基于Raw I/Q和深度学习的射频指纹识别方法综述

硬件差异会形成辐射源的独有指纹，并附加在无线电信号上，利用辐射源的这一独特属性可进行射频指纹识别。在非合作条件下，由于信道环境未知、信号调制方案等先验知识匮乏，基于特征工程的射频指纹识别方法面临巨大挑战，而基于深度学习的射频指纹识别方法，尤其是能够直接处理Raw I/Q的方法表现出了很大潜力，但是该方向的研究成果较为零散，妨碍了研究者对关键问题的把握。该文首先从先验知识的利用上，对基于深度学习的射频指纹识别方法进行了分类对比，将问题聚焦到基于Raw I/Q和深度学习的射频指纹识别方法。然后，该文重点对使用Raw I/Q进行射频指纹识别的深度神经网络模型进行了分类和讨论，并对射频指纹识别相关的开源数据集、数据表示方法和数据增强方法进行了整理和归纳。最后，该文讨论了基于深度学习的射频指纹识别方法所面临的难题和值得关注的研究方向，以期对射频指纹识别的研究与应用有所帮助。 

A Review of Radio Frequency Fingerprinting Methods Based on Raw I/Q and Deep Learning

The hardware imperfection can generate a unique fingerprint of the trasmitter, and it is attached to the radio signal. The unique attribute of transmitter can be used for Radio Frequency Fingerprinting (RFF). Due to the unknown channel conditional and the lack of prior information such as modulation scheme, the traditional method of RFF faces huge challenges to non-cooperative conditions. On the contrary, RFF methods based on Deep Learning (DL), especially those that can directly process raw I/Q, show great potential. However, the research results of this direction are scattered, which seriously hinders researchers from grasping the key issues. This paper first classifies and compares the RFF methods based on DL according to the utilization of prior knowledge, and focuses on the RFF methods based on raw I/Q and DL. Then, this paper focuses on the classification and discussion of the deep neural network model of RFF using raw I/Q, and summarizes the open source data sets, data representation methods and data augmentation methods related to RFF. Finally, this paper discusses the difficulties and research directions of the RFF based on DL, hoping to help the research and application of the RFF.