Have You Merged My Model? On the Robustness of Large Language Model IP Protection Methods Against Model Merging

Abstract. Model merging techniques enable combining the capabilities of multiple fine-tuned models without retraining, which has important implications for intellectual property (IP) protection of large language models. Various IP protection methods, including watermarking, model fingerprinting, and ownership verification, have been proposed to safeguard model owners’ rights. In this paper, we systematically evaluate the robustness of state-of-the-art LLM IP protection methods against model merging attacks. We show that many existing IP protection mechanisms can be evaded when the protected model is merged with other models using popular merging techniques such as model averaging, task arithmetic, and TIES-merging. We analyze the factors that affect the robustness of different protection methods and propose guidelines for designing IP protection that is resilient against model merging. Our work highlights a significant vulnerability in current LLM IP protection and provides insights for developing more robust protection mechanisms.