A radio frequency fingerprint recognition method based on LTE DMRS

Abstract

This invention discloses a radio frequency fingerprinting method based on LTE DMRS, comprising: (1) acquiring the DMRS transmitted by the terminal on the LTE physical uplink shared channel and demodulating it into an I / Q complex sequence; (2) extracting DMRS features to extract the first-order and second-order feature complex sequences of the DMRS sequence; (3) reconstructing the DMRS features to convert them into amplitude and phase feature reconstruction term sequences as DMRS radio frequency fingerprint features; (4) feeding the DMRS radio frequency fingerprint features into a multi-input convolutional neural network for training and saving the trained network as a DMRS radio frequency fingerprinting network; (5) when a DMRS to be identified is received, extracting its DMRS radio frequency fingerprint features and using the trained DMRS radio frequency fingerprinting network to identify the radio frequency fingerprint features. This invention is applicable to the uplink of the LTE system, realizes terminal identification based on LTE DMRS, and has a high identification accuracy and is easy to implement.

Description

Technical Field

[0001] This invention relates to a communication radiation source identification technology, belonging to the field of wireless communication technology, and particularly to radio frequency fingerprinting technology for LTE DMRS (Demodulation Reference Signal, DMRS). Background Technology

[0002] With the widespread application of wireless communication technology, communication between users is no longer limited by time and space, enabling faster and more convenient free communication and information resource sharing. Users here include individuals accessing systems or using wireless resources, as well as electronic devices. However, while wireless communication is driving rapid development across various industries, its openness also brings significant security risks. Users can access the network from anywhere with sufficient signal strength, making it impossible to eliminate potential security vulnerabilities through physical isolation. This makes them more susceptible to attacks such as illegal eavesdropping, information tampering, impersonation, and replay attacks. Furthermore, most commonly used security authentication methods are based on passwords and protocols, making them vulnerable to cracking.

[0003] Based on the information source used for identification, methods for individual terminal identification can be broadly categorized into three types: First, based on the physical carrier possessed by the user or device, such as a smart card; second, based on the user's or device's digital identity information, such as a username and password; and third, based on the user's or device's physical characteristics, such as the physical fingerprint of a radio frequency device. The first method requires a certain spatial distance between the identification device and the smart card being identified, and is not widely applicable to various mobile communication scenarios. The second method is based on password or protocol-based authentication and identification, but this requires a robust key management framework, which is costly, and because the user's authentication information or password is primarily purely digital, it is easily leaked or cracked, posing certain security risks. The third method is based on physical characteristics, which traditionally refer to features such as fingerprints, irises, and vein patterns in humans, and for devices, the inherent physical characteristics of the hardware. These characteristics, due to their uniqueness and inherent nature, are difficult to crack and can efficiently distinguish wireless communication devices. This inherent hardware physical characteristic of the device is called radio frequency fingerprinting (RFF).

[0004] Radio frequency fingerprints (RFFs) are typically generated during the fabrication and production of radio frequency (RF) devices. They are inherent in hardware manufacturing and involve a degree of randomness. These inherent minute deviations in electronic devices, carried on the signal, are collected and analyzed by the receiver with sufficient precision, generating a unique RF fingerprint feature that can be used for identification. RFFs utilize features carried in wireless signals that reflect inherent defects in the transmitter's hardware to uniquely identify the transmitter, offering advantages such as difficulty in cloning and forgery. Therefore, RF fingerprinting technology can stand independently of existing cryptographic mechanisms, serving as a powerful supplement to existing device identification mechanisms, and possesses broad development prospects and enormous application potential.

[0005] Furthermore, in the LTE Physical Uplink Shared Channel (PUSCH), the DMRS occupies the fourth SC-FDMA symbol in each uplink slot under normal cyclic prefix (CP) and the third SC-FDMA symbol under extended CP. The DMRS is transmitted on every RB scheduled by the base station for the user and occupies the same symbol position; therefore, the sequence length of the DMRS is equal to the number of subcarriers allocated to the UE for PUSCH transmission. The regular transmission of the DMRS and the one-to-one correspondence between the DMRS and the RBs allocated to each terminal are beneficial for the extraction and classification of the terminal's radio frequency fingerprint features. Summary of the Invention

[0006] Objective: This invention addresses existing LTE physical layer security issues by providing a radio frequency fingerprinting method based on LTE DMRS. This method processes and extracts features from the DMRS transmitted by the terminal on the PUSCH, using this as the terminal's DMRS radio frequency fingerprint. A multi-input convolutional neural network is trained based on the obtained DMRS radio frequency fingerprint features to achieve terminal identification and authentication.

