Scholarlink

<p>The rapid expansion of IoT ecosystems introduces significant security challenges due to high-dimensional traffic data, severe class imbalance, and heterogeneous resource constraints across IoT devices, edge gateways, and centralized servers. To address these challenges, this study proposes two correlation-based feature selection models designed to improve the stability, interpretability, and performance balance of anomaly-based Intrusion Detection Systems (IDS) in IoT environments. Model A preserves stable correlated feature clusters derived from multicollinearity patterns, while Model B prioritizes non-redundant features with strong discriminative power that capture behavioral characteristics of network traffic. Both models automate dimensionality reduction through a correlation-guided thresholding mechanism, enabling efficient IDS deployment under diverse computational conditions. Experimental evaluation on the Bot-IoT dataset using Decision Tree, k-Nearest Neighbors, and Random Forest classifiers under multiple feature configurations (10, 9, 6, and 3 features) demonstrates that the proposed models consistently outperform benchmark feature subsets derived from Principal Component Analysis, Genetic Algorithms, and K-best selection, particularly in maintaining balanced detection for normal IoT traffic. Model A achieves its best performance with six features, reaching a macro F1-score of 0.9975 and an MCC of 0.9951. Meanwhile, Model B maintains strong detection performance with only three features, achieving a macro F1-score of 0.9925 and an MCC of 0.9851. Model A is most effective in intermediate feature spaces by preserving stable correlated feature clusters, whereas Model B achieves near-perfect detection with only three features by prioritizing highly discriminative behavioral traffic attributes, making the proposed framework suitable for efficient and reliable intrusion detection in resource-constrained IoT environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Improving balance in intrusion detection for IoT networks using multicollinearity and discriminative power-based feature selection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月15日

查看全文
  • Prajanto Wahyu Adi,
  • Aris Sugiharto,
  • Satriyo Adhy,
  • Etna Vianita,
  • Gohar Rahman

摘要

The rapid expansion of IoT ecosystems introduces significant security challenges due to high-dimensional traffic data, severe class imbalance, and heterogeneous resource constraints across IoT devices, edge gateways, and centralized servers. To address these challenges, this study proposes two correlation-based feature selection models designed to improve the stability, interpretability, and performance balance of anomaly-based Intrusion Detection Systems (IDS) in IoT environments. Model A preserves stable correlated feature clusters derived from multicollinearity patterns, while Model B prioritizes non-redundant features with strong discriminative power that capture behavioral characteristics of network traffic. Both models automate dimensionality reduction through a correlation-guided thresholding mechanism, enabling efficient IDS deployment under diverse computational conditions. Experimental evaluation on the Bot-IoT dataset using Decision Tree, k-Nearest Neighbors, and Random Forest classifiers under multiple feature configurations (10, 9, 6, and 3 features) demonstrates that the proposed models consistently outperform benchmark feature subsets derived from Principal Component Analysis, Genetic Algorithms, and K-best selection, particularly in maintaining balanced detection for normal IoT traffic. Model A achieves its best performance with six features, reaching a macro F1-score of 0.9975 and an MCC of 0.9951. Meanwhile, Model B maintains strong detection performance with only three features, achieving a macro F1-score of 0.9925 and an MCC of 0.9851. Model A is most effective in intermediate feature spaces by preserving stable correlated feature clusters, whereas Model B achieves near-perfect detection with only three features by prioritizing highly discriminative behavioral traffic attributes, making the proposed framework suitable for efficient and reliable intrusion detection in resource-constrained IoT environments.