A multi-scale based open set OFDM communication signal radio frequency fingerprinting method

Abstract

The application discloses a kind of based on multi-scale open set OFDM communication signal radio frequency fingerprint identification method, solve the problem of low open set OFDM communication signal radio frequency fingerprint identification rate, belong to communication field;Including: cutting m-n transmitter data set into training set, test set and verification set, randomly extracting n transmitter data set and adding in test set;The training set is input into multi-scale network model, the spectral feature is enhanced by Fourier frequency domain enhancement preprocessing module, after multi-scale time series feature extraction fusion is carried out by HSMSSD multi-scale convolution feature extraction module, the m-n+1 transmitter categories corresponding to feature extraction fusion signal are identified by kernel clustering open set identification module based on m-n known categories and 1 preset unknown category;Get trained multi-scale network model;Test and verify by test set and verification set;The application improves the open set OFDM communication signal radio frequency fingerprint identification rate.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of communication, and relates to a multi-scale-based open set OFDM communication signal radio frequency fingerprint identification method in network security services. BACKGROUND

[0002] In the field of communication technology, from the long-term evolution LTE communication system enabled by the 3G system to the currently familiar 5G and 6G communication technologies, they are widely used in Internet of Things, smart city, automatic driving and other technical applications. In the process of communication technology transmission, in order to meet the characteristics of high speed and low delay, the communication technology has openness, which leads to the explosive growth of air interface radio frequency, and brings great challenges to efficient spectrum allocation. At the same time, malicious attackers can easily obtain and use low-cost, easy-to-program software-defined radio platforms to launch network attacks, thereby bringing potential risks to the security of Internet of Things.

[0003] In cooperative LTE communication, the sending end completes the access request through the initial access request, that is, the PRACH channel (Physical Random Access CHannel). In non-cooperative communication, the receiving end cannot determine whether the signal is a malicious attack by using the PRACH. The reasons are as follows:

[0004] First, if the malicious attack signal is not the LTE system, it does not have the PRACH channel to perform access request interaction, that is, it cannot access;

[0005] Second, if the malicious attack signal is the LTE system, and the malicious attack signal adds malicious attack elements in the PRACH channel, it can realize the destruction of the communication system.

[0006] Therefore, in order to ensure the security of the communication system, it is necessary to bypass the application meaning contained in the signal, and to determine the device through the characteristics of each domain before the signal access of the device, so as to realize the radio frequency fingerprint identification RFFI, so as to perform spectrum monitoring and ensure the orderly use of wireless spectrum.

[0007] Figure 1 The existing communication radio frequency fingerprint identification process system schematic diagram is shown in the figure. After the signal is collected by the antenna and collected into the computer of the receiving end, after signal detection and preprocessing, the preprocessed signal is obtained. The preprocessed signal is input into the RFFI model, the characteristics of the signal are extracted, and the signal is identified, so as to achieve the purpose of identifying the radio frequency device of the signal. Among them, Figure 1 The backup in the formula represents a transmitter of a known category, and the foreigner represents a transmitter of an unknown category.

[0008] In the prior art, the communication signal RFFI method is divided into two kinds: model-driven RFFI method and data-driven RFFI method. RFFI is divided into two kinds of open set RFFI and closed set RFFI, and the closed set RFFI mainly faces the task that the number of input training set categories is more than or equal to the number of output test set categories; the open set RFFI mainly faces the task that the number of input training set categories is less than the number of output test set categories. In the real environment of RFFI for communication signals, the scene has the characteristics of strong noise, multiple radio frequency machines and limited energy supply. Among them, the model-driven RFFI method cannot meet the above characteristics because it relies too much on prior design; the data-driven RFFI method needs a large amount of data for model training for deep learning, so as to achieve the function of extracting signal features; and the existing RFFI data set is diverse, including the ORACLE data set based on USRP; it is an IQ signal data set generated by using 17 USRP SDR and MATLAB WLAN system toolbox to meet the IEEE802.11a standard. Among them, 16 USRPs are used as transmitters, and one is used as a receiver. The center frequency of the transmitted signal is 2.45GHz, the transmitted content contains random payload, and the same address field is used. The receiver samples at a rate of 5MS / s and is in the same working mode as the transmitter. It also includes the LTE communication protocol represented by the OFDM communication system and the data set generated based on HackRF; the above data set has the following defects:

[0009] Firstly, the communication data set is divided into digital sampling points, radio frequency equipment and distance between equipment, mainly focusing on the specific application level, and has considerable theoretical conversion to application practice ability in the increasingly mature radio frequency fingerprint closed set identification at present, but lacks certain accuracy in the open set identification which needs to focus on the characteristics of the signal itself at present and conduct theoretical experiments;

[0010] Secondly, the data set generated based on HackRF can grasp the characteristics of the generated signal itself, but mainly focuses on single carrier, and it is difficult to simulate the multi-carrier communication system, especially the LTE communication protocol represented by the OFDM system.

[0011] Nowadays, the network for communication RFFI is gradually diversified, the accuracy is continuously improved, and various deep learning structures are gradually introduced to improve the accuracy of signal recognition.

