A Privacy-preserving LLM Fine-tuning Method for Cross-domain Scenarios

• 1、School of Computer Science(National Pilot Engineering School), Beijing University of Posts and Telecommunications, Beijing, 100876

Abstract：Large language models are natural language processing models based on deep learning, and their performance on specific tasks can be enhanced through fine-tuning with domain-specific data. However, small and medium-sized enterprises often lack the capability to pre-train large models and need to collaborate with other large model owners to conduct cross-domain fine-tuning. Meanwhile, the fine-tuning data may contain privacy information, and its direct use could lead to privacy leakage. Currently, most relevant studies achieve cross-domain fine-tuning with privacy protection by having the fine-tuning data holder send fine-tuning data satisfying differential privacy to the pre-trained large model holder, but the fine-tuned model obtained in this way will not be controlled by the fine-tuning data provider. Furthermore, differential privacy requires adding random noise to the embeddings, which can reduce the model\'s utility. Therefore, this paper proposes a privacy-preserving large language model fine-tuning method for cross-domain scenarios, PTCDS. In this method, to address the issue that the fine-tuned model is not controlled by the fine-tuning data provider, a privacy-preserving cross-domain fine-tuning framework based on split learning is designed; to address the issue of reduced model utility caused by differential privacy, a parameter update denoising method based on embedding availability is designed. Experimental results show that, compared with existing relevant work, PTCDS can effectively protect the privacy of fine-tuning data while achieving high accuracy.

Keywords： Artificial Intelligence Privacy Protection Differential Privacy Large Language Model Prompt Tuning