Scholarlink

<p>The strategy of IoT security is based on a cybersecurity strategy in protecting IoT devices and the vulnerable networks they connect to from cyber-attacks. Traditional intrusion detection systems (IDS) cannot understand the complexity and volume of IoT network traffic and need advanced solutions for such a system. The proposed paper enforces a multi-phase model for IDS in IoT systems with considerations of the collection of datasets, preprocessing of data, feature extraction, hybrid optimization, ensemble Deep Learning (DL) models, and evaluation. The preprocessing phase involves all those essential data-cleaning techniques to handle missing values, duplicate records, encoding categorical features, and standardizing the numerical data. Feature extraction extracts the statistical and frequency-based features, such as flow-based metrics and N-gram analysis, that can highlight patterns in abnormal traffic. Correlation analysis, removing redundancy and improving informative power, while DL methods such as CNN are used for spatial feature extraction. A new hybrid approach is proposed in this paper, Hybrid Waterwheel Plant Algorithm and a Mother Optimization Algorithm (HWPAMO), for selecting the best features and improving performance. Finally, design an ensemble model involving various DL architectures: InceptionV3, VGG16, and Long Short-Term Memory (LSTM) networks with autoencoders with attention mechanisms and residual blocks. The developed technique is validated with other prevailing techniques in terms of kappa score, accuracy, MCC, recall, and precision. The proposed model demonstrates superior performance across all evaluated metrics for both 70% and 80% training data splits. With 70% training data, it achieves an accuracy of 99.15%, a precision of 98.06%, a specificity of 99.41%, and a Kappa score of 96.23%. When trained on 80% of the dataset, it further improves its performance, achieving an accuracy of 99.33%, a precision of 98.45%, a specificity of 99.73%, and a Kappa score of 97.07%. These results confirm the robustness, precision, and generalization capability of the proposed model in IDS. </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

Hybrid Optimization and Ensemble Deep Learning Framework for Efficient Anomaly-Based Intrusion Detection in IoT Networks

  • 期刊论文

  • 发表日期：

    2025年7月12日

查看全文
  • Arifa Javid Shikalgar,
  • Shikalgar Shabanam Khalid ,
  • Sunita Sunil Shinde

摘要

The strategy of IoT security is based on a cybersecurity strategy in protecting IoT devices and the vulnerable networks they connect to from cyber-attacks. Traditional intrusion detection systems (IDS) cannot understand the complexity and volume of IoT network traffic and need advanced solutions for such a system. The proposed paper enforces a multi-phase model for IDS in IoT systems with considerations of the collection of datasets, preprocessing of data, feature extraction, hybrid optimization, ensemble Deep Learning (DL) models, and evaluation. The preprocessing phase involves all those essential data-cleaning techniques to handle missing values, duplicate records, encoding categorical features, and standardizing the numerical data. Feature extraction extracts the statistical and frequency-based features, such as flow-based metrics and N-gram analysis, that can highlight patterns in abnormal traffic. Correlation analysis, removing redundancy and improving informative power, while DL methods such as CNN are used for spatial feature extraction. A new hybrid approach is proposed in this paper, Hybrid Waterwheel Plant Algorithm and a Mother Optimization Algorithm (HWPAMO), for selecting the best features and improving performance. Finally, design an ensemble model involving various DL architectures: InceptionV3, VGG16, and Long Short-Term Memory (LSTM) networks with autoencoders with attention mechanisms and residual blocks. The developed technique is validated with other prevailing techniques in terms of kappa score, accuracy, MCC, recall, and precision. The proposed model demonstrates superior performance across all evaluated metrics for both 70% and 80% training data splits. With 70% training data, it achieves an accuracy of 99.15%, a precision of 98.06%, a specificity of 99.41%, and a Kappa score of 96.23%. When trained on 80% of the dataset, it further improves its performance, achieving an accuracy of 99.33%, a precision of 98.45%, a specificity of 99.73%, and a Kappa score of 97.07%. These results confirm the robustness, precision, and generalization capability of the proposed model in IDS.