[0012] For example, patent publication number: CN120224198A, patent name: a lightweight real-time radio frequency fingerprint identification method based on flow jump connection, which replaces the two-dimensional convolution layer in ResNet network with one-dimensional convolution layer, the ADC information sampling throughput in the edge device is greater than the radio frequency fingerprint model inference throughput, multiple input branches are added at different depths of the radio frequency fingerprint model, and the radio frequency fingerprint information of the same type at different times within the preset time is removed. Distortion characteristics of receiver, channel noise characteristics. Finally, the vector output by the full connection layer is mapped to a probability value through the Softmax function, the cross-entropy loss function and the SGD optimizer are used to train the neural network, and the trained neural network is used to identify the radio frequency fingerprint. This method is mainly for closed set identification. In the application of real cooperative scene, when a large number of radio frequency devices are accessed, there is a lack of capture and early warning function for unauthorized users, and there is a lack of security protection for communication systems; in the non-cooperative scene, the communication equipment of the non-cooperative party is sensitive to changes, and misjudgment is easy to occur, which is specifically represented as Figure 2

[0013] For example, patent publication number: CN120128285A, patent name: a communication radiation source target countermeasure decision method based on unsupervised learning, which inputs the IQ data features into the ResNet network for closed set individual classification, and outputs the closed set score; Based on the OpenMax algorithm, the closed set score is corrected to obtain the open set score; According to the open set score, the radiation source individual is identified to realize accurate open set identification. It is based on the openMax idea, which first detects the abnormality through the threshold value to determine whether the signal is the same class in the training set, and if it is the same class, the open set identification is realized through closed set identification. This method highly depends on the set threshold value, and cannot get rid of the original closed set identification idea.

[0014] The above two patents have the following defects:

[0015] First, the traditional deep learning-based signal, such as communication, radar, and measurement and control, etc. Radio frequency fingerprint identification mostly focuses on the recognition rate of closed set. In the application of real cooperative scene, when a large number of radio frequency devices are accessed, there is a lack of capture and early warning function for unauthorized users, and there is a lack of security protection for communication systems; in the non-cooperative scene, the communication equipment of the non-cooperative party is sensitive to changes, and misjudgment is easy to occur.

[0016] Second, for the signal, such as communication, radar, and measurement and control, etc. Radio frequency fingerprint open set identification mostly focuses on setting threshold value, setting "abnormal signal detection judgment-closed set signal identification" open set identification process, which belongs to linear process in the process, and has the problem of high identification training delay.

[0017] ​Third, in the image field, the multi-scale based deep learning recognition method, such as face recognition, can only grasp the scattered features of multiple convolution kernel sizes when used in the signal level, and it is difficult to accurately grasp the time sequence characteristics of the signal, which is not conducive to the extraction of the time sequence characteristics of the whole signal. SUMMARY

[0018] In order to solve the above technical problems, the present application provides a multi-scale based open set OFDM communication signal radio frequency fingerprint recognition method, for problem one, the present application distinguishes the signals of known categories and unknown categories through the constructed multi-scale network model, realizes the capture function of unauthorized users when accessing radio frequency equipment, and improves the security protection of the communication system. For problem two, the present application optimizes the (2-N classification) linear open set identification process of “abnormal signal detection and judgment-closed set signal identification” in the prior art to parallel open set identification (2*N) through the preset unknown category in the kernel clustering open set identification module, and realizes the optimization of training time. For problem three, the present application proposes an HSMSSD multi-scale convolution feature extraction module, realizes the control of the signal time sequence characteristics, and makes up for the shortcomings of multi-scale features and extraction modules in the field of signal recognition. The present application improves the open set OFDM communication signal radio frequency fingerprint recognition rate and realizes the security protection of the Internet.

[0019] The object of the present application is realized by the following technical solutions:

[0020] The present application discloses a multi-scale based open set OFDM communication signal radio frequency fingerprint recognition method, which comprises:

[0021] Step one, receiving m USRP transmitters of known categories sending preset modulation type LTE communication signals outdoors; based on the signal-to-noise ratio, generating and adding Gaussian white noise from the power of the LTE communication signal to obtain the OFDM communication signal; storing the OFDM communication signal added with the preset label as a pkl format data set;

[0022] Step two, cutting the data set of m-n categories of transmitters received into a training set, a test set and a validation set according to a preset proportion, and randomly extracting the data set of n categories of transmitters received according to a preset proportion and adding it into the test set; wherein m is the maximum value of the number of categories of transmitters, n is the serial number of the number of categories of transmitters, and n

[0023] Step three, sequentially compose a multi-scale network model by the Fourier frequency domain enhancement preprocessing module, the HSMSSD multi-scale convolution feature extraction module and the core clustering open set identification module; input the training set into the multi-scale network model, obtain the enhanced IQ signal after enhancing the spectral feature by the Fourier frequency domain enhancement preprocessing module, output the feature extraction fusion signal after multi-scale time sequence feature extraction and fusion by the HSMSSD multi-scale convolution feature extraction module, and identify m-n+1 transmitter categories corresponding to the feature extraction fusion signal based on m-n known categories and 1 preset unknown category by the core clustering open set identification module; obtain the trained multi-scale network model;

[0024] Step four, input the test set and the validation set into the trained multi-scale network model for testing and validation, and obtain the identification result.

