Radio frequency fingerprint identification method of graph convolutional neural network based on Tuck decomposition

Abstract

The invention provides a radio frequency fingerprint identification method of a graph convolutional neural network based on Tuck decomposition. The problem that an existing deep learning method is low in identification performance under the condition of minimum training samples is solved. According to the method, tuck decomposition and a graph convolutional neural network are applied to radio frequency fingerprint identification, a graph adjacency matrix of a graph signal is searched by utilizing the tuck decomposition, an optimal model parameter is determined by using the graph adjacency matrix and a double-layer graph convolutional network, and finally a final classification result is output by using a full connection layer. Experimental results show that the method obtains better recognition performance.

Description

Technical Field

[0001] The present invention relates to the field of communication technologies, and more specifically, to a radio frequency fingerprint recognition method based on Tucker decomposition of graph convolutional neural networks. Background Art

[0002] With the advent of the fifth-generation mobile communication technology era and the Internet of Things era, the number of wireless communication devices has increased sharply. The security issue between wireless communication devices has naturally become an urgent problem to be solved. Device identification and authentication are important means to ensure the security of the Internet of Things system. In the existing literature, using the special radio frequency fingerprint features of wireless communication devices to distinguish between legal and illegal users can effectively ensure the security of communication. By extracting the radio frequency characteristics of wireless communication devices that are difficult to modify in their wireless waveforms, the unique radio frequency fingerprint feature (RFF) feature can be obtained, and then the category to which the device belongs can be effectively identified.

[0003] Essentially, radio frequency fingerprint recognition can be regarded as a classification problem. The traditional methods for RFF feature extraction are statistical and machine learning methods, such as multi-layer perceptrons and linear Bayesian classifiers. However, the artificial feature extraction in traditional methods highly depends on expert knowledge, which has certain limitations in modern Internet of Things systems with complex application environments. Due to the progress of hardware and software, deep learning has developed vigorously. In the field of radio frequency fingerprint recognition, more and more researchers use the powerful feature extraction ability of deep learning to extract high-resolution features of radio frequency signals. For example, Riyaz et al. proposed a method to optimize convolutional neural networks and demonstrated its recognition accuracy in noisy multipath wireless channels at different distances through experiments. Shen et al. constructed a deep learning model composed of three signal representations in the time domain, frequency domain, and time-frequency domain, further improving the classification accuracy. In addition, Zhang Weifeng et al. designed a data augmentation method and applied it to the designed dual-attention convolutional layer. The results show that the accuracy of this method for 56 ADS-B devices reaches 95.7%.

[0004] Similarly, there are a large number of deep learning methods in the field of radio frequency fingerprint identification (RFFI). However, these deep learning methods in the RFFI field are highly dependent on the number of training samples. In addition, considering that the existing deep learning methods are difficult to fit and lack robustness under the condition of extremely small training samples, these deep learning methods have limitations in the actual working scenarios of RFFI. In order to better solve the RFFI problem under the condition of extremely small samples, it is necessary to find a radio frequency fingerprint recognition method applicable to the condition of extremely small training samples. Summary of the Invention

[0005] The present invention provides a radio frequency fingerprint recognition method based on Tucker decomposition of graph convolutional neural network, which solves the problem of low recognition performance of existing deep learning methods under extremely small training samples. This method applies Tucker decomposition and graph convolutional neural network to radio frequency fingerprint recognition, proposes to use Tucker decomposition to find the graph adjacency matrix of graph signals, and uses this graph adjacency matrix and a two-layer graph convolutional network to determine the optimal model parameters, and finally uses a fully connected layer to output the final classification result. Experimental results show that this method obtains better recognition performance compared with the comparison scheme.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] A radio frequency fingerprint recognition method based on Tucker decomposition of graph convolutional neural network, comprising the following steps:

[0008] S1: Collect the radio frequency signals of each individual radio device;

[0009] S2: Preprocess the collected radio frequency signals to obtain a radio frequency fingerprint data set;

[0010] S3: Divide the training set, validation set and test set based on the radio frequency fingerprint data set;

[0011] S4: Select a specified number of samples in the training set to train the graph convolutional neural network based on Tucker decomposition, and update the training parameters in the graph convolutional neural network based on Tucker decomposition

[0012] S5: After training is completed, use the above-mentioned validation set to test the performance of the trained graph convolutional neural network based on Tucker decomposition;

[0013] S6: In the deployment and application stage, read the parameters of the graph convolutional neural network based on Tucker decomposition;

[0014] S7: Realize radio frequency fingerprint recognition under extremely small sample conditions;

[0015] Preferably, the radio frequency signals of the individual radio devices in step S1 are collected by a data acquisition card.

