Systematic review of trends in deep learning for UAV cybersecurity

Unmanned Aerial Vehicles (UAVs) operate in navigation, sensing, and communication environments that are frequently degraded or adversarial. Their attack surface spans flight-control and payload software, radio links, and swarm coordination. This PRISMA-aligned systematic review synthesizes peer-reviewed studies published between 2015 and 2025 and organizes the evidence using an OSI-inspired threat taxonomy that maps spoofing, jamming, intrusion, and malware to system touchpoints and observable anomalies. We compare deep learning architectures, training targets, feature representations, evaluation practice, and deployment constraints relevant to single UAVs and swarms. Across the literature, convolutional and recurrent models dominate intrusion and anomaly detection pipelines, while attention-based, graph, and generative models appear in newer work targeting multi-agent settings and limited labels. Evidence most often relies on protocol traffic and onboard telemetry, whereas RF inputs are used less frequently and are typically represented as raw samples or spectrograms when datasets allow. Studies increasingly report efficiency-oriented deployment using pruning, quantization, distillation, or split inference to meet onboard compute and energy limits. Federated and multi-agent approaches are evaluated for scalability and robustness under poisoned updates, and blockchain-integrated designs are discussed under bandwidth and power constraints. Key gaps persist in shared datasets, repeatable adversarial stress testing, uncertainty and explainability reporting, privacy preservation, and certification-ready assurance cases for aviation regulation.