Scholarlink

<p>The widespread deployment of Internet of Things (IoT) devices has significantly increased the attack surface for cyber threats, making Intrusion Detection Systems (IDS) a critical component of modern network security. Traditional IDS approaches based on centralized machine learning face notable challenges, including data privacy concerns, high communication overhead, and degraded performance under data imbalance. These challenges are especially critical in real-world distributed environments, where rare attack types are sparsely represented across edge devices. This paper proposes FedWID (Federated Wasserstein-based Intrusion Detection), a hybrid IDS framework that integrates <i>Federated Learning</i> (FL) with <i>Wasserstein Generative Adversarial Networks</i> (WGANs) to address these issues in distributed IoT environments. In the FedWID architecture, each client independently generates synthetic intrusion samples for minority classes using WGANs, augments its local dataset, and trains a deep neural network classifier. Model updates are aggregated via the FedAvg algorithm, ensuring that raw data remains localized and confidential throughout the training process. Additionally, ANOVA-based feature selection is employed to reduce input dimensionality, thereby improving computational efficiency without compromising detection performance. The proposed framework is evaluated on two benchmark IDS datasets, NSL-KDD and CIC-IDS, under both binary and multi-class classification tasks. Extensive experiments show that FedWID consistently outperforms existing FL-based IDS approaches across accuracy, precision, recall, and F1-score. These findings validate the effectiveness of combining federated optimization with generative data augmentation, positioning FedWID as a scalable and high-performing solution for intrusion detection in next-generation IoT networks.</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

Federated Intrusion Detection in IoT with Local Data Augmentation Using Wasserstein GANs

  • 期刊论文

  • 发表日期：

    2025年12月17日

查看全文
  • Huu-Hoa Nguyen,
  • Minh-Tuan Thai,
  • The-Phi Pham

摘要

The widespread deployment of Internet of Things (IoT) devices has significantly increased the attack surface for cyber threats, making Intrusion Detection Systems (IDS) a critical component of modern network security. Traditional IDS approaches based on centralized machine learning face notable challenges, including data privacy concerns, high communication overhead, and degraded performance under data imbalance. These challenges are especially critical in real-world distributed environments, where rare attack types are sparsely represented across edge devices. This paper proposes FedWID (Federated Wasserstein-based Intrusion Detection), a hybrid IDS framework that integrates Federated Learning (FL) with Wasserstein Generative Adversarial Networks (WGANs) to address these issues in distributed IoT environments. In the FedWID architecture, each client independently generates synthetic intrusion samples for minority classes using WGANs, augments its local dataset, and trains a deep neural network classifier. Model updates are aggregated via the FedAvg algorithm, ensuring that raw data remains localized and confidential throughout the training process. Additionally, ANOVA-based feature selection is employed to reduce input dimensionality, thereby improving computational efficiency without compromising detection performance. The proposed framework is evaluated on two benchmark IDS datasets, NSL-KDD and CIC-IDS, under both binary and multi-class classification tasks. Extensive experiments show that FedWID consistently outperforms existing FL-based IDS approaches across accuracy, precision, recall, and F1-score. These findings validate the effectiveness of combining federated optimization with generative data augmentation, positioning FedWID as a scalable and high-performing solution for intrusion detection in next-generation IoT networks.