Transferring Troubles: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning

arXiv:2404.19597v1 Announce Type: new
Abstract: The implications of backdoor attacks on English-centric large language models (LLMs) have been widely examined - such attacks can be achieved by embedding malicious behaviors during training and activated under specific conditions that trigger malicious outputs. However, the impact of backdoor attacks on multilingual models remains under-explored. Our research focuses on cross-lingual backdoor attacks against multilingual LLMs, particularly investigating how poisoning the instruction-tuning data in one or two languages can affect the outputs in languages …

abstract arxiv attacks backdoor cross-lingual cs.cl cs.cr embedding english however impact instruction tuning language language models large language large language models llms training type