Scholarlink

Network Intrusion Detection Systems (NIDS) are essential for safeguarding Industrial Internet of Things (IIoT) environments, yet they grapple with significant challenges, including the sheer volume of data, extreme class imbalance, and limited computational resources. This study tackles these hurdles using the large-scale CICADA-IIoT2024 dataset (21.6M records) within the confines of a Kaggle notebook environment. We demonstrate how Explainable AI (XAI) techniques, specifically SHAP and LIME, can guide the iterative refinement of an LSTM model to dramatically enhance precision in the face of extreme class imbalance (a mere 0.01% anomaly rate). Our key contributions include a memory-optimized data processing pipeline that enables handling large-scale IIoT datasets within resource-constrained environments; XAI-driven feature engineering, leading to the identification and removal of 5 features with low impact or detrimental effects on model performance; and a simplified LSTM architecture, which, through XAI guidance and threshold tuning, achieves over 50% precision at 52.5% recall (at a threshold of 0.9993), resulting in a 97% reduction in false positives compared to the baseline model.This research bridges the gap between theoretical XAI applications and practical model refinement strategies for resource-constrained cybersecurity applications in IIoT.

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 IIoT Anomaly Detection Precision via XAI-Guided Feature Engineering and LSTM Tuning on Imbalanced Data Under Resource Constraints

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wassim Ahmad

摘要

Network Intrusion Detection Systems (NIDS) are essential for safeguarding Industrial Internet of Things (IIoT) environments, yet they grapple with significant challenges, including the sheer volume of data, extreme class imbalance, and limited computational resources. This study tackles these hurdles using the large-scale CICADA-IIoT2024 dataset (21.6M records) within the confines of a Kaggle notebook environment. We demonstrate how Explainable AI (XAI) techniques, specifically SHAP and LIME, can guide the iterative refinement of an LSTM model to dramatically enhance precision in the face of extreme class imbalance (a mere 0.01% anomaly rate). Our key contributions include a memory-optimized data processing pipeline that enables handling large-scale IIoT datasets within resource-constrained environments; XAI-driven feature engineering, leading to the identification and removal of 5 features with low impact or detrimental effects on model performance; and a simplified LSTM architecture, which, through XAI guidance and threshold tuning, achieves over 50% precision at 52.5% recall (at a threshold of 0.9993), resulting in a 97% reduction in false positives compared to the baseline model.This research bridges the gap between theoretical XAI applications and practical model refinement strategies for resource-constrained cybersecurity applications in IIoT.