BIPIA
收藏资源简介:
BIPIA数据集由中国科学技术大学和微软公司联合开发,旨在评估大型语言模型在间接提示注入攻击下的风险。该数据集包含多种应用场景和攻击目标,总计626,250个训练样本和86,250个测试样本。数据集通过模拟真实世界中的攻击场景,帮助研究人员和开发者理解和防御此类攻击。BIPIA的应用领域包括电子邮件QA、网络QA、表格QA、摘要生成和代码QA等,旨在提高语言模型在处理外部内容时的安全性和可靠性。
The BIPIA dataset, jointly developed by the University of Science and Technology of China and Microsoft Corporation, is designed to assess the risks faced by large language models (LLMs) under indirect prompt injection attacks. This dataset covers diverse application scenarios and attack targets, with a total of 626,250 training samples and 86,250 test samples. By simulating real-world attack scenarios, it enables researchers and developers to understand such attacks and develop corresponding defense strategies. The applicable domains of BIPIA include email QA, web QA, table QA, text summarization, and code QA, among others, with the goal of enhancing the security and reliability of these LLMs when processing external content.

- 1Benchmarking and Defending Against Indirect Prompt Injection Attacks on Large Language Models中国科学技术大学 · 2024年



