A radio frequency fingerprinting method based on a residual network model

Abstract

The present application belongs to the technical field of Internet of Things security, and particularly relates to a radio frequency fingerprint identification method based on a residual network model. The present application first converts a wireless device signal into a time-frequency graph of a fixed size, extracts a radio frequency fingerprint feature with higher distinguishability, and solves the problem that the sample length will affect the subsequent algorithm structure and processing complexity when a signal sample sequence is used as the extraction object of the radio frequency fingerprint. Then, a deep learning model CSAM-ResNet is designed and implemented for the time-frequency graph of the wireless device signal, automatic extraction of the radio frequency fingerprint is realized, details and key feature information in the time-frequency graph are fully mined, and the identification capability of the wireless device in a complex electromagnetic environment is further improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of Internet of Things security, and particularly relates to a radio frequency fingerprint identification method based on a residual network model. BACKGROUND

[0002] The first evolution standard of 5G was released in 2020, which added two application scenarios of high-speed low-latency communication and large-scale device connection on the basis of the original, further promoting the rapid development of the Internet of Things. The GSM Association estimates that more than 1.7 billion people will use 5G networks by 2025, which means that in the future, there will be a large number of wireless devices connected to provide services for various new Internet of Things applications. As the Internet of Things applications become more common, security vulnerabilities hidden in device communication also come with them. How to provide reliable identification and verification mechanisms to ensure that the communication between wireless devices is trustworthy is one of the challenges that the Internet of Things technology must face.

[0003] Most devices in wireless networks communicate in an open environment, and the information transmitted in the air can be received and obtained by all users in the network, so how to effectively protect the communication security of wireless networks has always been a topic of discussion and research by experts and scholars. The traditional implementation method of high-level security protection of wireless networks is mostly complex and resource-intensive, which is not suitable for all devices in the Internet of Things, especially those small in size and limited in resources and power. Therefore, the radio frequency fingerprint identification technology based on the physical layer has been widely studied and concerned since it was proposed. By extracting the radio frequency fingerprint contained in the electromagnetic wave emitted by the transmitter, the identity authentication of wireless devices can be realized. The radio frequency fingerprint is a unique feature of each device, which is generated by the inherent randomness of electronic components during manufacturing and use, especially the errors of analog components in the radio transmission chain, which are the basis for generating radio frequency fingerprints. By calculating and analyzing these errors, the physical layer unique features of wireless communication devices can be extracted as their own identity credentials, thereby ensuring the security and reliability of communication.

[0004] Nowadays, with the integration of artificial intelligence technology and radio frequency fingerprint technology, deep learning algorithms can be used to automatically extract the pattern features of radio waves, avoiding the experience-based manual feature extraction link. Currently, the radio frequency fingerprint technology based on deep learning mostly uses time-domain signal sample sequences as model inputs, but this technology has two problems. First, some radio frequency fingerprint features are not obvious in the time domain. Second, when using signal sample sequences as input for model training, a large amount of storage space resources are needed for data storage, and the training is also difficult and time-consuming due to the large input. SUMMARY

[0005] In view of the above problems, the application provides a radio frequency fingerprint identification method based on a residual network model, which takes a signal image of a wireless device with a fixed size as an input of a deep learning model, and solves the problems of single signal time domain radio frequency fingerprint feature and increasing model complexity with increasing signal sample sequence.

[0006] The technical scheme of the application is:

[0007] A radio frequency fingerprint identification method based on a residual network model comprises the following steps:

[0008] S1, converting the obtained original signal sample into a time-frequency graph, specifically, dividing the original signal sample with a length of N into N / L sub-sequences with a length of L, and performing Choi-Williams transformation on each sub-sequence to obtain a time-frequency graph;

