An MGF radio frequency fingerprinting method for LFM radar

Abstract

This invention discloses an MGF (Multi-Fingerprint Generator) radio frequency fingerprinting method for LFM radar. It uses the MFCC (Multi-Fingerprint Capacitor) speech feature extraction method to generate Mel features for each LFM pulse data. The feature matrices of different lengths are standardized to ensure dimensionality consistency. The standardized feature matrix is ​​the signal fingerprint feature. A single GRU cell is then used to model the signal fingerprint feature, and a fully connected layer is used to identify and classify the modeled features, achieving radio frequency fingerprint recognition and improving the recognition accuracy for difficult-to-classify samples. This invention can effectively extract the physical fingerprint features of LFM pulse signals under low signal-to-noise ratio and multi-signal multi-operation modes, effectively solving the problem of poor LFM signal recognition in practical application environments.

Description

Technical Field

[0001] This invention belongs to the field of information security technology, specifically relating to an MGF radio frequency fingerprinting method for LFM radar. Background Technology

[0002] With the continuous development of wireless communication technology, the number of IoT devices and wireless communication devices is increasing year by year. However, the open wireless network communication environment is vulnerable to various malicious attacks, which seriously hinders the development and application of wireless network communication technology. Ensuring the security of wireless networks is becoming increasingly important. Physical layer-based security protection mechanisms, such as classifying and identifying linear frequency modulation (LFM) signals, have been widely used in wireless communication technology.

[0003] Classical signal identification methods classify signals based on Pulse Descriptor Words (PDWs), including features such as the angle of arrival (DOA) and time of arrival (TOA) of each pulse. In wireless communication systems, complex signal waveforms mean that different signals, under the same configuration, can exhibit similar electromagnetic characteristics, increasing the difficulty of signal classification. Using PDWs alone is insufficient for effective signal classification. A device's radio frequency fingerprint, as a hardware characteristic of the device itself, is like a fingerprint—unique and difficult to clone. Physical fingerprint features are generally considered unique characteristics of wireless devices. Using these features as device fingerprints can relatively accurately identify different devices, improving wireless network security.

[0004] However, existing publicly available radar identification methods based on radio frequency fingerprinting (RF fingerprinting) primarily operate under high signal-to-noise ratio (SNR) conditions. In practical communication applications, the SNR is often low, resulting in poor identification performance of RF fingerprinting methods. Furthermore, the differences between different operating modes of the same transmitter are small, making identification and classification difficult. Therefore, exploring RF fingerprinting methods for LFM radar under low SNR and multiple operating modes is urgently needed. Summary of the Invention

[0005] This invention addresses the problem of low efficiency in LFM radar signal recognition under low signal-to-noise ratio and multiple operating modes in existing technologies. It provides an MGF (Multi-Fingerprint Generator) radio frequency fingerprinting method for LFM radar. The method uses the MFCC (Multi-Fingerprint Capacitor) speech feature extraction method to generate Mel features for each LFM pulse data. Feature matrices of different lengths are standardized to ensure dimensionality consistency. The standardized feature matrix is ​​the signal fingerprint feature. A single GRU cell is then used to model the signal fingerprint feature, and a fully connected layer is used to identify and classify the modeled features, achieving radio frequency fingerprint recognition and improving the recognition accuracy for difficult-to-classify samples. This invention can effectively extract the physical fingerprint features of LFM pulse signals under low signal-to-noise ratio and multiple signal / operating modes, effectively solving the problem of poor LFM signal recognition in practical application environments.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: an MGF radio frequency fingerprinting method for LFM radar, which uses the speech feature extraction method MFCC to generate Mel features for each LFM pulse data, unifies the length of feature matrices of different lengths to ensure dimensional consistency, and the feature matrix after unification of length is the signal fingerprint feature, uses a single GRU cell to model the signal fingerprint feature, and uses a fully connected layer to identify and classify the modeled features to achieve radio frequency fingerprint recognition.

