Scholarlink

Rapid expansion of IoT devices has presented significant vulnerabilities, making these systems prime targets for cyberattacks. Detecting and classifying IoT network intrusions is critical to ensuring these systems’ security, integrity, and availability. However, the inherent imbalance in IoT datasets, where attack instances are underrepresented, poses a challenge for IDS based on machine learning. In this paper, we propose a lightweight ensemble deep learning model that combines Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Deep Neural Networks (DNN) to address these challenges. Our methodology integrates XGBoost-based feature selection and Synthetic Minority Over-sampling Technique (SMOTE) to tackle class imbalance. Ensemble methods, including hard, soft, and weighted voting, are employed to enhance detection accuracy. Extensive evaluations on the RT_IoT2022 dataset demonstrate the model’s efficacy, achieving 99.4%, 99.4%, and 99.3% accuracy, respectively, with superior precision, recall, and F1-scores across various assault types. The proposed approach highlights its potential for real-time IoT security, especially in circumstances with limited resources.

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 Lightweight Ensemble Approach Using Deep Learning for IoT Attack Detection on Imbalanced Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Mohammad Zahid,
  • Taran Singh Bharati

摘要

Rapid expansion of IoT devices has presented significant vulnerabilities, making these systems prime targets for cyberattacks. Detecting and classifying IoT network intrusions is critical to ensuring these systems’ security, integrity, and availability. However, the inherent imbalance in IoT datasets, where attack instances are underrepresented, poses a challenge for IDS based on machine learning. In this paper, we propose a lightweight ensemble deep learning model that combines Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Deep Neural Networks (DNN) to address these challenges. Our methodology integrates XGBoost-based feature selection and Synthetic Minority Over-sampling Technique (SMOTE) to tackle class imbalance. Ensemble methods, including hard, soft, and weighted voting, are employed to enhance detection accuracy. Extensive evaluations on the RT_IoT2022 dataset demonstrate the model’s efficacy, achieving 99.4%, 99.4%, and 99.3% accuracy, respectively, with superior precision, recall, and F1-scores across various assault types. The proposed approach highlights its potential for real-time IoT security, especially in circumstances with limited resources.