Intrusion detection model of UAV system based on machine learning and neural network

Abstract

Introduction: With the growing complexity and scale of cyber attacks, intrusion detection for unmanned aerial vehicle (UAV) systems has become a critical challenge in modern network security. UAVs have unique constraints including limited battery life, restricted data-transmission distance, and small data-storage capacity, while malicious activities can disrupt their power usage, communication, and data storage—highlighting the need for dedicated intrusion-detection solutions. Traditional traffic detection methods lack efficient modeling of local and global features, making it difficult to capture complex data patterns. Methods: We propose an intrusion detection model integrating machine learning and neural networks. First, UAV data is cleaned, and traditional feature selection techniques (filtering, packaging, embedding) are used to separate key and non-key features. Non-key features are mapped to the key feature subspace via CNN + LSTM for feature fusion, and the fused features serve as model inputs. Machine learning and neural networks are then combined to detect UAV network traffic. Results: Testing on public datasets ISCXVPN2016, CICIDS2018, TON IoT, and CIC IoT 2023 shows that our method improves accuracy by up to 3%, F1 score by up to 4%, and recall by up to 3% compared to the three major feature selection techniques. Discussion: The integration of CNN + LSTM enables effective modeling of local and global features, addressing the limitations of traditional methods. The model’s optimization for feature fusion and UAV-specific constraints ensures it is suitable for resource-constrained UAV systems, providing reliable intrusion detection.