[0007] As an improvement of the present invention, an MGF radio frequency fingerprinting method for LFM radar includes the following steps:

[0008] S1, Pulse Extraction: After receiving the LFM radar signal, the pulse range is determined using the short-time average energy and the short-time average zero-crossing rate.

[0009] E v (i)=sum(S m (:,i). 2 )

[0010] Z v (i)=sum(S m (1:end-1).*S m (2:end-1))

[0011] EZ V (i)=E v (i)*Z v (i)

[0012] In the formula, S m For the received radar signal, E v Z represents the short-time energy of the signal. v EZ represents the short-time return-to-zero rate of the signal. V This indicates the product of the two, sum represents the summation operation, and i is the current pulse time;

[0013] S2, Feature Extraction: Extract the Mel feature of each pulse using the MFCC feature extraction method;

[0014] F = MFCC(S) m );

[0015] S3, Unify feature length: After obtaining the Mel feature matrix of each pulse in step S2, use the Resize operation to unify the feature matrix length, which is the hardware RF fingerprint of the pulse.

[0016] X = Resize(F);

[0017] S4, GRU model modeling: The radio frequency fingerprint obtained in step S3 is modeled using a single GRU Cell, and a fully connected layer is used for identification and classification.

[0018] S5, Identification: Input LFM signals with different signal-to-noise ratios from the same batch and model as those in step S1 into the model in step S4 to complete the signal identification and classification process.

[0019] As an improvement of the present invention, step 3 further includes: the method for extracting hardware radio frequency fingerprint features includes:

[0020] S31, Transpose the Mel characteristic matrix obtained in step S2:

[0021] M = [F n×d ] T

[0022] In the formula, F is the obtained Mel feature matrix, M is the transposed matrix, the original feature matrix has a dimension of n×d, n is the feature sequence length, and d is the RFF feature dimension;

[0023] S32, For the feature matrix after the transpose of step S31, calculate the length of its longest sequence and the length of the sequence to be copied and truncated:

[0024]

[0025] c = lt × n

[0026] In the formula, l is set as the longest sequence length, t represents the number of complete feature matrices required, and c represents the length of the feature sequence to be truncated.

[0027] S33, Perform a dimension unification operation on the feature matrix:

[0028] X d×l =M×t+M[0:c]

[0029] In the formula, X represents the final hardware RF fingerprint feature, M is the transposed matrix, t represents the number of complete feature matrices required, and c represents the length of the feature sequence to be extracted.

[0030] As another improvement of the present invention, step S4 further includes:

[0031] S41, the obtained hardware RF fingerprint is split by dividing the fingerprint along the feature dimension d:

[0032] X = X1 + X2 + ... + X d

[0033] In the formula, X1 represents the first dimension of the feature matrix, X2 represents the second dimension of the feature matrix, and so on up to d dimensions;

[0034] S42, 36 matrices are used as inputs to the GRU Cells, and are sequentially input into each GRU Cell. This process is repeated until all features are fed into the model, resulting in the final feature h. d :

[0035] h1 = grucell(X1, h0)

[0036] h2 = grucell(X2, h1)

[0037] ...

[0038] h d =grucell(X d ,h d-1 )

[0039] In the formula, h1 represents the first hidden layer feature obtained, X1 represents the first dimension feature matrix, h0 is the initialized hidden layer feature, h0 is a matrix of all zeros, and so on;

[0040] S43, the final hidden layer features h d The data is then fed into a fully connected layer for classification.

[0041] output = Linear(h) d ,num_classes)

[0042] In the formula, output is the final prediction result, Linear represents the fully connected layer, and h d For the final hidden layer features, num_classes represents the number of signals.

