北京信息科技大学学报（自然科学版）2026，Vol.41Issue(4)：75-82,后插1,9.DOI:10.16508/j.cnki.11-5866/n.2026.04.009
基于混合架构的SLM-LLM协同垃圾邮件检测方法
An SLM-LLM collaborative spam detection method based on hybrid architecture
刘志豪 1郭师光1
作者信息
- 1. 北京信息科技大学 信息与通信工程学院,北京 100192
- 折叠
摘要
Abstract
Given the escalating threat of spam evolving from"mass-dissemination"to"semantic confrontation"driven by artificial intelligence generated content(AIGC),traditional small language models(SLMs)encounter cognitive blind spots when facing semantic de-featurization attacks,whereas the direct application of large language models(LLMs)is hindered by high computational costs and inherent judgment biases.Regarding this issue,a SLM-LLM collaborative spam detection method based on hybrid architecture was proposed.An in-depth investigation into the efficacy boundaries of SLMs and LLMs within the context of Chinese spam detection was conducted,in which the"politeness bias"induced by safety alignment mechanisms in general-purpose LLMs and its associated mechanism for false negatives was analyzed,and a structured chain-of-thought(CoT)compensation strategy was proposed.Consequently,an SLM-LLM asynchronous collaborative detection architecture was designed and implemented,leveraging a confidence window mechanism to facilitate dynamic traffic shunting and expert-level deep arbitration.Experimental results indicate that while maintaining millisecond-level average response latency,the proposed architecture reduces comprehensive computational costs by approximately 84.78%,improves the F1 score to 0.950 7,and significantly outperforms single-model baselines,such as LoRA-fine-tuned LLMs and distilled SLMs,thereby providing an efficient and cost-effective technical paradigm for email security governance in the AIGC era.关键词
垃圾邮件检测/大语言模型/思维链推理/双模型混合架构/安全对齐偏见Key words
spam detection/large language model(LLM)/chain-of-thought(CoT)reasoning/dual-model hybrid architecture/safety alignment bias分类
信息技术与安全科学引用本文复制引用
刘志豪,郭师光..基于混合架构的SLM-LLM协同垃圾邮件检测方法[J].北京信息科技大学学报（自然科学版）,2026,41(4):75-82,后插1,9.