[0025] In step one, based on the signal-to-noise ratio, the method for obtaining the OFDM communication signal after adding Gaussian white noise to the power of the LTE communication signal is:

[0026] ;

[0027] In the formula, is the signal-to-noise ratio, is the power of the LTE communication signal, is the noise power; the amplitude of the noise is adjusted by the noise power to generate Gaussian white noise, and the Gaussian white noise is added to the LTE communication signal to obtain the OFDM communication signal; wherein,

[0028] The preset modulation type of the LTE communication signal is BPSK, 16QAM and 64QAM; the preset label is a label composed of the number of sampling points, the USRP transmitter and the signal-to-noise ratio.

[0029] In step three, the method for obtaining the enhanced IQ signal after the Fourier frequency domain enhancement preprocessing module enhances the spectral feature is:

[0030] The Fourier frequency domain enhancement preprocessing module is composed of a fast Fourier transform unit FFT, a multi-layer perception unit MLP and an inverse fast Fourier transform unit IFFT;

[0031] The OFDM communication signal is input into the FFT and is decomposed into a real part signal and an imaginary part signal in the frequency domain;

[0032] The MLP enhances the spectral feature of the real part signal and the imaginary part signal based on the hyperbolic tangent activation function, and recombines them into an IQ signal;

[0033] The residual error between the IQ signal and the OFDM communication signal is calculated by the IFFT, and the enhanced IQ signal is output.

[0034] In step three, the method for decomposing the OFDM communication signal into a real part signal and an imaginary part signal in the frequency domain after inputting the OFDM communication signal into the FFT is:

[0035] ;

[0036] wherein, represents the OFDM communication signal represents the signal after the fast Fourier transform, represents represents the real part signal decomposed in the frequency domain after the fast Fourier transform, and j is an imaginary unit, represents represents the imaginary part signal decomposed in the frequency domain after the fast Fourier transform.

[0037] The method for recombining the real part signal and the imaginary part signal into an IQ signal after the MLP enhances the spectral features of the real part signal and the imaginary part signal based on the hyperbolic tangent activation function is:

[0038] ;

[0039] wherein, represents the IQ signal, is a hyperbolic tangent activation function, represents a learnable weight parameter for enhancing the spectral features of the real part signal, represents a learnable weight parameter for enhancing the spectral features of the imaginary part signal.

[0040] The method for calculating the residual error between the IQ signal and the OFDM communication signal through the IFFT and outputting an enhanced IQ signal is:

[0041] ;

[0042] wherein, R is the enhanced IQ signal after the residual error connection, is an inverse fast Fourier transform operation on the IQ signal.

[0043] The method for outputting a feature extraction fusion signal after the multi-scale time series feature extraction and fusion through the HSMSSD multi-scale convolution feature extraction module in step three is:

[0044] ;

[0045] ;

[0046] ;

[0047] wherein, is the feature extraction fusion signal output by the HSMSSD multi-scale convolution feature extraction module, is the projection parameter of the feature extraction fusion signal connecting the hidden state and the output; represents the hidden state function in the HSMSSD multi-scale convolution feature extraction module, u is the hierarchical sequence number, and U is the total number of hierarchical levels, is the weight of the hidden state between different hierarchical levels, is the hidden state equation; represents the full connection layer function, is the sigmoid function, and e is the natural constant.

[0048] In step three, the hidden state equation is:

[0049] ;

[0050] In the formula, represents the sequence number of the scene position, L represents the total number of scene positions, is the specific value of the weighting parameter a at u and , represents the projection of the enhanced IQ signal in the first scene position to the state space of the hierarchical level u, is the expansion quantity at u and , T represents transposition, represents the projection of the enhanced IQ signal to the state space tensor.

[0051] In step three, the core cluster open set identification module identifies m-n+1 transmitter categories corresponding to the feature extraction fusion signal based on m-n known categories and 1 preset unknown category; the method for obtaining the trained multi-scale network model comprises:

[0052] The feature extraction fusion signal is judged based on the preset categories of the transmitters through the Softmax activation function, the preset category that the feature extraction fusion signal conforms to is judged, the prediction probability distribution of the transmitter corresponding to the preset category is output, and the preset category corresponding to the maximum prediction probability is taken as the identification result; wherein the preset categories are m-n+1 preset categories, including m-n known categories marked in the training set and 1 preset unknown category;

[0053] Based on the identification result, the trained multi-scale network model is obtained by jointly training the loss function and the closed set classifier.

[0054] In step three, the calculation method of the prediction probability distribution is:

[0055] ;

[0056] In the formula, the predicted probability distribution of the transmitters corresponding to the preset category to which the feature-extracted fused signal conforms, i is the label sequence number of m-n+1 preset categories, including m-n known categories of transmitter labels and 1 preset unknown category, the probability output by the binary boundary outputer corresponding to i.

[0057] In step three, the method of taking the preset category corresponding to the maximum predicted probability as the recognition result is:

[0058] The predicted probability distribution of the transmitters corresponding to the preset category to which the feature-extracted fused signal conforms is a 2*(m-n+1) dimensional vector;

[0059] The 2*(m-n+1) dimensional vector is uniformly cut into sub-vectors in order, and each sub-vector includes two adjacent vectors;

[0060] The maximum predicted probability of the two adjacent vectors in each sub-vector is recognized, and the m-n+1 preset categories corresponding to the maximum predicted probability are output as the recognition result.

