A radio frequency fingerprint recognition method based on image features

Abstract

The present invention belongs to the field of Internet of Things security technology, and specifically relates to a radio frequency fingerprint recognition method based on image features. The present invention first converts the sample sequence of the signal into a time-frequency image, and expands the fingerprint library by extracting image features, which can improve the recognition accuracy of the device. Secondly, the "brightness and darkness" of the pixel points in the time-frequency image obtained after the time-frequency conversion reflects the characteristics of the signal in the time domain and frequency domain, and also reflects the energy intensity of the signal at any time. Moreover, the time-frequency images of wireless device signals have obvious differences in geometric shape and texture features, so it is of great value to study and improve the shape feature and texture feature extraction method of the time-frequency image.

Description

Technical Field

[0001] The present invention belongs to the field of Internet of Things security technology, and specifically relates to a radio frequency fingerprint recognition method based on image features. Background Art

[0002] Over the past two decades of development in communications technology, with the increasing civilian adoption of advanced technologies, wireless communications have become a vital part of everyone's lives. They are used in a wide range of fields, including smart homes, autonomous driving, intelligent communications, and the Internet of Everything. Furthermore, wireless communications are also a key area of ​​focus for military applications, such as collaborative warfare and synchronized combat plans. This trend has made communications security increasingly important. Traditional approaches to ensuring communications security are primarily based on security protocols implemented at the network and transport layers. However, this approach is highly vulnerable to attacks from unauthorized users. Consequently, the wireless communications industry urgently needs new security authentication technologies to ensure user safety.

[0003] In 1995, Choe proposed using the inherent characteristics of communication signals to identify devices. In 2003, the concept of radio frequency fingerprint (RFF) was formally proposed. These characteristics are similar to human fingerprints and are unique characteristics of devices. Even if different devices transmit the same radio frequency signal, the actual transmitted signals will be different due to hardware differences in wireless communication devices. By analyzing the differences in these radio frequency signals, the device sending the signal can be determined. This technology is established at the physical layer and can serve as a powerful supplement to network layer security defense methods.

[0004] Classification methods designed for RF fingerprinting primarily include traditional machine learning and deep learning. The former extracts RF fingerprint features based on expert experience, then feeds them into a corresponding classifier. This trained classifier is then used to confirm the identity of wireless communication devices. The latter typically uses raw IQ signal data directly as model input, applying deep learning models to train large datasets. Existing RF fingerprinting technologies often focus on signal sample sequences, but these sequences only contain feature information in the time dimension of the signal; some RF fingerprint features are not readily apparent in the time domain. Furthermore, as the length of the signal sample sequence increases, both the storage space required to store the signal data and the processing structure of subsequent algorithms are impacted, increasing the complexity of RF fingerprint feature extraction and recognition. Summary of the Invention

[0005] This paper proposes an image-based RF fingerprint analysis method. This method converts a sequence of signal samples into a fixed-size time-frequency image. This method expands the RF fingerprint from the perspective of the time-frequency image, improving device recognition accuracy. Image edges, corners, and regional structures can all be used as feature representations, enriching the RF fingerprint and providing a new approach for device recognition.

[0006] The technical solution of the present invention is:

[0007] A radio frequency fingerprint recognition method based on image features comprises the following steps:

[0008] S1. Convert the acquired radio equipment I / Q signal into a time-frequency diagram using the Choi-Williams distribution.

