Hybrid deep neuro-bagging with attention-driven ensemble optimization for enhanced security threat detection

This article presents Adaptive Deep Neuro Ensemble bagging (ADNE-bag), a Hybrid Deep Neuro-Bagging framework with Attention-Driven Ensemble Optimization, developed to enhance the accuracy, robustness, and interpretability of intrusion detection systems (IDS). The proposed method integrates multiple deep neural network (DNN) base learners trained on stratified bootstrap samples, with architectural and hyperparameter diversity to reduce error correlation. An attention module assigns dynamic, instance-specific weights to each learner based on feature saliency, historical accuracy, and class-specific recall, ensuring optimal aggregation for each input. A feedback loop reinforces under-attended yet discriminative features, improving minority-class attack detection. Experiments on two complementary benchmark datasets—NSL-KDD and CICIDS2017—demonstrate superior performance over state-of-the-art methods. On NSL-KDD, ADNE-Bag achieved 97.84% accuracy, 97.43% F1-score, 97.51% precision, 97.36% recall, a false positive rate of 2.11%, and an Area Under the Receiver Operating Characteristic curve (AUC) of 0.972. On CICIDS2017, it reached 98.92% accuracy, 98.88% F1-score, 98.86% precision, 98.91% recall, a false positive rate of 1.14%, and an AUC of 0.986. Ablation studies confirm that removing the attention mechanism reduces F1-score by over 2%, validating its key role in performance gains. These results highlight ADNE-Bag’s capability to deliver high detection rates, low false alarm rates, and strong generalization across both legacy and modern intrusion scenarios.