【作者】 周新宇；

【导师】 胡爱群；

【作者基本信息】 东南大学 ， 网络空间安全， 2022， 博士

【摘要】 近年来,无线通信技术迅速发展,在军事和民用上都发挥了重要的作用,深入的改变了各行各业。然而,无线通信的开放性使其更容易遭受设备仿冒、信息篡改和数据伪造等非法攻击,带来了严峻的安全问题。无线通信安全的重要内容是无线接入安全,而身份认证是无线接入安全的首要任务。射频指纹技术是一种有效的身份认证解决方案。由于设计和制造过程的容差,不同设备具有不同的硬件差异,并反映在其发射的射频信号中。射频指纹技术通过对射频信号进行分析,提取相关的硬件差异作为特征形成射频指纹,对设备进行识别和接入认证,提高无线通信应用的安全性。因此,开展无线通信物理层安全的射频指纹研究在军事和民用领域都有着深刻的理论意义和重大的实用价值。本文围绕多制式无线通信射频指纹的提取与识别进行了研究,主要包括多通信制式下通用的稳定射频指纹的提取与识别方法、复杂信道下鲁棒的射频指纹的提取与识别方法、样本不充足时的射频指纹的提取与识别方法以及射频指纹的开集识别研究,并在ZigBee、GSM、Wi-Fi、LoRa和LTE五种通信制式上进行了验证。本文的创新性研究成果包括:1.提出了一种多通信制式下通用的稳定射频指纹的提取与识别方法。为了提高射频指纹技术的在多通信制式下的通用性和稳定性,本文提出了一种长时重复信号叠加算法。该算法利用通信制式中广泛存在的重复信号,通过长时叠加,降低时变因素的影响,从而提升射频指纹的稳定性。另外,理论推导证明了该算法可以带来较大的信噪比提升,减少信噪比波动对指纹稳定性的影响。该算法在ZigBee、Wi-Fi和LTE的多通信制式的实验中,都取得了100%的识别率。另外,在54台ZigBee持续18个月的长时间数据集上,其性能损失低于1%。2.提出了一种低信噪比下鲁棒的射频指纹的提取与识别方法。为了提高射频指纹技术在低信噪比下的鲁棒性,本文提出了一种人工噪声添加算法。该算法通过对训练信号添加人工噪声,使得训练信号和测试信号拥有相似的信噪比,从而提取该信噪比下具有区分度的硬件特性形成射频指纹。传统的鲁棒射频指纹算法需要在不同信噪比下进行训练,而噪声添加算法根据接收信号的信道状况,通过理论计算实现自适应的噪声添加和鲁棒的指纹提取,消耗资源低,更具有实际意义。实验结果显示,在0d B的AWGN信道场景下,54台ZigBee设备的识别率为91.82%。在信噪比为10d B-26d B时,识别率能够保持在100%附近。3.提出一种基于马氏距离信道均衡的抗多径射频指纹提取算法。无线通信设备在实际应用中需要工作在不同的信道环境下,其中多径信道的衰落会严重影响指纹的提取与识别。基于此,本文提出了一种基于马氏距离信道均衡的抗多径射频指纹提取算法。线性滤波器难以均衡非线性的射频指纹,因此,该算法通过控制滤波器的参数,使用马氏距离作为损失函数,在均衡信道衰落的同时保持射频指纹的区分度,从而实现多径信道下的设备识别。54台ZigBee设备的多场景信道仿真实验表明该算法可以有效的提高多径信道下射频指纹的鲁棒性。在进一步的实际环境实验中,在长距离的视距信道下的ZigBee数据集上,与不考虑多径的方案相比,该算法有16.21%的识别率提升。另外,在16台Wi-Fi设备的多位置数据集上,识别率最高提升了82.86%。4.提出一种基于频域商的抗多径射频指纹提取算法。当信道较为复杂时,线性滤波器可能无法有效的均衡信道。基于此,本文提出了一种基于频域商的抗多径射频指纹提取算法。不同的数据符号从不同的角度表达了设备的硬件特性,相邻的数据符号在信道相干时间内拥有相似的信道特征。因此,理论上可以利用具有不同的数据的相邻符号,构建频域商,在消除信道的同时,保留一定的指纹。本文通过20台Wi-Fi设备在不同位置的数据集对该算法进行了验证,实验结果表明,即使训练信号和测试信号的信道差异较大时,该算法也能够达到99%以上的识别率。5.提出一种基于子序列聚合的射频指纹提取与识别方法。受限于实际的资源,对于一些复杂的通信制式,难以收集到充分的训练样本,这可能会导致相应的射频指纹方案在处理未收集到的数据时出现性能不稳定。循环移位序列在通信制式中得到了广泛的应用,本文提出了一种基于子序列聚合的射频指纹提取与识别方法来解决循环移位序列的样本不充足问题。该方法利用循环移位序列的数据特性,生成耦合的子序列。耦合的子序列是硬件在循环移位域对相同数据的响应,因此具有一定的相似性。利用这种耦合关系,该方法只需要部分的序列样本,就可以识别全部的循环移位序列。在ZigBee、LTE前导和LoRa三种采用循环移位序列的通信制式上的实验结果表明,该方法只需要收集少量数据,就能够达到与收集全部数据时相似的性能,降低了对数据采集的要求,提高了数据利用率,具有实际意义。6.提出了一种基于生成模型的射频指纹开集识别技术。射频指纹技术的实际应用中需要面对未知设备的接入和干扰。对此,本文研究了射频指纹的开集识别问题,借鉴其他方向的研究成果,提出了一种基于生成模型的射频指纹开集识别技术。该技术将接收信号的特征向量分解为设备身份向量和噪声向量,对其生成过程建模,使得未知设备的识别被转化成简单的假设检验。在6台合法ZigBee设备与6台未知设备的开集实验中,该技术的等错误率只有0.63%,远低于基于深度学习的指纹开集识别技术。更多还原

【Abstract】 In recent years,wireless communication technology has developed rapidly,playing an important role in both military and civilian applications,and has profoundly changed all walks of life.However,the openness of wireless communication makes it more vulnerable to illegal attacks such as device counterfeiting,information tampering and data forgery,which brings serious security problems.An important element of wireless communication security is wireless access security,and identity authentication is the first task of wireless access security.Radio frequency fingerprint(RFF)technology is an effective authentication solution.Due to the tolerance of design and manufacturing process,different devices have different hardware differences and are reflected in the radio frequency(RF)signals.RFF technology has enhanced the security of wireless communication applications by analyzing RF signals and extracting relevant hardware differences as features to form fingerprints for device identification and access