Scholarlink

<p>Identifying zero-day attacks-new threats that have not been encountered during training continues to pose a significant challenge for machine learning-driven network threat detection systems. This paper introduces a deep learning framework that integrates Convolutional Neural Networks (CNNs) and Bidirectional Encoder Representations from Transformers (BERT) to recognize previously undetected network intrusions. The suggested method utilizes CNNs to capture local feature patterns within network traffic and BERT to comprehend the contextual and semantic relationships among traffic characteristics. These models collectively create a two-phase pipeline: initially, transforming raw traffic into significant representations, and subsequently deducing malicious behavior by pinpointing anomalies from established patterns. To assess effectiveness, we propose a novel metric, Zero-day Detection (ZDR), which evaluates the model’s capability to accurately identify unknown attack classes. Recent studies conducted on network traffic datasets show that our technique surpasses conventional approaches in detecting zero-day attacks, especially in confined data environments like IoT networks. This study demonstrates the potential of merging hybrid methods of deep learning to create resilient and adaptive threat detection systems.</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

Zero-shot learning for IoT attack detection using CNN and BERT

  • 期刊论文

  • 发表日期：

    2025年11月21日

查看全文
  • T. N. V. S. Praveen,
  • Rama Devi Odugu,
  • Sumalatha Lingamgunta

摘要

Identifying zero-day attacks-new threats that have not been encountered during training continues to pose a significant challenge for machine learning-driven network threat detection systems. This paper introduces a deep learning framework that integrates Convolutional Neural Networks (CNNs) and Bidirectional Encoder Representations from Transformers (BERT) to recognize previously undetected network intrusions. The suggested method utilizes CNNs to capture local feature patterns within network traffic and BERT to comprehend the contextual and semantic relationships among traffic characteristics. These models collectively create a two-phase pipeline: initially, transforming raw traffic into significant representations, and subsequently deducing malicious behavior by pinpointing anomalies from established patterns. To assess effectiveness, we propose a novel metric, Zero-day Detection (ZDR), which evaluates the model’s capability to accurately identify unknown attack classes. Recent studies conducted on network traffic datasets show that our technique surpasses conventional approaches in detecting zero-day attacks, especially in confined data environments like IoT networks. This study demonstrates the potential of merging hybrid methods of deep learning to create resilient and adaptive threat detection systems.