【作者】 李浩；

【导师】 汤羽；

【作者基本信息】 电子科技大学 ， 软件工程， 2022， 硕士

【摘要】 随着5G通信时代的到来,保证设备间的通信安全显得格外重要。通过分析射频信号的微小差异而提取出的硬件特征,可以作为每台设备独一无二的射频指纹,从而使得射频指纹识别技术能够为设备在物理层提供更加安全可靠的身份认证机制。目前已有的射频指纹识别技术都是在密集型数据集上实现的,在面临小样本问题时显得有些无力。特别是在某些特殊环境下,如无人机等处于较为隐蔽环境的通信设备,无法获取其大量信号进行射频指纹识别。为解决射频指纹小样本问题,本文提出将小样本学习方法应用到射频指纹识别技术中,研究无线电射频指纹小样本识别方法。本文通过分析现有射频指纹识别技术的不足之处,发现在面临小样本数据集时容易出现过拟合问题。而目前已有的小样本学习方法如元学习、迁移学习等等,主要应用于图像与文本领域。因此,基于以上问题,本文对已有元学习网络模型进行改进,设计实现了两个模型。一是对匹配网络模型的改进。为提取更有效的射频指纹特征,结合I/Q信号特点,修改了匹配网络的嵌入函数结构。并采用欧式距离公式作为计算样本特征相似度的注意力核函数,使模型更加能够区分I/Q信号样本之间的差异,从而保证模型识别准确率。二是结合迁移学习与元学习优点,构建元迁移学习模型。该模型将深度神经网络通过迁移学习引入到元学习模型中,并对模型参数进行了缩放,以适应射频指纹小样本任务,从而提高模型识别准确率。实验结果证明,小样本学习方法能够应用于射频指纹小样本识别。比起改进前,本文改进后的匹配网络模型在识别效果上表现良好,且在训练速度上也较快。同时,也验证了可以将深度神经网络作用于元学习模型上,并且元迁移学习模型的射频指纹小样本识别效果表现得更好。最后基于上述两个模型设计实现设备信号分析系统,以帮助快速识别射频信号所属设备。并为后续网络攻防下的无线通信设备的身份认证奠定基础,具有一定的战略意义。更多还原

【Abstract】 With the advent of the 5G communication era,it is very essential to ensure the communication security between devices.The hardware features extracted by analyzing the small differences of radio frequency(RF)signals can be used as the unique RF fingerprint of each device,so that the RF fingerprint recognition technology can provide a more secure and reliable authentication mechanism for the device at the physical layer.The existing RF fingerprint identification technologies are all implemented on dense data sets,which are somewhat helpless when faced with the problem of small samples.Especially in some special environments,such as drone and other equipment in a relatively concealed environment,it is impossible to obtain numerous of their signals for RF fingerprint identification.In order to solve the problem of small samples of RF fingerprints,this thesis proposes to apply the few-shot learning methods to RF fingerprint recognition technology,and studies the identification methods of small samples for RF fingerprints.In this thesis,by analyzing the shortcomings of the existing RF fingerprint recognition technology,it is found that the problem of overfitting is easy to occur when faced with small sample data sets.At present,the existing few-shot learning methods include meta-learning,transfer learning,etc.,which are mainly used in the field of image and text.Therefore,in view of the above issues,this thesis improves the existing meta-learning network model,and designs and implements two models.The first is the improvement of the matching network model.In order to extract more effective RF fingerprint features,combined with the characteristics of I/Q signal,the embedding function structure of the matching network is modified.And the Euclidean distance formula is used as the attention kernel function to calculate the similarity of sample features,so that the model can better distinguish the difference between I/Q signal samples,thereby ensuring the accuracy of model recognition.The second is to combine the advantages of transfer learning and meta-learning to build a meta-transfer learning model.This model introduces a deep neural network into a meta-learning model through transfer learning and scales the model parameters to suit the RF fingerprinting few-shot task,thereby improving the model recognition accuracy.The experimental results show that the few-shot learning method can be applied to the small sample recognition of RF fingerprints.Compared with before improvement,the improved matching network model in this thesis has well performance in recognition effect and faster training speed.At the same time,it is also verified that the deep neural network can be applied to the meta-learning model,and the meta-transfer learning model performs better in the recognition of small samples of RF fingerprints.Finally,an equipment signal analysis system is designed and implemented based on the above two models to help quickly identify the equipment to which the RF signal belongs.It also lays the foundation for the identity authentication of wireless communication equipment under the subsequent network attack and defense,which has certain strategic significance.更多还原