Scholarlink

<p>Intrusion detection in UAV swarms operating in Urban Air Mobility (UAM) environments is challenged by non-IID data distributions, privacy constraints, and vulnerability to adversarial and Byzantine attacks. To address these issues, this paper proposes FL-GIIDS-AR, a robust and privacy-preserving federated intrusion detection framework that integrates personalized federated learning (FedPer), Byzantine-resilient aggregation, and transferability-aware adversarial training. The framework is evaluated using both UAV telemetry data and the public CICIDS2017 benchmark under a unified federated setup. Experimental results show that the proposed method achieves up to 96.3% detection accuracy, an AUC of 0.978, and an F1-score of 93.8%, while maintaining a low false positive rate of 3.7%. Under adversarial settings, the model demonstrates strong robustness, achieving over 90% robust accuracy against FGSM and PGD attacks. Additionally, the proposed approach reduces communication overhead by approximately 30–35% compared to standard federated learning methods, while maintaining stable convergence even with up to 30% Byzantine clients. These results demonstrate that FL-GIIDS-AR effectively balances robustness, efficiency, and adaptability, making it suitable for secure and scalable UAV swarm 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

Federated learning-enabled intrusion detection with adversarial robustness for secure and scalable communications in heterogeneous UAV swarms

  • 期刊论文

  • 发表日期：

    2026年7月2日

查看全文
  • Ruofan Wang,
  • Vladimir Y. Mariano,
  • Fei Gao

摘要

Intrusion detection in UAV swarms operating in Urban Air Mobility (UAM) environments is challenged by non-IID data distributions, privacy constraints, and vulnerability to adversarial and Byzantine attacks. To address these issues, this paper proposes FL-GIIDS-AR, a robust and privacy-preserving federated intrusion detection framework that integrates personalized federated learning (FedPer), Byzantine-resilient aggregation, and transferability-aware adversarial training. The framework is evaluated using both UAV telemetry data and the public CICIDS2017 benchmark under a unified federated setup. Experimental results show that the proposed method achieves up to 96.3% detection accuracy, an AUC of 0.978, and an F1-score of 93.8%, while maintaining a low false positive rate of 3.7%. Under adversarial settings, the model demonstrates strong robustness, achieving over 90% robust accuracy against FGSM and PGD attacks. Additionally, the proposed approach reduces communication overhead by approximately 30–35% compared to standard federated learning methods, while maintaining stable convergence even with up to 30% Byzantine clients. These results demonstrate that FL-GIIDS-AR effectively balances robustness, efficiency, and adaptability, making it suitable for secure and scalable UAV swarm environments.