Enhanced equipment fingerprint identification system and method based on time-frequency domain feature fusion

Abstract

The invention discloses an enhanced equipment fingerprint identification system and method based on time-frequency domain feature fusion. The system is composed of a signal processing module, a feature extraction module, a feature fusion module and an equipment identification module. The method comprises the following steps: firstly, splitting IQ into three time domain features, converting the IQ into a frequency domain and a time-frequency domain by using corresponding FFT and STFT formulas, and extracting corresponding features; respectively inputting the features of the time domain, the frequency domain and the time-frequency domain into a feature extraction module for feature extraction; the extracted features are sent to a feature fusion module, and final fusion features are generated; and finally, the fusion features are transmitted to a classification model, and fingerprint matching and identification of the equipment are completed. By adopting the enhanced equipment fingerprint identification system and method based on time-frequency domain feature fusion, the security, accuracy and robustness of the Internet of Things equipment are improved, the sensitivity of the system to noise interference is reduced, the equipment fingerprint extraction cost is reduced, and the higher identification accuracy is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of device fingerprint identification, and particularly relates to an enhanced device fingerprint identification system and method based on time-frequency domain feature fusion. BACKGROUND

[0002] With the continuous development of wireless network and mobile communication technologies such as electronic devices, Internet of Things (IOT) technology is widely used and plays an increasingly important role in transportation, medical treatment, home and other fields. With the vigorous development of the Internet of Things, billions of Internet of Things devices are deployed everywhere to support computing and communication. However, these devices are extremely vulnerable to malicious attacks when operating in a wireless environment, making wireless communication security a serious challenge.

[0003] Traditional identity verification methods usually rely on software addresses (such as Internet Protocol IP or Media Access Control MAC) and use classical cryptography to achieve identity authentication, but such methods often require complex calculations and protocols, which are not ideal for Internet of Things devices with limited power consumption and computing resources. Therefore, device fingerprint identification technology obtains the unique physical characteristics of wireless devices by analyzing the defects of radio frequency circuit waveforms, which is used to identify and distinguish different device identities. Specifically, the hardware of wireless devices will produce slight differences in oscillators, power amplifiers, and analog filters during the manufacturing process, so even devices of the same model and from the same production line will have their own hardware characteristics.

[0004] However, due to the complex and variable electromagnetic environment of the wireless space, a single device fingerprint feature is difficult to effectively identify the device identity and has low robustness. At the same time, it is also very complex to analyze the relationship between multiple fingerprint features, and traditional neural networks are difficult to fully learn and fuse multiple fingerprint features.

[0005] Therefore, in order to overcome the above technical defects, the present application provides an enhanced device fingerprint identification system and method based on time-frequency domain feature fusion, which aims to enhance the extraction and fusion capabilities of device fingerprint features and improve the robustness and accuracy of device fingerprint identification. SUMMARY

[0006] The purpose of the present application is to provide an enhanced device fingerprint identification system and method based on time-frequency domain feature fusion, which aims to enhance the extraction and fusion capabilities of device fingerprint features and improve the robustness and accuracy of device fingerprint identification.

[0007] To achieve the above purpose, the present application provides an enhanced device fingerprint identification system based on time-frequency domain feature fusion, which is composed of a signal processing module, a feature extraction module, a feature fusion module and a device identification module.

[0008] The feature extraction module is constructed by a spatial channel attention convolutional layer SCAConv.

[0009] Preferably, the signal processing module is configured to perform data preprocessing on the raw IQ data, i.e., the in-phase component I and the quadrature component Q;

[0010] First, the signal processing module uses a moving window to segment the raw IQ data into three time-domain features: I, Q, and IQ, with a fixed size. The window moving step is set to half the window size. Then, the IQ data is converted to the frequency domain and the time-frequency domain using the fast Fourier transform (FFT) and the short-time Fourier transform (STFT) formula, and the corresponding features are extracted in these two domains.

[0011] The frequency-domain feature reflects the spectral distribution of the signal. The IQ signal is subjected to discrete Fourier transform using the fast Fourier transform (FFT), and the spectral information obtained is as follows:

[0012]

[0013] where Y FFT (f) represents the signal feature in the frequency domain; N s represents the number of signal samples; represents the IQ signal;

[0014] The time-frequency domain feature is calculated for the frequency variation of the signal within a local time range, and is suitable for signals whose frequency changes over time.

[0015] Let w(·) be a window function with a length of L w ; L s is the moving step of the window; after windowing, the IQ signal is subjected to discrete Fourier transform (STFT) to the frequency domain, as follows:

[0016]

[0017] where Y STFT (m,f) represents the STFT feature at discrete time component m and frequency component f.

[0018] Preferably, the feature extraction module is constructed by a spatial channel attention convolutional layer (SCAConv) to extract device fingerprint features in the time domain, the frequency domain, and the time-frequency domain, including I signal features, Q signal features, IQ signal features, FFT features, and STFT features.

