Enhanced device fingerprinting system and method based on time-frequency domain feature fusion

Abstract

The application discloses an enhanced device fingerprint identification system and method 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; first, IQ is split into three time domain features, and the corresponding FFT and STFT formulas are used to convert IQ into the frequency domain and the time-frequency domain and extract the corresponding features; the features of the time domain, the frequency domain and the time-frequency domain are respectively input into the feature extraction module for feature extraction; the extracted features are sent to the feature fusion module to generate the final fusion features; finally, the fusion features are transmitted to the classification model to complete the fingerprint matching and identification of the device. The application adopts the above-mentioned enhanced device fingerprint identification system and method based on time-frequency domain feature fusion, improves the security, accuracy and robustness of the Internet of Things device, reduces the sensitivity of the system to noise interference, reduces the cost of device fingerprint extraction, and realizes high recognition accuracy.

Description

Technical Field

[0001] This invention relates to the field of device fingerprint recognition technology, and in particular to an enhanced device fingerprint recognition system and method based on time-frequency domain feature fusion. Background Technology

[0002] With the continuous development of mobile communication technologies such as wireless networks and electronic devices, the Internet of Things (IoT) technology has been widely applied and is playing an increasingly important role in fields such as transportation, healthcare, and homes. Along with the booming development of IoT, billions of IoT devices are deployed everywhere, supporting computing and communication. However, these devices are highly vulnerable to malicious attacks when operating in wireless environments, making wireless communication security a serious challenge.

[0003] Traditional authentication methods typically rely on software addresses (such as Internet Protocol (IP) or Media Access Control (MAC) and classical cryptography for authentication. However, these methods often require complex computations and protocols, making them unsuitable for IoT devices with limited power consumption and computing resources. To address this, device fingerprinting technology analyzes defects in radio frequency circuit waveforms to obtain unique physical characteristics of wireless devices, enabling the identification and differentiation of different devices. Specifically, the hardware of wireless devices exhibits minute differences in components such as oscillators, power amplifiers, and analog filters during the manufacturing process. Even devices of the same model and from the same production line will display their own unique hardware characteristics.

[0004] However, due to the complex and variable electromagnetic environment in wireless space, a single device fingerprint feature is insufficient for effective device identification and exhibits low robustness. Furthermore, analyzing the relationships between multiple fingerprint features is also highly complex, making it difficult for traditional neural networks to fully learn and integrate various fingerprint features.

[0005] Therefore, in order to overcome the above-mentioned technical defects, this invention proposes an enhanced device fingerprint recognition 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 recognition. Summary of the Invention

[0006] The purpose of this invention is to provide an enhanced device fingerprint recognition 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 recognition.

[0007] To achieve the above objectives, the present invention provides an enhanced device fingerprint recognition system based on time-frequency domain feature fusion, which consists of a signal processing module, a feature extraction module, a feature fusion module and a device recognition module;

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

[0009] Preferably, the signal processing module is used to preprocess 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 original IQ data into three time-domain features: I, Q, and IQ. The window movement step size is set to half the window size. Then, the IQ data is transformed into the frequency domain and time-frequency domain using the Fast Fourier Transform (FFT) and Discrete Fourier Transform (STFT) formulas, and the corresponding features are extracted in these two domains.

[0011] Frequency domain characteristics reflect the spectral distribution of a signal. The spectral information obtained by performing a Discrete Fourier Transform (FFT) on the IQ signal is shown below:

[0012]

[0013] Among them, Y FFT (f) represents the signal Features in the frequency domain; N s Represents the number of signal samples; Indicates the IQ signal;

[0014] Time-frequency domain characteristics calculate the frequency change of a signal over a local time range and are suitable for signals whose frequency changes over time.

[0015] Let w(·) be a window function with length L. w L s The window's movement step size; after windowing, the IQ signal is transformed into the frequency domain via Discrete Fourier Transform (STFT), as shown below:

[0016]

[0017] Among them, Y STFT (m,f) represents the STFT characteristics under discrete time component m and frequency component f.

[0018] Preferably, the feature extraction module is constructed by a spatial channel attention convolutional 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;

[0019] Five parallel channels were designed for five signal characteristics: 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] In terms of network structure, both the time-domain and frequency-domain channels employ one-dimensional spatial channel attention convolutional layers; the time-frequency domain channels employ two-dimensional spatial channel attention convolutional layers; each channel consists of four cascaded spatial channel attention convolutional layers; the fingerprint features after passing through the four spatial channel attention convolutional layers are denoted as F. I F Q F IQ F FFT F STFT .

