研究人员发现了与部署过程中使用的优化技术相关的大型语言模型（LLM）的新漏洞。一项研究表明，旨在提高效率的编译过程可能被利用来植入隐藏的后门，这些后门在特定的编译条件下触发，绕过标准的安全性检查，并在开源LLM上实现高攻击成功率。另一篇理论论文探讨了，与直觉相反的是，在后门攻击中更强的触发器有时可以在高维环境中帮助防御者，攻击成功率在有限的触发器强度下达到峰值。 AI
影响 新研究强调了LLM部署管道中的关键安全漏洞，可能影响AI系统的安全性和可靠性。
排序理由 多篇学术论文发表在arXiv上，详细介绍了关于LLM漏洞和后门攻击理论方面的新研究。
AI 生成摘要 · Google Gemini · 来自 5 个来源。 我们如何撰写摘要 →
arXiv:2605.20641v1 Announce Type: cross Abstract: Inference optimization is a vital technique for deploying LLMs at scale. Compilation is the most widely adopted optimization technique for LLMs. While it assumes semantic equivalence between the original and compiled graphs, we fi…
arXiv:2605.21780v1 Announce Type: new Abstract: Randomized smoothing is a powerful tool for certifying robustness to adversarial perturbations, including poisoning attacks via randomized training and evasion attacks via randomized inference. Extending these guarantees to backdoor…
arXiv:2605.22481v1 Announce Type: new Abstract: Backdoor poisoning attacks behave counter-intuitively in high dimensions: stronger training triggers can help the defender. We study regularised generalised linear models on Gaussian-mixture data in the proportional regime ($p/n \to…
Backdoor poisoning attacks behave counter-intuitively in high dimensions: stronger training triggers can help the defender. We study regularised generalised linear models on Gaussian-mixture data in the proportional regime ($p/n \to κ$), varying the training trigger strength $α$ …
Inference optimization is a vital technique for deploying LLMs at scale. Compilation is the most widely adopted optimization technique for LLMs. While it assumes semantic equivalence between the original and compiled graphs, we first uncover its numerical side effects can be mali…