Zero-shot attack detection in UAV networks using foundation models

Traditional intrusion detection systems require extensive training data for each attack type, creating vulnerabilities in UAV networks where novel threats frequently emerge. We present ZeroUAV, a zero-shot attack detection framework using foundation models to identify unseen attack patterns without labeled examples. Our transformer-based architecture employs cross-modal attention to align network traffic with semantic attack descriptions, enabling classification based on conceptual similarity. The framework incorporates attack ontology, contrastive learning, and meta-learning for rapid adaptation. Evaluation on UAV-NIDD dataset shows 87.3 % zero-shot accuracy and 94.6 % accuracy with five examples per attack type, significantly outperforming supervised methods. The system achieves under 10 ms inference latency for real-time UAV deployment. Our contributions include the first foundation model for UAV cybersecurity, a novel zero-shot learning framework, and validation demonstrating practical viability for evolving threat landscapes.