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 2310.06987cs.CLcs.AIcs.CR
收藏

通过利用生成过程使开源大语言模型发生灾难性越狱

Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation

Princeton University(普林斯顿大学)

Yangsibo Huang, Samyak Gupta, Mengzhou Xia, Kai Li, Danqi Chen

AI总结:

本研究提出生成利用攻击,仅通过操纵解码方法变化,将11个开源大语言模型的错位率从0%提升至95%以上,并提出了降低错位率的对齐方法,揭示了当前安全评估的缺陷。

摘要:

开源大语言模型(LLMs)的快速发展正在显著推动人工智能的发展。在模型发布之前,人们已经做出了大量努力来使其行为与人类价值观保持一致,主要目标是确保其有用性和无害性。然而,即使是经过精心对齐的模型也可能被恶意操纵,导致意外行为,即所谓的“越狱”。这些越狱通常由特定的文本输入触发,通常称为对抗性提示。在这项工作中,我们提出了生成利用攻击,这是一种极其简单的方法,仅通过操纵解码方法的变化来破坏模型对齐。通过利用不同的生成策略,包括改变解码超参数和采样方法,我们在包括LLaMA2、Vicuna、Falcon和MPT系列在内的11个语言模型上将错位率从0%提高到超过95%,以低30倍的算力成本超越了最先进的攻击。最后,我们提出了一种有效的对齐方法,探索多样化的生成策略,这可以在我们的攻击下合理降低错位率。总的来说,我们的研究强调了当前开源LLM安全评估和对齐流程中的重大缺陷,强烈主张在发布此类模型之前进行更全面的红队测试和更好的对齐。我们的代码可在https://github.com/Princeton-SysML/Jailbreak_LLM获取。

英文摘要:

The rapid progress in open-source large language models (LLMs) is significantly advancing AI development. Extensive efforts have been made before model release to align their behavior with human values, with the primary goal of ensuring their helpfulness and harmlessness. However, even carefully aligned models can be manipulated maliciously, leading to unintended behaviors, known as "jailbreaks". These jailbreaks are typically triggered by specific text inputs, often referred to as adversarial prompts. In this work, we propose the generation exploitation attack, an extremely simple approach that disrupts model alignment by only manipulating variations of decoding methods. By exploiting different generation strategies, including varying decoding hyper-parameters and sampling methods, we increase the misalignment rate from 0% to more than 95% across 11 language models including LLaMA2, Vicuna, Falcon, and MPT families, outperforming state-of-the-art attacks with $30\times$ lower computational cost. Finally, we propose an effective alignment method that explores diverse generation strategies, which can reasonably reduce the misalignment rate under our attack. Altogether, our study underscores a major failure in current safety evaluation and alignment procedures for open-source LLMs, strongly advocating for more comprehensive red teaming and better alignment before releasing such models. Our code is available at https://github.com/Princeton-SysML/Jailbreak_LLM.

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。

↑