🏆 FIDL 排行榜 🏆

统一排名模型在各领域的泛化能力

本排行榜遵循 ForensicHub 中提出的 IFF-Protocol 的简化版本：所有模型仅在 📂OpenMMSec 上训练。该数据集由 ForensicHub 与蚂蚁集团联合发布，设计理念与 IFF-Protocol 相似。OpenMMSec 包含真实图像、从公开数据集中采样的篡改图像，以及部分私有伪造图像。数据集组成的详细信息可在页面底部查看。

你可以通过以下方式下载 OpenMMSec：百度网盘或Google Drive

在排名评测中，我们从每个领域中选择了最具挑战性的数据集，不进行微调（finetuning），以强调模型的跨域泛化能力。每个领域的 AUC 会先进行平均，然后再取四个领域平均值的整体均值作为最终排序依据。

用于评测的数据集如下：

Deepfake 检测：FF++， DF40
图像篡改检测与定位（IMDL）： IMD2020， Autosplice
AI 生成内容检测（AIGC）： DiffusionForensics， Chameleon
文档篡改检测与定位（Document）： RealTextManipulation, T-SROIE

🏆 Rank	Model	Deepfake 🖼️	IMDL 📝	AIGC 🤖	Doc 📄	Avg ⭐
🥇 1	Effort	0.614	0.587	0.410	0.788	0.600
🥈 2	Segformer-b3	0.629	0.576	0.339	0.724	0.567
🥉 3	Clip-ViT-L/14	0.664	0.543	0.317	0.724	0.562
4	ConvNeXT	0.662	0.573	0.337	0.669	0.560
5	Mesorch	0.541	0.562	0.460	0.591	0.538
6	UnivFD	0.442	0.486	0.463	0.734	0.531
7	IML-ViT	0.581	0.562	0.325	0.626	0.523
...

更多超参数设置可参考 ForensicHub。
各模型在不同数据集上的详细性能如下：

点击展开查看结果

[
    {
        "model": "Effort",
        "deepfake": {"DF40_CollabDiff": 0.7686, "DF40_deepfacelab": 0.4292, "DF40_heygen": 0.7061, "FF++c40": 0.5506},
        "imdl": {"IMD2020": 0.5704, "Autosplice": 0.6035},
        "aigc": {"Chameleon": 0.4898, "DiffusionForensics": 0.3304},
        "doc": {"RealTextManipulation": 0.6439, "T-SROIE": 0.9326}
    },
    {
        "model": "Segformer-b3",
        "deepfake": {"DF40_CollabDiff": 0.8503, "DF40_deepfacelab": 0.5125, "DF40_heygen": 0.685, "FF++c40": 0.4677},
        "imdl": {"IMD2020": 0.543, "Autosplice": 0.6098},
        "aigc": {"Chameleon": 0.411, "DiffusionForensics": 0.2676},
        "doc": {"RealTextManipulation": 0.5695, "T-SROIE": 0.8795}
    },
    {
        "model": "ConvNeXT",
        "deepfake": {"DF40_CollabDiff": 0.9572, "DF40_deepfacelab": 0.6103, "DF40_heygen": 0.5281, "FF++c40": 0.5512},
        "imdl": {"IMD2020": 0.5512, "Autosplice": 0.5945},
        "aigc": {"Chameleon": 0.3944, "DiffusionForensics": 0.2794},
        "doc": {"RealTextManipulation": 0.5235, "T-SROIE": 0.8138}
    },
    {
        "model": "UnivFD",
        "deepfake": {"DF40_CollabDiff": 0.7458, "DF40_deepfacelab": 0.3962, "DF40_heygen": 0.1665, "FF++c40": 0.4610},
        "imdl": {"IMD2020": 0.4887, "Autosplice": 0.4831},
        "aigc": {"Chameleon": 0.5727, "DiffusionForensics": 0.3537},
        "doc": {"RealTextManipulation": 0.5554, "T-SROIE": 0.9136}
    },
    {
        "model": "IML-ViT",
        "deepfake": {"DF40_CollabDiff": 0.9783, "DF40_deepfacelab": 0.2938, "DF40_heygen": 0.6297, "FF++c40": 0.4224},
        "imdl": {"IMD2020": 0.5229, "Autosplice": 0.6008},
        "aigc": {"Chameleon": 0.3707, "DiffusionForensics": 0.2799},
        "doc": {"RealTextManipulation": 0.5307, "T-SROIE": 0.7207}
    },
    {
        "model": "Mesorch",
        "deepfake": {"DF40_CollabDiff": 0.7139, "DF40_deepfacelab": 0.3324, "DF40_heygen": 0.6478, "FF++c40": 0.4699},
        "imdl": {"IMD2020": 0.5331, "Autosplice": 0.5905},
        "aigc": {"Chameleon": 0.3991, "DiffusionForensics": 0.5214},
        "doc": {"RealTextManipulation": 0.522, "T-SROIE": 0.6599}
    },
    {
        "model": "Clip-ViT-L/14",
        "deepfake": {"DF40_CollabDiff": 0.999, "DF40_deepfacelab": 0.352, "DF40_heygen": 0.8067, "FF++c40": 0.4984},
        "imdl": {"IMD2020": 0.5567, "Autosplice": 0.5295},
        "aigc": {"Chameleon": 0.3371, "DiffusionForensics": 0.2962},
        "doc": {"RealTextManipulation": 0.5836, "T-SROIE": 0.8651}
    },
]

OpenMMSec 数据集构成如下：

真实图像：COCO, Object365
Deepfake：DeepFakeFace 与 SFHQ 两种伪造类型
IMDL：基于 COCO 与 Object365 的小物体修补（inpainting）
AIGC：Community Forensics
Doc：OSTF及部分来自天池 2024 比赛的图像