To address the problem that the traffic data in the actual network environment shows significant non-independent and identical distribution (Non-IID) characteristics, a classification method that integrates large language models (LLM) with instruction tuning is proposed. In this method, the understanding capability of LLM for structured data is used to achieve encrypted traffic classification, and traffic data is converted into graph structures to explicitly represent logical relationships among flows. Based on this, a two-stage training framework is constructed. Firstly, the LLM is enabled to comprehend the semantics of traffic graphs by using graph structures instruction tuning. Subsequently, the reasoning capability of LLM is guided to the classification task through the traffic task instruction tuning. These two stages form a collaborative relationship with progressive capability enhancement. Experiments on USTC-TFC2016 and CIC-IDS-2017 datasets show that macro-average F1 scores reach 0.945 and 0.968 by the using proposed method, respectively, demonstrating significant improvement over those of mainstream baseline models. The results indicate that complex dependencies in traffic graphs can be effectively learned by using the proposed method, and a feasible approach for handling Non-IID traffic data is provided through the hierarchical learning mechanism.
针对上述问题,本研究旨在探索一种能够有效处理非独立同分布特性的加密流量分类方法。具体研究内容包括:提出一种将网络流量转化为图结构的数据表示方法,以显式刻画流内时序关系和流间逻辑关联;构建基于大语言模型(Large Language Models, LLM)的指令调优框架,通过分阶段的调优策略使模型逐步掌握流量图结构理解能力和分类任务执行能力;设计系统的实验方案,验证所提方法在典型应用场景下的有效性。
DARWISHR, ROYK. Comparative analysis of federated learning, deep learning, and traditional machine learning techniques for IoT malware detection[C]//Proceedings of the 2025 IEEE 4th International Conference on AI in Cybersecurity. Piscataway, USA: IEEE, 2025. DOI: 10.1109/ICAIC63015.2025.10849203 .
[3]
MELVINA A R, KATHRINEJ W, JEYABOSEA, et al. A deep learning model leveraging time-series system call data to detect malware attacks in virtual machines[J]. International Journal of Computational Intelligence Systems, 2025,18(1):No.58.
GORSHENINA, KOZLOVSKAYAA, GORBUNOVS, et al. Mobile network traffic analysis based on probability-informed machine learning approach[J]. Computer Networks, 2024,247:No.110433.
LINX J, XIONGG, GOUG P, et al. Respond to change with constancy: instruction-tuning with LLM for non-I.I.D. network traffic classification[J]. IEEE Transactions on Information Forensics and Security, 2025,20:5758-5773.
[12]
PLETTENBERGP, KHÖLERD, SICKB, et al. Flow-attentional graph neural networks[DB/OL]. (2025-09-27)[2025-10-05].
[13]
FENGR J, PEARCEH, LIGUORIP, et al. CGP-tuning: structure-aware soft prompt tuning for code vulnerability detection[J]. IEEE Transactions on Software Engineering, 2025,51(9):2533-2548.
[14]
ZHANGX X, WANGQ, QINM Y, et al. Enhanced few-shot malware traffic classification via integrating knowledge transfer with neural architecture search[J]. IEEE Transactions on Information Forensics and Security, 2024,19:5245-5256.
[15]
LIUZ, LIT, YANGY, et al. Encrypted malicious traffic detection based on spatial-temporal graph learning[C]//Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security. New York, USA: ACM, 2023:345-357.