[0061] In step three, the loss function is:

[0062] ; wherein,

[0063] ;

[0064] ;

[0065] ;

[0066] ;

[0067] ;

[0068] In the formula, is the loss function value, is the cross-entropy loss function value, is the negative sample bias loss function value, is the loss function weight, is the entropy minimization loss function value, z is the label sequence number of the real category; N k is the total number of labels of preset categories, is an indicator function, is the closed set classifier for is the predicted probability of the transmitters, is the real probability distribution of the transmitters corresponding to the real category to which the feature-extracted fused signal conforms, is the probability output by the binary boundary outputer corresponding to z.

[0069] The beneficial effects of the present application are:

[0070] 1. Receiving m USRP transmitters of known categories of signals of a preset modulation type of LTE communication signals as I and Q two-way steady-state baseband signals outdoors; based on the signal-to-noise ratio, a Gaussian white noise is generated from the power of the LTE communication signal, and an OFDM communication signal is obtained, realizing the reduction of the signal-to-noise ratio of the control signal, containing a low signal-to-noise ratio of less than 5dB, used to simulate the fading of the analog signal in the communication process; at the same time, in order to control the variable, ensure that different transmitters are not distinguished because of the modulation type, add a preset label to the OFDM communication signal and store it as a pkl format data set; this scheme simulates the real scene of transceiving communication, provides multiple signal-to-noise ratio levels for the data set, enhances the ability to evaluate the robustness of the model, and solves the problem of lack of training data set in real scene transceiving communication;

[0071] 2. The data set of m-n categories of transmitters received is cut into a training set, a test set and a validation set according to a preset proportion, and the data set of n categories of transmitters received is randomly extracted according to a preset proportion and added to the test set; the real scene when the base station receives users without backup can be further simulated, the sample is insufficient, and the real scene does not participate in training;

[0072] 3. The signal of known category and unknown category is distinguished through the constructed multi-scale network model, realizing the capture function of unauthorized users when accessing radio frequency equipment, and improving the security protection of the communication system;

[0073] 4. The Fourier frequency domain enhancement preprocessing module enhances the spectral features, suppresses the multipath effect and Doppler interference, optimizes the spectral enhancement, and thus improves the overall quality of the signal;

[0074] 5. The HSMSSD multi-scale convolution feature extraction module performs multi-scale time sequence feature extraction fusion, which effectively extracts local features by using multi-scale, and further extracts global features by using HSDSSM, separates out subtle hardware defects from the received signal, realizes the control of the time sequence features of the signal, makes up for the shortcomings of the multi-scale feature and extraction module in the field of signal recognition, and effectively improves the overall recognition rate of the model;

[0075] 6、Nuclear cluster open set identification module identifies m-n+1 transmitter categories corresponding to the feature extraction and fusion signal based on m-n known categories and 1 preset unknown category; an m-n classification problem is converted into m-n yes / no problems, m-n backups, and is identified as m-n classified specific equipment; n strangers are identified as 1 classified unknown equipment, the problem of high false alarm rate of open set identification is solved, through the preset unknown category, the (2-N classification) linear open set identification process of "abnormal signal detection and determination-closed set signal identification" in the prior art is optimized to parallel open set identification (2*N), and training time optimization is realized;

[0076] 7、The application compresses the intra-class distance of known categories through a loss function, so that the samples of the known categories are compact in the feature space, and the samples of unknown categories fall outside these compact regions, which is beneficial to open set identification and further improves the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0077] The application will be further described in detail below with reference to the drawings and embodiments.

[0078] Figure 1 It is an application diagram of radio frequency fingerprint open set identification.

[0079] Figure 2 It is a full connection layer open set identification design diagram proposed by the application.

[0080] Figure 3 It is a flow chart of a multi-scale-based open set OFDM communication signal radio frequency fingerprint identification method proposed by the application.

[0081] Figure 4 It is a Fourier frequency domain enhancement preprocessing module diagram proposed by the application.

[0082] Figure 5 It is a HSMSSD multi-scale convolution feature extraction module diagram proposed by the application.

[0083] Figure 6 It is a deployment application environment diagram in the embodiment of the application.

[0084] Figure 7 It is a recognition rate comparison diagram in the embodiment of the application.

[0085] Figure 8 It is a CCB open set identification module confusion matrix at 20dB in the embodiment of the application.

[0086] Figure 9 It is an existing OpenMax open set identification module confusion matrix at 20dB in the embodiment of the application.

[0087] Figure 10is the confusion matrix of the technical solution disclosed in the present application in the example under 20dB.

[0088] Figure 11 is the confusion matrix of the existing OpenMax open set identification module adding the multi-scale HSMSSD feature extraction module of the present application under 20dB. DETAILED DESCRIPTION

[0089] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0090] As shown in Figure 3 The embodiment of the present application provides a multi-scale based open set OFDM communication signal radio frequency fingerprint identification method, which comprises:

[0091] Step one, receiving m USRP transmitters of known categories sending LTE communication signals of a preset modulation type outdoors; based on the signal-to-noise ratio, generating and adding Gaussian white noise from the power of the LTE communication signal to obtain an OFDM communication signal; adding a preset label to the OFDM communication signal and storing it as a data set in pkl format;

