RAID
收藏资源简介:
RAID数据集是一个用于测试AI生成图像检测器对抗鲁棒性的大型数据集,包含72k个多样且高度可迁移的对抗性示例。该数据集通过使用多种对抗性攻击方法对七个最先进的检测器和四种不同的文本到图像模型生成的图像进行攻击而创建。实验结果表明,这些对抗性图像在迁移到未见过的检测器上时具有较高的成功率,可用于快速评估检测器的对抗鲁棒性。数据集的创建旨在解决当前AI生成图像检测器易受对抗性示例欺骗的问题,强调了开发更鲁棒方法的重要性。
The RAID dataset is a large-scale benchmark dataset designed for evaluating the adversarial robustness of AI-generated image detectors. It contains 72,000 diverse and highly transferable adversarial examples. This dataset is constructed by applying multiple adversarial attack methods to images generated by four distinct text-to-image models, targeting seven state-of-the-art AI-generated image detectors to generate these adversarial samples. Experimental results demonstrate that these adversarial images exhibit high success rates when transferred to unseen detectors, making them suitable for rapid assessment of the adversarial robustness of detectors. The dataset was developed to address the prevalent issue that current AI-generated image detectors are easily deceived by adversarial examples, underscoring the significance of developing more robust detection methodologies.
数据集概述
基本信息
- 数据集名称: RAID
- 托管平台: Hugging Face
- 许可证: MIT
许可证说明
- 许可证类型: MIT

- 1RAID: A Dataset for Testing the Adversarial Robustness of AI-Generated Image DetectorsUniversity of Cagliari, Italy; Ruhr University Bochum, Germany; University of Modena and Reggio Emilia, Italy; University of Pisa, Italy; Sapienza University of Rome, Italy; CINI, Italy · 2025年



