基于微调与检索增强生成的混合式智能协议理解方法

A Hybrid Intelligent Protocol Understanding Method Based on Fine-tuning and Retrieval Augmented Generation

摘要:

针对5G与天地一体化网络等新兴环境中快速演进的网络协议带来的传统解析方法在实时性与可解释性方面难以满足现代网络运维与安全分析需求的问题,对一种面向智能协议理解的检索增强生成系统进行了研究。该系统采用了协议感知的知识预处理、多模态混合索引构建与协议感知生成器微调等关键技术,经在涵盖12种主流协议、近2000个问题的测试集上进行的实验评估,该系统实现了77.97%的软召回率,显著优于Modular RAG与GraphRAG等主流方法；此外,十六进制推理类问题的端到端准确率由58.24%提升至79.63%,同时其他类型问题性能保持稳定。经实际应用验证,该技术满足了智能协议分析与自解释网络等工程场景对高精度、可追溯且无需外部解析器的协议理解需求,为相关领域提供了有效的技术路径。

Abstract:

To address the challenge that traditional parsing methods struggle to meet the real-time and interpretability requirements of modern network operation and security analysis in emerging environments such as 5G and integrated space-ground networks, which are driven by rapidly evolving network protocols, this paper studies a retrieval augmented generation system for intelligent protocol understanding. This system employs key technologies such as protocol-aware knowledge preprocessing, multimodal hybrid index construction, and protocol-aware generator fine-tuning. Experimental evaluation on a test set covering 12 mainstream protocols and nearly 2000 questions demonstrates a soft recall rate of 77.97%, significantly outperforming mainstream methods such as Modular RAG and GraphRAG. Furthermore, the end-to-end accuracy for hexadecimal reasoning questions improved from 58.24% to 79.63%, while performance remained stable for other question types. Practical application verification confirms that this technology meets the requirements of high-precision, traceable, and external-requirement-free protocol understanding in engineering scenarios such as intelligent protocol analysis and self-explaining networks, providing an effective technical path for related fields.