[0043] As another improvement of the present invention, the loss function used during model training in step S4 is FocalLoss:

[0044]

[0045] FL(p t )=-α t (1-p t ) γ log(p t )

[0046] In the formula, p t The value p represents the difficulty of classification, where p is the probability value output by the model. p is the probability value when the model predicts correctly (y = 1). t =p, p when the model prediction is wrong t = 1-p, where α represents the weight between categories, γ represents the weight of difficult-to-classify samples in the loss function, and t is the specific category.

[0047] As another improvement of the present invention, in step S1, the threshold values ​​for short-time energy and short-time zero-crossing rate are set to 5.4e+5.

[0048] As a further improvement of the present invention, in the MFCC feature extraction process of step S2, the MFCC dimension is 12, and the pulse feature includes MFCC features and energy, first-order difference and energy, second-order difference and energy, a total of 36 dimensions, that is, the feature dimension is 36.

[0049] Compared with existing technologies: This invention proposes an MGF radio frequency fingerprinting method for LFM radar. (1) The speech feature extraction method MFCC is used to generate Mel features for each pulse data, which reduces the amount of computation and speeds up the computation while retaining the temporal information of the original pulse. (2) Only a single GRU cell is used to model the features, which reduces the model size. Fully connected layers are used to identify and classify the modeled features. Focal Loss is used on the loss function to train the model, which improves the recognition accuracy for samples that are difficult to classify.

[0050] For LFM signals under low signal-to-noise ratio (SNR) conditions, this invention can effectively acquire their hardware RF fingerprint features. The recognition performance is less affected by the SNR, ensuring the robustness of the model and preventing excessive reduction in recognition accuracy due to lower SNR. For signals with multiple operating modes, which are easily confused into the same category by other methods, this invention can effectively distinguish between them, achieving better recognition results. This invention effectively solves the problem of poor LFM signal recognition in practical application environments. Attached Figure Description

[0051] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0052] Figure 2 This is a comparison of the confusion matrix of the recognition results of the Focal Loss loss function used in this invention and the traditional cross-entropy at a signal-to-noise ratio of 0dB.

[0053] Figure 3 The line graph shows the comparison of the recognition rates of the Focal Loss loss function used in this invention with traditional cross-entropy at different signal-to-noise ratios.

[0054] Figure 4 This is a schematic diagram illustrating the recognition accuracy of the present invention under different single working modes. Detailed Implementation

[0055] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0056] Example 1

[0057] A radio frequency fingerprint recognition method for low signal-to-noise ratio and multiple operating modes, the process of which is as follows: Figure 1 As shown, it includes the following steps.

[0058] Step S1: Receive the radio frequency signal, denoted as S, and determine its pulse range using short-time average energy and short-time average zero-crossing rate;

[0059] E v (i)=sum(S m (:,i). 2 )

[0060] Z v (i)=sum(S m (1:end-1).*S m (2:end-1))

[0061] EZ V (i)=E v (i)*Z v (i)

[0062] In the formula, S m For the received radar signal, E v Z represents the short-time energy of the signal. v EZ represents the short-time return-to-zero rate of the signal. V This represents the product of the two, sum represents the summation operation, and i is the current pulse time. The thresholds for short-time energy and short-time zero-crossing rate are set to 5.4e+5.

[0063] Step S2: Extract the Mel feature of each pulse using the MFCC feature extraction method;

[0064] F = MFCC(S)

[0065] The MFCC feature extraction method includes preprocessing operations such as pre-emphasis, framing, and windowing. After preprocessing, a 36-dimensional MFCC feature matrix is ​​obtained through FFT, Mel filter, logarithmic operation, DCT, and dynamic feature extraction. In the MFCC feature extraction algorithm, the MFCC dimension is 12, and the impulse feature includes MFCC features and energy, first-order difference and energy, and second-order difference and energy, totaling 36 dimensions, i.e., d=36.

[0066] After receiving the LFM pulse signal, the receiver uses the speech feature extraction method MFCC to generate Mel features for each pulse's data. This reduces the computational load and speeds up the computation process while preserving the timing information of the original pulse.

