UnsafeBench
收藏资源简介:
UnsafeBench是由CISPA亥姆霍兹信息安全中心创建的一个大型数据集,包含10,000张真实世界和AI生成的图像,这些图像根据11种不安全类别(如暴力、性、仇恨等)进行标注。该数据集旨在评估图像安全分类器的效果和鲁棒性,特别是在生成AI时代。数据集的创建过程涉及从公共数据库中精心挑选图像,并通过三位作者的标注来确定图像的安全性。UnsafeBench的应用领域包括帮助研究社区更好地理解图像安全分类的现状,并开发更有效的图像内容审核工具。
UnsafeBench is a large-scale dataset created by the CISPA Helmholtz Center for Information Security. It contains 10,000 real-world and AI-generated images, which are annotated under 11 unsafe categories such as violence, sexual content, hate speech, and others. This dataset aims to evaluate the performance and robustness of image safety classifiers, particularly in the era of generative AI. The construction of UnsafeBench involves carefully curating images from public databases, with annotations completed by three authors to determine the safety level of each image. The applications of UnsafeBench include helping the research community better understand the current state of image safety classification and developing more effective image content moderation tools.




