Scholarlink

<p>Federated learning provides a privacy-preserving training paradigm for distributed intrusion detection in the Internet of Things. However, its performance has been shown to degrade in fine-grained classification tasks that require precise identification of specific attack types, mainly due to non-independent and identically distributed (non-IID) client data and severe class imbalance. To address this issue, FGF-IDS is proposed as a federated framework for fine-grained intrusion detection under extreme imbalance. The framework integrates (1) a loss-based importance sampling strategy that emphasizes hard and minority instances during local training, (2) a composite local objective with feature-alignment regularization to reduce representation drift and encourage cross-client consistency, and (3) an adaptive gradient-balancing aggregation scheme that increases the influence of clients containing rare-class samples. Experiments on the UNSW-NB15 dataset under three Dirichlet-based non-IID partitioning settings demonstrated robust overall performance and substantially improved rare-class detection. The code is available at <a href="https://github.com/ShoeMaker-pixel/FGF-IDS">https://github.com/ShoeMaker-pixel/FGF-IDS</a>.</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

FGF-IDS: a federated framework for fine-grained intrusion detection

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • PeiCheng Yang,
  • Xiaohui Zhang,
  • Fenhua Bai,
  • Kai Zeng,
  • Tao Shen

摘要

Federated learning provides a privacy-preserving training paradigm for distributed intrusion detection in the Internet of Things. However, its performance has been shown to degrade in fine-grained classification tasks that require precise identification of specific attack types, mainly due to non-independent and identically distributed (non-IID) client data and severe class imbalance. To address this issue, FGF-IDS is proposed as a federated framework for fine-grained intrusion detection under extreme imbalance. The framework integrates (1) a loss-based importance sampling strategy that emphasizes hard and minority instances during local training, (2) a composite local objective with feature-alignment regularization to reduce representation drift and encourage cross-client consistency, and (3) an adaptive gradient-balancing aggregation scheme that increases the influence of clients containing rare-class samples. Experiments on the UNSW-NB15 dataset under three Dirichlet-based non-IID partitioning settings demonstrated robust overall performance and substantially improved rare-class detection. The code is available at https://github.com/ShoeMaker-pixel/FGF-IDS.