An adaptive multi-stage graph-based intrusion detection framework for securing internet of things networks
Keywords:
Internet of Things, intrusion detection system, graph neural networks, adaptive framework, GraphSAGE, GAT, IoTID20, edge computing, cybersecurityAbstract
This paper proposes an adaptive multi-stage intrusion detection framework for Internet of Things (IoT) networks based on graph neural networks (GNNs). Unlike record-level formulations, devices are modelled as graph nodes and traffic flows between them as edges, and the graph is decomposed into dynamic snapshots whose window length adapts to the observed traffic density. In the first stage a lightweight GraphSAGE model performs binary "normal/attack" classification with a decision threshold that is re-estimated online from the recent false-positive rate; the second stage, a Graph Attention Network (GAT), is invoked only for the flagged subgraph and identifies the attack type. A periodic incremental update step keeps the framework aligned with concept drift. Experiments on the public IoTID20 dataset yielded 98.4 % accuracy for the first stage and 96.9 % for the second. The adaptive cascade reduced computation by about 38 % compared with a single-stage multi-class model and processed one snapshot in 12.7 ms on a Raspberry Pi 4 class gateway, which makes the framework deployable on resource-constrained IoT edge devices
References
1. Ahanger A., Khan S., Masoodi F., Ahmad J. Design of an Efficient Multi-Stage Graph-Based Intrusion Detection System for the Internet of Things // Emerging Trends in Smart Computing, Communication and Automation. — CRC Press, 2026. — P. 177–181. — DOI: 10.1201/9781042026333-42.
2. Koroniotis N., Moustafa N., Sitnikova E., Turnbull B. Towards the development of realistic botnet dataset in the Internet of Things networks: Bot-IoT dataset // Future Generation Computer Systems. — 2019. — Vol. 100. — P. 779–796.
3. Ullah I., Mahmoud Q.H. A scheme for generating a dataset for anomalous activity detection in IoT networks // Canadian Conference on Artificial Intelligence. — Springer, 2020. — P. 508–520. (IoTID20)
4. Kipf T.N., Welling M. Semi-supervised classification with graph convolutional networks // ICLR. — 2017.
5. Hamilton W., Ying Z., Leskovec J. Inductive representation learning on large graphs // NeurIPS. — 2017. — P. 1024–1034.
6. Veličković P., Cucurull G., Casanova A., Romero A., Liò P., Bengio Y. Graph attention networks // ICLR. — 2018.
7. Moustafa N. A new distributed architecture for evaluating AI-based security systems at the edge: Network TON_IoT datasets // Sustainable Cities and Society. — 2021. — Vol. 72. — 102994.