[0016] Preferably, step S2 preprocesses the collected radio frequency signals to obtain a radio frequency fingerprint data set. The specific preprocessing method used is as follows:

[0017] S2.1: Extract the turn-on transient signal, that is, first use the Bayesian step change detection method to calculate the posterior probability of the change point in the search window and determine the starting point, and then take a fixed-length radio frequency signal from the starting point as the turn-on transient signal, and its formula is:

[0018]

[0019] Wherein, d is the data within the search window, Υ is the signal model and noise statistic, N d is the length of d, and n d is the index of this point in d.

[0020] S2.2: Regularization, that is, dividing the extracted transient signal by its root mean square value to eliminate the influence caused by the change in the received signal power. The formula is:

[0021]

[0022] Wherein, a[ν] and a′[ν] are the values of the original transient signal and the normalized transient signal V th respectively, and L is the signal length of the transient signal.

[0023] S2.3: Extract features from the envelope of the regularized transient signal. This feature extraction is realized by the Hilbert transform. The specific formula is:

[0024]

[0025] Preferably, in step S3, the radio frequency fingerprint data set is divided into a training set, a validation set, and a test set, and the division ratio can be flexibly adjusted according to the task.

[0026] Preferably, in step S4, a specified number of samples are selected from the training set to train the graph convolutional neural network based on Tucker decomposition, and the training parameters in the graph convolutional neural network based on Tucker decomposition are updated. Specifically:

[0027] S4.1: First, perform Tucker decomposition on the input data. Tucker Decomposition (TD) is a high-order data decomposition technique that can capture multi-dimensional relationships in high-order data. The formula is:

[0028]

[0029] It should be noted that this formula is for the Tucker decomposition of two-dimensional data. In addition, in order to enable the graph neural network to learn more "effective features" from the core matrix obtained by Tucker decomposition, we set the negative elements in the core matrix to zero and use this as the graph adjacency matrix.

[0030] S4.2: Then perform a normalized graph Laplacian operation on the core matrix. The formula is:

[0031]

[0032] Wherein A is the adjacency matrix.

[0033] S4.3: Next, the normalized Laplacian matrix is adapted, and its formula is:

[0034] LA = Laplacian ⊙ W A

[0035] In the formula, W A represents the weight matrix, and ⊙ represents the Hadamard product.

[0036] S4.4: Then, a two-layer graph convolutional network is applied to the adapted Laplacian matrix, and its formula is:

[0037]

[0038] H 1 = AX × tanh(W 1 ) + LA

[0039] H 2 = LA × H 1 ⊙ W 2

[0040] output = regressor(H 2 )

[0041] In the formula, × represents matrix multiplication, H 1 represents the first-layer graph convolutional network, H 2 represents the second-layer graph convolutional network, regressor represents the fully connected layer, W 1 and W 2 are the weight matrices of the corresponding dimensions respectively. It should be noted that the fully connected layer ensures that the dimension of the finally output data is the total number of categories set in our experiment.

[0042] Preferably, after the training in step S5 is completed, the performance of the trained graph convolutional neural network based on Tucker decomposition is tested using the above-mentioned validation set;

[0043] Preferably, in the deployment and application stage of step S6, the parameters of the graph convolutional neural network based on Tucker decomposition are read first;

[0044] Preferably, step S7 realizes radio frequency fingerprint recognition under the condition of extremely small samples; BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a schematic flowchart of the method of the present invention.

[0046] Figure 2 is a schematic diagram of the method of the present invention.

[0047] Figure 3 is a detailed schematic diagram of the method of the present invention.

[0048] Figure 4 It is a schematic diagram of the source data feature distribution.

[0049] Figure 5 It is a schematic diagram of the method feature extraction of the present invention.

[0050] Figure 6 It is a schematic diagram of the method confusion matrix of the present invention.

[0051] Figure 7 It is a schematic diagram of the recognition accuracy of the method of the present invention based on 5 types of devices.

[0052] Figure 8 It is a schematic diagram of the recognition accuracy of the method of the present invention based on 10 types of devices.

[0053] Figure 9 It is a schematic diagram of the ablation experiment of the method of the present invention.

[0054] Figure 10 It is a schematic diagram of the robustness experiment of the method of the present invention. Detailed implementation manners

[0055] The attached drawings are only for illustrative purposes and should not be construed as a limitation of this patent;

[0056] To better illustrate this embodiment, some components in the attached drawings will be omitted, enlarged or reduced, which do not represent the dimensions of the actual product;

[0057] For those skilled in the art, it is understandable that some well-known structures and their descriptions in the attached drawings may be omitted.

