Hybrid Convolutional–Bidirectional LSTM Model for Intrusion Detection in Unmanned Aerial Vehicle Cyber-Physical Networks

Unmanned Aerial Vehicles (UAVs) are increasingly deployed in applications ranging from surveillance to logistics, but their integration into cyber-physical systems exposes them to diverse cyber and physical intrusions. Traditional intrusion detection systems (IDS) often fail to capture both spatial and temporal dependencies in UAV telemetry, and they struggle with class imbalance and deployment constraints. To address these challenges, we propose a hybrid IDS based on Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM) networks. The CNN layers extract low-level spatial representations from UAV communication data, while BiLSTM layers model long-range temporal dependencies. To mitigate class imbalance, Synthetic Minority Oversampling Technique (SMOTE) and a class-weighted cross-entropy loss were employed. Experimental evaluation on a multi-class UAV intrusion dataset demonstrates that the proposed model achieves a test accuracy of 94.09% and a macro-F1 score of 0.94, outperforming prior IDS architectures on minority attack classes. Furthermore, the model shows potential for edge deployment with low latency and a compact memory footprint, making it suitable for FPGA/SoC implementations. Future work will extend this framework to federated and swarm-based UAV intrusion detection systems, ensuring robust and scalable protection in dynamic UAV networks.