摘要
Abstract
[Purpose]Given the bottlenecks of traditional hyperparameter optimization methods in threat intelligence entity recognition—such as low search efficiency and the difficulty in balancing performance and time consumption—along with the statutory requirements for data security and high-quality entity data specified in the Cybersecurity Law of the People's Republic of China,this study aims to propose a highly efficient and collaborative hyperparameter optimization solution,which provides technical support for network security situation awareness,attack chain tracing,and risk prediction.[Method]This study proposes a hyperparameter optimization method for threat intel-ligence entity recognition based on multi-agent collaboration,denoted as TIER-MACH.With the Qwen2.5 large language model as the foundational architecture,a dual-agent framework is constructed,comprising a"parameter generator"and a"parameter executor".The generator agent generates and outputs candidate hyperparameters by leveraging two core capabilities:in-depth semantic parsing of the tar-get task and dynamic parameter generation driven by historical optimization data.The executor agent undertakes three key functions:effi-cient execution of the generated hyperparameters,full-cycle monitoring of model training processes,and timely feedback on model per-formance.These two agents operate in synergy to form a closed-loop mechanism:"parameter generation → model training → perform-ance feedback → iterative parameter optimization",thereby achieving precise hyperparameter tuning.[Result/Conclusion]Experimental results demonstrate that TIER-MACH enables performance improvements across all baseline models,including deep learning models and traditional machine learning models(SVM,RF).When validated on two datasets(DNRTI and CDTier),the method stably increases model precision,recall,and F1-score by 1.6％to 3.0％.Furthermore,TIER-MACH effectively resolves the inherent trade-off in tradi-tional methods—namely,"high performance inevitably entails high time consumption"or"low time consumption is coupled with perform-ance degradation".It not only enhances recognition accuracy,but also controls optimization time within a more efficient range,fully satis-fying the requirements of precision and real-time responsiveness for threat intelligence entity recognition tasks.关键词
多智能体/超参数优化/大语言模型/威胁情报/威胁情报实体识别方法Key words
multi-agent/hyperparameter optimization/large language model/threat intelligence/TIER-MACH分类
社会科学
