Systematic review of trends in deep learning for UAV cybersecurity

Area of Science:

• Cybersecurity
• Artificial Intelligence
• Aerospace Engineering

Background:

• Unmanned Aerial Vehicles (UAVs) face complex operational environments prone to cyber threats.
• The attack surface of UAVs includes flight control, payload software, communication links, and swarm coordination.
• Existing research often lacks a structured approach to classifying and addressing these diverse threats.

Purpose of the Study:

• To systematically review and synthesize peer-reviewed literature on UAV cybersecurity from 2015-2025.
• To organize evidence using an OSI-inspired threat taxonomy mapping threats to system vulnerabilities.
• To compare deep learning approaches for intrusion and anomaly detection in single and multi-UAV systems.

Main Methods:

• Conducted a PRISMA-aligned systematic review of UAV cybersecurity studies.
• Utilized an OSI-inspired taxonomy to categorize threats (spoofing, jamming, intrusion, malware).
• Compared deep learning architectures, data representations, and evaluation metrics for UAV security.

Main Results:

• Convolutional and recurrent neural networks are prevalent for intrusion/anomaly detection.
• Attention-based, graph, and generative models are emerging for multi-agent systems with limited data.
• Protocol traffic and onboard telemetry are common data sources; RF data is less frequent.
• Efficiency techniques (pruning, quantization) are increasingly used for onboard deployment.
• Federated learning and blockchain integration are explored for scalability and security.

Conclusions:

• Significant advancements in deep learning for UAV cybersecurity exist, particularly in intrusion detection.
• Gaps remain in standardized datasets, adversarial testing, explainability, privacy, and regulatory assurance.
• Future work should focus on robust, certifiable security solutions for autonomous and swarmed UAV operations.