Scholarlink

With the increasing deployment of UAV swarms in complex and adversarial environments, accurate detection of network anomalies has become essential to ensure system reliability and security. However, conventional centralized anomaly detection methods face critical challenges regarding privacy, data heterogeneity, and communication overhead. This paper presents a federated learning-based anomaly detection framework using LSTM neural networks for UAV systems, where multiple clients collaboratively train models without sharing raw traffic data. The framework integrates FedAvg and FedProx aggregation strategies with dynamic parameter tuning and a dual-threshold decision mechanism to enhance robustness in non-IID environments. Simulation results and ablation studies demonstrate that the proposed method achieves high accuracy (86.5%), precision (99.2%), recall (86.4%), and F1-score (92.4%), while reducing communication bandwidth by 62%. The system also supports automated batch experiments and real-time visualization, demonstrating its potential for practical deployment in real-world scenarios.

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

A Federated Learning Framework with LSTM-RNN for UAV Network Intrusion Detection

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yu Cai,
  • Zijian Zhu,
  • Yujiang Zhong,
  • Haoyu Jiang,
  • Jingyuan Sun,
  • Haoran Tang

摘要

With the increasing deployment of UAV swarms in complex and adversarial environments, accurate detection of network anomalies has become essential to ensure system reliability and security. However, conventional centralized anomaly detection methods face critical challenges regarding privacy, data heterogeneity, and communication overhead. This paper presents a federated learning-based anomaly detection framework using LSTM neural networks for UAV systems, where multiple clients collaboratively train models without sharing raw traffic data. The framework integrates FedAvg and FedProx aggregation strategies with dynamic parameter tuning and a dual-threshold decision mechanism to enhance robustness in non-IID environments. Simulation results and ablation studies demonstrate that the proposed method achieves high accuracy (86.5%), precision (99.2%), recall (86.4%), and F1-score (92.4%), while reducing communication bandwidth by 62%. The system also supports automated batch experiments and real-time visualization, demonstrating its potential for practical deployment in real-world scenarios.