[0009] S2, constructing a residual network model, specifically, replacing the first standard convolutional layer in the original residual block with a dilated convolutional layer with an expansion rate of 2 to construct a dilated residual block, and inserting a hybrid attention module after each dilated residual block; the hybrid attention module comprises a channel attention module and a spatial attention module, wherein the channel attention module first performs average pooling and maximum pooling on the input feature map to obtain two channel feature descriptions, then adds and fuses the two channel feature descriptions to obtain a fused feature description, inputs the fused feature description into a multilayer perceptron for training and learning, and obtains a one-dimensional channel attention matrix after multilayer perceptron operation, and multiplies the channel attention matrix with the corresponding position elements of the input feature map to obtain a channel attention feature map; the spatial attention module is used to make up for the information ignored by the channel attention, first compresses the input feature map through a convolutional layer, then uses a channel dimension global average pooling to divide two branches corresponding to time and frequency feature vectors in the horizontal and vertical directions, then processes the feature maps in two dimensions through a convolutional layer to obtain attention vectors in the horizontal and vertical directions, and uses a Sigmoid function to normalize the spatial weight coefficient, and finally performs element-wise point multiplication operation on the input feature map and the two attention vectors to obtain a weighted spatial attention feature map;

[0010] S3, training the residual network model constructed in S2 using the time-frequency graph obtained in S1 to obtain a trained residual network model;

[0011] S4, converting the collected radio frequency fingerprint signal into a time-frequency graph and inputting it into the trained residual network model for identification.

[0012] Further, the size of the time-frequency graph is 64x64.

[0013] Further, the kernel size of the dilated convolution layer in the hole residual block is 3x3, the expansion rate is 2, the receptive field size is 7, the kernel size of the standard convolution layer is 3x3, the padding is 1, and the step is 1; After each convolution operation, normalization and nonlinear activation operation will be performed; the dilated convolution model is:

[0014]

[0015] wherein G i represents the feature map of the i-th layer, F i represents the convolution layer of the i-th layer, p represents the point in the generated feature map after convolution, a represents the point in the original feature map G i , and b represents the point in the convolution layer F i , and m represents the expansion rate.

[0016] Further, the channel attention matrix is:

[0017] M(X)=Sigmoid(MLP(AvgPool(X))+MLP(MaxPool(X)))

[0018] wherein X represents the input feature map, and MLP represents a multi-layer perceptron.

[0019] Further, the one-dimensional attention vector in the horizontal direction obtained by the spatial attention module is:

[0020]

[0021] The one-dimensional attention vector in the vertical direction is:

[0022] M H (X)=Sigmoid(GAvgpool(W θ X)W α )

[0023] wherein W θ , W α respectively represent the kernel weight parameters of 3x3, 1x7 and 7x1;

[0024] The weighted spatial attention feature map X" is:

[0025] X"=XM W (X)M H (X).

[0026] The beneficial effects of the present application are that the present application firstly converts the wireless device signal into a fixed-size time-frequency graph, extracts a radio frequency fingerprint feature with higher discrimination, solves the problem that the sample length will affect the subsequent algorithm structure and processing complexity when the signal sample sequence is taken as the radio frequency fingerprint extraction object. Then, the deep learning model CSAM-ResNet is designed and implemented for the wireless device signal time-frequency graph, realizing the automatic extraction of the radio frequency fingerprint, fully mining the detailed information and key feature information in the time-frequency graph, and further improving the recognition ability of the wireless device in the complex electromagnetic environment. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The residual network main framework of the present application.

[0028] Figure 2 The structure diagram of the hollow residual block I in the present embodiment.

[0029] Figure 3 The structure diagram of the hollow residual block II in the present embodiment.

[0030] Figure 4 The structure diagram of the channel attention module in the present embodiment.

[0031] Figure 5 The structure diagram of the spatial attention module in the present embodiment.

[0032] Figure 6 The structure diagram of the mixed attention module in the present embodiment. DETAILED DESCRIPTION

[0033] The present application will be described in detail below in combination with the drawings and embodiments.

