Scholarlink

<p>Intrusion Detection Systems (IDS) are essential for defending against modern cyber threats, yet traditional deep learning-based IDS often struggle with transparency, computational demands, and data privacy concerns. To address these limitations, we propose a novel IDS framework that integrates Federated Learning (FL), Explainable Artificial Intelligence (XAI), and Hyperband-based hyperparameter optimization within a Transformer architecture. FL enables privacy-preserving, decentralized model training. SHAP and LIME provide interpretable insights into predictions and Hyperband enhances efficiency through automated feature selection. Evaluated on CICIDS2017 and CICIDS2018 datasets, the proposed federated Transformer model achieves up to 99.03% and 98.75% accuracy respectively, with corresponding F1-scores of 0.99 and 0.98. Notably, training time is reduced by over 80% compared to centralized models, without sacrificing performance or interpretability. These results demonstrate the framework’s scalability, robustness, and transparency, making it highly suitable for real-world, privacy-sensitive intrusion detection scenarios.</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 learning and explainable AI-driven intrusion detection with hyperband optimization

  • 期刊论文

  • 发表日期：

    2025年8月21日

查看全文
  • Harshitha C,
  • Sendil Vadivu D,
  • Narendran Rajagopalan

摘要

Intrusion Detection Systems (IDS) are essential for defending against modern cyber threats, yet traditional deep learning-based IDS often struggle with transparency, computational demands, and data privacy concerns. To address these limitations, we propose a novel IDS framework that integrates Federated Learning (FL), Explainable Artificial Intelligence (XAI), and Hyperband-based hyperparameter optimization within a Transformer architecture. FL enables privacy-preserving, decentralized model training. SHAP and LIME provide interpretable insights into predictions and Hyperband enhances efficiency through automated feature selection. Evaluated on CICIDS2017 and CICIDS2018 datasets, the proposed federated Transformer model achieves up to 99.03% and 98.75% accuracy respectively, with corresponding F1-scores of 0.99 and 0.98. Notably, training time is reduced by over 80% compared to centralized models, without sacrificing performance or interpretability. These results demonstrate the framework’s scalability, robustness, and transparency, making it highly suitable for real-world, privacy-sensitive intrusion detection scenarios.