Scholarlink

Unmanned Aerial Vehicles (UAVs) play a crucial role in surveillance, disaster response, and military operations. However, their dependence on wireless communication exposes them to cyber threats such as GPS jamming and spoofing. Traditional intrusion detection systems (IDS) struggle to adapt to evolving attack patterns while ensuring data privacy. Federated Learning (FL) offers a promising approach by enabling distributed model training across UAVs without sharing sensitive data. This study presents a comparative analysis of two FL-based IDS models: a standard Neural Network (NN) and a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model. Using the UAV ATTACK dataset, we evaluate their performance across multiple federated clients. Experimental results reveal that the CNN-LSTM model outperforms the NN model, achieving superior accuracy (99.1% vs. 81.2%), higher recall, and fewer false positives. Additionally, the CNN-LSTM model demonstrates faster convergence, lower validation loss, and enhanced generalization, making it more suitable for real-time UAV security applications.

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

Robust UAV Intrusion Detection via Federated Learning: A Comparison of NN and CNN-LSTM Models

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sayed Muqayyad Hussain,
  • Madiha Haider Syed,
  • Adeel Anjum,
  • Muhammad Javed,
  • Ankit Raj

摘要

Unmanned Aerial Vehicles (UAVs) play a crucial role in surveillance, disaster response, and military operations. However, their dependence on wireless communication exposes them to cyber threats such as GPS jamming and spoofing. Traditional intrusion detection systems (IDS) struggle to adapt to evolving attack patterns while ensuring data privacy. Federated Learning (FL) offers a promising approach by enabling distributed model training across UAVs without sharing sensitive data. This study presents a comparative analysis of two FL-based IDS models: a standard Neural Network (NN) and a hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model. Using the UAV ATTACK dataset, we evaluate their performance across multiple federated clients. Experimental results reveal that the CNN-LSTM model outperforms the NN model, achieving superior accuracy (99.1% vs. 81.2%), higher recall, and fewer false positives. Additionally, the CNN-LSTM model demonstrates faster convergence, lower validation loss, and enhanced generalization, making it more suitable for real-time UAV security applications.