PEFT-As-An-Attack, Jailbreaking Language Models For Malicious Prompts

The report examines PEFT-as-an-Attack (PaaA) in federated parameter-efficient fine-tuning (FedPEFT), showing that malicious clients can inject toxic data to bypass safety alignment and generate harmful outputs. Across multiple PLMs and PEFT methods on domain-specific QA tasks, robust aggregation schemes fail under heterogeneous data, while post-PEFT safety alignment mitigates attacks at a significant accuracy cost; LoRA achieves the best accuracy but is more vulnerable. The findings highlight the need for new defenses that jointly preserve safety and model performance.