[0007] Technical solution: The LTE DMRS-based radio frequency fingerprint recognition method of the present invention includes the following steps:

[0008] 1) Receive the DMRS transmitted by the terminal on the LTE PUSCH and demodulate it using I / Q symbols. Preprocess the demodulated DMRS sequence to extract the first-order and second-order feature complex sequences of the DMRS.

[0009] Furthermore, step 1 includes the following steps:

[0010] 1.1 For the received DMRS time-domain signal, perform CP removal, down-conversion, FFT transformation, etc., demodulate to I / Q symbols, and obtain the DMRS complex sequence;

[0011] 1.2 The DMRS complex sequence is transformed into a first-order characteristic complex sequence of DMRS with ideal properties such as constant modulus and linear phase by performing the division operation between the preceding and following terms;

[0012] 1.3 The first-order characteristic complex sequence of DMRS is divided again by the terms before and after to obtain the second-order characteristic complex sequence of DMRS with an ideal fixed phase. The phase of the second-order complex sequence represents the phase change step size of the first-order characteristic complex sequence;

[0013] 2) Extract the amplitude reconstruction feature term and phase reconstruction feature term of DMRS from the first-order and second-order feature complex sequences of DMRS, respectively, as DMRS radio frequency fingerprint features.

[0014] Furthermore, step 2 includes the following steps:

[0015] 2.1 The first-order characteristic complex sequence of DMRS is processed as follows: the phase and amplitude of the first-order characteristic complex sequence of DMRS are transformed; the sequence is reordered according to the phase; and a fixed-point difference reconstruction is performed. The result is a sequence with dimension 1×N. F The real sequence is denoted as the amplitude reconstruction feature term of DMRS. F This indicates the desired size of the reconstructed dimension.

[0016] 2.2 The second-order characteristic complex sequence of DMRS is processed as follows: the phase of the second-order characteristic complex sequence of DMRS is formally transformed; the phases of the first-order characteristic complex sequence are reordered; and a fixed-point difference reconstruction is performed. The resulting sequence has a dimension of 1×N. F The real sequence is denoted as the phase reconstruction feature term of DMRS.

[0017] 3) The extracted DMRS RFID fingerprint features from multiple terminals are fed into a multi-input convolutional neural network for training. The trained neural network is then stored as a DMRS RFID fingerprint recognition network model.

[0018] Furthermore, step 3 includes the following steps:

[0019] 3.1 Construction of a Multi-Input Convolutional Neural Network. The network is trained using a multi-input convolutional neural network to recognize the extracted amplitude reconstruction features and phase reconstruction features of the DMRS. The two sets of features are fed into two different input branches of the multi-input convolutional neural network. Each branch contains 5 convolutional layers and 2 attention mechanism modules. The features extracted by the branch networks are concatenated and then fed into the backbone network, which contains 6 convolutional layers, for radio frequency fingerprint recognition.

[0020] 3.2 The constructed majority-input convolutional neural network is used to train the network to recognize the extracted DMRS amplitude reconstruction features and phase reconstruction features from different terminals. The two sets of reconstruction features are represented by Msyms×N. F The sample input dimension is fed into two different input branches of the multi-factor convolutional neural network for training. Here, syms represents the SC-FDMA symbol unit; M represents the number of symbols in the DMRS reconstructed feature terms contained in a sample.

[0021] 3.3 The network was trained according to different numbers of input symbols M, and the trained network was stored as a DMRS radio frequency fingerprinting network model. In the network training, the network model stored when M=10 symbols was used as the pre-training model of the network. The network was trained for other values ​​of M to significantly reduce the training time of the network.

[0022] 4) When a terminal transmits an LTE DMRS that requires identification and authentication, the DMRS radio frequency fingerprint features are extracted based on the received DMRS, sent to the DMRS radio frequency fingerprint identification network model for identification, and the identification result is returned to realize individual terminal identification.

[0023] Furthermore, step 4 includes the following steps:

[0024] 4.1 For the received DMRS of a certain terminal, extract the amplitude / phase reconstruction feature of the DMRS as the radio frequency fingerprint feature of the DMRS.

[0025] 4.2 The radio frequency fingerprint features of the DMRS signal are fed into the trained and saved DMRS radio frequency fingerprint recognition network model for recognition, and the recognition results are returned to realize the identification of individual terminals.

