Research On Deep Learning Based Key Generation For FDD Systems

With the rapid development of wireless communication technology,massive heterogeneous devices transmit a large amount of sensitive and confidential data on wireless channels.As a highly secure,convenient and practical key distribution method,the wireless channel-based physical-layer key generation(PKG)method has received extensive attention in recent years.At this stage,many PKG methods in the time division duplex(TDD)mode have appeared,and their feasibility has been verified in low-to-medium-rate mobile scenarios.Frequency division duplex(FDD)mode,as another multiplexing method,is widely used in existing cellular communications,such as LTE and NB-IoT.However,because the two communicating parties work in different frequency bands under the FDD system,the channel reciprocity is not ideal,and there are few PKG methods suitable for the FDD system at present.Therefore,it has profound research value and practical significance to study the key generation method based on FDD systems.Aiming at the problem of poor channel reciprocity and difficulty in generating consistent keys in FDD systems due to different operating frequencies of the two communicating parties,this thesis studies the channel model and its reciprocity under the FDD system,and proves that there is a frequency band feature mapping function in FDD systems.On this basis,this thesis proposes a set of deep learning-based single-scenario and multi-scenario FDD key generation methods,and verifies the feasibility of the proposed method in simulation and experiments.The main research work and innovative research results of this paper are as follows:1.The existence of a frequency band feature mapping function in FDD systems is proved.Aiming at the problem that the relationship between the uplink and downlink channels in the FDD system is not clear,and it is difficult to determine whether the relationship between the uplink and downlink feature mapping exists,this thesis firstly studies the channel reciprocity in the FDD system,and models the channels in the FDD system in detail.Then,from the perspective of location and surrounding environment,this thesis proves that in a given environment,if the mapping function of user location to channel is bijective,there is a frequency band feature mapping function that can map the uplink and downlink channel features.In real wireless scenarios,the probability of this assumption being true tends to be 1.This conclusion shows that there is a frequency band feature mapping relationship that can be used for reciprocal feature construction in FDD systems,which lays a theoretical foundation for deep learning-based FDD key generation.2.A deep learning-based FDD key generation method in a single scenario is proposed.Aiming at the problem that the frequency band feature mapping function is complex and difficult to be represented by formulas,this thesis proposes a deep learning-based FDD key generation method in a single scenario.The method obtains the band feature mapping function by training the key generation network(KGNet)model,and then uses the KGNet model to construct reciprocal channel features and generate keys.Simulation results show that KGNet can achieve better fitting and generalization performance than several other benchmark networks.When the signal-to-noise ratio(SNR)is higher than 20 dB,the key error rate(KER)is lower than 0.1.In addition,the complexity analysis shows that the resources consumed by the method proposed in this paper are completely acceptable,and the proposed method can be applied to resource-constrained terminal devices.3.The transfer learning and meta learning-based FDD key generation methods in multi scenarios are proposed.Aiming at the problems of incompatibility of tasks and poor performance of KGNet caused by scene changes,this thesis proposes the transfer learning and meta learning-based FDD key generation methods respectively to achieve rapid adaptation in new scenarios.Simulation results show that both methods can effectively improve the performance of KGNet in new scenarios.In the indoor corridor environment,when the SNR=20 dB,the KERs of the keys generated by the transfer learning and meta learning-based methods are reduced by 38.8%and 52.9%,respectively,compared with the method without adaptation in the new scenario.In the outdoor environment,when SNR=20 dB,the KERs of the keys generated by the transfer learning and meta learning-based methods are reduced by 73.14%and 77.29%,respectively,compared with the methods without adaptation in the new scenario.In addition,the complexity analysis shows that the meta learning-based method consumes less time and lower CPU and GPU resources than the transfer learning-based method.4.A set of FDD channel acquisition systems is built and the proposed method is preliminarily verified.In order to further prove the effectiveness of the proposed method,this thesis builds a set of GNURadio and universal software radio peripheral(USRP)based FDD channel acquisition systems,and collected a large amount of experimental data in indoor and outdoor environments to conduct preliminary verification of the deep learning-based FDD key generation method.The results show that this method can effectively generate keys for FDD systems,and the KER of keys generated in an indoor environment is only 0.0283,which is 87.8%lower than that of keys generated without performing frequency band feature mapping.The KER of the key generated in an outdoor environment is 0.1317,which is 67.5%lower than the KER of the key generated without performing frequency band feature mapping.