A radio frequency fingerprint recognition method based on chaotic weighted random fusion

Abstract

The present invention discloses a radio frequency fingerprint recognition method based on chaotic weighted random fusion, comprising the following steps: S1: collecting radio frequency signals from individual radio stations; S2: preprocessing the collected radio frequency signals to obtain a radio frequency fingerprint dataset; S3: dividing the radio frequency fingerprint dataset into a training set and a test set; S4: selecting a specified number of original samples from the training set, weightedly fusing the original samples using a chaotic sequence to generate new samples and obtain an enhanced training set; S5: training a neural network using the enhanced dataset to obtain a trained neural network; S6: performing a performance test on the trained neural network using the test set; and S7: implementing radio frequency fingerprint recognition using the neural network that passes the performance test. The present invention combines a chaotic sequence with excellent random characteristics with the RI algorithm to increase the diversity and maximum number of generated samples, resulting in improved recognition accuracy and robustness.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and more particularly to a radio frequency fingerprint recognition method based on chaotic weighted random fusion. Background Art

[0002] Due to individual hardware variations and manufacturing process factors, subtle differences are inevitable between any two radios. Even radios of the same model and produced on the same production line will have unique individual characteristics in their transmitted signals, known as RF fingerprints. As modern warfare rapidly evolves towards information warfare, various electronic devices, represented by radios, undertake crucial tasks such as communications, surveillance, and electronic jamming. Therefore, the ability to proactively identify enemy radios is a crucial aspect of electronic warfare and, to a certain extent, determines the initiative on the battlefield. Consequently, RF fingerprinting technology has attracted widespread attention from researchers worldwide in recent years.

[0003] Generally speaking, RF fingerprinting can be viewed as a classification problem, and most state-of-the-art RF fingerprinting classifiers are based on deep learning techniques. Thanks to their powerful feature extraction and pattern recognition capabilities, deep learning-based classifiers can achieve fairly high recognition accuracy. However, they require a large number of training samples during offline training to ensure satisfactory performance and avoid overfitting. In practical applications, especially in non-cooperative communication scenarios such as electronic countermeasures, obtaining sufficient RF fingerprinting data samples is a challenge. Currently, research on RF fingerprinting under small sample sizes is lacking. To address this issue, some literature has proposed applying random integration (RI) data augmentation algorithms to RF fingerprinting, achieving some success. However, conventional RI algorithms are limited by fixed sample fusion weights, resulting in insufficient diversity in the generated new samples and limiting the maximum number of new samples that can be generated. Therefore, data augmentation schemes suitable for RF fingerprinting under very small sample sizes still have significant room for improvement.

[0004] Existing technology involves a deep learning-based method for identifying radio frequency fingerprints of frequency-using devices. This method involves collecting and analyzing radio frequency signals, storing them as I / Q data, performing modulation-based shaping preprocessing on the I / Q data to construct training and test datasets, building and training a deep RCAN-RFF network, and inputting test data into the RCAN-RFF network to obtain the radio frequency fingerprint identification results of the frequency-using devices. This solution cannot be applied to small sample sizes. Summary of the Invention

[0005] The present invention provides a radio frequency fingerprint recognition method based on chaotic weighted random fusion, which effectively improves the radio frequency fingerprint recognition accuracy and robustness of deep learning classifiers under extremely small sample conditions.

[0006] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0007] A radio frequency fingerprint recognition method based on chaotic weighted random fusion includes the following steps:

[0008] S1: collects the radio frequency signals of each radio station;

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

[0010] S3: Divide the radio frequency fingerprint dataset into a training set and a test set;

[0011] S4: selecting a specified number of original samples from the training set, performing weighted fusion on the original samples using a chaotic sequence to generate new samples, and obtaining an enhanced training set, wherein the chaotic sequence is generated by a chaos generating function;

[0012] S5: Using the enhanced data set to train a neural network to obtain a trained neural network, wherein the neural network outputs a radio station label corresponding to the radio frequency fingerprint according to the input radio frequency fingerprint;

[0013] S6: Using the test set to perform a performance test on the trained neural network;

[0014] S7: Implement radio frequency fingerprint recognition using a neural network that has passed performance verification.

[0015] Preferably, the radio frequency fingerprint is a power-on transient signal of a radio frequency signal.