[0092] Step two, cutting the data set of m-n categories of transmitters received into a training set, a test set and a validation set according to a preset ratio, and the preferred ratio is 7:2:1; randomly extracting the data set of n categories of transmitters received according to a preset ratio and adding it to the test set, and the preferred ratio is 10% of the data set of n categories of transmitters; wherein m is the maximum number of categories of transmitters, n is the serial number of the number of categories of transmitters, and n

[0093] Step three, sequentially comprising a Fourier frequency domain enhancement preprocessing module, an HSMSSD multi-scale convolution feature extraction module and a kernel clustering open set identification module to form a multi-scale network model; inputting the training set into the multi-scale network model, enhancing the frequency spectrum features by the Fourier frequency domain enhancement preprocessing module to obtain an enhanced IQ signal, performing multi-scale time sequence feature extraction and fusion by the HSMSSD multi-scale convolution feature extraction module to output a feature extraction and fusion signal, and identifying m-n+1 categories of transmitters corresponding to the feature extraction and fusion signal by the kernel clustering open set identification module based on m-n known categories and one preset unknown category; obtaining a trained multi-scale network model; m-n "backups", identified as m-n classified specific devices; n "foreigners", identified as one classified unknown device.

[0094] Step four, input the test set and the validation set into the trained multi-scale network model for testing and validation to obtain the recognition result.

[0095] In step one, based on the signal-to-noise ratio, the method for generating and adding Gaussian white noise to the power of the LTE communication signal to obtain the OFDM communication signal is:

[0096] ;

[0097] In the formula, The signal-to-noise ratio is, The power of the LTE communication signal is, The noise power is; the amplitude of the noise is adjusted by the noise power to generate Gaussian white noise, and the Gaussian white noise is added to the LTE communication signal to obtain the OFDM communication signal, which contains a low signal-to-noise ratio of less than 5dB, which is different from the signal-to-noise ratio of more than 5dB in the prior art, and better simulates the actual signal; wherein,

[0098] The preset modulation type of the LTE communication signal is BPSK, 16QAM, and 64QAM, which is beneficial to extracting the essential features of the transmitter radio frequency fingerprint; the preset label is a label composed of the number of sampling points, the USRP transmitter, and the signal-to-noise ratio.

[0099] As shown in Figure 4 , in step three, the method for enhancing the frequency spectrum characteristics by the Fourier frequency domain enhancement preprocessing module to obtain the enhanced IQ signal is:

[0100] The Fourier frequency domain enhancement preprocessing module is composed of a fast Fourier transform unit FFT, a multilayer perception unit MLP, and an inverse fast Fourier transform unit IFFT;

[0101] The OFDM communication signal is input into the FFT and is decomposed into real part signals and imaginary part signals in the frequency domain;

[0102] After the MLP enhances the frequency spectrum characteristics of the real part signals and the imaginary part signals based on the hyperbolic tangent activation function, the signals are recombined into IQ signals;

[0103] The residual error between the IQ signals and the OFDM communication signals is calculated by the IFFT, and the enhanced IQ signals are output.

[0104] In step three, the method for inputting the OFDM communication signal into the FFT and decomposing it into real part signals and imaginary part signals in the frequency domain is:

[0105] ;

[0106] In the formula, The OFDM communication signal is the signal after fast Fourier transform, represents the real part signal decomposed in the frequency domain after fast Fourier transform, j is the imaginary unit, represents the imaginary part signal decomposed in the frequency domain after fast Fourier transform;

[0107] The method for recombining the IQ signal after the MLP enhances the spectral features of the real part signal and the imaginary part signal based on the hyperbolic tangent activation function is:

[0108] ;

[0109] In the formula, represents the IQ signal, is the hyperbolic tangent activation function, is the hyperbolic tangent activation function value after linear enhancement of the real part signal spectrum, is the hyperbolic tangent activation function value after linear enhancement of the imaginary part signal spectrum, represents the learnable weight parameter for spectral feature enhancement of the real part signal, represents the learnable weight parameter for spectral feature enhancement of the imaginary part signal, wherein, ;

[0110] ;

[0111] In the formula, e is the natural constant;

[0112] The calculation method for outputting the enhanced IQ signal by calculating the residual error between the IQ signal and the OFDM communication signal through IFFT is:

[0113] ;

[0114] In the formula, R is the enhanced IQ signal after residual connection, is the inverse fast Fourier transform operation on the IQ signal.

[0115] As shown in Figure 5 , in step three, the method for outputting the feature extraction fusion signal after multi-scale time series feature extraction and fusion by the HSMSSD multi-scale convolution feature extraction module is:

[0116] ;

[0117] ;

[0118] ;

[0119] In the formula, is the feature extraction fusion signal output by the HSMSSD multi-scale convolution feature extraction module, is the projection parameter connecting the hidden state and the output feature extraction fusion signal; denotes the hidden state function in the HSMSSD multi-scale convolution feature extraction module, u is the hierarchical sequence number, and U is the total number of hierarchical levels, is the weight of the hidden state between different levels, is the hidden state equation; denotes the full connection layer function, is the sigmoid function, denotes using the weighted sum and calculation of of different levels.

[0120] In step three, the hidden state equation is:

[0121] ;

[0122] In the formula, denotes the sequence number of the scene position, L denotes the total number of scene positions, is the specific value of the weighting parameter a at u and , which is used for weighted aggregation of input features, denotes the projection of the enhanced IQ signal in the first scene position to the state space of the level u, is the expansion quantity at u and , and each layer has an independent input matrix to capture features of different scales, and T denotes transposition, denotes the projection of the enhanced IQ signal to the state space tensor.

