DFFD：多样的假脸数据集

DFFD数据集由多个公开可用的数据集和图像组成，这些数据集和图片使用公开可用的方法进行合成/处理。通过结合真实图像的多个源，我们能够包括真实图像和合成/操纵图像的不同分辨率和图像质量。 Figure 1: Sample images in the DFFD dataset. 我们将DFFD数据集划分为训练（50%）、验证（5%）和测试（45%）分区，同时注意确保一个分区中的标识不会出现在任何其他分区中。在训练时，当我们观察到模型在验证分区上的收敛时，我们冻结模型并在测试分区上进行评估。 Figure 2: The baseline performance on the DFFD dataset.

The DFFD dataset comprises multiple publicly available datasets and images, which are synthesized or processed via publicly accessible methods. By integrating multiple sources of real-world images, we can encompass diverse resolutions and image qualities of both real and synthetic/manipulated images.
Figure 1: Sample images from the DFFD dataset.
We split the DFFD dataset into training (50%), validation (5%), and testing (45%) partitions, while carefully ensuring that identities present in one partition do not appear in any other partitions. During training, once we observe model convergence on the validation partition, we freeze the model and conduct evaluation on the test partition.
Figure 2: Baseline performance on the DFFD dataset.

搜集汇总

背景与挑战

背景概述

DFFD数据集是一个多样假脸数据集，包含真实和合成/操纵的图像，覆盖不同分辨率和质量。数据集已划分为训练、验证和测试分区，适用于人脸伪造检测相关研究。