[0034] Embodiment

[0035] The present example includes the following steps:

[0036] Firstly, the original signal sample with a length of N needs to be divided into N / L sub-sequences with a length of L before generating the signal image. Then, Choi-Williams transformation is performed on each sample sub-sequence (s1, s2…s L ), and a time-frequency graph is drawn, and then a color image of the sub-sequence sample is obtained. During image conversion, the size of each time-frequency graph is fixed to 64x64.

[0037] The convolutional neural network in deep learning can extract sufficient high-level semantic features of the image through multi-layer convolution structure, and has excellent performance in pattern recognition and feature fitting. Wireless device recognition is essentially a kind of pattern recognition, and through the conversion of the signal sequence into the form of time-frequency graph, a radio frequency fingerprint feature with higher discrimination can be extracted.

[0038] In this embodiment, a network model CSAM-ResNet suitable for wireless device signal time-frequency diagram classification and identification is designed, and the model structure is as shown in Figure 1 The specific network structure parameters are shown in Table 1 below.

[0039] Table 1 CSAM-ResNet network structure parameters

[0040]

[0041] The network model replaces the ordinary convolution in the residual module with a hollow convolution, expands the receptive field, and enables the network model to extract more effective radio frequency fingerprint features. In addition, a hybrid attention mechanism is added, so that the network model can pay attention to more useful feature information in the wireless device signal time-frequency diagram. The specific structure is: Conv+MaxPool+DBlock1+DBlock1+CSAM+DBlock2+DBlock1+CSAM+DBlock2+DBlock1+CSAM+DBlock1+DBlock1+CSAM+AvgPool+FC; wherein, DBlock is a hollow convolution residual block, the input data of the hollow convolution residual block is sequentially subjected to a hollow convolution layer, a batch normalization layer, a ReLU layer, a convolution layer, a batch normalization layer, and then fused with the input data, and then output after the ReLU layer; CSAM is a channel attention mechanism module.

[0042] The hollow convolution residual block structure of the application is as shown in Figure 2 and Figure 3 The hollow residual block contains two convolution layers, the first convolution layer is a hollow convolution layer with a convolution kernel size of 3x3 and a dilation rate of 2; the second convolution layer is a standard convolution layer with a convolution kernel size of 3x3. The difference between the hollow residual block II and the hollow residual block I is that the input of the hollow residual block II is also subjected to a 1x1 convolution and a BN layer before being fused with the output of the 3x3 standard convolution layer.

[0043] In a neural network, the receptive field of each neuron refers to its response range to the local area of the input data. In a convolutional neural network, the neurons in each convolutional layer usually apply a fixed-size convolution kernel, thereby only processing the information in a specific area of the input image. The hollow convolution adjusts the receptive field by adding a hollow, which does not need to introduce additional parameters to increase the calculation amount, and preserves the detail information of the image, and can conveniently extract features of different sizes in the feature map. The convolution kernel size of the first layer of the hollow residual module is 3x3, and the dilation rate is 2, which is equivalent to a 5x5 convolution kernel, and the receptive field size is 7. The convolution kernel size of the second layer of the standard convolution is 3x3, the padding is 1, and the step is 1. After each convolution operation, a normalization and nonlinear activation operation are performed, and the hollow convolution model is as follows:

[0044]

[0045] Where G i represents the feature map of the i-th layer, F i Represents the i-th convolution layer, p represents the point in the feature map generated after convolution, and a represents the original feature map G i Points inside, b represents the convolutional layer F i The points within the convolution kernel are represented by m, and m represents the dilation rate. The calculation formula for dilated convolution shows that by increasing the spacing in the convolution kernel, dilated convolution can cause the convolution kernel to skip some pixels on the input feature map for convolution operations. When the dilation rate is larger, the sampling range of the dilated convolution is larger, and the number of sampling points per unit range is smaller.

[0046] Dilated convolution achieves a larger receptive field with smaller parameters and computational complexity, playing a key role in image recognition tasks. By adding dilated convolution to the residual module, it is possible to extract key RF fingerprint features from the signal's time-frequency graph while preserving image details.