[0016] Preferably, in step S2, the collected radio frequency signal is preprocessed to obtain a radio frequency fingerprint data set, specifically:

[0017] S2.1: Extract the power-on transient signal of the RF signal as the RF fingerprint;

[0018] S2.2: performing regularization processing on the extracted startup transient signal to obtain a regularized startup transient signal;

[0019] S2.3: Extract the envelope of the regular power-on transient signal as the RF fingerprint dataset.

[0020] Preferably, in step S2.1, the power-on transient signal of the radio frequency signal is extracted as the radio frequency fingerprint, specifically:

[0021] The starting point is determined by calculating the posterior probability that each sampling point in the sliding window is the starting point of the power-on transient signal, and a radio frequency signal of a fixed length L starting from the starting point is taken as the power-on transient signal.

[0022] Preferably, in step S2.2, the extracted startup transient signal is subjected to regularization processing, specifically:

[0023]

[0024] Where a[v] and a ′ [v] are the vth instantaneous values ​​of the initial startup transient signal and the regular startup transient signal, respectively, and L is the length of the startup transient signal.

[0025] Preferably, in step S2.3, the envelope of the regular startup transient signal is extracted, specifically:

[0026]

[0027] Where, is the Hilbert transform function, and j represents the imaginary unit.

[0028] Preferably, in step S4, a second-order Chebyshev polynomial function is used to generate the chaotic sequence x, specifically:

[0029]

[0030] Where x n Represents the nth element in the chaotic sequence x.

[0031] Preferably, in step S4, the chaotic sequence is used to perform weighted fusion on the original samples to generate new samples, specifically:

[0032] Determine the original sample set S o The size M and the number of original samples N used in each fusion, and then according to the o Select all combinations of N samples to construct a fusion label table

[0033] When a new sample is generated for the i-th time, a chaotic sequence x=[x1,…, N ], and then by The standardized chaotic weight y=[y1,…, N ];

[0034] According to the jth item T in the fusion label table j From the original sample set S o =[S o,1 ,…, o,M ] select N samples from in And perform chaotic weighted fusion to obtain new samples

[0035] Repeat the above operation of generating new samples until the number of generated samples is greater than or equal to the value pre-specified by the user.

[0036] Preferably, each time a new sample is generated, the generated chaotic sequence is different.

[0037] Preferably, in step S5, the enhanced data set is used to train the neural network to obtain a trained neural network, specifically:

[0038] Through a preset loss function, the loss function value between the estimated radio station label probability output by the neural network and the actual radio station label of the training sample is calculated. Subsequently, through the back-propagation algorithm, the partial derivative of the loss function value with respect to each trainable parameter in the neural network is calculated, so that the neural network parameters are updated using a preset optimization algorithm.

[0039] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0040] The present invention combines a chaotic sequence with good random characteristics with the RI algorithm to generate dynamically changing fusion weights, thereby increasing the diversity and maximum number of generated samples, thereby improving the neural network training effect, and further effectively improving the accuracy and robustness of radio frequency fingerprint recognition of deep learning-based classifiers under extremely small sample conditions, thereby alleviating the problem of difficulty in sample acquisition in actual application scenarios of radio frequency fingerprint recognition based on deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the method of the present invention.

[0042] Figure 2 It is a schematic diagram of the framework of the present invention.

[0043] Figure 3 A schematic diagram of a neural network structure provided for an embodiment.

[0044] Figure 4 The following is a schematic diagram comparing the recognition accuracy of different algorithms under different numbers of original samples provided in the embodiment.

[0045] Figure 5 This is a schematic diagram comparing the recognition accuracy of different algorithms under different neural network parameters provided in the embodiment.

[0046] Figure 6 This is a schematic diagram comparing the recognition accuracy of different algorithms under test sets with different interference levels provided in the embodiment. DETAILED DESCRIPTION

[0047] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0048] In order to better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size;

[0049] It is understandable to those skilled in the art that some well-known structures and descriptions thereof may be omitted in the drawings.

[0050] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0051] Example 1

[0052] A radio frequency fingerprint recognition method based on Chaotic Weighting Random Integration (CWRI), such as Figure 1 and Figure 2 As shown, the following steps are included:

[0053] S1: collects the radio frequency signals of each radio station;

[0054] S2: Preprocess the collected radio frequency signals to obtain a radio frequency fingerprint dataset;

[0055] S3: Divide the radio frequency fingerprint dataset into a training set and a test set;

[0056] S4: selecting a specified number of original samples from the training set, performing weighted fusion on the original samples using a chaotic sequence to generate new samples, and obtaining an enhanced training set, wherein the chaotic sequence is generated by a chaos generating function;

[0057] S5: Using the enhanced data set to train a neural network to obtain a trained neural network, wherein the neural network outputs a radio station label corresponding to the radio frequency fingerprint according to the input radio frequency fingerprint;

[0058] S6: Using the test set to perform a performance test on the trained neural network;

[0059] S7: Implement radio frequency fingerprint recognition using a neural network that has passed performance verification.

