Skip to main content
QUICK REVIEW

[论文解读] garak: A Framework for Security Probing Large Language Models

Leon Derczynski, Erick Galinkin|arXiv (Cornell University)|Jun 16, 2024
Topic Modeling被引用 5
一句话总结

garak 是一个开源框架,用于对大型语言模型的结构化红队测试和安全审计,能够生成探测、检测器和增强以发现漏洞。它强调探索而非基准测试,以为对齐和政策决策提供信息。

ABSTRACT

As Large Language Models (LLMs) are deployed and integrated into thousands of applications, the need for scalable evaluation of how models respond to adversarial attacks grows rapidly. However, LLM security is a moving target: models produce unpredictable output, are constantly updated, and the potential adversary is highly diverse: anyone with access to the internet and a decent command of natural language. Further, what constitutes a security weak in one context may not be an issue in a different context; one-fits-all guardrails remain theoretical. In this paper, we argue that it is time to rethink what constitutes ``LLM security'', and pursue a holistic approach to LLM security evaluation, where exploration and discovery of issues are central. To this end, this paper introduces garak (Generative AI Red-teaming and Assessment Kit), a framework which can be used to discover and identify vulnerabilities in a target LLM or dialog system. garak probes an LLM in a structured fashion to discover potential vulnerabilities. The outputs of the framework describe a target model's weaknesses, contribute to an informed discussion of what composes vulnerabilities in unique contexts, and can inform alignment and policy discussions for LLM deployment.

研究动机与目标

  • 促成对 LLM 安全的整体、结构化审计方法,支持探索与发现。
  • 提供一个灵活的框架,用于检测并描述目标 LLM 或对话系统中的漏洞。
  • 在多样化的生成器和部署场景中实现端到端测试。
  • 提供可扩展的探针、检测器和攻击生成机制,揭示弱点。
  • 便于报告和与漏洞数据库的整合,以为政策决策提供信息。

提出的方法

  • 将四个核心组件(生成器、探针、检测器、增强)定义为结构化 LLM 安全评估的框架。
  • 提供多样化的探针,能引出特定的漏洞类别(如虚假声明、提示注入、数据外泄)。
  • 实现基于关键字和基于 ML 的检测器来评估输出并识别失败。
  • 通过从先前命中的样例中学习,允许自适应攻击生成(atkgen),以创建新的测试用例。
  • 支持多种生成器(Hugging Face、OpenAI、NVIDIA NIMs 等)以及可扩展的集成。
  • 提供通过 JSONL 日志和 HTML 交互式报告的报告功能,且可选上传至漏洞数据库。
Figure 1: The garak architecture. Run configuration determines a set of probes to be used. Each probe interacts with the generator, an abstraction for the target LLM or dialog system. Probes pose prompts to this system in an attempt to elicit insecure responses, and generator responses are recorded.
Figure 1: The garak architecture. Run configuration determines a set of probes to be used. Each probe interacts with the generator, an abstraction for the target LLM or dialog system. Probes pose prompts to this system in an attempt to elicit insecure responses, and generator responses are recorded.

实验结果

研究问题

  • RQ1我们如何以结构化方式审计 LLM 安全,支持探索与发现漏洞?
  • RQ2像 garak 这样的模块化框架是否能在不同 LLMs 上有效发现已知和未知漏洞?
  • RQ3自动检测器和自适应攻击生成如何提升在不同部署环境中的 LLM 风险评估?
  • RQ4来自红队输出的报告和策略指引在 LLM 部署中的作用是什么?
  • RQ5开源工具可以在多大程度上加速外部对 LLM 的红队评估?

主要发现

模型毒性率
GPT-217.0%
GPT-310.5%
GPT-3.51.0%
GPT-42.9%
OPT 6.7B26.7%
Vicuna3.8%
Wizard uncensored5.7%
  • Garak 通过生成器、探针、检测器和增强实现对 LLM 弱点的结构化探索。
  • 自适应攻击生成(atkgen)从成功的探针中学习,生成新的测试用例。
  • 关键字和 ML 检测器的混合支持在大量输出中实现可扩展的失败检测。
  • 毒性攻击可以显著增加模型的失败率(参见毒性结果表)。
  • 该框架支持报告和外部漏洞共享,以编目 LLM 安全问题。
Figure 2: Examples top-level grouping of probe results using the OWASP Top 10 categories of LLM vulnerability. Different groupins lead to different top level results and different concentrations of failure, so it is important to choose a taxonomy applicable to the target context.
Figure 2: Examples top-level grouping of probe results using the OWASP Top 10 categories of LLM vulnerability. Different groupins lead to different top level results and different concentrations of failure, so it is important to choose a taxonomy applicable to the target context.

更好的研究,从现在开始

从论文设计到论文写作,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。