BadEdit: Implanting backdoors into large language models via model editing

Researchers propose BadEdit, which frames backdoor injection as a lightweight knowledge editing problem, implanting backdoors by directly modifying large language model parameters. The method requires only 15 samples to complete injection and adjusts only some parameters, thereby reducing time consumption. Experiments show that the attack success rate on pretrained large language models reaches up to 100%, while maintaining the model's performance on benign inputs, and the backdoor can persist after subsequent fine-tuning or instruction fine-tuning.