[0067] Step S3: After obtaining the Mel feature matrix of each pulse, the Resize operation is used to unify the length of the feature matrix to ensure that the input feature dimensions are consistent, which is the hardware RF fingerprint of the pulse.

[0068] X = Resize(F)

[0069] The operation of unifying the length of the feature matrix in step S3 specifically includes:

[0070] Step S31: For the obtained Mel feature matrix, we first transpose it to ensure the temporal sequence of the model input:

[0071] M = [F n×d ] T

[0072] In the formula, F is the obtained Mel feature matrix, M is the transposed matrix, the original feature matrix has a dimension of n×d, n is the feature sequence length, and d is the RFF feature dimension;

[0073] Step S32: For the transposed feature matrix, calculate the length of its longest sequence and the length of the sequence to be copied and truncated.

[0074]

[0075] c = lt × n

[0076] In the formula, l is set as the longest sequence length, t represents the number of complete feature matrices required, and c represents the length of the feature sequence to be truncated.

[0077] Step S33: After obtaining the number of copies and the truncation length, we perform a dimension unification operation on the feature matrix:

[0078] X d×l =M×t+M[0:c]

[0079] In the formula, X represents the final hardware RF fingerprint feature, M is the transposed matrix, t represents the number of complete feature matrices required, and c represents the length of the feature sequence to be extracted.

[0080] Step S4: In the modeling stage, the obtained radio frequency fingerprint is modeled using a single GRU Cell and classified using a fully connected layer.

[0081] Step S4 of this section involves using a deep learning model to model, identify, and classify LFM signals. Specifically, this includes:

[0082] Step S41: The obtained hardware RF fingerprint is split by dividing the fingerprint along the feature dimension d:

[0083] X = X1 + X2 + ... + Xd

[0084] In the formula, X1 represents the first dimension of the feature matrix, X2 represents the second dimension of the feature matrix, and so on up to d dimensions.

[0085] In step S42, the 36 matrices are used as inputs to the GRU Cells, and are sequentially input into each individual GRU Cell. This process is repeated until all features are fed into the model, resulting in the final feature h. d :

[0086] h1 = grucell(X1, h0)

[0087] h2 = grucell(X2, h1)

[0088] ...

[0089] h d =grucell(X d ,h d-1 )

[0090] In the formula, h1 represents the first hidden layer feature obtained, X1 represents the first dimension feature matrix, h0 is the initialized hidden layer feature, and h0 is a matrix of all zeros.

[0091] Step S43, obtain the final hidden layer features h d Then, we pass it into a fully connected layer for classification:

[0092] output = Linear(h) d ,num_classes)

[0093] In the formula, output is the final prediction result, Linear represents the fully connected layer, and h d For the final hidden layer features, num_classes represents the number of signals.

[0094] During training, the loss function used is Focal Loss:

[0095]

[0096] FL(p t )=-α t (1-p t ) γ log(p t )

[0097] In the formula, p t The value p represents the difficulty of classification, where p is the probability value output by the model. p is the probability value when the model predicts correctly (y = 1). t =p, p when the model prediction is wrongt = 1-p, where α represents the weight between categories, γ represents the weight of difficult-to-classify samples in the loss function, and t is the specific category.

[0098] By using only a single GRU cell to model features, the model size is reduced. Fully connected layers are used to identify and classify the modeled features. Focal Loss is used to train the model on the loss function, which improves the recognition accuracy for samples that are difficult to classify.

[0099] Step S5 involves inputting the LFM signals from the same batch and model as those in Step S1, under different signal-to-noise ratios and multiple operating modes, into the model in Step S4 to complete the signal identification and classification process. This invention can effectively extract the physical fingerprint features of LFM pulse signals under low signal-to-noise ratio and multiple signal operating modes, effectively solving the problem of poor LFM signal identification in practical application environments.

