DeepShield: A Dataset of Real and AI-Generated Images also Shared on Social Networks

The DeepShield dataset is a large-scale benchmark for evaluating the robustness of fake image detection systems. It contains 100,000 images, divided between real and AI-generated content produced using advanced generative models, including StyleGAN, StyleGAN2, StyleGAN3, Stable Diffusion 1.5, 2.1, 3, and XL, as well as Flux 1.0. To simulate real-world distortions, 30,000 images were shared on Facebook, X (formerly Twitter), and Telegram, then re-collected to include platform-induced compression and artifacts. This approach ensures that the dataset captures authentic distribution noise and artifacts encountered in real-world scenarios. The DeepShield dataset supports rigorous benchmarking of fake image detectors under realistic and challenging conditions. The dataset was created by Truebees Srl and is made available under the Creative Commons Attribution 4.0 International License. Visit https://truebees.eu/site/DeepShield数据集（DeepShield dataset）是一个用于评估伪造图像检测系统鲁棒性的大规模基准测试集。它包含10万张图像，分为真实图像与使用先进生成模型生成的AI合成图像两类，涵盖StyleGAN、StyleGAN2、StyleGAN3、Stable Diffusion 1.5、2.1、3及XL，以及Flux 1.0等模型。 为模拟真实世界中的图像畸变，研究团队将3万张图像发布至Facebook、X（原Twitter）及Telegram平台，随后重新采集以纳入平台诱导产生的压缩失真与伪影。该设计可确保该数据集能够捕捉真实场景中存在的真实分布噪声与伪影。 DeepShield数据集支持在贴近现实且极具挑战性的条件下，对伪造图像检测器开展严谨的基准测试。 该数据集由Truebees Srl制作，并依据知识共享署名4.0国际许可协议（Creative Commons Attribution 4.0 International License）进行发布。 请访问 https://truebees.eu/site/