| 摘 要: 针对毫米波大规模阵列场景下传统深度学习物理层认证方法对标注样本依赖较强、少样本条件下易过拟合的问题，提出一种基于波束域功率图样与余弦相似度孪生网络的少样本物理层认证方法。依据毫米波信道角域稀疏特性，对波束域信道矩阵进行频谱中心化、对数功率映射、动态范围裁剪和归一化处理，构建能够表征用户空间位置差异的波束域功率图样指纹；结合参数共享孪生卷积神经网络完成特征嵌入，并利用余弦相似性实现注册参考样本与待测样本间的相似性匹配与身份判别。仿真结果表明，该方法在不同 N-way K-shot 设置及不同信噪比条件下均具有较好的认证性能和鲁棒性，且余弦相似度优于欧氏距离。结果说明，该方法能够有效提升少样本毫米波物理层认证的准确性与稳定性。 |
| 关键词: 毫米波通信 物理层认证 少样本学习 孪生网络 余弦相似度 |
|
中图分类号:
文献标识码:
|
| 基金项目: 江苏省研究生科研与实践创新计划项目（KYCX24_1207） |
|
| Few-Shot Identity Authentication Method for Millimeter-Wave Beamspace Based on Cosine-Similarity Siamese Network |
|
FengYi, YANG Lijun
|
School of Internet of Things, Nanjing University of Posts and Telecommunications
|
| Abstract: To address the issues of heavy dependence on labeled samples and the susceptibility to overfitting under few-shot conditions in traditional deep learning-based physical layer authentication methods for millimeter-wave(mmWave) massive array scenarios, a few-shot physical layer authentication method based on beam-domain power patterns and a Cosine Similarity Siamese Network is proposed. Leveraging the angular domain sparsity of mmWave channels, the beam-domain channel matrix undergoes spectrum centering, log-power mapping, dynamic range clipping, and normalization to construct beam-domain power pattern fingerprints that characterize spatial location differences of users. A parameter-sharing Siamese Convolutional Neural Network is employed for feature embedding, while cosine similarity is utilized to achieve similarity matching and identity discrimination between registered reference samples and test samples. Simulation results demonstrate that the proposed method maintains superior authentication performance and robustness across various N-way K-shot settings and signal-to-noise ratios. Furthermore, cosine similarity is shown to outperform Euclidean distance in this context. The results indicate that the proposed method effectively enhances the accuracy and stability of few-shot mmWave physical layer authentication. |
| Keywords: Millimeter-wave (mmWave) communications Physical layer authentication Few-shot learning Siamese network Cosine similarity |
