AS-Net: A Lightweight Network for Robust UAV RF Signal Identification

Abstract

Benefiting from the convergence of embedded systems and micro-avionics, unmanned aerial vehicles (UAVs) have achieved various applications in fields such as aerial photography and logistics. However, accompanied by the increase in non-cooperative UAVs, the efficiency of management and identification of these UAVs has become a significant challenge. Among current UAV identification methods, radio frequency (RF) detection offers lower power consumption and longer detection range over radar, visual, and acoustic systems, making it a key technology for efficient UAV surveillance. Despite the physical benefits of RF detection, the high computational overhead and long inference latency of existing algorithms impede their deployment on resource-constrained edge devices. Therefore, this paper proposes Adaptive Suppression Network (AS-Net), a lightweight and robust network designed for efficient UAV RF signal identification across various signal-to-noise ratios (SNRs). Specifically, a channel dispersion regulation (CDR) module is designed to serve as a soft gating mechanism, dynamically allocating weights to individual channels based on their statistical dispersion to extract reliable RF fingerprints under low SNR scenarios. Furthermore, by integrating a multi-scale dilated backbone and a simplified transformer encoder with the CDR module, the proposed model achieves a lightweight architecture while enhancing representational capacity and robustness. The experimental results show that the proposed AS-Net utilizes only $\mathbf{0. 1 9 M}$ parameters and exhibits an average recognition accuracy of 81.15% from -20 to 5 dB, demonstrating a superior trade-off between model computational overhead and recognition capability.