[0026] Beneficial Effects: This invention combines signal processing and deep learning methods to achieve efficient radio frequency fingerprint recognition based on LTE DMRS. The DMRS used in this method corresponds to the subcarriers occupied by each terminal in the frequency domain and occupies a fixed position of OFDM symbols in the time domain, making it easier to acquire and extract features, and offering good versatility for LTE system terminals. The method first performs signal processing and feature extraction on the acquired DMRS to remove some interference features and enhance useful features. Then, a deep learning convolutional neural network is used for further radio frequency fingerprint feature extraction and terminal device identification to improve recognition accuracy. The combination of DMRS signal processing and deep learning ensures high recognition accuracy even at low signal-to-noise ratios. Furthermore, this method can identify DMRS with different symbol lengths, allowing for the selection of different symbol numbers M based on actual real-time performance and recognition rate requirements. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the present invention;

[0028] Figure 2 This is a schematic diagram of the process for extracting DMRS radio frequency fingerprint features from DMRS in this invention;

[0029] Figure 3 This is a schematic diagram of the architecture of a multi-input convolutional neural network;

[0030] Figure 4 This is a schematic diagram of radio frequency fingerprint extraction and individual identification of the received DMRS to be identified;

[0031] Figure 5 This is an example diagram of a constellation diagram of the first-order characteristic sequence of DMRS;

[0032] Figure 6 This is an example diagram of a constellation diagram of second-order characteristic sequences of DMRS;

[0033] Figure 7 This is a schematic diagram of the experimental results based on the method of this invention, where noise is artificially added to the measured data. Detailed Implementation

[0034] This embodiment discloses a radio frequency fingerprint recognition method based on LTE DMRS, such as... Figure 1 As shown, it includes the following steps:

[0035] 1) The DMRS transmitted by the receiving terminal on the LTEPUSCH is demodulated using I / Q symbols. The demodulated DMRS sequence is preprocessed to extract the first-order feature complex sequence a. 1st_DMRS and second-order characteristic complex sequence a 2nd_DMRS .

[0036] The specific extraction steps are as follows: Figure 2 The DMRS signal acquisition section and DMRS feature extraction section are shown in the figure, including the following steps:

[0037] Step 1.1 as follows Figure 2 The DMRS signal acquisition section is shown in the diagram. First, the uplink signal transmitted by the terminal on the PUSCH is acquired, and the DMRS time-domain frequency band signal is separated. Then, the acquired DMRS digital time-domain signal undergoes operations such as CP removal, digital down-conversion, and FFT to demodulate it to I / Q symbols, resulting in a length of [missing information]. DMRS complex sequence: in The number of subcarriers allocated to this terminal.

[0038] Step 1.2 as follows Figure 2As shown in the DMRS feature extraction section, the first-order feature complex sequence of DMRS is calculated: in

[0039]

[0040] Step 1.3 as follows Figure 2 As shown in the DMRS feature extraction section, the second-order feature complex sequence of DMRS is calculated.

[0041] in

[0042]

[0043] Figure 5 The image shown is an example of a constellation diagram of a first-order feature complex sequence of DMRS. Figure 6 The image shown is an example of a constellation diagram of second-order characteristic complex sequences in DMRS.

[0044] 2) Extract the amplitude reconstruction feature term (feat) of DMRS from the first-order and second-order feature complex sequences of DMRS respectively. DMRS_am Phase reconstruction feature terms of DMRS DMRS_angle As a DMRS radio frequency fingerprint feature.

[0045] The specific extraction steps are as follows: Figure 2 The DMRS feature reconstruction section, as shown in the figure, includes the following steps:

[0046] Step 2.1 Formal Transformation (First-Order Terms): Transform the term a obtained in Step 1... 1st_DMRS After formal transformation, we obtain the real number vector group A. am_DMRS ={arg(a 1st_DMRS ), am(a 1st_DMRS )}. Where arg(x) represents taking the phase of each complex number in the complex vector x, and the dimension after the operation is the same as that of x; am(x) represents taking the magnitude of each complex number in the complex vector x, and the dimension after the operation is the same as that of x.

[0047] Step 2.2 Formal Transformation (Second-Order Terms): Transform the term a obtained in Step 1... 1st_DMRS After formal transformation, we obtain the real number vector group A. angle_DMRS The conversion steps are as follows:

[0048]

[0049] The mod operation represents the remainder operation.

[0050] Step 2.3 Reordering (First-order terms): Calculate column vector group A sort_am =sortrows(A am_DMRS,1), where sortrows(X,1) means sorting the column vector group X in ascending order according to the value of the first column.

[0051] Step 2.4 Reordering (Second-order terms): Calculate column vector group A sort_angle =sortrows(A angle_DMRS ,1).

