ForgeryNet

真实感合成技术的快速发展已经达到了一个临界点，真实图像和操纵图像之间的边界开始模糊。因此，基准测试和推进数字伪造分析已成为一个紧迫的问题。然而，现有的人脸伪造数据集要么多样性有限，要么只支持粗粒度分析。为了应对这种新出现的威胁，我们构建了 ForgeryNet 数据集，这是一个非常大的人脸伪造数据集，在图像和视频级数据中具有统一的注释，跨越四个任务：1）图像伪造分类，包括双向（真/假）、三个-way（真/假，带有身份替换的伪造方法/带有身份保留的伪造方法的假）和 n-way（真实和 15 种各自的伪造方法）分类。 2）空间伪造定位，将伪造图像的操纵区域与其相应的源真实图像进行分割。 3）视频伪造分类，它重新定义了视频级别的伪造分类，其中操纵帧位于随机位置。这项任务很重要，因为现实世界中的攻击者可以自由地操纵任何目标帧。 4) Temporal Forgery Localization，对被操作的时间段进行定位。 ForgeryNet 在数据规模（290 万张图像，221,247 个视频）、操作（7 个图像级方法、8 个视频级方法）、扰动（36 个独立和更多混合扰动）方面是迄今为止最大的公开可用的深度人脸伪造数据集) 和注释（630 万个分类标签、290 万个操作区域注释和 221,247 个临时伪造片段标签）。我们对现有的人脸取证方法进行了广泛的基准测试和研究，并获得了一些有价值的观察结果。

The rapid development of photorealistic synthesis technologies has reached a critical point, where the boundary between real and manipulated images has begun to blur. Therefore, benchmarking and advancing digital forgery analysis has become an urgent issue. However, existing face forgery datasets either have limited diversity or only support coarse-grained analysis. To address this emerging threat, we construct the ForgeryNet dataset, a large-scale face forgery dataset with unified annotations for both image and video-level data, spanning four tasks: 1) Image forgery classification, including two-way (real/fake), three-way (real, fake via identity-replaced forgery methods, and fake via identity-preserved forgery methods), and n-way (real and 15 respective forgery methods) classification. 2) Spatial forgery localization, which segments the manipulated regions of forged images against their corresponding source real images. 3) Video forgery classification, which redefines video-level forgery classification where manipulated frames are placed at random positions. This task is critical because real-world attackers can freely manipulate any target frame. 4) Temporal forgery localization, which locates manipulated time periods. ForgeryNet is by far the largest publicly available deep face forgery dataset in terms of data scale (2.9 million images, 221,247 videos), manipulation methods (7 image-level methods, 8 video-level methods), perturbations (36 independent and numerous mixed perturbations), and annotations (6.3 million classification labels, 2.9 million manipulated region annotations, and 221,247 temporal forgery segment labels). We conduct extensive benchmarking and research on existing face forensics methods, and obtain several valuable observations.

搜集汇总

背景与挑战

背景概述

ForgeryNet是一个大规模的人脸伪造数据集，旨在基准测试和推进数字伪造分析，包含290万张图像和22.1万个视频，支持图像和视频级别的伪造分类、空间定位及时间定位任务。它具有多样化的操作方法（7种图像级和8种视频级方法）和扰动（36种独立和混合扰动），并提供详细的注释（如分类标签和操作区域），是目前最大的公开深度人脸伪造数据集，适用于多粒度伪造检测研究。