Scholarlink

With the widespread use of encrypted transmission and increasingly sophisticated network attacks, ensuring encrypted traffic security has become a major challenge. This paper proposes a deep contrastive learning-based anomaly detection model to address limitations in multi-scale feature extraction, adversarial robustness, and imbalanced traffic. The model combines a Transformer with channel and sequence attention, and a multi-scale 1D dilated convolution module to capture both local and global traffic features. To defend against adversarial samples, we introduce an optimization framework integrating FreeLB adversarial training and Barlow Twins contrastive learning with Double Positive Loss. Moreover, combining CB Loss and Focal Loss improves performance on imbalanced data. Experiments on the CICIDS-2017 dataset show the model achieves an F1-score of 97.72%, a 1.45% gain over the baseline. With CB and Focal Loss, the F1-score improves by 0.42%, and the false alarm rate is reduced by 10% under strong adversarial conditions, demonstrating the model’s robustness and effectiveness.

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

An Encrypted Anomaly Traffic Detection Method Integrating Adversarial Training and Multi-scale Contrastive Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Gang Wang,
  • Yunpeng Gao,
  • Yi Chen,
  • Peng Wang,
  • Yong Ding

摘要

With the widespread use of encrypted transmission and increasingly sophisticated network attacks, ensuring encrypted traffic security has become a major challenge. This paper proposes a deep contrastive learning-based anomaly detection model to address limitations in multi-scale feature extraction, adversarial robustness, and imbalanced traffic. The model combines a Transformer with channel and sequence attention, and a multi-scale 1D dilated convolution module to capture both local and global traffic features. To defend against adversarial samples, we introduce an optimization framework integrating FreeLB adversarial training and Barlow Twins contrastive learning with Double Positive Loss. Moreover, combining CB Loss and Focal Loss improves performance on imbalanced data. Experiments on the CICIDS-2017 dataset show the model achieves an F1-score of 97.72%, a 1.45% gain over the baseline. With CB and Focal Loss, the F1-score improves by 0.42%, and the false alarm rate is reduced by 10% under strong adversarial conditions, demonstrating the model’s robustness and effectiveness.