Research on Lightweight AI Models for UAV Identity Authentication and Intrusion Detection in the Low-Altitude Economy

This paper proposes a TinyML-based lightweight deep learning model to address the risks of UAV identity forgery and illegal intrusion, enabling real-time identity authentication and abnormal behavior detection at the drone edge. To further optimize computational efficiency, we explore pruning and dynamic sparsity techniques alongside quantization-aware training, reducing FLOPs by 40% without accuracy loss. By combining depthwise separable convolution with an attention mechanism, the proposed network architecture significantly reduces computational complexity while maintaining recognition accuracy. Future work will integrate multimodal data (e.g., acoustic/lidar) to enhance robustness against environmental variations. Experimental results demonstrate that the model achieves 98.7% accuracy in UAV identity authentication tasks, an F1-score of 96.3% for anomaly detection, and a compact size of only 1.2MB, enabling real-time inference on mainstream drone edge computing chips. This research provides an effective lightweight solution for low-altitude economy security, with scalability to diverse regulatory frameworks.