Scholarlink

<p>This paper proposes a hybrid deep meta-learning framework to detect denial-of-service (DoS) attacks in wireless sensor networks (WSNs), tackling the challenges of limited labeled data and the need for real-time threat detection. The approach combines Siamese networks for binary classification and Prototypical networks for multi-class classification, utilizing few-shot learning to achieve high performance with minimal training data. Evaluations on the WSN-DS dataset yield impressive results, the Siamese network achieves 99.86% accuracy and a 98.91% F1-score, while the Prototypical network records 97.99% accuracy and a 98.16% F1-score, both using only 3,000 samples in a 2-shot setting. On the DV1-WSN dataset, the framework delivers 98.33% accuracy and a 98.01% F1-score for binary classification, and 97.95% accuracy and a 96.83% F1-score for multi-class detection. These results outperform traditional detection methods and recent intrusion detection systems, highlighting the framework’s efficiency and adaptability. Statistical analysis validates the consistency of both models across datasets. This research advances WSN security by offering a highly accurate, resource-efficient solution for real-time DoS attack mitigation in data-scarce environments.</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

Adaptive meta-learning for denial-of-service defense

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Salmi Salim,
  • Oughdir Lahcen

摘要

This paper proposes a hybrid deep meta-learning framework to detect denial-of-service (DoS) attacks in wireless sensor networks (WSNs), tackling the challenges of limited labeled data and the need for real-time threat detection. The approach combines Siamese networks for binary classification and Prototypical networks for multi-class classification, utilizing few-shot learning to achieve high performance with minimal training data. Evaluations on the WSN-DS dataset yield impressive results, the Siamese network achieves 99.86% accuracy and a 98.91% F1-score, while the Prototypical network records 97.99% accuracy and a 98.16% F1-score, both using only 3,000 samples in a 2-shot setting. On the DV1-WSN dataset, the framework delivers 98.33% accuracy and a 98.01% F1-score for binary classification, and 97.95% accuracy and a 96.83% F1-score for multi-class detection. These results outperform traditional detection methods and recent intrusion detection systems, highlighting the framework’s efficiency and adaptability. Statistical analysis validates the consistency of both models across datasets. This research advances WSN security by offering a highly accurate, resource-efficient solution for real-time DoS attack mitigation in data-scarce environments.