[0052] Step 2.5 Calculate the interpolation coordinates: Calculate the interpolation coordinate vector z = (z(0), ..., z(N)). F -1)) T Where z(n) = -π + 2π(n + 0.5) / N F ,0≤n≤N F -1. Where N is in the formula. F This indicates the desired size of the reconstructed dimension.

[0053] Step 2.6 Interpolation Reconstruction (First-Order Term): Using z(n) as the new coordinate node, perform interpolation and calculate the DMRS amplitude reconstruction feature term. Its dimension is N F ×1. Wherein, This indicates that y is linearly interpolated using x as the original coordinate and z as the new coordinate.

[0054] Step 2.7 Interpolation Reconstruction (Second-Order Term): Using z(n) as the new coordinate node, perform interpolation and calculate the DMRS phase reconstruction feature term. Its dimension is N F ×1.

[0055] 3) The extracted DMRS RFID fingerprint features from multiple terminals are fed into a multi-input convolutional neural network for training. The trained neural network is then stored as a DMRS RFID fingerprint recognition network model.

[0056] The constructed multi-input convolutional neural network, such as Figure 3 As shown in the diagram. In the diagram, x1 represents a single-branch path, x2 represents a double-branch path; Conv represents a convolutional layer, each convolutional layer is followed by a ReLU activation layer by default, Maxpooling represents a max pooling layer, Concatenate represents a connection layer, Dropout represents a dropout layer, Flatten represents data flattening, and Dense represents a fully connected layer; Channel attention represents a channel attention mechanism layer, and Spatial attention represents a spatial attention mechanism layer. The numbers after the layer names indicate the layer number for the same type.

[0057] The training and storage of a network includes the following steps:

[0058] Step 3.1 Constructing a Multi-Input Convolutional Neural Network. The network is trained using a multi-input convolutional neural network to recognize the extracted amplitude reconstruction features and phase reconstruction features of the DMRS. The two sets of features are fed into two different input branches of the multi-input convolutional neural network. Each branch contains 5 convolutional layers and 2 attention mechanism modules. The features extracted by the branch networks are concatenated and then fed into the backbone network containing 6 convolutional layers for radio frequency fingerprint recognition.

[0059] Step 3.2 uses the constructed majority-input convolutional neural network to train the network to recognize the extracted DMRS amplitude / phase reconstruction features from different terminals. The two sets of features are represented by Msyms×N. F The sample input dimension is fed into two different input branches of the multi-factor convolutional neural network for training. Here, syms represents the SC-FDMA symbol unit; M represents the number of symbols in the DMRS reconstructed feature terms contained in a sample.

[0060] Step 3.3 Train the network according to different input symbol numbers M, and store the trained network as a DMRS radio frequency fingerprint recognition network model. In the network training, the network model stored when M=10 symbols is used as the pre-trained model of the network. The network is trained for other values ​​of M to significantly reduce the training time of the network.

[0061] 4) When a terminal transmits an LTE DMRS that requires identification and authentication, the DMRS radio frequency fingerprint features are extracted based on the collected DMRS, sent to the DMRS radio frequency fingerprint recognition network model for identification, and the identification result is returned to realize individual terminal identification.

[0062] The steps for RF fingerprint extraction and individual identification of the received DMRS to be identified are as follows: Figure 4 The steps shown are as follows:

[0063] Step 4.1 Collect the DMRS of the terminal to be identified, according to Figure 2 The steps are to extract the amplitude / phase reconstruction features of the DMRS as the radio frequency fingerprint features of the DMRS.

[0064] Step 4.2 The radio frequency fingerprint features of the DMRS signal are fed into the trained and saved DMRS radio frequency fingerprint recognition network model for recognition, and the recognition result is returned to realize the individual terminal identification.

[0065] The method of this invention is then tested in a practical system, specifically on 15 terminals from 5 different brands (Vivo, Honor, Huawei, Xiaomi, and Iqoo). The experiment first collected DMRS signals transmitted by the 15 terminals, totaling 57,880 syms (SC-FDMA symbols). Half of these signal samples were used as the training set, and DMRS amplitude / phase reconstruction features were extracted as DMRS RF fingerprints. These DMRS RF fingerprints were then fed into a multi-input convolutional neural network for training with different numbers of input symbols. The trained network was saved as a DMRS RF fingerprint recognition network. The other half of the collected signal samples were used as the DMRS to be identified, i.e., the test set. DMRS RF fingerprints were extracted from these samples and fed into the trained and saved DMRS RF fingerprint recognition network for identification, outputting the terminal's identification result.