[0019] For the five signal features, five parallel channels are designed: I channel, Q channel, IQ channel, FFT channel, and STFT channel. Among them, the I, Q, and IQ channels are used to extract time-domain features; the FFT channel is used to extract frequency-domain features; and the STFT channel is used to extract time-frequency domain features.

[0020] On the network structure, the time domain and frequency domain channels adopt one-dimensional spatial channel attention convolutional layers; the time-frequency domain channels adopt two-dimensional spatial channel attention convolutional layers; each channel is sequentially cascaded by four spatial channel attention convolutional layers; the fingerprint features after passing through the four spatial channel attention convolutional layers are respectively denoted as F I , F Q , F IQ , F FFT , F STFT .

[0021] Preferably, the spatial channel attention convolutional layer is composed of a fingerprint feature convolution branch, a channel attention branch, a spatial attention branch and an attention mechanism; wherein the fingerprint feature convolution branch is used to extract the fingerprint features of the input signal; the channel attention branch is used to express the importance of each fingerprint feature; the spatial attention branch is used to express the local importance of each fingerprint feature;

[0022] The specific implementation process of the spatial channel attention convolutional layer is as follows:

[0023] First, the outputs of the fingerprint feature convolution branch are multiplied with the outputs of the channel attention branch and the spatial attention branch respectively; then, the fingerprint features at the three levels are jointly input into the attention mechanism to enhance the overall feature representation of the device fingerprint.

[0024] Preferably, the fingerprint feature convolution branch is used to extract the fingerprint features from the input signal, and is composed of a single convolutional layer containing multiple convolution kernels; the convolution kernel performs element-wise multiplication and addition operation on each sub-region of the signal under the control of the specified stride, thereby generating a feature value; after traversing all sub-regions, the complete fingerprint features of the input signal are obtained;

[0025] The calculation formulas of the one-dimensional fingerprint feature convolution branch and the two-dimensional fingerprint feature convolution branch are as follows:

[0026]

[0027] Wherein, out(B i ,C j ,L) represents the feature map output by the one-dimensional fingerprint feature convolution branch, and the size of the feature map is determined by the convolution kernel size, the stride and the padding; bias(C j ) represents the bias of the jth convolution kernel; out(B i ,C j ,H,W) represents the feature map output by the two-dimensional fingerprint feature convolution branch; B i represents the batch size of the training data; L represents the signal sequence length; C j represents the output channel number; C nrepresents the number of channels of the input signal; k represents the input channel index; * represents the cross-correlation operation; H represents the frequency component of the signal feature map; W represents the time component of the signal feature map; k h and k w respectively represent the height and width of the convolution kernel;

[0028] wherein the calculation formulas of L, H and W are as follows:

[0029]

[0030] wherein L out and L in respectively represent the length of the output feature map and the length of the input signal; W out and W in respectively represent the length of the time component of the output signal feature map and the length of the time component of the input signal feature map; H out and H in respectively represent the length of the frequency component of the output signal feature map and the length of the frequency component of the input signal feature map; K represents the size of the convolution kernel; S is the convolution step; and P is the padding size.

[0031] Preferably, the channel attention branch assigns a corresponding weight to each fingerprint feature, distinguishing the importance of different channels in the identification task.

[0032] A. The one-dimensional channel attention branch is composed of two fully connected layers FC, and uses a Tanh activation function between layers, and finally outputs the weight through a Sigmoid function;

[0033] The first fully connected layer reduces the dimension of the fingerprint feature in the channel dimension, and converts it into a one-dimensional vector; then, the Tanh activation function performs a nonlinear change on the reduced data, and the range of the Tanh function is [-1, 1]; the second fully connected layer further increases the dimension of the vector to be consistent with the original channel number, and finally the Sigmoid function is used to scale the weight to [0, 1], and the channel attention branch is multiplied with the output of the fingerprint feature convolution branch;

[0034] B. The two-dimensional channel attention branch is composed of one convolution layer, one pooling layer and two fully connected layers.

[0035] Firstly, the channel number of the channel attention branch is adjusted to match the channel number of the fingerprint feature convolution branch by a two-dimensional convolution layer; then, average pooling and max pooling are used to highlight the overall and significant areas of the fingerprint feature map and reduce the spatial dimension; the first fully connected layer reduces the dimension of the fingerprint feature map in the channel dimension, and then the data is subjected to nonlinear transformation by the Tanh activation function; the second fully connected layer increases the dimension of the fingerprint feature map in the channel dimension, which is the same as the first convolution layer, and matches the channel number of the fingerprint feature convolution branch again; finally, the Sigmoid function is used to scale the weight to [0, 1], and multiplied with the fingerprint feature convolution branch by channel.

[0036] Preferably, the role of the spatial attention branch is to assign weights to each spatial position of the fingerprint feature, highlight the key area, and suppress the irrelevant and noise area.

