Scholarlink

<p>The widespread deployment of IoT and heterogeneous networks poses significant challenges to traditional Network Intrusion Detection Systems (NIDS), especially in terms of cross-domain generalization and computational efficiency. To address these issues, we propose QuCAD-IDS, a lightweight cross-domain intrusion detection framework that integrates hierarchical contrastive autoencoding and queue-based adaptive distillation. First, a teacher model (HiMSR-CAE) is trained on a source domain using both reconstruction and attack-aware contrastive losses to learn transferable feature representations. Then, a queue-based contrastive distillation mechanism (QuCAD) transfers the structural knowledge of the teacher’s embedding space to a compact GhostNet student network. Finally, Maximum Mean Discrepancy (MMD) alignment and target-domain fine-tuning adapt the model to heterogeneous target domains. Extensive experiments on the UNSW-NB15, NSL-KDD, and CIC-IDS2017 datasets show that QuCAD-IDS achieves detection accuracy (e.g., 99.70% F1 on NSL-KDD multi-class) and cross-domain adaptability, while reducing model parameters by &#xa0;98% and computational cost by two orders of magnitude. This work provides a practical, lightweight solution for deploying effective NIDS in edge and IoT 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

QuCAD–IDS: cross–domain network intrusion detection via hierarchical contrastive autoencoding and queue-based adaptive distillation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月25日

查看全文
  • Mingqi Wang,
  • Yu Yang,
  • Jinliang Yuan

摘要

The widespread deployment of IoT and heterogeneous networks poses significant challenges to traditional Network Intrusion Detection Systems (NIDS), especially in terms of cross-domain generalization and computational efficiency. To address these issues, we propose QuCAD-IDS, a lightweight cross-domain intrusion detection framework that integrates hierarchical contrastive autoencoding and queue-based adaptive distillation. First, a teacher model (HiMSR-CAE) is trained on a source domain using both reconstruction and attack-aware contrastive losses to learn transferable feature representations. Then, a queue-based contrastive distillation mechanism (QuCAD) transfers the structural knowledge of the teacher’s embedding space to a compact GhostNet student network. Finally, Maximum Mean Discrepancy (MMD) alignment and target-domain fine-tuning adapt the model to heterogeneous target domains. Extensive experiments on the UNSW-NB15, NSL-KDD, and CIC-IDS2017 datasets show that QuCAD-IDS achieves detection accuracy (e.g., 99.70% F1 on NSL-KDD multi-class) and cross-domain adaptability, while reducing model parameters by  98% and computational cost by two orders of magnitude. This work provides a practical, lightweight solution for deploying effective NIDS in edge and IoT environments.