[0060] Example 2

[0061] This embodiment, based on the first embodiment, further discloses the following contents:

[0062] The radio frequency fingerprint is a power-on transient signal of the radio frequency signal.

[0063] In step S2, the collected radio frequency signal is preprocessed to obtain a radio frequency fingerprint data set, specifically:

[0064] S2.1: Extract the power-on transient signal of the RF signal as the RF fingerprint;

[0065] S2.2: performing regularization processing on the extracted startup transient signal to obtain a regularized startup transient signal;

[0066] S2.3: Extract the envelope of the regular power-on transient signal as the RF fingerprint dataset.

[0067] In step S2.1, the power-on transient signal of the RF signal is extracted as the RF fingerprint, specifically:

[0068] The starting point is determined by calculating the posterior probability that each sampling point in the sliding window is the starting point of the power-on transient signal, and a radio frequency signal of a fixed length L starting from the starting point is taken as the power-on transient signal.

[0069] In step S2.2, the extracted startup transient signal is regularized, specifically:

[0070]

[0071] Where a[v] and a ′ [v] are the vth instantaneous values ​​of the initial startup transient signal and the regular startup transient signal, respectively, and L is the length of the startup transient signal.

[0072] In step S2.3, the envelope of the regular startup transient signal is extracted, specifically:

[0073]

[0074] Where, is the Hilbert transform function, and j represents the imaginary unit.

[0075] In step S4, a second-order Chebyshev polynomial function is used to generate a chaotic sequence x, specifically:

[0076]

[0077] Where x n Represents the nth element in the chaotic sequence x.

[0078] In step S4, the chaotic sequence is used to perform weighted fusion on the original samples to generate new samples, specifically:

[0079] Determine the original sample set S o The size M and the number of original samples N used in each fusion, and then according to the o Select all combinations of N samples to construct a fusion label table

[0080] When a new sample is generated for the i-th time, a chaotic sequence x=[x1,…, N ], and then by The standardized chaotic weight y=[y1,…, N ];

[0081] According to the jth item T in the fusion label table j From the original sample set S o =[S o,1 ,…, o,M ] select N samples from in And perform chaotic weighted fusion to obtain new samples

[0082] Repeat the above operation of generating new samples until the number of generated samples is greater than or equal to the value pre-specified by the user.

[0083] Each time a new sample is generated, the generated chaotic sequence is different.

[0084] In step S5, the enhanced data set is used to train the neural network to obtain a trained neural network, specifically:

[0085] Through a preset loss function, the loss function value between the estimated radio station label probability output by the neural network and the actual radio station label of the training sample is calculated. Subsequently, through the back-propagation algorithm, the partial derivative of the loss function value with respect to each trainable parameter in the neural network is calculated, so that the neural network parameters are updated using a preset optimization algorithm.

[0086] In this implementation, the neural network structure used is as follows Figure 3 The figure shows a one-dimensional convolutional neural network for radio frequency fingerprint recognition and speech signal processing, which includes three one-dimensional convolutional (1D convolutional) layers and two fully connected (FC) layers. Each convolutional layer is followed by a batch normalization (BN) layer and a maximum pooling layer. The size of the transient feature as input is 1×2 after downsampling. 11 , where the channel dimension size is 1 and the time dimension size is 2 11. Each convolutional layer is mainly responsible for extracting RF fingerprint features at different scales. The output channel dimensions are Q, 2, and 4, the convolution kernels are 17, 9, and 5, and the padding sizes are 8, 4, and 2, respectively. The BN layer following the convolutional layer is combined with the ReLU activation function to alleviate the gradient vanishing and gradient exploding problems during training and accelerate the convergence of the neural network. The maximum pooling layer after the BN layer is used to reduce the time channel size while retaining key features, thereby reducing complexity and accelerating convergence. The filter sizes of the three maximum pooling layers are 8, 16, and 16, respectively. Finally, the two FC layers are combined with the Softmax activation function to integrate the extracted features and output a probability vector p. Each element in p represents the probability that the input signal belongs to the Z radio stations to be identified. During the training process, the loss function used by the neural network is the cross entropy loss function, the optimization algorithm is the Adam algorithm, and the learning rate is set to 1×10 -3 .

