DeepMoiréFake (DMF) dataset
收藏资源简介:
DeepMoiréFake数据集是一个包含12,832个视频,总时长35.64小时的数据集,它由来自Celeb-DF、DFD、DFDC、UADFV和FF++数据集的视频组成。该数据集在多种现实条件下捕获视频,包括不同的屏幕、智能手机、照明设置和相机角度,以模拟真实世界的Moiré模式对深度伪造检测的影响。数据集的创建过程包括从五个著名的深度伪造数据集中选择子集,并使用两种不同的智能手机在四种屏幕和两种照明条件下手动捕获视频。DeepMoiréFake数据集旨在为评估深度伪造检测器的鲁棒性提供一个现实世界的基准,并推动未来研究朝着缩小控制实验和实际深度伪造检测之间的差距方向发展。
The DeepMoiréFake dataset is a collection of 12,832 videos with a total duration of 35.64 hours, comprising videos extracted from five existing deepfake datasets: Celeb-DF, DFD, DFDC, UADFV, and FF++. The videos in this dataset were captured under diverse real-world conditions, including varying screens, smartphones, lighting settings, and camera angles, to simulate the impact of real-world Moiré patterns on deepfake detection. The construction of the DeepMoiréFake dataset involves first selecting subsets from these five renowned deepfake datasets, then manually capturing videos using two different smartphones across four screen types and two lighting conditions. The DeepMoiréFake dataset is designed to provide a real-world benchmark for evaluating the robustness of deepfake detectors, and to promote future research aimed at narrowing the gap between controlled laboratory experiments and real-world deepfake detection scenarios.

- 1Through the Lens: Benchmarking Deepfake Detectors Against Moiré-Induced DistortionsSungkyunkwan University, South Korea · 2025年