[0037] A, the one-dimensional spatial attention branch is composed of a convolution layer, a max pooling layer, a deconvolution layer and an up-sampling layer;

[0038] The convolution layer reduces the dimension of the data, the max pooling layer is used to extract local responses and highlight the key features of the input tensor; the deconvolution layer is used to restore the input fingerprint and strengthen its significant features; the up-sampling layer is used to restore the fingerprint feature to the size matching the fingerprint feature convolution branch; the Sigmoid function is used to scale the weight to [0, 1], which is convenient for the spatial attention branch to be multiplied with the fingerprint feature convolution branch.

[0039] B, the two-dimensional spatial attention branch is composed of a max pooling layer, an average pooling layer and a convolution layer;

[0040] The max pooling and average pooling are executed in parallel, retaining the maximum local features of the feature map and the overall characteristics of the original feature map, and realizing the weighting of different spatial areas; the convolution layer extracts features in the local area, captures the context information in the spatial dimension, highlights the important area and suppresses the unimportant area; finally, the Sigmoid function is used to scale the weight to [0, 1], which is convenient for the spatial attention branch to be multiplied with the fingerprint feature convolution branch.

[0041] Preferably, the feature fusion module is used to fuse the features output by the I channel, Q channel, IQ channel, FFT channel and STFT channel, thereby generating the final enhanced fingerprint feature.

[0042] The feature fusion module adopts a shared attention strategy. First, average pooling and maximum pooling are respectively performed on the output of each channel to aggregate spatial information and form two different context features. Second, the two context features are concatenated in the channel dimension and convolved through a first one-dimensional convolution, activated by a Relu function. Then, a second one-dimensional convolution is performed, and finally, a Sigmoid function is used to scale the weights to [0, 1]. The important parts are enhanced and the unimportant parts are weakened by element-wise multiplication of the attention map and the fingerprint feature. Then, the obtained features are denoted as: F r I , F r Q , F r IQ , F r FFT and F r STFT ; The obtained neural feature maps are concatenated in the channel dimension to form a fused feature, as follows:

[0043] F r = [F r I , F r Q , F r IQ , F r FFT , F r STFT ];

[0044] Finally, the fused feature is input into the device recognition module to perform device fingerprint recognition.

[0045] Preferably, the device recognition module is composed of two fully connected layers. First, the fingerprint feature enhanced by feature extraction and attention mechanism is input into the first fully connected layer to reduce the dimension of the input feature and perform nonlinear transformation through the ReLU activation function. Then, the second fully connected layer projects the mapped vector to the device category and outputs the final recognition result.

[0046] The enhanced device fingerprint recognition method based on time-frequency domain feature fusion is applied to the above system and includes the following steps:

[0047] Step S1, using USRP to collect signals represented in the form of in-phase component and quadrature component, i.e. IQ symbol;

[0048] Step S2, in the signal processing module, the collected signal data is preprocessed, first IQ is divided into three time domain characteristics: I, Q and IQ; then the corresponding FFT and STFT formula is used, IQ is converted to the frequency domain and the time-frequency domain, and the corresponding characteristics are extracted in the two domains;

[0049] Step S3, the characteristics of time domain, frequency domain and time-frequency domain are input into the feature extraction module for feature extraction;

[0050] Step S4, the characteristics extracted from the feature extraction module are sent to the feature fusion module to generate the final fusion characteristics;

[0051] Step S5, finally, the fusion characteristics are transmitted to the classification model, the fingerprint matching and identification of the device are completed, and the identity recognition of the physical layer of the wireless device is realized.

[0052] Therefore, the enhanced device fingerprint identification system and method based on time-frequency domain feature fusion are adopted, the security of the Internet of Things device is improved, and the system can be deployed on the Internet of Things device with limited power consumption or old device without encrypted transmission; the accuracy and robustness of device identification are improved, the sensitivity of the system to noise interference is significantly reduced by supplementing the frequency domain characteristics, and high identification accuracy can be maintained even under low signal-to-noise ratio conditions; the cost of device fingerprint extraction is reduced, high identification accuracy can be achieved for single physical layer fingerprint (such as IQ signal), and high feature extraction overhead is avoided.

[0053] The technical solutions of the present application will be further described in detail below with the help of drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 It is a schematic diagram of the enhanced device fingerprint identification system based on time-frequency domain feature fusion of the present application.

[0055] Figure 2 It is a schematic diagram of the model framework of the enhanced device fingerprint identification system based on time-frequency domain feature fusion of the present application.

[0056] Figure 3 It is a structural schematic diagram of the one-dimensional spatial channel attention convolutional layer of the present application.

[0057] Figure 4 It is a structural schematic diagram of the two-dimensional spatial channel attention convolutional layer of the present application.

[0058] Figure 5 It is a structural schematic diagram of the shared attention strategy of the present application.

[0059] Figure 6 It is a flowchart of the enhanced device fingerprint identification method based on time-frequency domain feature fusion of the present application. DETAILED DESCRIPTION

[0060] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0061] like Figure 1 and Figure 2 As shown in the figure, the enhanced device fingerprint recognition system based on time-frequency domain feature fusion of the present invention consists of a signal processing module, a feature extraction module, a feature fusion module and a device recognition module; among them, the feature extraction module is constructed by a spatial channel attention convolution layer (Spatial Channel Attention Convolution, SCAConv).