[0087] Example 3

[0088] This embodiment provides the following specific embodiments based on Embodiment 1 and Embodiment 2:

[0089] a)

[0090] In this specific embodiment, five identical DMR radios, model Kolixun DP485-01, were used as the radio equipment to be identified. To minimize the impact of differences in radio parameter settings on RF fingerprint identification, each radio's transmit signal center frequency was uniformly set to 141.825 MHz, bandwidth to 12.5 kHz, and modulation to 4FSK. The radio's transmit signal was transmitted via a shielded cable to a data acquisition card for acquisition. The card used a sampling rate of 400 MS / s and a sampling accuracy of 14 bits. Each radio collected 60 signal blocks, and 60 transient power-on signals were preprocessed to form the RF fingerprint dataset. Of the sampled and preprocessed dataset, 50 samples were used as the test set, and M of the remaining 10 samples were used as the original samples in the training set. In the RI and CWRI data augmentation process, the number of original samples N used in each fusion was set to 3. Nine new samples were generated, which, together with the M original samples, formed an augmented training set of 10 for neural network training. Considering that a system that does not use data enhancement and a system that uses a general RI enhancement method may not be able to generate 9 new samples, in order to ensure fairness, the samples in its largest training set will be copied so that the total number of training samples reaches the specified value of 10.

[0091] In order to test the performance of the CWRI method of the present invention and the general RI method and the no-data enhancement scheme as the control scheme, first use each data enhancement method or copy samples to obtain an enhanced training set of size 10M from the original sample set of size M, and then use the training set to train the neural network models with the same hyperparameter settings. Finally, use a test set of size 50 to test the accuracy of each classifier to verify the impact of each data enhancement method and the no-data enhancement scheme on the system recognition performance. In Example a), the performance of the proposed CWRI data enhancement method and the general RI method and the no-data enhancement scheme under different original sample numbers is mainly tested, and a comparative analysis is performed. The original sample number M is 3, 4, 5, 6, and 7, respectively, and the hyperparameter Q that controls the number of neural network parameters is 10. The experimental results are as follows. Figure 4 As shown in the figure, it can be seen that with the increase in the number of original samples, the recognition accuracy of each system increases. However, the CWRI method proposed in the present invention has better sample diversity, and the classifier trained can always achieve better recognition accuracy, which proves the effectiveness of the method described in the present invention.

[0092] b)

[0093] In this specific embodiment, the system construction and workflow, the individual radio stations and hardware parameters used, and the neural network model used are the same as those in specific embodiment a). The only difference lies in the hyperparameter settings during the training process. In Example 2, the performance of the proposed CWRI data enhancement method, the general RI method, and the no-data enhancement scheme are mainly tested under different numbers of trainable parameters of the neural network, and a comparative analysis is performed. The number of original samples M is fixed to 5, and the hyperparameter Q that controls the number of neural network parameters is set to 20, 30, 40, and 50, respectively. Generally speaking, as the parameters of the neural network increase, its learning ability will also increase accordingly, but it will also cause increased complexity and training difficulties, and may lead to overfitting problems. The experimental results of Example b) are as follows. Figure 5 As shown. According to the results, as the neural network parameters increase, the recognition accuracy of the classifier trained using the general RI method remains basically unchanged, while the accuracy of the classifier without data enhancement gradually decreases. This is because the control scheme has insufficient data diversity under extremely small sample conditions, which makes it easy to have overfitting problems during training. The CWRI data enhancement method described in the present invention can effectively improve data diversity and the maximum number of samples that can be generated, so it can make full use of the learning ability improvement brought about by the increase in neural network parameters, thereby achieving better and gradually increasing recognition accuracy compared to the control scheme under the condition of gradually increasing neural network parameters.

[0094] c)

[0095] In this specific embodiment, the system construction and workflow, the individual radio stations and hardware parameters used, and the neural network model used are the same as those in Example a) and Example b). The only difference is that after the transient power-on signal is extracted, additive white Gaussian noise (AWGN) of different intensities is added to the extracted signal to obtain data sets with different interference levels to test the robustness performance of each scheme. The training signal-to-noise ratio (SNR) is set to 30dB, the test SNR is 20-50dB, the number of original samples M is fixed to 5, and the hyperparameter Q that controls the number of neural network parameters is fixed to 50. The experimental results of Example 3 are shown in Figure 3. Figure 6 As shown in the figure, the results show that due to the different levels of noise interference between the training and test sets, the feature distributions of the training and test set samples are different, making overfitting more likely to occur during training. This results in a decrease in the recognition accuracy of both the method of the present invention and the control scheme, especially for systems without data enhancement schemes. However, under these conditions, the CWRI data enhancement method of the present invention can still achieve better recognition accuracy at various test SNRs, indicating that the system using the method of the present invention has better robustness.

