@Article{cmc.2026.084490,
AUTHOR = {Ye Lu, Haoyang Hu, Wenyi Chang, Wanbin Liu, Wenyuan Zhang},
TITLE = {A Compact Hybrid TCN-BiGRU-TinyTransformer Framework for Fine-Grained Multiclass Intrusion Detection},
JOURNAL = {Computers, Materials \& Continua},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/cmc/online/detail/27668},
ISSN = {1546-2226},
ABSTRACT = {Fine-grained multiclass intrusion detection over flow-level traffic remains difficult, largely because class boundaries are often entangled, temporal dependence is non-negligible, and the label distribution is heavily long-tailed. In this study, a compact temporal convolutional network (TCN)-bidirectional gated recurrent unit (BiGRU)-TinyTransformer framework is developed to bring these issues into a single modeling pipeline: the TCN branch focuses on short-range anomalous patterns, the BiGRU branch captures bidirectional temporal structure, and the TinyTransformer branch complements them with broader contextual interaction learning. To reduce the bias induced by extreme imbalance, training is not driven by a single correction mechanism, but by a coordinated strategy that combines class-balanced (CB) sampling and weighting, label-distribution-aware margin (LDAM)-Focal loss, and staged warmup plus deferred re-weighting (DRW) optimization. On the NetFlow-based UNSW-NB15-v2 dataset (NF-UNSW-NB15-v2), the resulting model reaches 98.06% Accuracy, 98.05% Weighted-F1, and 80.19% Macro-F1. These results point to a practical improvement in fine-grained attack discrimination, especially for minority classes, without weakening overall detection performance.},
DOI = {10.32604/cmc.2026.084490}
}