Swarm-Optimized Ensemble Learning for Intrusion Detection using CICIDS2018 and UNSW-NB15

Abstract

This paper presents a comprehensive framework for enhancing the accuracy of intrusion detection systems (IDS) by combining multiple machine learning classifiers with swarm-based optimization algorithms. We use different classification models (Logistic Regression, Decision Tree, Extra Trees, Random Forest, and XGBoost) for evaluating the impact of the proposed approach on two benchmark cybersecurity datasets CICIDS2018 and UNSW-NB15. To address shortcomings pertaining to the detection precision and model stability, three metaheuristic optimization algorithms i.e., Particle Swarm Optimization (PSO), Ant Colony Optimization (ACO) and Bat Algorithm are employed for feature selection and hyperparameter optimization. Empirical results indicate that the proposed CRF-CAPS obtains significant performance improvements in all evaluation criteria and can achieve as large as 7.5% of accuracy improvement over baseline models. The best accuracy of 97.6% for the improved model based on UNSW-NB15 and 90.9% for the CICIDS2018. In addition, the optimization resulted in a decrease in the inference time of many models, which enables real-time operation. These findings demonstrate the efficacy of hybrid optimization to narrow the performance gaps observed in the recent IDSII literature. The proposed model achieves higher overall performance than recent IDS studies between 2020 and 2025, which showed accuracy in the range of 84–91%. In addition, the swarm-based optimization could reduce features by around 30%, which translated into significant improvement in inference speed and model efficiency.