Research on Data Privacy and Information Security Evaluation and Optimization for Large Language Models

With the widespread application of large language models (LLMs), issues concerning data privacy and information security have become increasingly prominent. To address the limitations of existing evaluation methods in terms of comprehensiveness, adaptability to Chinese scenarios, and automation, this study proposes an integrated evaluation and optimization framework. The framework constructs a multi-dimensional evaluation system through three core modules: data layer, model layer, and evaluation metrics. It develops an automated toolkit supporting four types of attack tests and introduces a three-tier collaborative defense scheme combining differential privacy fine-tuning, dynamic desensitization, and prompt reinforcement. Experimental results demonstrate that the framework significantly outperforms existing methods in Chinese PII recognition accuracy (92.3%) and cross-domain stability (variance reduced by 35%). The combined defense solution reduces privacy leakage rates by 60-75% while keeping model utility loss below 15%. This research provides a systematic approach to balancing the performance and security of large language models.