[0058] The technical solutions of the present invention will be further described below with reference to the attached drawings and embodiments.

[0059] Embodiment 1

[0060] In Embodiment 1, 5 DMR walkie-talkies of the same model, model number Kelixun DP485-01, are used as the radio station equipment to be recognized. To avoid the influence of the walkie-talkie parameter settings on the radio frequency fingerprint recognition, the center frequency of the transmitted signal of each walkie-talkie is uniformly set to 141.825 MHz, the bandwidth is 12.5 KHz, and the modulation method is 4FSK. The transmitted signal of the walkie-talkie is transmitted to the data acquisition card through a shielded wire for acquisition. The sampling rate of the used data acquisition card is 400-MS / s, and the sampling accuracy is 14-bit. 78 signal blocks are collected for each walkie-talkie, and 775 transient power-on signals obtained through preprocessing are used as the radio frequency fingerprint data set. In the data set obtained by sampling and preprocessing, 25 samples are taken as the training set to train the graph convolutional neural network based on Tucker decomposition, and 750 samples are taken as the test set.

[0061] This example provides a radio frequency fingerprint recognition method based on Tucker decomposition for graph convolutional neural networks, as Figure 1 shown, including the following steps:

[0062] S1: Collect the radio frequency signals of each individual radio device;

[0063] S2: Preprocess the collected radio frequency signals to obtain a radio frequency fingerprint dataset;

[0064] S3: Divide the training set, validation set, and test set based on the radio frequency fingerprint dataset;

[0065] S4: Select a specified number of samples in the training set to train the graph convolutional neural network based on Tucker decomposition, and update the training parameters in the graph convolutional neural network based on Tucker decomposition

[0066] S5: After training is completed, use the above-mentioned validation set to test the performance of the trained graph convolutional neural network based on Tucker decomposition;

[0067] S6: In the deployment and application stage, first read the parameters of the graph convolutional neural network based on Tucker decomposition;

[0068] S7: Achieve radio frequency fingerprint recognition under the condition of extremely small samples;

[0069] To test the feature extraction performance of the method of the present invention, we use the above 25 training set samples to train the graph convolutional neural network based on Tucker decomposition, and after training is completed, input 750 test set samples into the model network to obtain a feature set. Subsequently, principal component analysis (PCA) is performed on each feature set to reduce the dimension to a two-dimensional graph to visually display the feature extraction ability trained by the method of the present invention.

[0070] The experimental results are as Figure 4 and Figure 5 shown. Figure 4 shows the data feature distribution before training, where the data features of some types significantly overlap and are significantly mixed together. At the same time, Figure 5 is the visualization of the feature extraction of the method of the present invention. It can be seen from the feature map that due to the powerful feature extraction ability of the method of the present invention, when applying the proposed method, similar feature clusters can be significantly separated. It should be noted that the PCA visualization of data features will inevitably lead to a reduction in data dimensions and the loss of corresponding information. Therefore, in order to quantitatively evaluate the extracted features of the method of the present invention, the confusion matrix listed is as Figure 3 shown. Specifically, Figure 3 the horizontal axis of Figure 3Among them, except for the inaccurate predictions of Device 2 and Device 3, the predictions of most device metrics are accurate.

[0071] Example 2

[0072] Based on Example 1, the following content is further disclosed in this example:

[0073] In Example 2, the construction and working process of the system of the present invention, the individual radio stations and hardware parameters used, and the network model adopted are the same as those in Example 1. The difference is that in Example 2, the number of known devices is fixed at 5, and the number of test samples for each device is set to 50, 100, and 150 respectively. The control schemes are deep learning schemes in the field of radio frequency fingerprint recognition in recent years, namely CVCNN, ARFNet, ResNet, and TDNN.

[0074] The experimental results of Example 2 are as Figure 7 shown. Experiments were conducted on the four control schemes and the method of the present invention under the condition of a very small number of training samples. Among them, the number of test samples for each device increased from 50 to 100, and finally to 150. In the case of 50 test samples for each device, the method of the present invention achieved the highest device estimation accuracy value compared with the four control schemes. However, as the number of test samples for each device continued to increase, the device estimation accuracy of each method decreased slightly. Specifically, compared with the CVCNN scheme, the recognition accuracy of the method of the present invention increased by 0.40% - 0.53%. Among the other benchmark schemes, the recognition accuracy of the TDNN scheme was approximately 60%, but ResNet and ARFNet could only reach an estimation accuracy of 20% - 30%. Obviously, the recognition accuracy of these four control schemes under the condition of extremely small samples is lower than that of the method of the present invention.

