The dataset collects 45k real images and 15k AI-edited images from 8 sources, and generates 480k fake images via 12 state-of-the-art generative models based on 40k prompts from Flickr8k. DFBench enables bidirectional evaluation of detection accuracy and evasion capabilities for both detection models and generative models.
DFBench 数据集概述
数据集简介
- 名称:DFBench
- 用途:评估大型多模态模型(LMMs)的深度伪造图像检测能力
- 核心方法:MoA-DF(多模型聚合检测框架)
数据集内容
- 数据类型:包含真实图像和AI生成/编辑的伪造图像
- 特征分布:
- 无失真真实图像
- 带失真的真实图像
- AI编辑图像
- AI生成图像
- 数据文件:
img_train_shuffled.jsonimg_test.jsonimg_train_shuffled.jsonlimg_test.jsonl
下载方式
- Hugging Face地址:https://huggingface.co/datasets/IntMeGroup/DFBench/tree/main
- 下载命令: bash huggingface-cli download IntMeGroup/DFBench --repo-type dataset --local-dir ./DFBench
支持的模型
-
Qwen2.5-VL
- 模型权重下载: bash huggingface-cli download Qwen/Qwen2.5-VL-7B-Instruct --local_dir ./Qwen/Qwen2.5-VL-7B-Instruct
-
InternVL2.5
- 模型权重下载: bash huggingface-cli download OpenGVLab/InternVL2_5-8B --local_dir ./internvl25/OpenGVLab/InternVL2_5-8B
-
InternVL3
- 模型权重下载: bash huggingface-cli download OpenGVLab/InternVL3-9B --local_dir ./internvl3/OpenGVLab/InternVL3-9B
评估指标
- 提供logit概率计算和准确率评估脚本: bash python logit_calculation.py python process_results.py
可视化
- 特征分布可视化脚本: bash python feature_distribution.py python plot_features.py
引用信息
bibtex @misc{wang2025dfbenchbenchmarkingdeepfakeimage, title={DFBench: Benchmarking Deepfake Image Detection Capability of Large Multimodal Models}, author={Jiarui Wang and Huiyu Duan and Juntong Wang and Ziheng Jia and Woo Yi Yang and Xiaorong Zhu and Yu Zhao and Jiaying Qian and Yuke Xing and Guangtao Zhai and Xiongkuo Min}, year={2025}, eprint={2506.03007}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2506.03007}, }
联系方式
- 邮箱:wangjiarui@sjtu.edu.cn

- 1DFBench: Benchmarking Deepfake Image Detection Capability of Large Multimodal Models上海交通大学 · 2025年