authentication.Therefore,conducting RF fingerprinting research on wireless communication physical layer security has profound theoretical significance and significant practical value in both military and civilian fields.In this thesis,we focus on RFF extraction and recognition for multi-standard wireless communication protocols mainly including the general and stable RFF extraction and recognition in multi-standard wireless communication protocols,robust RFF extraction and recognition in complex channels,RFF extraction and recognition with insufficient samples,and the open set recognition problem.The proposed algorithms are verified with five communication protocols including ZigBee,GSM,Wi-Fi,LoRa and LTE.The innovative achievements of this thesis include:1.A general and stable RFF extraction and recognition method for multi-standard wireless communication protocols is proposed.In order to improve the generality and stability of RFF technology for multi-standard wireless communication protocols,this thesis proposes a long-term stacking of repetitive symbols(LSRS)algorithm.LSRS utilizes the widely existing repetitive signals for stacking,which has reduced the influence of time-varying factors to improve the stability of RFF.In addition,the theoretical derivation demonstrates that LSRS can lead to SNR improvement and reduce the impact of SNR fluctuation on fingerprint stability.The algorithm has achieved a recognition rate of 100% in the experiments of ZigBee,Wi-Fi and LTE.In addition,the performance loss is less than 0.8% in the 18 month data set of 54 ZigBee devices.2.A robust RF fingerprint extraction and recognition method with low SNR is proposed.In order to improve the robustness of RFF technique with low SNR in AWGN channel,an artificial noise adding(ANA)algorithm is proposed in this thesis.ANA makes the training signal and test signal have similar SNR by adding artificial noise to the training signal to extract the discriminative hardware features under this SNR to form the RFF.The traditional robust RFF algorithm needs to be trained under each SNR,while ANA uses theoretical calculation to add noise adaptively according to the channel condition of the received signal,which consumes low resources and has significant practical significance.The experimental results show that the recognition rate of 54 ZigBee devices is 91.82% in the 0d B AWGN channel scenario,and the recognition rate can be maintained near 100% when the SNR is in 10 d B-26 d B.3.A channel robust RFF extraction and recognition method based on Mahalanobis distance equalization is proposed.Wireless communication devices need to work in different channels in practical applications,where the fading of multipath channels can seriously affect the extraction and recognition of fingerprints.Based on this,this thesis proposes a channel robust RFF extraction and recognition method based on Mahalanobis distance equalization(MDE)algorithm.The linear filter is difficult to equalize the nonlinearity of RFF.Therefore,MDE can recognize devices under multipath channels by controlling the parameters of the filter and using the Mahalanobis distance as the loss function to equalize channel fading while maintaining the RFF features.Simulation experiments with 54 ZigBee devices