EUVIMEDEUVIMED entdecken
Open Access · CC BY-NC-ND 4.0

Distilling the Giant: A Lightweight Command Injection Detection Framework for Grid Edge Networks via LLM‐Guided Contrastive Learning

Hu Song, Yixin Ding, Yumingzhi Pan, Xinjian Zhao, Menghan Wang, Shaofeng Li, Guangchi Liu

IET Communications · 2026-01 · 20 (1)

DOI: 10.1049/cmu2.70188 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IET Communications, online erschienen am 08.07.2026. Verfasst von Hu Song und 6 weiteren Autor:innen. Die Arbeit nennt 23 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

In modern power systems, the communication networks connecting heterogeneous edge devices such as inspection drones and smart charging piles are increasingly targeted by sophisticated, unknown threats, including command injection. While large language models (LLMs) have demonstrated superior semantic capabilities in detecting these complex attacks, their prohibitive computational latency renders them unsuitable for online network defense. To bridge the gap between deep semantic understanding and operational feasibility, we introduce Sentill, a novel framework that distills the reasoning capabilities of giant LLMs into lightweight embedding models. Unlike traditional knowledge distillation that merely mimics output probabilities, Sentill employs a reasoning‐aware contrastive learning strategy, aligning the student model's embedding space with the teacher's chain‐of‐thought (CoT) explanations to capture the underlying logic of attack payloads. Our evaluation using real‐world traffic demonstrates that Sentill enables compact models (e.g., 0.6B parameters) to achieve state‐of‐the‐art detection performance, with a gte‐Qwen2 backbone reaching an F1‐score of 0.8662 and near‐zero false positives. Crucially, the framework delivers a throughput of 346.46 samples per second—representing a speedup over the teacher LLM—making robust, semantic‐aware detection deployable on communication gateways.

Zitieren

APA

Song, H., Ding, Y., Pan, Y., Zhao, X., Wang, M., Li, S., & Liu, G. (2026). Distilling the Giant: A Lightweight Command Injection Detection Framework for Grid Edge Networks via LLM‐Guided Contrastive Learning. IET Communications, 20(1). https://doi.org/10.1049/cmu2.70188

Vancouver

Song H, Ding Y, Pan Y, Zhao X, Wang M, Li S, et al. Distilling the Giant: A Lightweight Command Injection Detection Framework for Grid Edge Networks via LLM‐Guided Contrastive Learning. IET Communications. 2026;20(1). doi:10.1049/cmu2.70188

BibTeX
@article{Song2026,
  author = {Song, Hu and Ding, Yixin and Pan, Yumingzhi and Zhao, Xinjian and Wang, Menghan and Li, Shaofeng and Liu, Guangchi},
  title = {Distilling the Giant: A Lightweight Command Injection Detection Framework for Grid Edge Networks via LLM‐Guided Contrastive Learning},
  journal = {IET Communications},
  year = {2026},
  volume = {20},
  number = {1},
  doi = {10.1049/cmu2.70188}
}
Kostenlos starten →