Specific emitter identification (SEI) is a device identification technology that extracts radio frequency (RF) fingerprint from received signals. However, channel effects on RF fingerprint can vary between the training and testing stage, and SEI based on deep learning (DL) will be unable to withstand channel changes. To address this problem, we propose a channel-robust SEI scheme driven by causal learning. We analyze received signals from the causal perspective and construct a structural causal model (SCM) of SEI. In the SCM, received signals are considered as mixtures of the causal element and interference element, and only the former affects identification. Additionally, we design a new RF fingerprint feature representation called the centralized logarithmic power spectrum (CLPS) to reduce the impact of channel effects. Furthermore, we propose a causal purification network (CPNet) driven by causality to further alleviate channel effects. CPNet weakens the spurious associations between the channel and emitter labels through feature decorrelation and feature purification, strengthens the correlation between RF fingerprint and labels, and improves the generalization of SEI. Finally, our approach is evaluated extensively using 20 ZigBee devices under different channel environments. Experimental results demonstrate that our scheme can effectively alleviate channel effects, improve SEI performance under various channel environments, and exhibit good generalization and stability.

中文翻译：

---

通过因果学习对未知通道统计数据进行稳健的特定发射器识别

特定发射器识别 (SEI) 是一种从接收信号中提取射频 (RF) 指纹的设备识别技术。然而，在训练和测试阶段，信道对射频指纹的影响可能会有所不同，基于深度学习（DL）的SEI将无法承受信道变化。为了解决这个问题，我们提出了一种由因果学习驱动的通道鲁棒 SEI 方案。我们从因果角度分析接收到的信号，并构建SEI的结构因果模型（SCM）。在SCM中，接收信号被认为是因果元素和干扰元素的混合，并且只有前者影响识别。此外，我们设计了一种新的射频指纹特征表示，称为集中对数功率谱（CLPS），以减少通道效应的影响。此外，我们提出了一种由因果关系驱动的因果净化网络（CPNet），以进一步减轻通道效应。 CPNet通过特征去相关和特征纯化削弱了通道和发射器标签之间的虚假关联，加强了RF指纹和标签之间的相关性，提高了SEI的泛化性。最后，我们的方法在不同的信道环境下使用 20 个 ZigBee 设备进行了广泛的评估。实验结果表明，我们的方案可以有效缓解信道效应，提高各种信道环境下的SEI性能，并表现出良好的泛化性和稳定性。