[0100] Test case

[0101] The hardware configuration used in this test case is as follows:

[0102] Graphics Processing Unit (GPU): NVIDIA GeForce RTX 3090, with 24GB of video memory;

[0103] Software development environment: Ubuntu 20.04, Python 3.7, PyTorch 1.7, CUDA 11.3.

[0104] This test case uses an LFM signal dataset, which includes three linear frequency modulated signals of the same type and batch. Each pulse signal has five different operating modes, and 2000 data points were collected for each operating mode of a single pulse signal, for a total of 30,000 data points from fifteen signals as the experimental dataset. In terms of dataset partitioning, 70% of each pulse from each signal was selected as the training set, and the remaining 30% as the test set.

[0105] Due to the similarity between the feature matrix and the image, the model uses the SGD gradient optimization algorithm. The initial learning rate during the training phase is 0.1, and the training epochs are set to 100. When the epochs reach 10, 20, and 50, the learning rate is reduced to 0.1 of the current learning rate, respectively. The batch size is set to 128. The comparison results are shown in the table below.

[0106] Table 1 Test Results

[0107]

[0108] Comparative Example 1 in the table above uses the traditional cross-entropy loss function, with other parameters the same as the test case.

[0109] Comparative Example 2 uses the LAFS method (Reference: Yang Y, Hu A, Xing Y, Yu J, Zhang ZA data-independent radio frequency fingerprint extraction scheme. IEEE Wirel Commun Lett 2021. https: / / doi.org / 10.1109 / LWC.2021.3106396.1–1.)

[0110] Comparative Example 3 used the IMF-DNA method (Reference: Shen, G., Zhang, J., Marshall, A., Peng, L. and Wang, X., 2021, May. Radio frequency fingerprint identification for LoRa using spectrogram and CNN. In IEEE INFOCOM 2021-IEEE Conference on Computer Communications (pp. 1-10). IEEE).

[0111] Figure 2 This is a comparison of the confusion matrices of the recognition results using the Focal Loss method of this invention versus traditional cross-entropy at a signal-to-noise ratio of 0dB; from Figure 2 The confusion matrix reveals that this invention can significantly improve the overall recognition and classification results of radio frequency fingerprint recognition systems for pulse signals that are difficult to classify.

[0112] Figure 3 This is a line graph comparing the recognition rates of the Focal Loss loss function used in this invention with traditional cross-entropy at different signal-to-noise ratios. Figure 4 The figure represents the recognition accuracy of this invention under different single working modes, where the parameters of Focal Loss are α = 0.5 and γ = 1. From... Figure 3 As can be seen, compared to Comparative Example 1, the Example 1 outperforms Comparative Example 1 under different signal-to-noise ratio conditions; compared to Comparative Example 2, under the condition of larger dataset and the same signal-to-noise ratio, the Example 1 can effectively identify and classify low signal-to-noise ratio data, and the identification results are much higher than those of Comparative Example 2; compared to Comparative Example 3, which can only achieve 85.3% recognition accuracy in single working mode, the Example 1 can achieve 91.26% recognition accuracy in the more difficult multi-working mode recognition, and the Example 1 can achieve 100% recognition accuracy in single working modes 1, 3, and 5.

[0113] In summary, this invention provides a radio frequency fingerprint recognition method for low signal-to-noise ratio (SNR) and multiple operating modes. Compared to existing technologies, it can effectively acquire hardware radio frequency fingerprint features without significantly reducing recognition accuracy due to lower SNR. For signals in multiple operating modes, it can effectively distinguish between results that are easily confused with the same type, unlike other methods. Compared to existing methods, it solves the problem of poor recognition in radio frequency fingerprint systems under low SNR and multiple operating modes, effectively improving the recognition performance of LFM radar signals.

