DiffusionFace
收藏资源简介:
DiffusionFace是首个基于扩散模型的面部伪造分析数据集,由中国教育部多媒体可信感知与高效计算重点实验室和厦门大学联合创建。该数据集包含600,000张图像,涵盖无条件和文本引导的面部图像生成、图像到图像转换、修复和基于扩散的面部交换算法等多种伪造类别。数据集利用Multi-Modal-CelebA-HQ数据集的高质量、丰富注释的面部图像,通过11种扩散模型生成伪造面部。DiffusionFace不仅为开发高级检测模型提供了基础,还通过引入实用的测试协议,严格评估了区分模型在检测伪造面部图像方面的有效性,旨在增强面部图像认证过程的安全性。
DiffusionFace is the first facial forgery analysis dataset based on diffusion models, jointly created by the Key Laboratory of Multimedia Trusted Perception and Efficient Computing of the Ministry of Education of China and Xiamen University. This dataset contains 600,000 images, covering multiple forgery categories including unconditional and text-guided facial image generation, image-to-image translation, inpainting, and diffusion-based face swapping algorithms. Leveraging the high-quality and richly annotated facial images from the Multi-Modal-CelebA-HQ dataset, the dataset generates forged facial images via 11 distinct diffusion models. Beyond serving as a foundational resource for developing advanced detection models, DiffusionFace also introduces practical test protocols to rigorously evaluate the efficacy of discriminative models for detecting forged facial images, with the objective of strengthening the security of facial image authentication processes.

- 1DiffusionFace: Towards a Comprehensive Dataset for Diffusion-Based Face Forgery Analysis中国教育部多媒体可信感知与高效计算重点实验室,厦门大学,中国 · 2024年



