通过调整模型对指令的理解方式，有效防御越狱攻击和后门漏洞。
Latent Instruction Representation Alignment: defending against jailbreaks, backdoors and undesired knowledge in LLMs
- 在隐层层面对齐指令表示，让模型更准确理解指令意图。
- 阻断99%以上PEZ越狱攻击，清除复杂代码后门。
- 适合需安全可控的LLM部署场景，如医疗、金融领域。
我们针对大语言模型（LLMs）的越狱攻击、后门漏洞和知识遗忘问题提出解决方案。与以往通过模型行为训练不同，我们的方法直接训练模型改变对指令的解读方式。所提方法Latent Instruction Representation Alignment（LIRA）显著提升泛化能力，并通过内部对抗训练进一步增强效果。实验表明，该方法可阻断超过99%的PEZ越狱攻击，成功移除一个复杂的不安全代码后门，并在WMDP网络安全数据集上实现最优遗忘效果，同时对良性能力损失极小。
原文摘要 · Abstract (English)
We address jailbreaks, backdoors, and unlearning for large language models (LLMs). Unlike prior work, which trains LLMs based on their actions when given malign instructions, our method specifically trains the model to change how it interprets instructions. Our method, Latent Instruction Representation Alignment (LIRA), greatly improves generalization. We further boost generalization through an internally adversarial training algorithm. Our methods block over 99% of PEZ jailbreak attacks; remove a challenging insecure code backdoor; and achieve optimal forgetting on WMDP cyber with negligible loss of benign capabilities.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。