Adaptive Threat Recognition in UAV Swarm Communication Using Collaborative Pattern Learning

Abstract

UAV swarm networks depend on reliable communication among multiple aerial nodes, ground stations, and mission-control systems. However, the high mobility and dynamic topology of UAV swarms make them vulnerable to abnormal traffic caused by jamming, spoofing, flooding, route manipulation, and compromised nodes. This study proposes a SHAP-based ensemble anomaly detection framework for secure communication in UAV swarm networks. The framework combines CatBoost, Random Forest, and AdaBoost classifiers through an adaptive voting strategy to improve robustness under unstable wireless communication conditions. SHAP values are used to identify the most important traffic and topology features behind each anomaly classification. A UAV swarm communication dataset is generated using 60 UAV nodes and 6 ground-control nodes under simulated surveillance and delivery missions, producing 1.58 million labeled traffic records. The dataset includes normal communication, flooding attacks, GPS spoofing-related traffic changes, route disruption, and malicious command injection. After feature extraction, 43 features are retained, including packet loss ratio, route change frequency, control message interval, flow duration, retransmission count, signal-related delay variation, and source-destination traffic imbalance. The proposed model achieves 97.47% accuracy, 96.63% macro-F1, and 98.09% AUC across five traffic classes. Compared with standalone CatBoost and SVM models, the proposed ensemble improves macro-F1 by 1.94% and 5.21%, respectively. SHAP interpretation shows that route change frequency, abnormal control message intervals, retransmission spikes, and sudden traffic concentration are the most important indicators of compromised UAV communication. The results demonstrate that explainable ensemble learning can provide transparent and effective anomaly detection for UAV swarm network security.