[0096] The same or similar reference numerals correspond to the same or similar components;

[0097] The terms used in the drawings to describe positional relationships are for illustrative purposes only and should not be construed as limiting this patent;

[0098] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A radio frequency fingerprint recognition method based on chaotic weighted random fusion, characterized in that: The following steps are involved: S1: collects the radio frequency signals of each radio station; S2: Preprocess the collected radio frequency signals to obtain a radio frequency fingerprint dataset; S3: Divide the radio frequency fingerprint dataset into a training set and a test set; S4: selecting a specified number of original samples from the training set, performing weighted fusion on the original samples using a chaotic sequence to generate new samples, and obtaining an enhanced training set, wherein the chaotic sequence is generated by a chaos generating function; S5: Using the enhanced data set to train a neural network to obtain a trained neural network, wherein the neural network outputs a radio station label corresponding to the radio frequency fingerprint according to the input radio frequency fingerprint; S6: Using the test set to perform a performance test on the trained neural network; S7: Implementation of radio frequency fingerprint recognition using a neural network that has passed performance verification; The radio frequency fingerprint is a power-on transient signal of the radio frequency signal; In step S2, the collected radio frequency signal is preprocessed to obtain a radio frequency fingerprint data set, specifically: S2.1: Extract the power-on transient signal of the RF signal as the RF fingerprint; S2.2: performing regularization processing on the extracted startup transient signal to obtain a regularized startup transient signal; S2.3: Extract the envelope of the regular power-on transient signal as the RF fingerprint dataset; In step S2.3, the envelope of the regular startup transient signal is extracted, specifically: Where, is the Hilbert transform function, i represents the imaginary unit; In step S4, a second-order Chebyshev polynomial function is used to generate a chaotic sequence x, specifically: Where x n represents the nth element in the chaotic sequence x; In step S4, the chaotic sequence is used to perform weighted fusion on the original samples to generate new samples, specifically: Determine the original sample set S o The size M and the number of original samples N used in each fusion, and then according to the o Select all combinations of N samples to construct a fusion label table When a new sample is generated for the i-th time, a chaotic sequence x=[x1,…,x N ], and then by The standardized chaotic weight y=[y1,…,y N ]; According to the jth item T in the fusion label table j From the original sample set S o =[S o,1 ,…,S o,M ] select N samples from in And perform chaotic weighted fusion to obtain new samples Repeat the above operation of generating new samples until the number of generated samples is greater than or equal to the value pre-specified by the user.

2. The radio frequency fingerprint recognition method based on chaotic weighted random fusion according to claim 1 is characterized in that: In step S2.1, the power-on transient signal of the RF signal is extracted as the RF fingerprint, specifically: The starting point is determined by calculating the posterior probability that each sampling point in the sliding window is the starting point of the power-on transient signal, and a radio frequency signal of a fixed length L starting from the starting point is taken as the power-on transient signal.

3. The radio frequency fingerprint recognition method based on chaotic weighted random fusion according to claim 2 is characterized in that: In step S2.2, the extracted startup transient signal is regularized, specifically: Where a[v] and a ′ [v] are the vth instantaneous values ​​of the initial startup transient signal and the regular startup transient signal, respectively, and L is the length of the startup transient signal.

4. The radio frequency fingerprint recognition method based on chaotic weighted random fusion according to claim 3 is characterized in that: Each time a new sample is generated, the generated chaotic sequence is different.

5. The radio frequency fingerprint recognition method based on chaotic weighted random fusion according to any one of claims 1 to 4, characterized in that: In step S5, the enhanced data set is used to train the neural network to obtain a trained neural network, specifically: Through a preset loss function, the loss function value between the estimated radio station label probability output by the neural network and the actual radio station label of the training sample is calculated. Subsequently, through the back-propagation algorithm, the partial derivative of the loss function value with respect to each trainable parameter in the neural network is calculated, so that the neural network parameters are updated using a preset optimization algorithm.

Images