[0047] In this embodiment, a channel attention mechanism is used between feature maps to model the importance of each feature, thereby further avoiding the influence of background noise of the time-frequency map on feature extraction.

[0048] like Figure 4 The following is the structure diagram of the channel attention module. The channel attention matrix expression is as follows:

[0049] M(X)=Sigmoid(MLP(AvgPool(X))+MLP(MaxPool(X)))

[0050] First, perform average pooling and maximum pooling on the input feature map to obtain two feature descriptions of size 3×3 about the channel. During average pooling, the input feature map is divided into several sub-regions, and then the average value of each sub-region is calculated. The expression is as follows:

[0051]

[0052] Average pooling can extract useful information from images and aggregate feature information.

[0053] Max pooling also divides the input into several sub-regions, but it takes the maximum value of each sub-region. Its definition is as follows:

[0054]

[0055] Max pooling extracts the maximum value in each sub-region and inputs it into the next layer, highlighting the important information in the feature map, removing a large amount of redundant information, and achieving space compression.

[0056] The feature descriptors obtained by maximum pooling and average pooling are then added and fused to obtain a feature description of size C×1×1, which combines the important features of the input image and useful background information. This feature description is then input into a multi-layer perceptron for training and learning. After the multi-layer perceptron operation, a one-dimensional channel attention matrix can be obtained:

[0057] M(X)=Sigmoid(MLP(AvgPool(X))+MLP(MaxPool(X)))

[0058] Where X represents the feature map. After obtaining the feature attention matrix through the feature channel, it is normalized using the Sigmoid function to obtain the channel weight coefficient. Finally, the feature map input to the channel attention module is multiplied element-wise by this channel weight coefficient to obtain the refined channel attention feature map.

[0059] The channel attention mechanism acts between feature maps, which can model the importance of each feature, further avoid the influence of background noise in the time-frequency map on feature extraction, and allow the model to pay more attention to the effective information in the time-frequency map.

[0060] In this embodiment, a spatial attention mechanism is used to weight features at different locations in the neural network, allowing the network to focus more on task-relevant areas and improving its performance. Spatial attention can compensate for information overlooked by channel attention. Its main principle is to exploit the correlation between pixels at different locations in the image to extract key information in the spatial dimension.

[0061] The time-frequency diagram represents the change of time in the horizontal direction and the change of signal frequency in the vertical direction. In the design of the spatial attention module, in order to take into account the feature information of both time and frequency dimensions, two branches are used to process them separately, so that the network model can pay more attention to the key features of the time-frequency diagram.

[0062] like Figure 5 The following is the structure of the spatial attention module. First, the input feature map X is compressed using a 3×3 convolution layer. Then, global average pooling of one channel dimension is used in the horizontal and vertical directions to separate the feature vectors into two branches corresponding to time and frequency. The calculation formula of global average pooling is as follows:

[0063]

[0064] wherein y k represents the output feature vector, and |R| represents the total number of elements on the feature map.

[0065] Then, the two-dimensional feature map is input into a 1*7 and 7*1 convolution layer respectively to obtain the attention vectors in the horizontal and vertical directions, and a Sigmoid function is used for normalization to obtain the spatial weight coefficient.

[0066] The expression of the one-dimensional attention vector in the horizontal direction is:

[0067]

[0068] The expression of the one-dimensional attention vector in the vertical direction is:

[0069] M H (X) = Sigmoid(GAvgpool(W θ X)W α )

[0070] wherein W θ , W α respectively represent the convolution kernel weight parameters of 3*3, 1*7 and 7*1.

[0071] After obtaining the spatial attention vectors in the two directions through the above calculation operation, the input feature map X is element-wise multiplied with the two attention vectors to obtain the weighted spatial attention feature map X'', and the expression is:

[0072] X'' = XM W (X)M H (X)

[0073] The spatial attention mechanism weights the features in different positions in the neural network, so that the network can pay more attention to the regions related to the task, improve the performance of the network, and extract radio frequency fingerprint features with higher discrimination.

