From Efficiency to Leakage: Privacy Backdoor in Federated Language Model Fine-Tuning

arXiv·medium signal

Shows a malicious parameter server can stealthily turn a PEFT adapter into a privacy backdoor (NeuroImprint) by assigning dedicated neurons to memorize individual training samples without degrading model performance. It recovers 59-79% of fine-tuning samples across BERT, GPT-2, and Llama 3.2. A concrete threat model for federated or outsourced fine-tuning.