Research on Data Integrity Protection and Encryption Technologies for Large Language Models

Addressing critical challenges in data integrity and privacy protection for large language models (LLMs) during training and inference, this paper introduces a ChatGLM-3-based security framework designed to overcome limitations in current semantic integrity verification and homomorphic encrypted inference methods. Our approach begins by constructing a data lifecycle risk model using STRIDE threat analysis and Attack Tree modeling, precisely identifying key tampering vulnerabilities. We then develop a semantic hash embedding mechanism leveraging Sentence-BERT combined with Merkle Tree structures, enabling robust semantic-level integrity verification. For secure inference, we implement an end-to-end encrypted pipeline based on the CKKS homomorphic encryption scheme. Furthermore, we integrate differential privacy gradient perturbation with comprehensive training audit logs, establishing a closed-loop defense against data leakage. Experimental validation on the LLaMA 2-13B platform demonstrates the effectiveness of our solution: the semantic hashing showed high robustness with minimal latency, homomorphic inference maintained accuracy at acceptable latency while effectively thwarting reconstruction attacks, and the combined DP-audit mechanism significantly reduced leakage probability enabling effective traceability. These results confirm that our integrated framework strengthens data integrity and privacy safeguards for LLMs without severely compromising utility, offering a practical path for deployment in security-sensitive domains.