[0123] In step three, the core cluster open set identification module identifies m-n+1 transmitter categories corresponding to the feature extraction fusion signal based on m-n known categories and 1 preset unknown category; the method for obtaining the trained multi-scale network model comprises:

[0124] The feature extraction fusion signal is judged based on the preset categories of the transmitters through the Softmax activation function, the preset category to which the feature extraction fusion signal conforms is judged, the prediction probability distribution of the transmitter corresponding to the preset category is output, and the preset category corresponding to the maximum prediction probability is taken as the identification result; wherein the preset categories are m-n+1 preset categories, including m-n known categories marked in the training set and 1 preset unknown category;

[0125] Based on the identification result, the trained multi-scale network model is obtained by joint training of the loss function and the closed set classifier.

[0126] In step three, the calculation method of the predicted probability distribution is:

[0127] ;

[0128] In the formula, is the predicted probability distribution of the transmitter corresponding to the preset category of the feature extraction fusion signal, i is the label sequence number of m-n+1 preset categories, including m-n transmitter marked known categories and 1 preset unknown category, whether to judge is carried out through the threshold value of the known category and the preset unknown category, and if all m-n categories output are "no", it is determined as "stranger"; is the probability output by the binary boundary outputer corresponding to i.

[0129] In step three, the method for taking the preset category corresponding to the maximum predicted probability as the recognition result is:

[0130] The predicted probability distribution of the transmitter corresponding to the preset category of the feature extraction fusion signal is a 2*(m-n+1) dimensional vector;

[0131] The 2*(m-n+1) dimensional vector is uniformly cut into sub-vectors in order, and each sub-vector includes two adjacent vectors;

[0132] The maximum predicted probability of the two adjacent vectors in each sub-vector is recognized, and the m-n+1 preset categories corresponding to the maximum predicted probability are output as the recognition result.

[0133] The 2*(m-n+1) outputs of the application further weaken the correlation between each other, so that although the output vector dimension increases, it brings a certain degree of network calculation complexity rise; but can reduce the false alarm rate to a certain extent.

[0134] In step three, the loss function is:

[0135] ;

[0136] The loss function is jointly trained with the closed set classifier, forcing the classifier to focus on the inter-class boundary features, thereby expanding the distance between the decision regions of different categories and avoiding clustering overlap.

[0137] ;

[0138] The cross-entropy loss function constructed by the application ensures that the known categories can be accurately classified.

[0139] ;

[0140] The first term of the negative sample bias loss function constructed by the application Forcing the probability of the real class to be close to 1; The minimum "negative sample probability" representing all non-real classes. The second term Requires the classifier to correctly reject the most difficult negative samples, avoids the decision boundary tilting towards the negative samples, and solves the negative sample bias problem in traditional binary classification.

[0141] ;

[0142] The entropy minimization loss function constructed in the application forces The prediction probability distribution to be more concentrated, enhancing the distinguishability of the features of the known classes.

[0143] ;

[0144] ;

[0145] In the formula, Loss function value, Cross-entropy loss function value, Negative sample bias loss function value, Loss function weight, Entropy minimization loss function value, z is the label serial number of the real class; N k Total number of preset labels, Indicator function, Closed set classifier for The transmitter transmits the prediction probability, The real probability distribution of the transmitter corresponding to the real class conforming to the feature extraction fusion signal, The probability output by the binary boundary output device corresponding to z.

[0146] In order to explain the technical solutions of the application in detail, a specific example is provided:

[0147] As shown in Figure 6 The data set in the verification example of the application collects data sets of 12 USRP transmitter radio frequency devices, sets the labels of 10 devices as known classes (device1-device10), and the remaining device11, 12 as unknown classes, that is, m=12, n=2; input the known 10 classes into the multi-scale network model, train through spectrum enhancement and time sequence feature extraction, and then assist the reverse propagation through the designed loss function to make the classes more compact, and then input all 12 classes for testing to obtain the recognition result as shown in Figure 7 to Figure 11 .

[0148] Among them, Figure 7In the figure, red is the multi-scale network model proposed by the application, Fourier frequency domain enhancement preprocessing module MSF, HSMSSD multi-scale convolution feature extraction module HSMSSD and core clustering open set identification module CCB (MSF+ HSMSSD+ CCB), blue is the Fourier frequency domain enhancement preprocessing module, HSMSSD multi-scale convolution feature extraction module and traditional classification module OpenMax of the application (spectrum enhancement+ multi-scale HSMSSD+ OpenMax); green and dark purple are to remove HSMSSD, only keep spectrum enhancement and multi-scale OpenMax and CCB, pink is the result obtained by the core clustering open set identification module CCB, and the rest are prior art.

[0149] Figure 8 to Figure 11 For the specific effect of the identification performed at 20dB, it can be found that after adding multi-scale HSMSSD (MSF-HSMSSD), the application performs well in recognition rate, can assist the open set identification module to identify by further improving the effectiveness of signal feature extraction, thereby effectively improving the decision accuracy of the model for unknown device access and reducing the false alarm rate.