[0009] S2. Extract shape features and texture features from the obtained time-frequency graph, specifically:

[0010] Extract shape features: Preprocess the time-frequency graph to obtain the first feature extraction time-frequency graph, and then calculate the Zernike moment of the first feature extraction time-frequency graph. The calculation formula of the Zernike moment is:

[0011]

[0012] Where * represents conjugation, n and m are both integers, n is greater than or equal to 0, and |m|≤n. nm The definition of (x,y) is as follows:

[0013]

[0014] Where r, θ, R nm (r) is defined as follows:

[0015]

[0016] By calculating the low-order moment Z 10 , Z 31 and high-order moments Z 43 , Z 53 , Z 66 The feature vector is used as the extracted shape feature, and its expression is:

[0017] Z=[Z 10 Z 31 Z 43 Z 53 Z 66 ] (4)

[0018] Extracting texture features: Preprocessing the time-frequency graph to obtain the second feature extraction time-frequency graph, and using the enhanced uniform local binary pattern to extract texture features from the second feature extraction time-frequency graph, specifically:

[0019] First, the second feature extraction time-frequency map is divided into 8×8 sub-regions;

[0020] Then, the LBP value of each pixel in each sub-region is calculated. With each pixel as the center, a local neighborhood block of size 3×3 is taken. The grayscale value of the central pixel is compared with the grayscale value of each pixel in the neighborhood block. The value greater than the central pixel is recorded as 1, and the value less than the central pixel is recorded as 0. In this way, the 8-bit LBP value of each pixel can be obtained. The calculation formula is as follows:

[0021]

[0022] Where P represents the pixel point, R represents the neighborhood radius, g C Indicates the gray value of the center pixel, g P represents the grayscale value of point P in the neighborhood, and s(·) is defined as follows:

[0023]

[0024] Secondly, in the process of calculating the LBP value of each sub-region pixel, the value of 0, 1 with a change frequency less than 2 is recorded and retained. Then the frequency of this value in the corresponding sub-region is calculated and expressed in the form of a frequency histogram. For n LBP values, the frequency histogram vector value of each value is represented by v1, v2…v n Represented by, and then normalized, with x i Represents the vector value of each frequency histogram after normalization. When normalizing, L2 normalization is used, and its calculation formula is:

[0025]

[0026] Finally, the vector value calculated above is amplified to obtain the EULBP feature vector. The EULBP vectors of each sub-region are connected to obtain the texture feature EULBP feature vector of the entire time-frequency graph. The EULBP feature vector is defined as:

[0027]

[0028] S3. Connect the Zernike moment features and EULBP features of the time-frequency graph horizontally along the columns to form a single feature matrix, which includes the shape features and texture features of the time-frequency graph.

[0029] S4. Use a support vector machine as a classifier to classify and identify wireless devices. In a support vector machine, the robustness and generalization of the classification results are guaranteed by maximizing the distance between the support vector and the hyperplane. The expression of the optimal hyperplane in the support vector machine is as follows:

[0030]

[0031] Where w represents the direction vector; c is the penalty factor, which indicates the tolerance to erroneous samples; ε k represents the slack variable.

[0032] When implementing a support vector machine, the penalty factor c is set to 10, the kernel parameter gamma is set to 1, and a radial basis kernel function is used to identify and classify wireless devices. For the problem of classifying K wireless devices, K binary classifiers are constructed. Each classifier considers one category as a positive example and the other K-1 categories as negative examples. A new RF fingerprint feature is then fed into each of the K classifiers for prediction. The prediction from the classifier with the highest score is selected as the final classification result.

[0033] Furthermore, the time-frequency graph preprocessing includes grayscale conversion and Wiener filtering denoising in sequence.

[0034] The beneficial effects of the present invention are as follows: first, the signal sample sequence is converted into a time-frequency image, and by extracting image features to expand the fingerprint library, the device's recognition accuracy can be improved. Secondly, the brightness and darkness of the pixels in the time-frequency image obtained after the time-frequency conversion reflects the characteristics of the signal in the time and frequency domains, and also reflects the energy intensity of the signal at any time. Moreover, the time-frequency images of wireless device signals have significant differences in geometric shape and texture features, so researching and improving the methods for extracting shape and texture features from time-frequency images is of great practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the radio frequency fingerprint recognition method of the present invention.

[0036] Figure 2 This is a flow chart of the preprocessing of time-frequency graph feature extraction of the present invention.

