网络安全与数据治理2025，Vol.44Issue(11)：1-6,6.DOI:10.19358/j.issn.2097-1788.2025.11.001
领域大语言模型的内容安全控制研究
Research on content safety control of domain-specific large language models
张欣欣 1李涛 1赵龙彪 1贾真真 2周衡广3
作者信息
- 1. 中国人民解放军92981 部队,北京 100161
- 2. 中国人民解放军91977 部队,北京 100036
- 3. 中国人民解放军91526 部队,广东 湛江 524064
- 折叠
摘要
Abstract
With the increasing adoption of large language models in specialized domains,these models have demonstrated signifi-cant potential in areas such as knowledge management,decision support,and secure information exchange.However,given the high level of specialization and sensitivity in these domains,ensuring the safety and compliance of generated content in specific scenarios presents a major challenge.Current approaches predominantly rely on model retraining or fine-tuning,which are re-source-intensive and lack flexibility.This study proposes a refined output control method that bypasses the need for model retrain-ing.By framing output control as a classification problem,classification algorithms are employed to evaluate generated content and determine its appropriateness for release.This mechanism combines mathematical modeling and feature engineering to strike a bal-ance between meeting business requirements and minimizing potential risks,thereby enhancing the safety and compliance of gen-erated outputs.关键词
大语言模型/安全控制/内容过滤/分类算法Key words
large language model/safety control/content filtering/classification algorithm分类
信息技术与安全科学引用本文复制引用
张欣欣,李涛,赵龙彪,贾真真,周衡广..领域大语言模型的内容安全控制研究[J].网络安全与数据治理,2025,44(11):1-6,6.
