Scholarlink

<p>In IoT intrusion detection, neural classifiers often produce overconfident predictions on novel or rare threats, undermining reliability and operational trust. This challenge highlights the urgent need for models that not only detect intrusions accurately but also express calibrated uncertainty to flag ambiguous or unfamiliar patterns. We propose GIBBON, a Graph-Based Intrusion Bayesian BNN with Gibbs, that integrates Graph Neural Networks (GraphSAGE) with Bayesian inference via Monte Carlo (MC) dropout and a blocked Gibbs sampling refinement for principled uncertainty estimation. GIBBON constructs relational graphs from raw network flow data to capture structural dependencies between hosts and connections. It applies MC dropout for variational posterior approximation and further refines the posterior using a Metropolis-within-Gibbs sampling scheme that iteratively samples classifier layer weights to improve uncertainty calibration. Evaluated on the complex NF-ToN-IoT-v3 dataset, our model achieves high predictive performance (accuracy = 0.8896, F1-score = 0.8845) while significantly outperforming prior IDS models in uncertainty calibration (ECE = 0.0055, MCE = 0.0217) and out-of-distribution detection (AUROC = 0.8832, AUPR = 0.4056, FPR@95%TPR = 0.2933). These results suggest that GIBBON’s uncertainty-aware predictions can reliably identify ambiguous or novel attacks and reduce overconfident misclassifications.</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

A graph-based Bayesian intrusion detection framework with Gibbs sampling for uncertainty-aware IoT security

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Franco Osei-Wusu,
  • Obed Appiah,
  • Patrick K. Mensah,
  • Peter Nimbe,
  • Elvis K. Donkoh

摘要

In IoT intrusion detection, neural classifiers often produce overconfident predictions on novel or rare threats, undermining reliability and operational trust. This challenge highlights the urgent need for models that not only detect intrusions accurately but also express calibrated uncertainty to flag ambiguous or unfamiliar patterns. We propose GIBBON, a Graph-Based Intrusion Bayesian BNN with Gibbs, that integrates Graph Neural Networks (GraphSAGE) with Bayesian inference via Monte Carlo (MC) dropout and a blocked Gibbs sampling refinement for principled uncertainty estimation. GIBBON constructs relational graphs from raw network flow data to capture structural dependencies between hosts and connections. It applies MC dropout for variational posterior approximation and further refines the posterior using a Metropolis-within-Gibbs sampling scheme that iteratively samples classifier layer weights to improve uncertainty calibration. Evaluated on the complex NF-ToN-IoT-v3 dataset, our model achieves high predictive performance (accuracy = 0.8896, F1-score = 0.8845) while significantly outperforming prior IDS models in uncertainty calibration (ECE = 0.0055, MCE = 0.0217) and out-of-distribution detection (AUROC = 0.8832, AUPR = 0.4056, FPR@95%TPR = 0.2933). These results suggest that GIBBON’s uncertainty-aware predictions can reliably identify ambiguous or novel attacks and reduce overconfident misclassifications.