[0037] Figure 3 The simulation signal flow chart generated by MATLAB in this embodiment.

[0038] Figure 4 This is the time-frequency diagram of the noisy MATLAB simulation signal in this embodiment.

[0039] Figure 5: These are the feature extraction preprocessing images in this embodiment, where (a) is the original time-frequency image, (b) is the grayscale time-frequency image, and (c) is the denoised time-frequency image.

[0040] Figure 6 This is a comparison chart of the recognition accuracy of the radio frequency fingerprint algorithm in the embodiment. DETAILED DESCRIPTION

[0041] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0042] The overall process of the present invention is as follows Figure 1 As shown in Figure 1, the signal samples are first converted into time-frequency images using the Choi-Williams distribution algorithm.

[0043] For the signal x(t), its Choi-Williams distribution is defined as:

[0044]

[0045] Where A x The definition of (τ,v) is as follows:

[0046]

[0047] In the above formula, φ(τ,v) is the kernel function, which is defined as follows:

[0048]

[0049] By using the Choi-Williams distribution algorithm to perform time-frequency analysis on the original I / Q signal, the signal can be converted from a one-dimensional time domain format to a two-dimensional time-frequency domain format, exposing more signal features.

[0050] Then, shape features and texture features are extracted based on the time-frequency graph:

[0051] First, the time-frequency image is preprocessed. According to the order of preprocessing, it can be divided into two processes: grayscale and Wiener filter denoising. The specific process is as follows: Figure 2 As shown. The weighted average method is used to grayscale the original time-frequency image. The calculation formula is as follows:

[0052] Y=0.11R+0.58G+0.3B (13)

[0053] After obtaining the grayscale image, the Wiener filter algorithm is used to smooth the noise and retain the effective information of the image. The Wiener filter minimizes the mean square error criterion between the original image and the denoised image, and achieves image denoising through linear transformation. The specific algorithm is:

[0054] H(f)=G(f)*(|G(f)|2 +K(f)) -1 (14)

[0055] where H(f) is the frequency response of the filter, G(f) is the Fourier transform, K(f) is the power spectral density of the noise, and * denotes the complex conjugate.

[0056] Then, the Zernike moment of the shape feature of the target image is extracted. The Zernike moment is defined as:

[0057]

[0058] By combining the low-order moments Z of the Zernike moments 10 , Z 31 and high-order moments Z 43 , Z 53 , Z 66 To form the shape feature vector, its low-order moment and high-order moment represent the overall shape of the image and the detailed information of the image respectively. The expression of the feature vector is:

[0059] Z=[Z 10 Z 31 Z 43 Z 53 Z 66 ] (15)

[0060] Secondly, extract the texture features of the target image. The specific steps are as follows:

[0061] First, the preprocessed time-frequency map is divided into 8×8 sub-regions.

[0062] Then, the LBP value of each pixel in each sub-region is calculated. With each pixel as the center, a local neighborhood block of size 3×3 is taken. The grayscale value of the central pixel is compared with the grayscale value of each pixel in the neighborhood block. The value greater than the central pixel is recorded as 1, and the value less than the central pixel is recorded as 0. In this way, the 8-bit LBP value of each pixel can be obtained. The calculation formula is as follows:

[0063]

[0064] Where P represents the pixel point, R represents the neighborhood radius, g C Indicates the gray value of the center pixel, g P represents the grayscale value of point P in the neighborhood, and s(·) is defined as follows:

[0065]

[0066] In the process of calculating the LBP value of each sub-region pixel, the value of 0, 1 with a change frequency less than 2 is recorded and retained. Then the frequency of this value in the corresponding sub-region is calculated and expressed in the form of a frequency histogram. For n LBP values, the frequency histogram vector value of each value is represented by v1, v2…v n Represents, and then normalizes it and uses x i Represents the vector value of each frequency histogram after normalization. When normalizing, L2 normalization is used, and its calculation formula is:

[0067]

[0068] Finally, the vector value calculated above is amplified to obtain the EULBP feature vector. The EULBP vectors of each sub-region are connected to obtain the texture feature EULBP feature vector of the entire time-frequency graph. The EULBP feature vector is defined as:

[0069]

[0070] After extracting the shape feature Zernike moment and texture feature EULBP of the time-frequency graph respectively, the two feature vectors are horizontally connected along the columns to form a single feature matrix. This feature matrix is ​​the target feature vector for wireless device classification and identification.

