X-AVDT
收藏资源简介:
该数据集(MMDF)是为论文“X-AVDT: 用于鲁棒深度伪造检测的音频-视觉交叉注意力”构建的,主要用于音频-视觉多模态深度伪造检测任务。数据集包含真实和伪造的视频片段。原始视频数据经过预处理,包括提取视频帧和音频文件,随后通过Hallo特征提取方法生成多种特征表示。最终用于训练的数据组织形式为:在根目录下,按照数据分割(如训练集、验证集)、标签(真实/伪造)、深度伪造模型ID和视频剪辑ID进行层级划分,每个剪辑目录下包含以.pt文件格式存储的多种预计算特征,例如原始特征、反转特征、重建特征、残差特征和注意力特征。这些特征文件以16帧为块进行切片,适用于训练基于音频-视觉交叉注意力机制的深度伪造检测模型。
This dataset (MMDF) is constructed for the paper X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection and is primarily used for audio-visual multimodal deepfake detection tasks. It contains both genuine and fake video clips. The original video data undergoes preprocessing, including extraction of video frames and audio files, followed by the Hallo feature extraction method to generate various feature representations. The data organization for training is structured as follows: under the root directory, it is hierarchically divided by data splits (e.g., training set, validation set), labels (genuine/fake), deepfake model IDs, and video clip IDs. Each clip directory contains multiple precomputed features stored in .pt file format, such as original features, inversion features, reconstruction features, residual features, and attention features. These feature files are sliced into blocks of 16 frames, suitable for training deepfake detection models based on audio-visual cross-attention mechanisms.
数据集概述:X-AVDT
数据集名称:MMDF(用于X-AVDT方法)
用途:用于训练和评估音频-视觉交叉注意力深度伪造检测模型X-AVDT,重点在于提升对未见深度伪造生成器的鲁棒性。
来源:由韩国科学技术院(KAIST)视觉媒体实验室(Visual Media Lab @ KAIST)的研究团队发布,相关论文被CVPR 2026接收。
关联论文:X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection
数据下载地址:https://huggingface.co/datasets/zaqxsw0526/MMDF
数据预处理与特征提取
- 视频转换:将原始视频转换为帧文件夹和wav文件,需使用
ffmpeg。 - Hallo特征提取:对每个片段提取
original.mp4、inverted.mp4、reconstructed.mp4、residual.mp4、attn_map.mp4以及attn_feat.pt特征文件。 - 特征打包:将片段级特征拆分为16帧的
.pt块文件,形成训练所需的目录结构。
训练与评估
- 训练:使用打包后的特征文件,运行训练脚本。
- 评估:需下载预训练的X-AVDT检测器权重(下载链接),然后运行评估脚本。
目录结构
特征提取后,训练数据应组织为如下格式(针对真实和伪造两个根目录):
<root>/<split>/<label>/<model_id>/<clip_id>/ original/.pt inverted/.pt reconstructed/.pt residual/.pt attn_feat/*.pt
引用
bibtex @article{kim2026x, title={X-AVDT: Audio-Visual Cross-Attention for Robust Deepfake Detection}, author={Kim, Youngseo and Yun, Kwan and Hong, Seokhyeon and Cha, Sihun and Koo, Colette Suhjung and Noh, Junyong}, journal={arXiv preprint arXiv:2603.08483}, year={2026} }




