返回全部动态
LLM 对抗性攻击综述
原标题：Adversarial Attacks on LLMs
AI 摘要
本文由 Lil'Log 发布，概述了针对大型语言模型（LLM）的对抗性攻击方法，包括威胁模型、白盒与黑盒攻击的区别，以及五种攻击类型：令牌操作、基于梯度的攻击、越狱提示、人工红队和模型红队。文章重点介绍了令牌操作攻击，如 TextFooler 和 BERT-Attack，并指出这些攻击旨在触发模型输出不安全内容。
以上摘要由 AI 生成，可能存在误差。事实请以原文为准。
正文节选
The use of large language models in the real world has strongly accelerated by the launch of ChatGPT. We (including my team at OpenAI, shoutout to them) have invested a lot of effort to build default safe behavior into the model during the alignment process (e.g. via RLHF). However, adversarial attacks or jailbreak prompts could potentially trigger the model to output something undesired. A large body of ground work on adversarial attacks is on images, and differently it operates in the continuo
发布时间：2023-10-25 08:00
抓取时间：2026-08-02 00:26
来源机构：Lilian Weng