[0071] Then support vector machine is used as a classifier to realize classification and identification of wireless devices.

[0072] The support vector machine uses a radial basis kernel function to identify and classify wireless devices. A combination of grid search and cross-validation is used to determine the penalty factor c and kernel function parameter g. The optimal recognition performance is achieved when c = 10 and g = 1, which is obtained by training the support vector machine using a time-frequency graph.

[0073] Example

[0074] The signals used in this example are simulated signals generated by modeling and simulating a group of transmitters with different parameters using MATLAB.

[0075] In this embodiment, in order to represent different wireless communication devices, an OFDM link is used in the simulation. The number of subcarriers is 512, the modulation mode is QPSK, the carrier frequency is set to 420 kHz, and the channel is assumed to be additive white Gaussian noise (AWGN). The RF impairment parameters are shown in Table 1:

[0076] Table 1 RF damage parameter setting table

[0077]

[0078] In real-world environments, wireless devices cannot avoid noise interference during signal transmission. To verify the reliability of RF fingerprint feature extraction based on time-frequency graphs, noise with different signal-to-noise ratios was added when generating I / Q samples to evaluate performance under different signal-to-noise ratios.

[0079] During the data preparation phase, MATLAB was used to generate signal datasets for four devices under varying signal-to-noise ratios. Each dataset consisted of 1,000,000 I / Q samples. To generate time-frequency plots for feature extraction, the Choi-Williams time-frequency transform was performed on groups of 1,000 samples. 1,000 64×64 time-frequency plots were generated for each device. In this example, the Choi-Williams distribution parameter σ was set to 0.1 to minimize the impact of cross-terms on signal recognition.

[0080] In order to better extract useful features from the time-frequency image, the time-frequency image needs to be preprocessed first.

[0081] In this embodiment, the preprocessing of time-frequency graph shape feature extraction still adopts the following method: Figure 2 The process shown in the figure makes the contour of the signal component in the image smoother. The result of the preprocessing process of the embodiment is as follows: Figure 5 Then, the Zernike moment feature of the time-frequency graph is extracted to obtain the shape feature information of the time-frequency image. The calculation formula of the Zernike moment feature is:

[0082]

[0083]

[0084] Subsequently, the texture feature extraction of the time-frequency graph is performed. First, the pre-processed time-frequency graph is divided into 8×8 sub-regions, and the LBP value of the pixel points in each sub-region is calculated. The calculation formula is as follows:

[0085]

[0086]

[0087] In the process of calculating the LBP value of each sub-region pixel, record and retain the value of 0, 1 with a change frequency less than 2. Then calculate the frequency of this value in the corresponding sub-region and express it in the form of a frequency histogram. For n LBP values, the frequency histogram vector value of each value is represented by v1, v2…v n Represented by, and then normalized, with x i Represents the vector value of each frequency histogram after normalization. When normalizing, L2 normalization is used, and its calculation formula is:

[0088]

[0089] Finally, the vector value calculated above is amplified to obtain the EULBP feature vector. The EULBP vectors of each sub-region are connected to obtain the texture feature EULBP feature vector of the entire time-frequency graph. The EULBP feature vector calculation formula is:

[0090]

[0091] In this embodiment, after extracting the shape feature Zernike moment of the time-frequency graph and the texture feature EULBP respectively, the two feature vectors are horizontally connected along the columns to form a single feature matrix. The feature matrix is ​​the target feature vector for wireless device classification and identification.