[0062] Example

[0063] 1. Signal processing module.

[0064] The signal processing module is used to perform data preprocessing on the acquired original IQ data, namely the in-phase component I and the orthogonal component Q.

[0065] First, IQ is split into three time-domain features: I, Q, and IQ. To meet the convolutional neural network's requirement for input tensor size consistency, the signal processing module uses a moving window to segment the original IQ data into fixed-size segments, and sets the moving step to half the window size. This ensures the coverage of data samples while effectively utilizing computing resources.

[0066] Then, the IQ is converted into the frequency domain and time-frequency domain using the corresponding fast Fourier transform (FFT) and discrete Fourier transform (STFT) formulas, and the corresponding features are extracted in these two domains.

[0067] 1. Time domain characteristics.

[0068] The signal's time-domain characteristics not only reflect its energy distribution and periodicity, but also reveal many subtle characteristics stemming from hardware manufacturing variations. Due to inevitable manufacturing variations in components such as the transmitter and receiver's local oscillators, digital-to-analog converters, or analog-to-digital converters, the modulated and demodulated signals—the in-phase component I(t) and the quadrature component Q(t)—are not identical, and their phases cannot maintain the ideal 90-degree difference.

[0069] Therefore, the time-domain features can well reflect the hardware characteristics of the transmitter, and the signals generated by the respective independent hardware of I(t) and Q(t) will embed different fingerprint information. By deeply analyzing the respective time-domain characteristics of the two signals, the unique fingerprint of the device can be more accurately excavated. Traditional radio frequency fingerprint extraction methods often regard IQ as a whole, ignoring the independent characteristics of each hardware, which reduces the efficiency of device fingerprint identification. Therefore, the comprehensive use of time-domain features can not only capture the overall characteristics, but also excavate the subtle differences implied by each signal, significantly improving the recognition effect and accuracy of the radio frequency fingerprint.

[0070] 2. Frequency-domain features.

[0071] Frequency-domain features mainly reflect the spectral distribution of the signal. Fast Fourier Transform (FFT) is used to perform discrete Fourier transform on the IQ signal, thereby obtaining more abundant spectral information, as follows:

[0072]

[0073] where Y FFT (f) represents the signal feature in the frequency domain; N s represents the number of signal samples; represents the IQ signal.

[0074] Therefore, the FFT feature Y FFT (f) is the frequency-domain representation of the IQ signal. The FFT feature can effectively represent the spectral characteristics of the signal, helping to capture the subtle differences in frequency components. This is particularly important for identifying the unique behavior patterns of devices at different frequencies, making the hardware defects of each device more pronounced at certain frequencies.

[0075] 3. Time-frequency domain features.

[0076] Unlike frequency-domain features, time-frequency domain features calculate the frequency changes of a signal within a local time range, which are suitable for signals with frequency changes over time.

[0077] Let w(·) be a window function with a length of L w ; L s is the moving step of the window. After windowing processing, the IQ signal is transformed to the frequency domain by the Short-Time Fourier Transform (STFT), as follows:

[0078]

[0079] where Y STFT (m,f) represents the STFT feature at the discrete time component m and the frequency component f.

[0080] Therefore, the use of STFT features YSTFT (m, f) as another frequency-domain representation of the IQ signal. STFT features reveal the frequency characteristics of a signal over time by segmenting the signal into short time frames and performing a Fourier transform within each frame. This is useful for analyzing transient changes and dynamic characteristics in non-stationary signals, helping to capture unique behavioral patterns of devices within different time windows.

[0081] II. Feature extraction module.

[0082] The feature extraction module is built by a spatial channel attention convolution layer (SCAConv) to extract device fingerprint features in the time domain, frequency domain, and time-frequency domain, including I signal features, Q signal features, IQ signal features, FFT features, and STFT features.

[0083] 1. Design of five parallel channels.

[0084] In the feature extraction module, to extract the above-mentioned five device fingerprint features, five parallel channels are designed: I channel, Q channel, IQ channel, FFT channel, and STFT channel, to avoid interference between features in each channel. Among them, the I, Q, and IQ channels are used to extract time-domain features; the FFT channel is used to extract frequency-domain features; and the STFT channel is used to extract time-frequency domain features.

[0085] In terms of network structure, both the time-domain and frequency-domain channels are built with one-dimensional spatial channel attention convolution layers (SCAConv); the time-frequency domain channel is built with a two-dimensional spatial channel attention convolution layer. Among them, one-dimensional and two-dimensional spatial channel attention convolution layers are shown in Figure 3 and Figure 4 Each channel is cascaded by four layers of SCAConv in turn, fully mining and representing the fingerprint features in their respective fields, ensuring that the network can fully learn the device fingerprint features contained in each channel.