[0066] During the experiment, the batch size used in network training was 64, the initial learning rate was 0.001, and the optimizer was Adam. The signal-to-noise ratio estimated by the collected DMRS was approximately 18.46 dB. Different levels of Gaussian white noise were artificially added to the signal to test the robustness of the proposed method. The experimental results are as follows: Figure 7 As shown. When the number of input symbols M = 1 sym, the method of the present invention achieves a recognition accuracy of 77.1% on 15 terminals of the original signal dataset. The recognition accuracy is 94.1% when M = 10 syms. When the number of input symbols M = 20 syms, the method of the present invention still achieves an accuracy of 84.9% at SNR = 5dB; and when M = 40 syms, the 5dB recognition rate still reaches 94.6%.

[0067] The above-disclosed embodiments are merely preferred examples of the present invention and should not be construed as limiting the scope of the invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A radio frequency fingerprint recognition method based on LTE DMRS, characterized in that: (1) The receiving terminal demodulates the DMRS transmitted on the LTE physical uplink shared channel with I / Q symbols; the DMRS sequence obtained after demodulation is subjected to DMRS feature extraction, which includes: dividing the DMRS sequence by terms to convert it into a first-order complex DMRS feature sequence with constant modulus and linear phase characteristics; dividing the first-order complex DMRS feature sequence by terms again to obtain a second-order complex DMRS feature sequence with ideal fixed phase. (2) Communication reconstruction of DMRS features; the communication reconstruction includes: extracting the amplitude reconstruction feature term of DMRS from the first-order feature complex sequence of DMRS, and extracting the phase reconstruction feature term of DMRS from the second-order feature complex sequence of DMRS, as DMRS radio frequency fingerprint features; (3) The extracted DMRS radio frequency fingerprint features from multiple terminals are fed into a multi-input convolutional neural network for training; the trained neural network is stored as a DMRS radio frequency fingerprint recognition network model. (4) When a terminal transmits an LTE DMRS that needs to be identified and authenticated, the DMRS radio frequency fingerprint features are extracted based on the received DMRS, sent to the DMRS radio frequency fingerprint identification network model for identification, and the identification result is returned to realize terminal identification.

2. The LTE DMRS-based radio frequency fingerprint recognition method according to claim 1, characterized in that: The DMRS radio frequency fingerprint feature extraction step described in step (2) includes: The first-order characteristic complex sequence of DMRS is processed as follows: the phase and amplitude of the first-order characteristic complex sequence of DMRS are transformed; the sequence is reordered according to the phase; a fixed-point difference reconstruction is performed; the result is a sequence with dimension 1×N. F The real sequence is denoted as the amplitude reconstruction feature term of DMRS, where N F Indicates the expected size of the reconstructed dimension; The second-order characteristic complex sequence of DMRS is processed sequentially as follows: the phase of the second-order characteristic complex sequence of DMRS is formally transformed; the phases of the first-order characteristic complex sequence are reordered; a fixed-point difference reconstruction is performed; the resulting dimension is 1×N. F The real sequence is denoted as the phase reconstruction characteristic term of DMRS.

3. The LTE DMRS-based radio frequency fingerprint recognition method according to claim 1, characterized in that: The architecture of the DMRS radio frequency fingerprint recognition network described in step (3) includes: using a multi-input convolutional neural network to train the extracted amplitude reconstruction feature terms and phase reconstruction feature terms of the DMRS for network recognition; the two sets of feature terms are respectively fed into two different input branches of the multi-input convolutional neural network; the two network branches each contain 5 convolutional layers and 2 attention mechanism modules; the features extracted by the branch networks are concatenated and then fed into the backbone network containing 6 convolutional layers for radio frequency fingerprint recognition.

4. The LTE DMRS-based radio frequency fingerprint recognition method according to claim 1, characterized in that: The training and storage steps of the DMRS radio frequency fingerprinting network described in step (3) include: The input to the DMRS radio frequency fingerprinting network consists of M DMRS amplitude and phase reconstruction feature terms corresponding to M symbols of DMRS, meaning that the input dimension of both branch networks is M×N. F ; For different numbers of input symbols M, the network is trained and the corresponding DMRS radio frequency fingerprinting network model is stored. In the network training, the network model stored when M=10 syms is used as the pre-training model of the network. The network is trained for other values ​​of M to significantly reduce the training time of the network.

5. The LTE DMRS-based radio frequency fingerprint recognition method according to claim 1, characterized in that: In step (4), when the DMRS of a terminal is received, the DMRS radio frequency fingerprint feature of the terminal is extracted using steps (1) and (2), and the DMRS radio frequency fingerprint recognition network trained in step (3) is used to identify and authenticate the terminal, and the identification result is output.

Images