Forensics-Bench
收藏资源简介:
Forensics-Bench是一个新的伪造检测评估基准套件,旨在评估大型视觉语言模型(LVLMs)在多种伪造检测任务中的表现,要求模型具备全面的识别、定位和推理能力。该数据集包含63,292个精心策划的多选视觉问题,涵盖112种独特的伪造检测类型,从5个角度进行分类:伪造语义、伪造模态、伪造任务、伪造类型和伪造模型。
Forensics-Bench is a novel benchmark suite for forgery detection evaluation, which aims to assess the performance of Large Vision-Language Models (LVLMs) across various forgery detection tasks and requires the models to possess comprehensive capabilities of recognition, localization and reasoning. This dataset includes 63,292 carefully curated multiple-choice visual questions, covering 112 unique forgery detection categories, which are classified from five perspectives: forgery semantics, forgery modalities, forgery tasks, forgery types and forgery models.
Forensics-Bench 数据集概述
基本信息
- 名称: Forensics-Bench
- 类型: 伪造检测评估基准套件
- 论文标题: Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models
- 作者: Jin Wang, Chenghui Lv, Xian Li, Shichao Dong, Huadong Li, Kelu Yao, Chao Li, Wenqi Shao, Ping Luo
- 发布日期: 2025-03-22
- 论文状态: 已被CVPR 2025接受
数据集内容
- 数据量: 63,292个多选视觉问题
- 覆盖范围:
- 112种独特的伪造检测类型
- 5个视角: 伪造语义、伪造模态、伪造任务、伪造类型和伪造模型
- 评估模型:
- 22个开源大型视觉语言模型(LVLM)
- 3个专有模型: GPT-4o, Gemini 1.5 Pro, Claude 3.5 Sonnet
数据集特点
- 专注于评估LVLM在伪造检测方面的综合能力
- 要求模型具备识别、定位和推理多种伪造内容的能力
- 旨在推动AIGC时代全方位伪造检测器的发展
获取方式
- Hugging Face地址: Forensics-bench/Forensics-bench
- 文件名称: ForensicsBench.tsv
评估框架
- 使用工具: VLMEvalKit
- 主要评估脚本: run.py
- 评估模式:
- 完整评估(推理+评估)
- 仅推理模式
系统要求
- Python包:
- transformers(不同模型需要不同版本)
- 其他依赖见安装说明
- API密钥: 评估专有模型需要配置相应API密钥
引用格式
bibtex @misc{wang2025forensicsbenchcomprehensiveforgerydetection, title={Forensics-Bench: A Comprehensive Forgery Detection Benchmark Suite for Large Vision Language Models}, author={Jin Wang and Chenghui Lv and Xian Li and Shichao Dong and Huadong Li and kelu Yao and Chao Li and Wenqi Shao and Ping Luo}, year={2025}, eprint={2503.15024}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2503.15024}, }




