Toward Trusted 6G Mobile Edge Computing: A Secure Batch Large Language Models Deployment Framework

Large Language Models (LLMs) sparked massive applications in 6G. However, the emerging 6G Mobile Edge Computing (MEC) based on LLM caching leaves model protection unconsidered. To protect the LLM assets under the extended Dolev-Yao (DY) threat model, a secure batch LLMs deployment framework is proposed for 6G MEC, which securely delivers the sanitized crafted (san-crafted) LLM and the crafted-random-values (CR-values) from 6G edge to the Rich Execution Environment (REE) and Trusted Execution Environment (TEE) of mobile devices, respectively. Firstly, the 6G MEC cached LLM is san-crafted by an efficiency-improved sanitizable signature to protect the integrity of the LLM during the entire deployment process. Then, a lightweight batch authentication protocol is proposed to improve the efficiency of verifying ultra-massive model requests. Finally, to be compatible with the state-of-the-art secure inference (i.e., Magnitude), the san-crafted LLM delivered into mobile device's REE is verified by sanitizable signatures, and then conducts secure inference with the corresponding CR-values provisioned into the TEE. Rigorous security proofs confirm that our framework meets the security requirements of LLM deployment. Compared to the benchmark, the proposed framework significantly improves both computational and communication efficiency, reducing computational overhead by 89.92% and communication overhead by 47.08%. This framework facilitates the TEE-based LLMs secure inference for ultra-massive mobile devices in 6G MEC.