Optimized intrusion detection using bees algorithm enhanced deep neural networks with perfect ROC separability

Intrusion Detection Systems (IDSs) continue to face significant challenges due to high-dimensional network traffic, rapidly evolving cyberattacks, and the need for accurate multi-class recognition. This study proposes a novel hybrid framework that integrates the Bees Algorithm (BA) with a Deep Neural Network (DNN) to enhance feature selection and improve intrusion detection performance. The BA efficiently reduces the original dataset to 22 optimal features, enabling the DNN to learn highly discriminative representations with reduced computational complexity. Model performance was evaluated using a holdout validation strategy with separate training, validation, and testing sets to ensure unbiased assessment. Experimental results demonstrate that the proposed BA–DNN model achieves outstanding classification performance across 12 intrusion classes, attaining an overall accuracy of 99.92%, precision of 97.96%, recall of 100%, F1-score of 98.97%, and zero false positive and false negative rates (FPR = 0.0%, FNR = 0.0%). The ROC curve further yields an AUC of 1.000, confirming perfect separability between normal and malicious traffic. Convergence analysis shows that BA reduces the fitness error to 0.02 within 100 iterations, outperforming PSO, GWO, and GA, whose final errors remain above 0.10. Comparative evaluation also reveals that the proposed approach surpasses several state-of-the-art IDS methods while using fewer features and handling more attack classes. These results demonstrate that the BA–DNN framework provides a highly accurate, efficient, and scalable solution for modern intrusion detection, offering substantial potential for deployment in real-world IoT and cybersecurity environments.