IoTSplit: Teacher-Guided Federated IoT Intrusion Detection Using Small Language Models
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1019-1026
DOI: 10.1609/aaaiss.v10i1.43521 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Abderrahman Elhajjout und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Internet of Things (IoT) networks produce large volumes of traffic, yet only a small portion may have reliable labels for training an intrusion detector. These labeled examples are distributed across locations that may observe different types of attacks. Federated Learning (FL) allows these locations to learn jointly, while a server teacher can provide predictions to help clients use their unlabeled traffic. However, transmitting every unlabeled flow to the teacher incurs substantial communication overhead. To address this issue, we introduce IoTSplit, a federated intrusion detection framework that consists of: (1) a fixed TinyLlama-based encoder that transforms local flow representations into compact numerical messages, reducing the amount of data transmitted to a server-side teacher; (2) a query-based training strategy that selects unlabeled flows within a query budget and uses the teacher's class probabilities together with available labels to train local detectors; and (3) a coordinator that aggregates local detector updates into a global model, enabling clients to classify new traffic locally without further interaction with the teacher. To evaluate IoTSplit, we conducted experiments using CICIoMT2024 dataset. The experimental results show that compact messages substantially reduce the amount of data transmitted to the teacher. Moreover, randomly selected flows receive more accurate teacher predictions than uncertainty-selected flows, highlighting the importance of teacher reliability when selecting flows for querying.
Zitieren
APA
Elhajjout, A., Moudoud, H., & Houda, Z. A. E. (2026). IoTSplit: Teacher-Guided Federated IoT Intrusion Detection Using Small Language Models. Proceedings of the AAAI Symposium Series, 10(1), 1019-1026. https://doi.org/10.1609/aaaiss.v10i1.43521
Vancouver
Elhajjout A, Moudoud H, Houda ZAE. IoTSplit: Teacher-Guided Federated IoT Intrusion Detection Using Small Language Models. Proceedings of the AAAI Symposium Series. 2026;10(1):1019-1026. doi:10.1609/aaaiss.v10i1.43521
BibTeX
@article{Elhajjout2026,
author = {Elhajjout, Abderrahman and Moudoud, Hajar and Houda, Zakaria Abou El},
title = {IoTSplit: Teacher-Guided Federated IoT Intrusion Detection Using Small Language Models},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {1019--1026},
doi = {10.1609/aaaiss.v10i1.43521}
}
EUVIMED