[0074] The present application provides a network model suitable for wireless device signal time-frequency map classification and recognition. In order to extract more effective radio frequency fingerprint features, the structure of the residual module is modified, the first layer standard convolution of the original residual module is replaced with a hollow convolution, the receptive field is expanded, and radio frequency fingerprint features with higher discrimination are extracted. Channel attention mechanism and spatial attention mechanism are added to enable the network model to pay attention to more useful feature information and improve the recognition accuracy of the network model. Compared with the traditional radio frequency fingerprint recognition algorithm, the improved CSAM-ResNet network model has better generalization ability and recognition effect.

Claims

1. A radio frequency fingerprinting method based on a residual network model, characterized in that, The method comprises the following steps: S1, converting the obtained original signal sample into a time-frequency graph, specifically: dividing the original signal sample with a length of N into NL sub-sequences with a length of L, and performing Choi-Williams transformation on each sub-sequence to obtain a time-frequency graph; S2, constructing a residual network model, specifically: on the basis of the residual network model ResNet18, replacing the first standard convolutional layer in the original residual block with a dilated convolutional layer with a dilation rate of 2 to construct a dilated residual block, and inserting a hybrid attention module after each dilated residual block; the hybrid attention module comprises a channel attention module and a spatial attention module, wherein the channel attention module first performs average pooling and maximum pooling on the input feature map to obtain two channel feature descriptions, then adds and fuses the two channel feature descriptions to obtain a fused feature description, inputs the fused feature description into a multilayer perceptron for training and learning, and obtains a one-dimensional channel attention matrix after multilayer perceptron operation, and multiplies the channel attention matrix with the corresponding position elements of the input feature map to obtain a channel attention feature map; the spatial attention module is used to make up for the information ignored by the channel attention, first compresses the input feature map through a convolutional layer, then uses a channel dimension global average pooling to divide two branches corresponding to time and frequency feature vectors in the horizontal and vertical directions, then processes the feature maps in two dimensions through a convolutional layer to obtain attention vectors in the horizontal and vertical directions, and uses a Sigmoid function to normalize the spatial weight coefficient, and finally performs element-wise point multiplication operation on the input feature map and the two attention vectors to obtain a weighted spatial attention feature map; S3, training the residual network model constructed in S2 using the time-frequency graph obtained in S1 to obtain a trained residual network model; S4, converting the collected radio frequency fingerprint signal into a time-frequency graph and inputting it into the trained residual network model for identification. 2.The RF fingerprinting method based on a residual network model according to claim 1, wherein, The size of the time-frequency graph is 64x64. 3.The RF fingerprinting method based on a residual network model according to claim 2, wherein, The kernel size of the dilated convolutional layer in the dilated residual block is 3x3, the dilation rate is 2, the receptive field size is 7, the kernel size of the standard convolutional layer is 3x3, the padding is 1, and the step is 1; after each convolution operation, normalization and nonlinear activation operation are performed; the dilated convolution model is: wherein G i represents the feature map of the i-th layer, F i represents the i-th layer convolution layer, p represents a point in the generated feature map after convolution, a represents a point in the original feature map G i , b represents a point in the convolution layer F i , and m represents the dilation rate.

4. The method of claim 3, wherein the residual network model is a ResNet-18 model. The channel attention matrix is: M(X) = Sigmoid(MLP(AvgPool(X)) + MLP(MaxPool(X))) Where X represents the input feature map, and MLP represents the multilayer perceptron.

5. The method of claim 4, wherein the residual network model is a ResNet-18 model. The one-dimensional attention vector in the horizontal direction obtained by the spatial attention module is: The one-dimensional attention vector in the vertical direction is: M H (X) = Sigmoid(GAvgpool(W θ X)W α ) wherein W θ , W α respectively represent 3x3, 1x7 and 7x1 convolution kernel weight parameters; The weighted spatial attention feature map X'' is: X" = XM W (X)M H (X).

Images