[0114] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. An MGF radio frequency fingerprinting method for LFM radar, characterized in that: The MFCC speech feature extraction method is used to generate Mel features for each LFM pulse data. Feature matrices of different lengths are standardized to ensure dimensionality consistency. The standardized feature matrices are the signal fingerprint features. A single GRU cell is used to model the signal fingerprint features, and a fully connected layer is used to identify and classify the modeled features, thus achieving radio frequency fingerprint recognition. The steps include: S1, Pulse Extraction: After receiving the LFM radar signal, the pulse range is determined using the short-time average energy and the short-time average zero-crossing rate; ; ； ; In the formula, For the received radar signal, This represents the short-time energy of the signal. Indicates the short-time return-to-zero rate of the signal. Then it represents the product of the two. This represents the summation operation. The current pulse time; S2, Feature Extraction: Extract the Mel feature of each pulse using the MFCC feature extraction method; ； S3, Unify feature length: After obtaining the Mel feature matrix of each pulse in step S2, use the Resize operation to unify the feature matrix length, which is the hardware RF fingerprint of the pulse. ； S31, Transpose the Mel characteristic matrix obtained in step S2: ; In the formula, F is the obtained Mel feature matrix, M is the transposed matrix, the original feature matrix has a dimension of n×d, n is the feature sequence length, and d is the RFF feature dimension; S32, For the feature matrix after the transpose of step S31, calculate the length of its longest sequence and the length of the sequence to be copied and truncated: ， ; ; In the formula, l is set as the longest sequence length, t represents the number of complete feature matrices required, and c represents the length of the feature sequence to be truncated. S33, Perform a dimension unification operation on the feature matrix: ; In the formula, X represents the final hardware RF fingerprint feature, M is the transposed matrix, t represents the number of complete feature matrices required, and c represents the length of the feature sequence to be extracted. S4, GRU model modeling: The radio frequency fingerprint obtained in step S3 is modeled using a single GRU Cell, and a fully connected layer is used for identification and classification. S5, Identification: Input LFM signals with different signal-to-noise ratios from the same batch and model as those in step S1 into the model in step S4 to complete the signal identification and classification process.

2. The MGF radio frequency fingerprinting method for LFM radar as described in claim 1, characterized in that: Step S4 further includes: S41, the obtained hardware RF fingerprint is split by dividing the fingerprint along the feature dimension d: ; In the formula, X1 represents the first-dimensional feature matrix, X2 represents the second-dimensional feature matrix, and so on up to d dimensions; S42, 36 matrices are used as inputs to the GRU Cells, and are sequentially input into each GRU Cell. This process is repeated until all features are fed into the model, resulting in the final feature h. d : ; ; …… ; In the formula, h1 represents the first hidden layer feature obtained, X1 represents the first-dimensional feature matrix, h0 is the initialized hidden layer feature, h0 is a matrix of all zeros, and so on; S43, the final hidden layer features h d The data is then fed into a fully connected layer for classification. ; In the formula, output is the final prediction result, Linear represents the fully connected layer, and h d For the final hidden layer features, num_classes represents the number of signals.

3. The MGF radio frequency fingerprinting method for LFM radar as described in claim 2, characterized in that: The loss function used during model training in step S4 is Focal Loss: ; ; In the formula, p t The value p represents the difficulty of classification, where p is the probability value output by the model. A correct prediction by the model results in... p t =p, p when the model prediction is incorrect t =1-p, Indicates the weight between categories, This represents the weight of the difficult-to-classify samples in the loss function. For specific categories.

4. The MGF radio frequency fingerprinting method for LFM radar as described in claim 3, characterized in that: In step S1, the threshold values ​​for short-time energy and short-time zero-crossing rate are set to 5.4e+5.

5. The MGF radio frequency fingerprinting method for LFM radar as described in claim 3, characterized in that: In step S2, during MFCC feature extraction, the MFCC dimension is 12, and the pulse feature includes MFCC features and energy, first-order difference and energy, and second-order difference and energy, totaling 36 dimensions, i.e., the feature dimension is 36.

Images