DFDM (Deepfake videos generated from different models)

我们创建了一个名为 DFDM 的新数据集，其中包含由不同自动编码器模型生成的 6,450 个 Deepfake 视频。具体来说，已经选择了五个分别具有编码器、解码器、中间层和输入分辨率变化的自动编码器模型来基于相同的输入生成 Deepfakes。我们首先观察到不同 Deepfake 之间可见但细微的视觉差异，证明了模型归因伪影的证据。然后，我们将 Deepfakes 模型归因作为一个多类分类任务，并提出了一种基于时空注意力的方法来探索 DFDM 数据集中 Deepfakes 之间的差异。
DFDM 数据集包括基于 Faceswap 软件的五个自动编码器模型生成的面部交换 Deepfakes 视频，包括 Faceswap、Lightweight、IAE、Dfaker 和 DFL-H128。三种 H.264 压缩率被认为可以获得不同质量的视频，包括恒定速率因子 (crf) 为 0 的无损、crf 为 10 的高质量和 crf 为 23 的低质量。总共 6,450 个 Deepfakes创建的。

We present a novel dataset named DFDM, which comprises 6,450 Deepfake videos generated by diverse autoencoder models. Specifically, five autoencoder models with variations in their encoder architectures, decoder structures, intermediate layers, and input resolutions were selected to generate Deepfake samples using identical input data. We first observed visible yet subtle visual differences across different Deepfakes, which provides evidence of model-attributed artifacts. Then, we framed the task of Deepfake model attribution as a multi-class classification problem, and proposed a spatio-temporal attention-based method to explore the differences among Deepfakes in the DFDM dataset.
The DFDM dataset consists of face-swapped Deepfake videos generated by five autoencoder models based on the Faceswap software framework, including Faceswap, Lightweight, IAE, Dfaker, and DFL-H128. Three H.264 compression settings were employed to produce videos of varying quality: lossless compression with a constant rate factor (crf) of 0, high-quality compression with crf=10, and low-quality compression with crf=23. In total, 6,450 Deepfake videos were generated for this dataset.

搜集汇总

背景与挑战

背景概述

DFDM数据集包含6,450个Deepfake视频，由五种基于Faceswap软件的自动编码器模型生成，并涵盖三种不同压缩质量。该数据集旨在支持模型归因研究，通过时空注意力方法分析不同Deepfake之间的视觉差异。