ASSESSING LEARNING-BASED MODELS FOR INTRUSION DETECTION IN UAV NETWORKS
Abstract
Unmanned Aerial Vehicle (UAV) networks are being used in various military, civilian and infrastructure-monitoring fields, making them vulnerable to an increasing number of cyber threats. To achieve strong security in these resource-limited, dynamic environments, lightweight and accurate intrusion detection systems (IDS) are necessary. In this paper, three Machine Learning (ML) classifiers: Logistic Regression, Random Forest, and Light Gradient Boosting Machine (LightGBM) and three Deep Learning (DL) architectures: Feed Forward Neural Network (FFNN), one-dimensional Convolutional Neural Network (1D-CNN), and Long Short-Term Memory (LSTM), are compared to find the best performance for intrusion detection in UAV networks. An 80:20 train-test split was used for training and evaluation, with the same 80:20 ratio applied to both of the two recently released, publicly available benchmark datasets, UAVIDS-2025 and UAV-GCS-IDS, and with weighted-average Accuracy, Precision, Recall, and F1-score as the evaluation metrics. The performance of the models shows that LightGBM has the highest overall performance with 100.0% accuracy and precision on UAVIDS-2025 and 99.04% and 98.93% accuracy and F1 score on UAV-GCS-IDS, outperforming Random Forest with 99.97% and 88.89% accuracy on UAVIDS-2025 and UAV-GCS-IDS, respectively, and the deep learning models with accuracy ranging from 88.2% to 99.6% on UAVIDS-2025 and 89.6% to 98.93% on UAV-GCS-IDS. The Logistic Regression always yielded the lowest scores with only 86.8%-87.0% accuracy in UAVIDS-2025 and around 85.1% in UAV-GCS-IDS, showing the capabilities of a linear decision boundary in these traffic patterns. The results show that when computational resources are available, deep learning architectures are also effective at learning complex, non-linear attack patterns, while gradient-boosted ensemble methods provide the best detection accuracy with the lowest computational burden for UAV intrusion detection. Finally, the paper outlines recommended deployment strategies for ensembles and hybrid ML-DL, which are applicable in the context of resource-constrained environments of UAV networks.