Robust Cross-Domain Drone RFFI Method Using Domain-Invariant Adversarial Learning and Manifold Regularization

The malicious use of unmanned aerial vehicles (UAVs) has raised serious concerns about public safety and personal privacy, necessitating an urgent need for safe and effective UAV identification technology. Radio frequency fingerprint identification (RFFI) is a key technology for physical layer authentication. Existing UAV RFFI methods are based on training and testing data collected under identical conditions, leading to performance degradation when applied to data collected from unknown environments. To alleviate these issues, we propose a robust UAV RFFI method that combines domain-invariant adversarial training and manifold regularization. First, we design a multi-scale feature extraction network and employ domain-invariant adversarial training to capture features that remain consistent across different domains. Next, we introduce a manifold regularization strategy to diversify the training data by performing linear interpolation on the hidden layer representations, thereby enhancing the model’s generalizability in cross-domain scenarios. A publicly available UAV RF dataset is utilized to evaluate the proposed method. Simulation results indicate that it achieves better performance than existing cross-domain identification methods and exhibits high stability across different hyper-parameter settings. Furthermore, it maintains robust performance even under low signal-to-noise ratio (SNR) scenarios. Specifically, when only 200 samples per domain, it achieves at least 92% identification accuracy.