[0150] In summary, the radio frequency fingerprint identification of the communication signal can not only be used in non-cooperative scenarios, but also be used in cooperative communication with high safety coefficient requirements and receiver sensitivity, and can achieve spectrum sensing to a certain extent. For example, in the traditional LTE communication system, the LTE sending end can perform initial access request through the PRACH channel (Physical Random Access CHannel) to perform access request. However, in non-cooperative communication, the receiving end faces two problems, so it cannot determine whether the signal is a malicious attack by using PRACH.

[0151] The beneficial effects of the embodiments of the application are:

[0152] 1. In outdoor reception, m USRP transmitters of known categories are used to send preset modulation type LTE communication signals as I and Q two-way stable baseband signals; based on the signal-to-noise ratio, Gaussian white noise is generated from the power of the LTE communication signal to obtain an OFDM communication signal, the signal-to-noise ratio of the control signal is reduced, and a low signal-to-noise ratio less than 5dB is contained to simulate signal fading in the communication process; at the same time, in order to control the variable and ensure that different transmitters are not distinguished because of the modulation type, the OFDM communication signal is added with a preset label and stored as a pkl format data set; this scheme simulates real scene transceiving communication, provides multiple signal-to-noise ratio levels for the data set, enhances the model robustness evaluation ability, and solves the problem of lack of training data set in real scene transceiving communication;

[0153] 2. The received datasets of transmitters in mn categories are divided into training, testing, and validation sets according to a preset ratio. The received datasets of transmitters in n categories are randomly selected according to a preset ratio and added to the testing set. This can further simulate the real-world scenario where the base station receives data from unbacked-up users, resulting in insufficient samples and no training participation.

[0154] 3. By constructing a multi-scale network model, signals of known and unknown categories are distinguished, enabling the capture of unauthorized users when accessing radio frequency devices, thereby improving the security protection of the communication system;

[0155] 4. The Fourier frequency domain enhancement preprocessing module enhances the spectral characteristics, suppresses multipath effects and Doppler interference, optimizes spectral enhancement, and thus improves the overall signal quality.

[0156] 5. The HSMSSD multi-scale convolutional feature extraction module performs multi-scale temporal feature extraction and fusion. By effectively extracting local features at multiple scales, and then further extracting global features using HSMSSD, subtle hardware defects are separated from the received signal, enabling control over the temporal features of the signal. This compensates for the shortcomings of the multi-scale feature extraction module in the field of signal recognition and effectively improves the overall recognition rate of the model.

[0157] 6. The kernel clustering open set recognition module identifies m-n+1 transmitter categories corresponding to the feature extraction and fusion signal based on mn known categories and 1 preset unknown category. It transforms an mn classification problem into mn yes / no questions, with mn backups identified as specific devices of the mn categories and n outsiders identified as 1 device of the unknown category. This solves the problem of high false alarm rate in open set recognition. By pre-setting the unknown category, it optimizes the linear open set recognition process of (2-N classification) in the existing technology of "abnormal signal detection and judgment - closed set signal recognition" into parallel open set recognition (2*N), thereby optimizing the training time.

[0158] 7. This invention uses a loss function to compress the intra-category spacing of known categories, making the samples of known categories compact in the feature space, while the samples of unknown categories will fall outside these compact regions, which is beneficial for open set recognition and further improves accuracy.

[0159] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-scale based open set OFDM communication signal radio frequency fingerprinting method, characterized in that, The method comprises: Step one, receiving m USRP transmitters of known categories transmitting preset modulation type LTE communication signals outdoors; based on the signal-to-noise ratio, generating and adding Gaussian white noise to the power of the LTE communication signals to obtain OFDM communication signals; adding preset labels to the OFDM communication signals and storing the data sets in pkl format; Step two, cutting the data sets of m-n categories of transmitters into a training set, a test set and a validation set according to a preset proportion, and randomly extracting the data sets of n categories of transmitters and adding them to the test set according to a preset proportion; wherein m is the maximum number of categories of transmitters, n is the serial number of the number of categories of transmitters, and n < m; Step three, sequentially comprising a Fourier frequency domain enhancement preprocessing module, an HSMSSD multi-scale convolution feature extraction module and a kernel clustering open set recognition module to form a multi-scale network model; inputting the training set into the multi-scale network model, enhancing the spectral features by the Fourier frequency domain enhancement preprocessing module to obtain enhanced IQ signals, performing multi-scale time sequence feature extraction and fusion by the HSMSSD multi-scale convolution feature extraction module to output feature extraction and fusion signals, and recognizing m-n+1 categories of transmitters corresponding to the feature extraction and fusion signals based on m-n known categories and one preset unknown category by the kernel clustering open set recognition module; obtaining the trained multi-scale network model; Step four, inputting the test set and the validation set into the trained multi-scale network model for testing and verification to obtain the recognition result.

2. The method of claim 1, wherein, In step one, the method for generating and adding Gaussian white noise to the power of the LTE communication signals to obtain the OFDM communication signals based on the signal-to-noise ratio is: ； In the formula, S / N is a signal-to-noise ratio, P is a power of the LTE communication signal, N is a noise power; a Gaussian white noise is generated by adjusting the amplitude of the noise according to the noise power, and the Gaussian white noise is added to the LTE communication signal to obtain an OFDM communication signal; wherein, The preset modulation type of the LTE communication signals is BPSK, 16QAM or 64QAM; and the preset label is a label composed of the number of sampling points, the USRP transmitter and the signal-to-noise ratio.

