260214 Enhance publishing | A Radio Frequency Fingerprint Open Set Identification Method Combining Multi-Scale Wavelets and Hypersphere Representation

Combining Multi-Scale Wavelets and Hypersphere Representation[J]. Journal of Electronics & Information Technology, in press. doi: 10.11999/JEIT260214 Authors: TIAN Xinyu, LI Zirui, ZHENG Qinghe*, ZHOU Fuhui, YU Lisu, HUANG Chongwen, JIANG Weiwei, SHU Feng, ZHAO Yizhe.Author：School of Intelligent Engineering, Shandong Management UniversityDOI：10.11999/JEIT260214Original：jeit.ac.cn/cn/article/doi/10.11999/JEIT260214Correspondents: ZHENG Qinghe，zqh@sdmu.edu.cnOpen source date: April 27(th), 2026Funds: National Natural Science Foundation of China (62401070), Shandong Provincial Natural Science Foundation (ZR2019ZD01, ZR2023QF125), Shandong Provincial Youth Innovation Team Plan of Higher Education Institutions (2024KJH005), Shandong Provincial Science and Technology Based Small and Medium sized Enterprises Innovation Capability Enhancement Project (2024TSGC0055) Open source content1 A Radio Frequency Fingerprint Open Set Identification

Method

Combining Multi-Scale Wavelets and Hypersphere Representation-Recurrence CodeAbstract:   Objective  To address the challenges of radio frequency fingerprint (RFF) identification in the open-set scenarios under low signal-to-noise ratio (SNR) conditions, where fingerprint features are easily masked by noise, multipath effects introduce nonlinear distortions, and existing methods struggle with feature extraction and unknown device detection, this paper proposes a deep learning framework that integrates multi-scale wavelet decomposition with hypersphere representation for robust open-set RFF identification.  Methods  The proposed method MS-RANet consists of three key components.

First, a multi-scale feature extraction front-end based on one-dimensional stationary wavelet transform is designed to perform full-resolution and multi-scale decomposition of I/Q signals, preserving discriminative fingerprint information while suppressing noise. Second, a multi-scale residual attention network is constructed by incorporating deep residual learning, global self-attention mechanism, and bidirectional LSTM to enhance the model’s ability to perceive subtle fingerprint features and model long-range temporal dependencies.

Finally, hyperspherical metric learning is introduced to constrain the feature space onto a unit hypersphere, optimizing angular margins to achieve compact intra-class and separable inter-class feature distributions. Unknown device detection is subsequently performed based on cosine similarity. Results and Discussions  Extensive experiments conducted on high-fidelity IEEE 802.11-based simulation dataset demonstrate the effectiveness of MS-RANet.

It achieves an average classification accuracy of 65.34% across SNR levels ranging from –5 dB to 20 dB, and an area under the curve (AUC) of 0.81 at –5 dB SNR, significantly outperforming existing methods such as DNN, GRU, CNN-LSTM, ResNet50, and DRSN-CA. Confusion matrices and receiver operating characteristic (ROC) curves further validate its robustness under extreme channel conditions. Feature visualization using t-SNE reveals that MS-RANet forms well-separated and compact clusters for known devices while effectively pushing unknown samples away from known class regions.

Ablation studies confirm the individual contributions of multi-scale wavelet front-end, global attention, BiLSTM, and hyperspherical metric learning modules.  Conclusions  This paper presents a robust open-set RFF identification method that effectively combines multi-scale wavelet decomposition with hyperspherical representation learning. The proposed framework exhibits strong noise resilience, superior feature discrimination, and reliable unknown device detection under low-SNR and multipath fading conditions.

The future work will focus on reducing computational complexity, improving inference speed, exploring generalization capability across different scenarios and protocols, and integrating the method with other physical-layer security mechanisms for collaborative authentication frameworks.