مسار بولونيا
EN

علي كريم عبد كامل

السنة عنوان البحث نشر البحث
2026 A Crow Search-Optimized K-Nearest Neighbors Framework for Intrusion Detection in Government Security Networks Academic International Journal of Engineering Science
Government security networks face an expanding range of cyber-attacks, and machine-learning intrusion detection systems (IDS) offer an adaptive defence, but their accuracy and cost depend heavily on input feature quality. We propose CSA-KNN, a wrapper feature-selection framework that couples the Crow Search Algorithm (CSA) with a K-Nearest Neighbors (KNN) classifier, using Macro-F1 as the search fitness to remain sensitive to rare attack classes. On the NSL-KDD benchmark, with categorical features one-hot encoded and minority classes (R2L, U2R) rebalanced with SMOTE, CSA selects 21.4 of 41 features on average and, across 5 independent random seeds on the official held-out KDDTest+ partition, achieves 75.07% ± 0.50% accuracy and 55.92% ± 3.01% Macro- F1, versus 75.73% ± 0.00% accuracy and 57.56% ± 0.00% Macro-F1 for a plain, all-feature KNN trained under an identical protocol (Wilcoxon signed-rank p = 0.438 on Macro-F1). These findings demonstrate that CSA-based feature selection can reduce KNN's input dimensionality while maintaining or improving detection quality and minority-class sensitivity on a standard intrusion-detection benchmark; we present this as evidence of potential applicability to government network monitoring, not as a validated real-world deployment, which remains future work.