Syntactic paraphrase-based synthetic data generation for backdoor attacks against Chinese language models

Man Hu, Yatao Yang, Deng Pan, Zhongliang Guo, Luwei Xiao, Deyu Lin, Shuai Zhao

Abstract: Highlights•We first attempt to implement Chinese backdoor attacks using syntactic structure.•We design a syntactic paraphrase model based on Qwen2.5 fine-tuned with LoRA.•We propose a backdoor attack method with three distinct attack strategies.•Our method enhances stealth and significantly improves poisoned sample quality.•Experiments show high attack success and strong resistance to defense mechanisms.