kontur-ai/rtfs-10k

我们提出了一个用于检测类似视频会议条件下的视频深度伪造的数据集。它包含收集的短肖像视频片段、用作面部交换来源的肖像照片以及使用面部交换模型生成的深度伪造视频。数据集包含1,636个目标视频片段、2,000张源照片和9,772个生成的深度伪造视频。所有材料都标注了描述外观和场景属性（照明、位置、性别、配饰、面部特征、动作）的标签，视频、照片和生成片段的元数据以CSV文件形式提供。

We present a dataset for detecting video deepfakes in video-conference–like conditions. It contains short portrait clips collected, portrait photos used as sources for face swapping, and deepfake videos generated with face-swap models. The dataset contains 1,636 target video clips, 2,000 source photos, and 9,772 generated deepfake videos. All materials are annotated with tags describing appearance and scene properties (lighting, location, gender, accessories, facial features, movement), and the metadata for videos, photos and generated clips is provided in CSV files.