NeuroRenderedFake: A Challenging Benchmark to Detect Fake Images Generated by Advanced Neural Rendering Methods

A Large Dataset of Neural-rendered Fake Images: Using NeRF- and 3DGS-based neural rendering techniques, we generate a variety of realistic 3D scene representations and render 2D fake images from them. Additionally, we synthesize fake images containing combined artifacts by integrating generative models with neural rendering technologies. In contrast to existing databases that focus exclusively on fake images from generative models like GAN and DM, our diverse collection expands the scope of fake image detection, being the first in its inclusion of images derived from neural rendering-based synthesized/edited 3D scenes.

神经渲染伪造图像大型数据集：本数据集采用基于神经辐射场（NeRF）与三维高斯溅射（3DGS）的神经渲染技术，生成多种逼真的三维场景表征，并从中渲染得到二维伪造图像。此外，本研究通过将生成式模型与神经渲染技术相结合，合成了带有复合伪影的伪造图像。与现有仅聚焦生成对抗网络（GAN）、扩散模型（DM）等生成式模型产出伪造图像的数据库不同，本数据集的多样化样本拓展了伪造图像检测的研究范畴，是首个收录基于神经渲染技术合成或编辑的三维场景衍生图像的数据集。

搜集汇总

背景与挑战

背景概述

NeuroRenderedFake是一个大规模数据集，包含由先进神经渲染技术生成的假图像，旨在为假图像检测方法提供评估基准。该数据集通过跨域评估协议和频谱能量分析，帮助研究者理解不同生成方法之间的差异和共同特征。