Privacy-Preserving Large Language Model in Terms of Secure Computing: A Survey

As large language models (LLMs) become increasingly embedded in diverse applications, from natural language processing to cybersecurity, the demand for robust privacy-preserving solutions has surged. This paper discusses privacy vulnerabilities in LLMs, identifying risks at system, application, and network levels. We categorize and evaluate current privacy-preserving methods, based on homomorphic encryption (HE), trusted execution environments (TEE), and secure multi-party computation (MPC), to assess their effectiveness in mitigating data exposure while supporting LLM performance. Building on these insights, we propose a novel privacy-preserving framework for fine-tuning and inference in LLMs. Our findings highlight existing gaps and propose future directions for developing secure, efficient, and scalable privacy-preserving LLM architectures.