Lightweight Memory Network for Drone Type RF Fingerprint Identification in Dynamic Environments

Abstract

Radio frequency fingerprint identification (RFFI) is a crucial physical layer security technique for unauthorized drone identification. However, current drone RFFI methods are primarily designed for line-of-sight (LOS) video transmission or remote control signals while suffering significant performance degradation under non-line-of-sight (NLOS) propagation, particularly in dynamic environments. To overcome this challenge, we propose a lightweight memory network framework for drone RFFI that enables reliable individual drone identification. First, a multi-stage feature extraction framework is proposed to adaptively capture discriminative multiscale spectrogram features while selecting salient channel and spatial information through attention mechanisms, and is then integrated with a dual path module with multi-head attention to adaptively learn global and local feature relevance. Furthermore, we incorporate a learnable memory slot module for cross-sample knowledge sharing and employ a gated weighted fusion mechanism for adaptive feature selection, improving model robustness. Finally, we enhance model recognition performance by combining spatial features from the dual-path network with the cross-sample temporal information from the memory module. We also construct a real-world drone dataset comprising nine classes for type identification. Extensive experiments demonstrate that the proposed method achieves an average recognition accuracy of 95.16% on this dataset, maintaining a favorable trade-off between model complexity and identification accuracy. Additionally, ablation studies confirm the effectiveness of each component in the framework.