arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2407.03876cs.CRcs.CL
收藏

自动化渐进式红队测试

Automated Progressive Red Teaming

College of Intelligence and Computing, Tianjin University(天津大学智能与计算学部);Du Xiaoman Finance(度小满金融)

Bojian Jiang, Yi Jing, Tianhao Shen, Tong Wu, Qing Yang, Deyi Xiong

AI总结:

本文提出自动化渐进式红队测试框架APRT,通过意图扩展、意图隐藏和恶意制造器三个模块渐进探索LLM漏洞,并引入攻击有效率(AER)指标,在开源与闭源LLM上均展现出强大的攻击与迁移能力。

摘要:

确保大型语言模型(LLMs)的安全性至关重要,但识别潜在漏洞具有挑战性。虽然手动红队测试有效,但它耗时、成本高且缺乏可扩展性。自动化红队测试(ART)提供了一种更具成本效益的替代方案,能够自动生成对抗性提示以暴露LLM漏洞。然而,在当前的ART研究中,缺乏一个稳健的框架将红队测试明确界定为一项可有效学习的任务。为弥补这一空白,我们提出了自动化渐进式红队测试(APRT)作为一个可有效学习的框架。APRT利用三个核心模块:一个生成多样化初始攻击样本的意图扩展LLM(Intention Expanding LLM),一个精心制作欺骗性提示的意图隐藏LLM(Intention Hiding LLM),以及一个管理提示多样性并过滤无效样本的恶意制造器(Evil Maker)。这三个模块通过多轮交互共同渐进地探索和利用LLM漏洞。除框架外,我们进一步提出了一种新指标——攻击有效率(Attack Effectiveness Rate, AER),以缓解现有评估指标的局限性。通过测量引发不安全但看似有用的响应的可能性,AER与人类评估高度一致。大量包含自动和人工评估的实验证明了APRT在开源和闭源LLM上的有效性。具体而言,APRT有效地从Meta的Llama-3-8B-Instruct中诱发了54%不安全但有用的响应,从GPT-4o(API访问)中诱发了50%,从Claude-3.5(API访问)中诱发了39%,展示了其强大的攻击能力和跨LLM的迁移性(尤其是从开源LLM到闭源LLM)。

英文摘要:

Ensuring the safety of large language models (LLMs) is paramount, yet identifying potential vulnerabilities is challenging. While manual red teaming is effective, it is time-consuming, costly and lacks scalability. Automated red teaming (ART) offers a more cost-effective alternative, automatically generating adversarial prompts to expose LLM vulnerabilities. However, in current ART efforts, a robust framework is absent, which explicitly frames red teaming as an effectively learnable task. To address this gap, we propose Automated Progressive Red Teaming (APRT) as an effectively learnable framework. APRT leverages three core modules: an Intention Expanding LLM that generates diverse initial attack samples, an Intention Hiding LLM that crafts deceptive prompts, and an Evil Maker to manage prompt diversity and filter ineffective samples. The three modules collectively and progressively explore and exploit LLM vulnerabilities through multi-round interactions. In addition to the framework, we further propose a novel indicator, Attack Effectiveness Rate (AER) to mitigate the limitations of existing evaluation metrics. By measuring the likelihood of eliciting unsafe but seemingly helpful responses, AER aligns closely with human evaluations. Extensive experiments with both automatic and human evaluations, demonstrate the effectiveness of ARPT across both open- and closed-source LLMs. Specifically, APRT effectively elicits 54% unsafe yet useful responses from Meta's Llama-3-8B-Instruct, 50% from GPT-4o (API access), and 39% from Claude-3.5 (API access), showcasing its robust attack capability and transferability across LLMs (especially from open-source LLMs to closed-source LLMs).

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