Raccoon

Raccoon数据集由杜克大学和麻省大学阿默斯特分校的研究团队创建，专注于评估大型语言模型（LLM）对提示提取攻击的脆弱性。该数据集包含42种不同的攻击策略，涵盖14个攻击类别，旨在通过模拟真实世界的攻击场景来测试模型的安全性。数据集的创建过程涉及专家设计和筛选，确保攻击的有效性和多样性。Raccoon数据集的应用领域主要集中在增强LLM的安全性，通过系统评估帮助研究人员和开发者理解和防范提示提取攻击，从而保护模型免受未授权访问和潜在的知识产权侵犯。

The Raccoon Dataset was developed by research teams from Duke University and the University of Massachusetts Amherst, with the core goal of evaluating the vulnerability of Large Language Models (LLMs) against prompt extraction attacks. This dataset includes 42 distinct attack strategies spanning 14 attack categories, designed to test model security by simulating real-world attack scenarios. The development process of the Raccoon Dataset involved expert design and screening to guarantee the effectiveness and diversity of the attacks. The primary application of the Raccoon Dataset focuses on enhancing LLM security: through systematic evaluation, it helps researchers and developers understand and mitigate prompt extraction attacks, thereby protecting models from unauthorized access and potential intellectual property infringement.