[0075] Example 3

[0076] In Example 3, the construction and working process of the system of the present invention, the individual radio stations and hardware parameters used, and the network model adopted are the same as those in Examples 1 and 2. The difference from Example 2 is that in Example 3, the number of known devices is fixed at 10, and the number of test samples for each device is set to 50, 100, and 150 respectively. The control schemes are deep learning schemes in the field of radio frequency fingerprint recognition in recent years, namely CVCNN, ARFNet, ResNet, and TDNN.

[0077] The experimental results of Example 2 are as Figure 8As shown in the figure. Under the conditions of different numbers of test samples, the method of the present invention has achieved the highest device estimation accuracy value compared with four control schemes. Specifically, compared with the CVCNN scheme, the recognition accuracy of the method of the present invention has increased by 6.62% - 7.76%. Among other benchmark schemes, the recognition accuracy is lower than 20%. Obviously, the recognition accuracy of these four control schemes under extremely small sample conditions is lower than that of the method of the present invention.

[0078] Example 4

[0079] In Example 4, the construction and working process of the system of the present invention, the individual radio stations and hardware parameters adopted, and the network model adopted are all the same as those in Examples 1, 2, and 3. In Example 4, the number of known devices is fixed at 5, and the number of test samples for each device is set to 50, 100, and 150 respectively. The control schemes are several ablation schemes of the method of the present invention. Specifically, compared with the method of the present invention, in ablation scheme 1, the graph adjacency matrix obtained by Tucker decomposition is replaced by a small part of the data matrix in the transient signal. Compared with the method of the present invention, in ablation scheme 2, the first graph convolutional layer is removed. Compared with the method of the present invention, in ablation scheme 3, the second graph convolutional layer is removed.

[0080] The experimental results of Example 4 are as Figure 9 shown. Under the conditions of different numbers of test samples, the method of the present invention has achieved the highest device estimation accuracy value compared with three ablation schemes. Specifically, in ablation experiment 1, without applying Tucker decomposition, the device recognition accuracy drops sharply. Obviously, Tucker decomposition plays a crucial role in the method of the present invention because the graph generation and graph adjacency matrix determine the effectiveness of the method of the present invention. Next, in ablation experiment 2, it can be seen that after removing the first graph convolutional layer, the device recognition accuracy still decreases. Then, in ablation experiment 3, it can be determined that the second graph convolutional layer still plays an important role in the aggregation and update of graph nodes. Finally, through the three groups of ablation experiments, it is verified that the device recognition accuracy is the highest after the components of the method of the present invention are fused, and the necessity and effectiveness of each component of the method of the present invention are verified.

[0081] Example 5

[0082] In Embodiment 5, the construction and working process of the system of the present invention, the individual radios and hardware parameters adopted, and the network model adopted are the same as those in Embodiments 1, 2, 3, and 4. In Embodiment 5, the number of known devices is fixed at 5, and the number of test samples for each device is 50. It is worth noting that in Embodiment 5, in order to simulate the impact of environmental changes on the model under real deployment conditions, the method of the present invention is tested and the device recognition accuracy of the control scheme under different proportions of damaged neurons is compared. The control schemes are deep learning schemes in the field of radio frequency fingerprint recognition in recent years, namely CVCNN, ARFNet, ResNet, and TDNN.

[0083] The experimental results of Embodiment 5 are as Figure 10 shown. When the number of test samples for each device is 50 and the probability of different neurons being damaged, the method of the present invention reaches the highest device estimation accuracy value compared with the four control schemes. In other words, the method of the present invention obtains relatively good reliability performance, which indicates that the method of the present invention has the best robustness compared with the four control schemes, thus significantly reducing the difficulty and cost of model retraining in practical applications. Therefore, it can be obtained that the method of the present invention has better ability to adapt to changing environments, which is an urgent requirement for the deployment of real-time radio frequency fingerprint recognition methods.

[0084] Embodiment 6

[0085] In Embodiment 6, the construction and working process of the system of the present invention, the individual radios and hardware parameters adopted, and the network model adopted are the same as those in Embodiments 1, 2, 3, 4, and 5. In Embodiment 6, the number of known devices is fixed at 5, and the number of test samples for each device is 50. It is worth noting that in Embodiment 6, we compared the time and space complexity of the method of the present invention with the four control schemes of CVCNN, ARFNet, ResNet, and TDNN, and the results are shown in the following table.