[0086] The fingerprint features after four layers of spatial channel attention convolution layers are denoted as F I , F Q , F IQ , F FFT , and F STFT .

[0087] 2. Spatial channel attention convolution layer.

[0088] In traditional deep learning methods, the same convolution kernel weight is used for all fingerprints by the convolutional neural network, however, the importance of each fingerprint itself is not the same, and different regions of the same fingerprint may also have different weights. Therefore, the spatial channel attention convolution layer introduces a channel attention branch and a spatial attention branch to determine the weight of each fingerprint as a whole and the weight of each part of the fingerprint, respectively.

[0089] Therefore, the spatial channel attention convolution layer is composed of a fingerprint feature convolution branch, a channel attention branch, a spatial attention branch, and an attention mechanism. Among them, the fingerprint feature convolution branch is used to extract the fingerprint features of the input signal; the channel attention branch is used to express the importance of each fingerprint feature; and the spatial attention branch is used to express the local importance of each fingerprint feature.

[0090] The specific implementation process of the spatial channel attention convolution layer is as follows:

[0091] First, the outputs of the fingerprint feature convolution branch are multiplied by the outputs of the channel attention branch and the spatial attention branch, respectively, to highlight the significance of the fingerprint in the channel and spatial dimensions.

[0092] Then, the fingerprint features at the three levels are jointly input into the attention mechanism to further emphasize the key elements and enhance the overall feature representation of the device fingerprint.

[0093] (1) Fingerprint feature convolution branch.

[0094] The fingerprint feature convolution branch is used to extract fingerprint features from the input signal and is composed of a single convolution layer, which contains multiple convolution kernels. The convolution kernel performs element-wise multiplication and addition operation on each sub-region of the signal under the control of the specified stride, thereby generating a feature value. When all sub-regions are traversed, the complete fingerprint feature of the input signal is obtained. In a deep learning model, the convolution kernel weight is automatically updated according to the data to extract the differences related to the hardware fingerprint in the signal.

[0095] The calculation formulas of one-dimensional fingerprint feature convolution branch and two-dimensional fingerprint feature convolution branch are as follows:

[0096]

[0097]

[0098] wherein out(B i ,C j ,L) represents the feature map output by the one-dimensional fingerprint feature convolution branch, the size of the feature map is determined by the convolution kernel size, stride and padding; bias(C j ) represents the bias of the jth convolution kernel; out(B i ,Cj H, W) denotes the feature map output by the two-dimensional fingerprint feature convolution branch; B i denotes the batch size of the training data; L denotes the signal sequence length; C j denotes the number of output channels; C n denotes the number of input signal channels; k denotes the input channel index; * denotes the cross-correlation operation; H denotes the frequency component of the signal feature map; W denotes the time component of the signal feature map; k h and k w denote the height and width of the convolution kernel, respectively.

[0099] wherein the calculation formula of L, H, and W is as follows:

[0100]

[0101] wherein L out and L in denote the length of the output feature map and the length of the input signal, respectively; W out and W in denote the length of the time component of the output signal feature map and the length of the time component of the input signal feature map, respectively; H out and H in denote the length of the frequency component of the output signal feature map and the length of the frequency component of the input signal feature map, respectively; K denotes the size of the convolution kernel; S is the convolution step; and P is the padding size.

[0102] (2) Channel attention branch.

[0103] The core goal of the channel attention branch is to assign a corresponding weight to each fingerprint feature, thereby distinguishing the importance of different channels in the identification task.

[0104] A, The one-dimensional channel attention branch is composed of two fully connected layers (FC) and uses a Tanh activation function between layers, and finally outputs the weight through a Sigmoid function.

[0105] The first fully connected layer reduces the dimension of the fingerprint feature in the channel dimension, converting it into a one-dimensional vector. Subsequently, the Tanh activation function performs a nonlinear change on the reduced data, and the range of the Tanh function is [-1, 1]. This distribution is beneficial to avoid gradient disappearance and enrich feature representation in neural networks. The second fully connected layer further increases the dimension of the vector to be consistent with the original number of channels, and finally the Sigmoid function is used to scale the weight to [0, 1], which is convenient for the channel attention branch to multiply the output of the fingerprint feature convolution branch.

[0106] B, The two-dimensional channel attention branch is composed of one convolution layer, one pooling layer, and two fully connected layers.

[0107] Firstly, the channel number of the channel attention branch is adjusted to match the channel number of the fingerprint feature convolution branch by a two-dimensional convolution layer. Then, average pooling and max pooling are used to highlight the overall and significant areas of the fingerprint feature map, respectively, and to reduce the spatial dimension to some extent, thereby reducing the risk of overfitting. The first fully connected layer reduces the dimension of the fingerprint feature map in the channel dimension, allowing the network to pay more attention to key channel information. Then, a Tanh activation function is used to perform nonlinear transformation on the data. The second fully connected layer increases the dimension of the fingerprint feature map in the channel dimension, which is the same as the first convolution layer, and matches the channel number of the fingerprint feature convolution branch. Finally, a Sigmoid function is used to scale the weight to [0, 1] so that it can be multiplied with the fingerprint feature convolution branch by channel.

