Hybrid Deep Architectures for Imbalanced WSN Attack Classification: Comparative Evaluation on the WSN-DS Dataset
The wireless sensor network (WSN) technology is getting increasingly vulnerable to highly advanced routing and MAC layer attacks. Most of the existing IDS solutions utilise only one paradigm-based model, which cannot consider multiple feature structures of spatial, temporal, and relational nature simultaneously. In this paper, we propose and evaluate two new hybrid deep learning approaches, namely, CNN–TCN–CapsuleNet–LSTM and Transformer–BiGRU, for five-class WSN-DS attack classification. Both hybrid models were compared with four single-paradigm baselines (LSTM, TCN, CapsuleNet, Transformer) and nine other previously proposed models. The CNN–TCN–CapsuleNet–LSTM model achieved the record-high accuracy of 98.42%, outperforming the previous state-of-the-art solution of 99.39% and six other internal baseline models. The findings show the potential of hybrid multi-stream architectures, offering significant performance advantages compared to traditional models, especially in the case of minority attack classes due to the class imbalance problem.
Authors
- Doaa Mhmood Abbas
Publication Details
- Journal
- Iraqi Journal of Intelligent Computing and Informatics (IJICI)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.52940/ijici.v5i2.758
- Primary Topic
- Security in Wireless Sensor Networks
- Type
- article
- Field-Weighted Citation Impact
- 0.00