EJ-ANet: A Lightweight Envelope-Jitter Asymmetric Network for Drone Radio Frequency Fingerprint Identification
Abstract
Drone radio frequency fingerprint identification (RFFI) provides a passive solution for fine-grained drone recognition. However, time-frequency image-based methods introduce additional preprocessing costs, while raw in-phase/quadrature (IQ) networks remain limited in modeling complementary temporal variations. This letter proposes a lightweight Envelope-Jitter Asymmetric Network (EJ-ANet) for raw-IQ drone RFFI. EJ-ANet extracts multi-scale base features, uses asymmetric high- and low-resolution pathways to capture complementary temporal variations, and adopts bidirectional cross-attention for feature fusion. A prototype-guided fine-tuning objective is further introduced to refine the embedding space. Under recording-file-disjoint evaluation on DroneDetect, EJ-ANet achieves 94.47% accuracy on the CLEAN subset with only 0.085 M parameters and 0.78 ms latency, while maintaining strong performance under multiple interference conditions.