[0108] (3) Spatial attention branch.

[0109] The role of the spatial attention branch is to assign weights to each spatial position of the fingerprint feature to highlight key areas and suppress irrelevant or noisy areas.

[0110] A. The one-dimensional spatial attention branch is composed of a convolution layer on a specific channel, a max pooling layer, a deconvolution layer, and an up-sampling layer.

[0111] The convolution layer reduces the dimension of the data, which can reduce the spatial weight distortion that may occur in the fingerprint and eliminate channel interference. The max pooling layer is used to extract the strongest local response to highlight the key features of the input tensor. The deconvolution layer is used to restore the input fingerprint and enhance its significant features. The up-sampling layer is used to restore the fingerprint feature to a size that matches the fingerprint feature convolution branch. The Sigmoid function is used to scale the weight to [0, 1] to facilitate the multiplication of the spatial attention branch and the fingerprint feature convolution branch.

[0112] B. The two-dimensional spatial attention branch is composed of a max pooling layer, an average pooling layer, and a convolution layer.

[0113] Max pooling and average pooling are performed in parallel, not only preserving the maximum local features of the feature map, but also embodying the overall characteristics of the original feature map, achieving weighting of different spatial regions. The convolution layer extracts features in the local region, which can better capture the context information in the spatial dimension, highlight important areas, and suppress unimportant areas. Finally, the Sigmoid function is used to scale the weight to [0, 1] to facilitate the multiplication of the spatial attention branch and the fingerprint feature convolution branch.

[0114] Three, feature fusion module.

[0115] The feature fusion module is used to fuse the features output by the I channel, Q channel, IQ channel, FFT channel, and STFT channel, thereby generating the final enhanced fingerprint feature.

[0116] As shown in Figure 5 , the feature fusion module adopts a shared attention strategy. First, average pooling and max pooling are respectively performed on the output of each channel to aggregate spatial information and form two different context features. Second, the two context features are concatenated in the channel dimension and convolved through the first one-dimensional convolution, activated by the Relu function. Then, the second one-dimensional convolution is performed, and finally the Sigmoid function is used to scale the weight to [0, 1], so that the important part is enhanced and the unimportant part is weakened by element-level multiplication of the attention map and the fingerprint feature. Then, the obtained features are denoted as: r I , F r Q , F rIQ , F r FFT and F r STFT . These refined neural feature maps are concatenated in the channel dimension to form the fusion feature, as follows:

[0117] F r = [F rI , F r Q , F rIQ , F r FFT , F r STFT ];

[0118] Finally, the fused feature is input into the device recognition module to perform device fingerprint recognition.

[0119] Four, device recognition module.

[0120] The device recognition module is composed of two fully connected layers.

[0121] First, the fingerprint feature enhanced by feature extraction and attention mechanism is input into the first fully connected layer to reduce the dimension of the input feature and perform nonlinear transformation through the ReLU activation function. This not only alleviates the gradient vanishing problem to some extent, but also helps to speed up the training convergence of the model. Then, the second fully connected layer projects the mapped vector to the device category and outputs the final recognition result.

[0122] As shown in Figure 6 , based on the above-mentioned enhanced device fingerprint recognition system based on time-frequency domain feature fusion, the present application also proposes an enhanced device fingerprint recognition method based on time-frequency domain feature fusion, which is applied to the above-mentioned system and includes the following steps:

[0123] Step S1, using USRP (Universal Software Radio Peripheral) to collect signals represented in the form of in-phase component and quadrature component, i.e. IQ symbols.

[0124] Step S2, in the signal processing module, the collected signal data is preprocessed, first, IQ is split into three time domain features: I, Q and IQ; then, IQ is converted to the frequency domain and time-frequency domain by using corresponding FFT and STFT formulas, and corresponding features are extracted in the two domains.

[0125] Step S3, the features in the time domain, frequency domain and time-frequency domain are respectively input into the feature extraction module for feature extraction.

[0126] Step S4, the features extracted in the feature extraction module are sent to the feature fusion module to generate the final fusion features.

[0127] Step S5, finally, the fusion features are transmitted to the classification model to complete the fingerprint matching and identification of the device, realize the identity recognition of the physical layer of the wireless device, and improve the security of the wireless device.

[0128] Therefore, the application adopts the above-mentioned enhanced device fingerprint identification system and method based on time-frequency domain feature fusion, improves the security of the Internet of Things device, can be deployed on the Internet of Things device with limited power consumption or old device that cannot transmit encrypted, improves the accuracy and robustness of device identification, through supplementing the frequency domain features, the sensitivity of the system to noise interference is significantly reduced, even in the condition of lower signal-to-noise ratio, the recognition accuracy can be maintained; reduces the cost of device fingerprint extraction, and high recognition accuracy can be achieved for single physical layer fingerprint (such as IQ signal), and the high feature extraction overhead is avoided.

