Scholarlink

<p>The rapid growth of the IoT has transformed connectivity between physical devices and the digital world, but it has also increased the vulnerability of IoT networks to DDoS attacks. Machine learning algorithms effectively detect DDoS attacks by learning from evolving patterns, but traditional ML approaches often fall short in IoT ecosystems, mainly due to privacy and security concerns. Traditional distributed ML algorithms require transmitting large amounts of data to a central system for processing. This process is inefficient and introduces delays that make real-time threat detection harder. This is a big problem when mitigating DDoS threats. Federated learning offers a promising solution by enabling IoT devices to train a global model collaboratively without sharing raw data. FL encounters challenges, especially in addressing the computational complexity associated with resource-constrained IoT devices. To tackle these challenges, this study improves FL local models by using deep autoencoders to reduce the computational load and boost performance. These autoencoders help with reducing data dimensionality, which is essential since IoT devices have limited processing power and memory. Additionally, IoT data is often non-IID, which adds complexity. We address this using a combination of autoencoders, retraining techniques, and partial selection methods, which together improve the accuracy, robustness, and performance of the FL model in detecting IoT-based DDoS attacks. The model's effectiveness is evaluated using FedAvg and FedAvgM aggregation algorithms, with metrics including true positive rate, false positive rate, F1-score, and AUC. Leveraging the non-IID N-BaIoT dataset, our evaluation shows that FedAvg achieves an average F1-score of 93.07% with an AUC of 93.34%. At the same time, FedAvgM slightly improves performance, with an F1-score of 93.10% and an AUC of 93.95%, and reduces execution time from 58.43 to 49.01&#xa0;s.</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

Mitigating DDoS attacks on the internet of things using federated learning

  • 期刊论文

  • 发表日期：

    2025年6月4日

查看全文
  • Ghazaleh Shirvani,
  • Saeid Ghasemshirazi,
  • Mohammad Abdollahi Azgomi

摘要

The rapid growth of the IoT has transformed connectivity between physical devices and the digital world, but it has also increased the vulnerability of IoT networks to DDoS attacks. Machine learning algorithms effectively detect DDoS attacks by learning from evolving patterns, but traditional ML approaches often fall short in IoT ecosystems, mainly due to privacy and security concerns. Traditional distributed ML algorithms require transmitting large amounts of data to a central system for processing. This process is inefficient and introduces delays that make real-time threat detection harder. This is a big problem when mitigating DDoS threats. Federated learning offers a promising solution by enabling IoT devices to train a global model collaboratively without sharing raw data. FL encounters challenges, especially in addressing the computational complexity associated with resource-constrained IoT devices. To tackle these challenges, this study improves FL local models by using deep autoencoders to reduce the computational load and boost performance. These autoencoders help with reducing data dimensionality, which is essential since IoT devices have limited processing power and memory. Additionally, IoT data is often non-IID, which adds complexity. We address this using a combination of autoencoders, retraining techniques, and partial selection methods, which together improve the accuracy, robustness, and performance of the FL model in detecting IoT-based DDoS attacks. The model's effectiveness is evaluated using FedAvg and FedAvgM aggregation algorithms, with metrics including true positive rate, false positive rate, F1-score, and AUC. Leveraging the non-IID N-BaIoT dataset, our evaluation shows that FedAvg achieves an average F1-score of 93.07% with an AUC of 93.34%. At the same time, FedAvgM slightly improves performance, with an F1-score of 93.10% and an AUC of 93.95%, and reduces execution time from 58.43 to 49.01 s.