Scholarlink

<p>Intrusion detection in Internet-of-Things (IoT) networks require models that recognize multiple attack types while remaining efficient and robust to noisy telemetry. This study presents an integrated three-phase framework for multiclass intrusion detection evaluated on the recent ToN_IoT benchmarks. Phase 1 establishes strong baselines by benchmarking 15 machine-learning classifiers across seven ToN_IoT subsets. Phase 2 improves accuracy and efficiency via hybrid feature selection, combining ANOVA (filter) with Boruta (wrapper) to remove redundant and weak signals. Phase 3 targets the more challenging Windows 10 (Win10) subset using Bayesian-optimized Weighted Soft Voting that jointly tunes ensemble weights and the feature subset size. The framework attains up to 99.59% accuracy (0.996 macro-F1) on the Train_Test_Network subset and 98.31% accuracy (0.982 macro-F1) on Win10 after optimization. Practical considerations—class weighting for imbalance, mean/mode imputation for missing values, and standardized preprocessing—contribute to stability and generalization. Results indicate that combining systematic baselining, complementary feature selection, and Bayesian ensemble tuning yields a scalable, reproducible IDS pipeline suitable for real-world IoT/IIoT deployments.</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

An integrated three-phase framework with feature selection and Bayesian optimization for multiclass intrusion detection on ToN_IoT datasets

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Waleed M. Bahgat,
  • Tamer Ahmed Farrag,
  • Amr E. Eldin Rashed

摘要

Intrusion detection in Internet-of-Things (IoT) networks require models that recognize multiple attack types while remaining efficient and robust to noisy telemetry. This study presents an integrated three-phase framework for multiclass intrusion detection evaluated on the recent ToN_IoT benchmarks. Phase 1 establishes strong baselines by benchmarking 15 machine-learning classifiers across seven ToN_IoT subsets. Phase 2 improves accuracy and efficiency via hybrid feature selection, combining ANOVA (filter) with Boruta (wrapper) to remove redundant and weak signals. Phase 3 targets the more challenging Windows 10 (Win10) subset using Bayesian-optimized Weighted Soft Voting that jointly tunes ensemble weights and the feature subset size. The framework attains up to 99.59% accuracy (0.996 macro-F1) on the Train_Test_Network subset and 98.31% accuracy (0.982 macro-F1) on Win10 after optimization. Practical considerations—class weighting for imbalance, mean/mode imputation for missing values, and standardized preprocessing—contribute to stability and generalization. Results indicate that combining systematic baselining, complementary feature selection, and Bayesian ensemble tuning yields a scalable, reproducible IDS pipeline suitable for real-world IoT/IIoT deployments.