[0092] In this embodiment, to identify and classify wireless devices in four categories, four binary classifiers are constructed. Each classifier uses one category as a positive example and the other three categories as negative examples. A feature vector is then fed into each of the four classifiers for prediction, and the prediction result from the classifier with the highest score is selected as the final classification result. Each classifier uses a radial basis kernel function to identify and classify wireless devices. The penalty factor c = 10 and the kernel function parameter g = 1 are used.

[0093] Finally, we simulated the time-frequency graphs under different signal-to-noise ratios. The four devices generated 4,000 time-frequency graphs under different signal-to-noise ratios. 80% of the samples were used to train the model and 20% were used for testing. The recognition accuracy comparison results are shown in the figure below. Figure 6 shown.

[0094] Through the above design, the present invention first converts a signal sample sequence into a time-frequency image and extracts the corresponding image features to expand the fingerprint library. Secondly, based on the characteristics of the wireless device signal time-frequency image, the time-frequency image is preprocessed to extract the shape and texture features of the time-frequency image. These shape and texture features are combined to construct a more discriminative RF fingerprint. Finally, this feature is input into a support vector machine for training to achieve wireless device recognition and classification, further improving the accuracy of wireless device identification.

Claims

1. A radio frequency fingerprint recognition method based on image features, characterized in that: The following steps are involved: S1. Convert the acquired radio equipment I / Q signal into a time-frequency diagram using the Choi-Williams distribution. S2. Extract shape features and texture features from the obtained time-frequency graph, specifically: Extract shape features: Preprocess the time-frequency graph to obtain the first feature extraction time-frequency graph, and then calculate the Zernike moment of the first feature extraction time-frequency graph. The calculation formula of the Zernike moment is: By calculating the low-order moment Z 10 , Z 31 and high-order moments Z 43 , Z 53 , Z 66 The feature vector is used as the extracted shape feature, and the expression is: Z=[Z 10 WITH 31 WITH 43 WITH 53 WITH 66 ] Extracting texture features: Preprocessing the time-frequency graph to obtain the second feature extraction time-frequency graph, and using the enhanced uniform local binary pattern to extract texture features from the second feature extraction time-frequency graph, specifically: The second feature extraction time-frequency map is divided into 8×8 sub-regions; Calculate the LBP value of the pixel in each sub-region: take each pixel as the center, take a local neighborhood block of size 3×3, compare the grayscale value of the central pixel with the grayscale value of each pixel in the neighborhood block, record the value greater than the central pixel as 1, and record the value less than the central pixel as 0, and obtain the 8-bit LBP value of each pixel: Where P represents the pixel point, R represents the neighborhood radius, g C Indicates the gray value of the center pixel, g P represents the grayscale value of point P in the neighborhood, and s(·) is defined as follows: In the process of calculating the LBP value of each sub-region pixel, record and retain the value of 0, 1 with a change frequency less than 2, then calculate the frequency of this value in the corresponding sub-region and express it in the form of a frequency histogram. For n LBP values, the frequency histogram vector value of each value is represented by v1, v2…v n Represented by, and then normalized, with x i Represents the normalized vector value of each frequency histogram; when normalizing, L2 normalization is used: The calculated vector value is amplified to obtain the EULBP feature vector, and the EULBP vectors of each sub-region are connected to obtain the texture feature EULBP feature vector of the entire time-frequency graph: S3, horizontally connecting the Zernike moment features and EULBP features of the time-frequency graph along the columns to form a single feature matrix, which includes the shape features and texture features of the time-frequency graph; S4. Use support vector machine as a classifier and use the obtained feature matrix to realize classification and identification of wireless devices.

2. The radio frequency fingerprint recognition method based on image features according to claim 1, characterized in that: The preprocessing includes grayscale conversion and Wiener filtering denoising in sequence.

Images