HADS-Net: A Hybrid Attention-Based Deep Security Network for Network Intrusion Detection

Abstract

Network intrusion detection remains a fundamental cybersecurity challenge due to the increasing diversity and sophistication of malicious network traffic. Conventional signature-based approaches exhibit limited capability in detecting previously unseen attacks, while many machine learning methods suffer from class imbalance, high false-positive rates, and limited adaptability across heterogeneous network environments. To address these challenges, this paper proposes HADS-Net (Hybrid Attention-based Deep Security Network), which integrates a feature-wise multi-head self-attention module with a stacked ensemble consisting of Random Forest and Gradient Boosting base learners combined through out-of-fold stacking and a logistic regression meta-learner. The attention mechanism adaptively emphasizes informative network features to improve discriminative learning, while the stacking strategy enhances generalization and reduces overfitting. Experiments were conducted on a class-stratified 15,000-record subsample derived from the NSL-KDD KDDTrain+ corpus, in which the minority R2L and U2R categories were deliberately enriched to obtain usable test support; the subsample was divided into 70% training and 30% testing partitions, giving a held-out test partition of 4,500 records. Because this constructed distribution departs from the native NSL-KDD proportions, the results reported here are not directly comparable with published KDDTest+ figures, and no such comparison is claimed. The proposed model achieved 96.20% accuracy, 97.57% precision, 94.32% recall, 95.92% F1-score, and 98.83% AUC-ROC on the binary Normal-versus-Attack task. A multi-layer perceptron baseline attained higher accuracy (98.11%) and F1-score (97.99%) than the proposed model on the same partition, and the ablation deltas attributable to the attention and stacking components lie below 0.5%; these deltas are reported as single-run point estimates. The principal contribution of this work is consequently architectural and diagnostic rather than a performance improvement: it documents that combining feature-wise attention with out-of-fold stacked generalization does not, in this setting, outperform a plain multi-layer perceptron, while incurring the highest memory footprint of the six models evaluated.