show that this method can effectively improve the robustness under multipath channels.In real-world experiments,MDE shows a 16.21% recognition rate improvement over the conventional scheme in the ZigBee dataset under long range line-of-sight channels.In addition,the recognition rate is improved by up to 82.86% on the multi-location dataset with 16 Wi-Fi devices.4.A channel robust RFF extraction and recognition method based on frequency domain quotient is proposed.When the channel is more complex,linear filters may not be able to equalize the channel effectively.Based on this,a channel robust RFF extraction and recognition method based on frequency domain quotient(FDQ)is proposed in this thesis.Different data symbols express the hardware characteristics from different perspectives,and adjacent data symbols have similar channel characteristics in the channel coherence time.Therefore,it is theoretically possible to construct frequency domain quotient using adjacent symbols with different data to retain certain fingerprints while eliminating the channel.In this thesis,this method is evaluated with a dataset of 20 Wi-Fi devices at different locations,and the experimental results show that the algorithm is able to achieve a recognition rate of more than 99% even when the channels of training signals and test signals are different.5.A sub-sequence aggregation based RFF extraction and recognition method is proposed.Limited by practical resources,it is difficult to collect sufficient training samples for some complex communication protocols,which may lead to unstable performance of the RFF scheme when processing uncollected data.Cyclic shift sequences are widely used in communication protocols,and this thesis proposes a sub-sequence aggregation based RFF extraction and recognition method for cyclic shift sequences with insufficient samples.This method utilizes the cyclic shift property to construct coupled sub-sequences.The coupled sub-sequences are similar as they are hardware responses to the same data in the cyclic shift domain.Using this coupling relationship,this method requires only part of sequence types to recognize all types of cyclic shift sequence.Experimental results in ZigBee,LTE and LoRa have proved that this method is of practical importance as it only needs to collect a small amount of data to achieve similar performance as when collecting all types of data,reducing the requirement for data collection and improving data utilization.6.A generative model based open set RFF recognition technique is proposed.The practical application of RFF needs to face the access and interference of unknown devices.In this regard,this thesis studies the open set identification problem,and proposes a generative model based open set RFF recognition technique with reference to the research results of other fields.The technique decomposes the feature vector of the received signal into device identity vector and noise vector and tries to model the generation process,so that the identification of unknown devices is transformed into a simple hypothesis test.In the open set experiments with six legitimate ZigBee devices and six unknown devices,the equal error rate(EER)is only 0.63%,which is much lower than the deep learning based technique.更多还原