文献类型：期刊文献

英文题名：HybridCrypt-LLM: Lightweight privacy for LLM training and inference

作者：Li, Te[1];Guo, Yi[1];Fu, Jiaojiao[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Meilong Rd 130, Shanghai 200237, Peoples R China

年份：2026

卷号：314

外文期刊名：EXPERT SYSTEMS WITH APPLICATIONS

收录：;EI(收录号：20261420411581);WOS:【SCI-EXPANDED(收录号:WOS:001695520900001)】；

语种：英文

外文关键词：LLMs; Diffie-Hellman; Model obfuscation; Privacy preservation

摘要：The rapid proliferation of Large Language Models (LLMs) has introduced severe privacy risks, as LLMs often require access to sensitive proprietary data during training and inference. Existing privacy-preserving techniques fail to balance security and utility: Federated Learning (FL) remains susceptible to gradient-based inversion attacks; Differential Privacy (DP) significantly degrades model accuracy due to noise injection; and Homomorphic Encryption (HE) incurs prohibitive computational overhead, rendering it impractical for real-time applications. This paper proposes HybridCrypt-LLM, a lightweight, end-to-end encryption framework designed to protect data confidentiality without compromising model performance. The framework integrates a vocabulary-level encryption scheme with randomized token substitution to introduce noise into n-gram statistics, thereby increasing the computational complexity required for semantic reconstruction. Furthermore, a key-driven model obfuscation mechanism leverages Diffie-Hellman (DH) key agreement, and encrypted output authentication is implemented to secure the entire computation pipeline. Our analysis demonstrates that the framework provides a practical layer of defense by increasing the complexity of data reconstruction to a computationally unfeasible level for common adversarial threats.