[0086] Radio Frequency Fingerprint Identification Scheme Recognition Time Consumption (seconds) Number of Neural Network Parameters CVCNN 0.11423 1079005 ARFNet 0.32927 814078 ResNet 3.62628 16123430 TDNN 0.02147 7105 Proposed Scheme 0.04480 104405

[0087] According to the results, when the number of test samples for each device is 50, the method of the present invention can simplify the neural network structure and reduce the scale of the neural network while ensuring excellent device recognition accuracy compared with the four control schemes, so as to significantly reduce the number of parameters of the neural network. Therefore, the method of the present invention has excellent real-time performance and practicality compared with the control scheme, and is more suitable for the actual application scenario of radio frequency fingerprint recognition.

[0088] The above embodiments show that the proposed method has better device accuracy under extremely small training samples.

[0089] The descriptions of the positional relationships in the drawings are only for illustrative purposes and should not be construed as limitations on this patent;

[0090] Obviously, the above examples of the present invention are merely illustrations for clearly explaining the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all implementation manners here. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A radio frequency fingerprint recognition method based on a graph convolutional neural network with Tucker decomposition, comprising the following steps: S1: Collect the radio frequency signals of each individual radio device; S2: pre-processing the collected radio frequency signals to obtain a radio frequency fingerprint data set; S3: Divide the training set, validation set and test set based on the RF fingerprint dataset; S4: Select a specified number of samples in the training set to train the graph convolutional neural network based on Tucker decomposition, and update the training parameters in the graph convolutional neural network based on Tucker decomposition S5: After the training is completed, the performance of the trained graph convolutional neural network based on Tucker decomposition is tested using the above validation set; S6: In the deployment and application phase, read the graph convolutional neural network parameters based on Tucker decomposition; S7: Realize RF fingerprint recognition under extremely small sample conditions; Preferably, the radio frequency signals of the individual radio devices in step S1 are collected by a collection card. Preferably, step S2 preprocesses the collected radio frequency signals to obtain a radio frequency fingerprint data set. The preprocessing method used is specifically: S2.1: Extraction of power-on transient signal, that is, first use the Bayesian step change detection method to calculate the posterior probability of the change point in the search window and determine the starting point, and then take the fixed length of the RF signal from the starting point as the power-on transient signal. The formula is: Where d is the data in the search window, Υ is the signal model and noise statistics, N d is the length of d, n d is the index of the point in d. S2.2: Regularization, that is, dividing the extracted transient signal by its RMS value to eliminate the influence of the received signal power variation. The formula is: Where a[ν] and a′[ν] are the original transient signal and the normalized transient signal V, respectively. th The value of , L is the signal length of the transient signal. S2.3: Extract features from the envelope of the regularized transient signal. This feature extraction is achieved by Hilbert transform. The specific formula is: Preferably, in step S3, the radio frequency fingerprint data set is divided into a training set, a validation set and a test set, and the division ratio can be flexibly adjusted according to the task. Preferably, in step S4, a specified number of samples are selected from the training set to train the graph convolutional neural network based on Tucker decomposition, and the training parameters in the graph convolutional neural network based on Tucker decomposition are updated, specifically: S4.1: First, perform Tucker decomposition on the input data. Tucker Decomposition (TD) is a high-order data decomposition technique that can capture multidimensional relationships in high-order data. Its formula is: It is worth noting that this formula is for Tucker decomposition of two-dimensional data. In addition, in order to enable the graph neural network to learn more "effective features" from the core matrix obtained by Tucker decomposition, we set the negative elements in the core matrix to zero and use it as the graph adjacency matrix. S4.2: Then the core matrix is ​​subjected to the normalized graph Laplacian operation, and its formula is: In the formula A is the adjacency matrix. S4.3: Then the normalized graph Laplacian matrix is ​​adapted, and its formula is: LA=Laplacian⊙W A Where W A represents the weight matrix, and ⊙ represents the Hadamard product. S4.4: Then apply a two-layer graph convolutional network to the adapted graph Laplacian matrix, as follows: H1=AX×tanh(W1)+LA H2=LA×H1⊙W2 output = regressor(H2) In the formula, × represents matrix product, H1 represents the first layer of graph convolutional network, H2 represents the second layer of graph convolutional network, regressor represents the fully connected layer, and W1 and W2 are weight matrices of corresponding dimensions. It is worth noting that the fully connected layer ensures that the final output data dimension is the total number of categories set in our experiment. Preferably, after the training is completed in step S5, the performance of the trained Tucker decomposition-based graph convolutional neural network is tested using the above-mentioned validation set; Preferably, in step S6, at the deployment and application stage, firstly read the graph convolutional neural network parameters based on Tucker decomposition; Preferably, step S7 implements radio frequency fingerprint recognition under extremely small sample conditions.

Images