[0129] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that: it can still modify or equivalently replace the technical solutions of the present application, and these modifications or equivalent replacements also cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present application.

Claims

1. An enhanced device fingerprint recognition system based on time-frequency domain feature fusion, characterized by: It consists of signal processing module, feature extraction module, feature fusion module and device identification module; Among them, the feature extraction module is constructed by the spatial channel attention convolution layer SCAConv.

2. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1 is characterized in that: The signal processing module is used to perform data preprocessing on the original IQ data, i.e., the in-phase component I and the orthogonal component Q; First, the signal processing module uses a moving window to segment the raw IQ data into three time-domain features: I, Q, and IQ. The window step size is set to half the window size. The IQ data is then converted into the frequency and time-frequency domains using the Fast Fourier Transform (FFT) and Discrete Fourier Transform (STFT) formulas, and features are extracted from these two domains. The frequency domain characteristics reflect the spectrum distribution of the signal. The spectrum information obtained by performing discrete Fourier transform on the IQ signal using fast Fourier transform FFT is as follows: Among them, Y FFT (f) indicates signal Features in the frequency domain; N s Represents the number of signal samples; represents the IQ signal; The time-frequency domain feature is to calculate the frequency change of the signal within a local time range, which is applicable to signals whose frequency changes with time; Let w(·) be the window function, whose length is L w ;L s is the moving step of the window; after the IQ signal is windowed, it is converted to the frequency domain through discrete Fourier transform STFT, as shown below: Among them, Y STFT (m,f) represents the STFT feature at the discrete time component m and frequency component f.

3. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1 is characterized in that: The feature extraction module is constructed by the spatial channel attention convolution layer SCAConv, which is used to extract device fingerprint features in the time domain, frequency domain, and time-frequency domain, including I signal features, Q signal features, IQ signal features, FFT features, and STFT features; According to the five signal characteristics, five parallel channels are designed: I channel, Q channel, IQ channel, FFT channel and STFT channel. Among them, I, Q and IQ channels are used to extract time domain features; FFT channel is used to extract frequency domain features; STFT channel is used to extract time-frequency domain features. In terms of network structure, both time domain and frequency domain channels use one-dimensional spatial channel attention convolution layer; time-frequency domain channel uses two-dimensional spatial channel attention convolution layer; each channel is cascaded by four layers of spatial channel attention convolution layer; the fingerprint features after the four layers of spatial channel attention convolution layer are recorded as F I 、F Q 、F IQ 、F FFT 、F STFT .

4. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 3 is characterized by: The spatial channel attention convolution layer consists of a fingerprint feature convolution branch, a channel attention branch, a spatial attention branch, and an attention mechanism. The fingerprint feature convolution branch is used to extract the fingerprint features of the input signal; the channel attention branch is used to express the importance of each fingerprint feature; and the spatial attention branch is used to express the local importance of each fingerprint feature. The specific implementation process of the spatial channel attention convolution layer is as follows: First, the output of the fingerprint feature convolution branch is multiplied with the output of the channel attention branch and the spatial attention branch respectively; then, the fingerprint features of these three levels are input into the attention mechanism together to enhance the overall feature representation of the device fingerprint.

5. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 4 is characterized in that: The fingerprint feature convolution branch is used to extract fingerprint features from the input signal. It consists of a single convolution layer containing multiple convolution kernels. The convolution kernel performs element-by-element multiplication and addition operations on each sub-region of the signal under the control of a specified stride to generate feature values. After traversing all sub-regions, the complete fingerprint feature of the input signal is obtained. The calculation formulas for the one-dimensional fingerprint feature convolution branch and the two-dimensional fingerprint feature convolution branch are as follows: Among them, out(B i ,C j ,L) represents the feature map output by the one-dimensional fingerprint feature convolution branch. The size of the feature map is determined by the convolution kernel size, stride and padding; bias(C j ) represents the bias of the jth convolution kernel; out(B i ,C j ,H,W) represents the feature map output by the convolution branch of the two-dimensional fingerprint feature; B i represents the batch size of training data; L represents the length of signal sequence; C j Indicates the number of output channels; C n represents the number of channels of the input signal; k represents the input channel index; * represents the cross-correlation operation; H represents the frequency component of the signal feature graph; W represents the time component of the signal feature graph; k h and k w Represent the height and width of the convolution kernel respectively; The calculation formulas for L, H, and W are as follows: Among them, L out and L in Respectively represent the length of the output feature map and the length of the input signal; W out and W in Respectively represent the time component length of the output signal characteristic graph and the time component length of the input signal characteristic graph; H out and H in They represent the frequency component length of the output signal feature map and the frequency component length of the input signal feature map respectively; K represents the size of the convolution kernel; S is the convolution step size; and P is the padding size.

6. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 4 is characterized in that: The channel attention branch assigns corresponding weights to each fingerprint feature to distinguish the importance of different channels in the recognition task; A. The one-dimensional channel attention branch consists of two fully connected layers FC, and uses the Tanh activation function between layers, and finally outputs the weight through the Sigmoid function; The first fully connected layer reduces the dimension of the fingerprint features in the channel dimension and converts them into a one-dimensional vector. Subsequently, the Tanh activation function performs nonlinear transformations on the reduced-dimensional data, and the range of the Tanh function is [-1, 1]. The second fully connected layer then increases the dimension of the vector to match the original number of channels. Finally, the Sigmoid function is used to scale the weights to [0, 1] and multiply the channel attention branch with the output of the fingerprint feature convolution branch. B. The 2D channel attention branch consists of one convolutional layer, one pooling layer, and two fully connected layers; First, the number of channels of the channel attention branch is adjusted to match the fingerprint feature convolution branch through a two-dimensional convolution layer; Afterwards, average pooling and maximum pooling are used to highlight the overall and significant areas of the fingerprint feature map respectively, and reduce the spatial dimension; the first fully connected layer reduces the dimension of the pooled fingerprint feature map in the channel dimension, and then the Tanh activation function performs nonlinear changes on the data; the second fully connected layer increases the dimension of the fingerprint feature map in the channel dimension, which is the same as the first convolution layer, and again matches the number of channels of the fingerprint feature convolution branch; finally, the Sigmoid function is used to scale the weight to [0,1] and multiply it with the fingerprint feature convolution branch by channel.

7. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 4 is characterized in that: The role of the spatial attention branch is to assign weights to each spatial position of the fingerprint feature, highlighting the key areas and suppressing irrelevant and noisy areas; A. The one-dimensional spatial attention branch consists of a convolutional layer, a max pooling layer, a deconvolution layer, and an upsampling layer; The convolution layer reduces the dimensionality of the data, and the maximum pooling layer is used to extract local responses and highlight the key features of the input tensor. The deconvolution layer is used to restore the input fingerprint and enhance its salient features. The upsampling layer is used to restore the fingerprint features to a size that matches the fingerprint feature convolution branch. The Sigmoid function is used to scale the weights to [0, 1] to facilitate the multiplication of the spatial attention branch and the fingerprint feature convolution branch. B. The 2D spatial attention branch consists of a max pooling layer, an average pooling layer, and a convolutional layer; Maximum pooling and average pooling are performed in parallel to retain the maximum local features of the feature map and the overall characteristics of the original feature map, thereby achieving weighted weighting of different spatial regions. The convolution layer extracts features within the local region, captures contextual information in the spatial dimension, highlights important regions, and suppresses unimportant regions. Finally, the weights are scaled to [0, 1] using the Sigmoid function to facilitate multiplication of the spatial attention branch with the fingerprint feature convolution branch.

8. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1 is characterized in that: The feature fusion module is used to fuse the features output by the I channel, Q channel, IQ channel, FFT channel and STFT channel to generate the final enhanced fingerprint features; The feature fusion module adopts a shared attention strategy. First, average pooling and maximum pooling are performed on the output of each channel to aggregate spatial information and form two different context features. Secondly, the two context features are spliced ​​in the channel dimension and convolved through the first layer of one-dimensional convolution and activated by the ReLU function. Then, the second layer of one-dimensional convolution is performed, and finally the Sigmoid function is used to scale the weights to [0,1]. The important parts are enhanced and the unimportant parts are weakened by multiplying the attention map and the fingerprint features at the element level. Then, the obtained features are recorded as: F rI 、F rQ 、F rIQ 、F rFFT and F rSTFT ; The obtained neural feature maps are spliced ​​together in the channel dimension to form fusion features, as shown below: F r =[F rI ,F rQ ,F rIQ ,F rFFT ,F rSTFT ]; Finally, the fused features are input into the device identification module to perform device fingerprint recognition.

9. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1 is characterized in that: The device identification module consists of two fully connected layers. First, the fingerprint features after feature extraction and attention mechanism enhancement are input into the first fully connected layer, which reduces the dimensionality of the input features and performs a nonlinear transformation using the ReLU activation function. Then, the second fully connected layer projects the mapped vector to the device category and outputs the final recognition result.

10. According to any one of claims 1 to 9, an enhanced device fingerprint recognition system based on time-frequency domain feature fusion is proposed, which is applied to the above system and is characterized in that: The following steps are involved: Step S1: using USRP to collect a signal represented by an in-phase component and a quadrature component, i.e., an IQ symbol; Step S2: In the signal processing module, the collected signal data is preprocessed. First, the IQ is split into three time-domain features: I, Q, and IQ; then, the corresponding FFT and STFT formulas are used to convert the IQ to the frequency domain and time-frequency domain, and the corresponding features are extracted in these two domains; Step S3: Input the features of time domain, frequency domain and time-frequency domain into the feature extraction module respectively for feature extraction; Step S4: Send the features extracted in the feature extraction module to the feature fusion module to generate the final fusion features; Step S5: Finally, the fused features are passed to the classification model to complete the fingerprint matching and identification of the device, thereby realizing the identity recognition of the physical layer of the wireless device.

Images