Char-mander 使用mBackdoor！多语言大语言模型中的跨语言后门攻击研究
Char-mander Use mBackdoor! A Study of Cross-lingual Backdoor Attacks in Multilingual LLMs
AI总结：
该研究聚焦多语言大语言模型的跨语言后门攻击问题，揭示单一语言植入的后门可借共享嵌入空间跨语言迁移，验证了单语言数据投毒即可攻破系统，暴露了模型架构的关键安全漏洞。
摘要：
我们探究多语言大语言模型（mLLMs）中的跨语言后门攻击（Cross-lingual Backdoor ATtacks，X-BAT），揭示了植入单一语言的后门如何通过共享嵌入空间自动迁移到其他语言。以毒性分类为案例研究，我们证明攻击者仅需投毒单一语言的数据就能攻破多语言系统，稀有词和高频词元可作为特定且有效的触发器。我们的发现暴露了一个影响模型架构的关键漏洞，会在信息流中产生隐蔽的后门效应。我们的代码和数据已公开于https://github.com/himanshubeniwal/X-BAT。
英文摘要：
We explore \textbf{C}ross-lingual \textbf{B}ackdoor \textbf{AT}tacks (X-BAT) in multilingual Large Language Models (mLLMs), revealing how backdoors inserted in one language can automatically transfer to others through shared embedding spaces. Using toxicity classification as a case study, we demonstrate that attackers can compromise multilingual systems by poisoning data in a single language, with rare and high-occurring tokens serving as specific, effective triggers. Our findings expose a critical vulnerability that influences the model's architecture, resulting in a concealed backdoor effect during the information flow. Our code and data are publicly available https://github.com/himanshubeniwal/X-BAT.
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。