3. The method of claim 2, wherein, In step three, the method for obtaining the enhanced IQ signals by the Fourier frequency domain enhancement preprocessing module is: The Fourier frequency domain enhancement preprocessing module comprises a fast Fourier transform unit FFT, a multi-layer perception unit MLP and an inverse fast Fourier transform unit IFFT; The OFDM communication signals are input into the FFT and are decomposed into real part signals and imaginary part signals in the frequency domain; The MLP enhances the spectral features of the real part signals and the imaginary part signals based on the hyperbolic tangent activation function and recombines them into IQ signals; The residual error between the IQ signals and the OFDM communication signals is calculated by the IFFT to output the enhanced IQ signals.

4. The method of claim 3, wherein, In step three, the method for inputting the OFDM communication signals into the FFT and decomposing them into real part signals and imaginary part signals in the frequency domain is: ； wherein represents an OFDM communication signal a signal after fast Fourier transform, represents a real part signal decomposed in a frequency domain after fast Fourier transform, j is an imaginary unit, represents an imaginary part signal decomposed in a frequency domain after fast Fourier transform; The method for recombining the real part signals and the imaginary part signals into IQ signals after the MLP enhances the spectral features of the real part signals and the imaginary part signals based on the hyperbolic tangent activation function is: ； wherein represents an IQ signal, is a hyperbolic tangent activation function, represents a learnable weight parameter for spectral feature enhancement of the real part signal, represents a learnable weight parameter for spectral feature enhancement of the imaginary part signal; The calculation method for outputting the enhanced IQ signals by calculating the residual error between the IQ signals and the OFDM communication signals by the IFFT is: ； where R is the residual connected enhanced IQ signal, is an inverse fast Fourier transform operation on the IQ signal.

5. The method of claim 4, wherein, In step three, the method for outputting the feature extraction and fusion signals after the HSMSSD multi-scale convolution feature extraction module performs multi-scale time sequence feature extraction and fusion is: ； ； ； In the formula, is a feature extraction fusion signal output by the HSMSSD multi-scale convolution feature extraction module, is a projection parameter for connecting the hidden state and the feature extraction fusion signal, represents a hidden state function in the HSMSSD multi-scale convolution feature extraction module, u is a layer number, and U is a total number of layers, is a weight of the hidden state between different layers, is a hidden state equation; represents a full connection layer function, is a sigmoid function, and e is a natural constant.

6. The method of claim 5, wherein, In step three, the hidden state equation is: ； wherein L is the total number of scene positions, is a specific value of the weighting parameter a at u and is a specific value of the weighting parameter a at u and denotes the projection of the enhanced IQ signal in the th scene position into the state space of level u, is the state space tensor of level u, is the extension of the state space tensor of level u at u and is the extension of the state space tensor of level u at u and T denotes the transpose, denotes the projection of the enhanced IQ signal into the state space tensor.

7. The method of claim 6, wherein, In step three, the m-n known classes and 1 preset unknown class are used by the core cluster open set identification module to identify m-n+1 transmitter classes corresponding to the feature extraction fusion signal; The method for obtaining the trained multi-scale network model comprises: The feature extraction fusion signal is judged based on the preset classes of the transmitters through a Softmax activation function, the preset class to which the feature extraction fusion signal conforms is judged, a prediction probability distribution of the transmitter corresponding to the preset class is output, and the preset class corresponding to the maximum prediction probability is taken as the identification result; wherein the preset classes are m-n+1 preset classes, including m-n known classes marked in the training set and 1 preset unknown class; Based on the identification result, the trained multi-scale network model is obtained by joint training of a loss function and a closed set classifier.

8. The method of claim 7, wherein, In step three, the calculation method of the prediction probability distribution is: ； In the formula, is the predicted probability distribution of the transmitter corresponding to the preset category to which the feature extraction fusion signal conforms, i is the label serial number of m-n+1 preset categories, including m-n known categories marked by the transmitter and 1 preset unknown category, is the probability output by the binary boundary output device corresponding to i.

9. The method of claim 8, wherein, In step three, the method for taking the preset class corresponding to the maximum prediction probability as the identification result is: The prediction probability distribution of the transmitter corresponding to the preset class to which the feature extraction fusion signal conforms is a 2*(m-n+1) dimensional vector; The 2*(m-n+1) dimensional vector is evenly cut into sub-vectors in order, and each sub-vector includes two adjacent vectors; The maximum prediction probabilities of the two adjacent vectors in each sub-vector are identified, and the m-n+1 preset classes corresponding to the maximum prediction probabilities are output as the identification result.

10. The method of claim 9, wherein, In step three, the loss function is: ; wherein, ； ； ； ； ； In the formula, is a loss function value, is a cross-entropy loss function value, is a negative sample bias loss function value, is a loss function weight, is an entropy minimization loss function value, z is a label number of a real category; N k is a total number of labels of a preset category, is an indicator function, is a closed set classifier pair is a predicted probability emitted by a transmitter, is a real probability distribution of a transmitter corresponding to a real category of a feature extraction fusion signal, is a probability output by a binary boundary output device corresponding to z.

Images