VideoForensicsHQ

有人担心，合成高质量面部视频的新方法可能会被滥用来操纵具有恶意意图的视频。因此，研究社区开发了用于检测修改后的素材的方法，并为此任务组装了基准数据集。在本文中，我们研究了伪造检测器的性能如何取决于人眼可以看到的人工制品的存在。我们引入了一个新的基准数据集，用于面部视频伪造检测，质量空前。它使我们能够证明现有的检测技术很难检测到可靠地愚弄人眼的假货。因此，我们引入了一系列新的检测器，这些检测器检查了时空特征的组合，并且在检测精度和泛化方面都优于现有方法。Concerns have emerged that novel methods for synthesizing high-quality facial videos could be misused to manipulate videos with malicious intent. To address this issue, the research community has developed methods for detecting altered media and assembled benchmark datasets for this task. In this paper, we investigate how the performance of forgery detectors depends on the presence of artifacts visible to the human eye. We introduce a novel benchmark dataset for facial video forgery detection with unprecedented quality, which enables us to demonstrate that existing detection techniques struggle to detect forgeries that reliably fool the human eye. Accordingly, we propose a series of new detectors that examine combinations of spatial-temporal features, outperforming existing methods in both detection accuracy and generalization.

搜集汇总

背景与挑战

背景概述

VideoForensicsHQ是一个2021年发布的面部视频伪造检测基准数据集，由萨尔大学等机构开发，旨在评估检测器在识别高质量合成视频方面的性能。该数据集支持研究新型时空特征检测方法，以提高检测精度和泛化能力。