Dimensionality Reduction for Intrusion Detection Using Crow Search Optimization and Grey Wolf Optimization

Abstract

In high dimensional intrusion detection data, repetitive and superfluous features may cause calculation durations to increase and classification efficiency to decrease. Therefore, one of the most important steps in creating effective detection models is to choose only the most beneficial features. Feature selection improves generalization, reduces noise, and improves classifier performance. In this study, the NSL-KDD and UNSWNB15 datasets are used to examine two nature-inspired metaheuristic feature selection techniques. While the Grey Wolf Optimizer (GWO) imitates the cooperative hunting strategy of grey wolves, the Crow Search Optimization (CSO) algorithm is based on the complex food concealment and following behavior of crows. Both datasets were preprocessed by removing superfluous attributes, converting symbolic fields into numerical values, and applying min-max normalization. Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), k -Nearest Neighbors (k-NN), and Extreme Gradient Boosting (XGBoost) were the five classifiers used to evaluate the compact subsets of features that were found using both approaches. The experimental analysis shows that both strategies maintain high detection accuracy while significantly reducing the number of characteristics. For NSL-KDD, GWO obtained $\mathbf{9 8. 7 1 \%}$ accuracy with Random Forest using 16 selected features, whereas CSO reached 99.33 % with Random Forest using 28 features. For UNSW-NB15, CSO achieved 94.17 % accuracy with Random Forest using 27 features, while GWO obtained 92.96 % with 16 features. This demonstrates that by decreasing dimensionality and boosting computational performance, meta-heuristic feature selection can greatly enhance intrusion detection systems across diverse network traffic datasets.