A New UAV Identification Method Based on Multi-Domain Prior Information Extraction and Cross-Environment Composite Loss Regularization

Unmanned aerial vehicle (UAV) identification is crucial for guaranteeing low-altitude security. Most existing radio frequency fingerprint (RFF)-based UAV identification methods only exploit single-domain features, underestimating the feature variations across different environments and the importance of features on other domains. In this paper, to identify UAVs in different environments, we propose a new UAV identification method that exploits valuable information of UAV signals in multiple domains, and incorporates composite loss regularization to exploit the invariant features across different environments. We first design a multi-domain prior information generation module, in which the physical priors of UAV signals are complementarily represented across the modulation, time-frequency, and micro-Doppler domains. We then propose a cross-environment feature extraction module (CEFEM). By introducing cross-environment residual (CeRes) blocks and a composite loss regularization term, the CEFEM can mitigate the distribution discrepancies of signals in different environments, and thus, improves UAV identification accuracy. Experimental results show that our proposed method can improve the UAV identification accuracy above 2% compared to the state-of-the-art on both self-collected and open-source datasets.