[0021] Preferably, the spatial channel attention convolutional layer consists of a fingerprint feature convolutional branch, a channel attention branch, a spatial attention branch, and an attention mechanism; wherein, the fingerprint feature convolutional branch is used to extract fingerprint features from 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.

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

[0023] First, the output of the fingerprint feature convolution branch is multiplied by the outputs 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 to enhance the overall feature representation of the device fingerprint.

[0024] Preferably, the fingerprint feature convolution branch is used to extract fingerprint features from the input signal. It consists of a single convolutional layer containing multiple convolutional kernels. The convolutional kernels perform element-wise multiplication and addition operations on each sub-region of the signal under the control of a specified stride, thereby generating feature values. After traversing all sub-regions, the complete fingerprint features of the input signal are obtained.

[0025] The calculation formulas for 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. The size of the feature map is determined by the kernel size, stride, and padding; bias(C) j ) represents the bias of the j-th convolutional kernel; out(B i C j (H,W) represents the feature map output by the two-dimensional fingerprint feature convolution branch; B i L represents the batch size of the training data; C represents the length of the signal sequence. j Indicates the number of output channels; C nThe input signal has [number] channels; k represents the input channel index; * indicates 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 These represent the height and width of the convolution kernel, respectively.

[0028] The formulas for calculating L, H, and W are as follows:

[0029]

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

[0031] Preferably, the channel attention branch assigns a corresponding weight to each fingerprint feature to distinguish the importance of different channels in the recognition task;

[0032] A. The one-dimensional channel attention branch consists of two fully connected (FC) layers, with the Tanh activation function used between the layers, and the weights are finally output through the Sigmoid function.

[0033] The first fully connected layer reduces the dimensionality of the fingerprint features in the channel dimension, converting them into a one-dimensional vector. Then, the Tanh activation function performs a non-linear transformation on the dimensionality-reduced data, and the range of the Tanh function is [-1,1]. The second fully connected layer then increases the dimensionality of the vector to match the original number of channels. Finally, the Sigmoid function is used to scale the weights to [0,1], and the output of the channel attention branch is multiplied by the output of the fingerprint feature convolution branch.

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

[0035] First, the number of channels in the channel attention branch is adjusted to match the number of channels in the fingerprint feature convolution branch using a two-dimensional convolutional layer. Then, average pooling and max pooling are used to highlight the overall and salient regions of the fingerprint feature map, respectively, and reduce the spatial dimension. The first fully connected layer reduces the dimensionality of the pooled fingerprint feature map in the channel dimension, and then the Tanh activation function performs non-linear transformation on the data. The second fully connected layer increases the dimensionality of the fingerprint feature map in the channel dimension, the same as the first convolutional layer, and again matches the number of channels in the fingerprint feature convolution branch. Finally, the Sigmoid function is used to scale the weights to [0,1] and multiply them by channels in the fingerprint feature convolution branch.

[0036] Preferably, the role of the spatial attention branch is to assign weights to each spatial location of the fingerprint feature, highlighting key regions and suppressing irrelevant and noisy regions;

[0037] A. The one-dimensional attention branch consists of a convolutional layer, a max pooling layer, a deconvolutional layer, and an upsampling layer;

[0038] The data is reduced in dimensionality through convolutional layers; max pooling layers are used to extract local responses and highlight key features of the input tensor; deconvolutional layers are used to recover the input fingerprint and enhance its salient features; upsampling layers are used to restore the fingerprint features to a size that matches the fingerprint feature convolutional branch; and the sigmoid function is used to scale the weights to [0,1] to facilitate the multiplication of the spatial attention branch with the fingerprint feature convolutional branch.

[0039] B. The two-dimensional spatial attention branch consists of max pooling layers, average pooling layers, and convolutional layers;

[0040] Max pooling and average pooling are performed in parallel to preserve the maximum local features of the feature map and the overall characteristics of the original feature map, thereby achieving weighting of different spatial regions. Convolutional layers extract features within local regions, capturing contextual information in the spatial dimension, highlighting important regions and suppressing unimportant regions. Finally, the weights are scaled to [0,1] using the Sigmoid function, which facilitates the multiplication of the spatial attention branch with the fingerprint feature convolutional branch.

[0041] Preferably, the feature fusion module is used to fuse the features output from the I channel, Q channel, IQ channel, FFT channel, and STFT channel to generate the final enhanced fingerprint feature;

[0042] The feature fusion module employs a shared attention strategy. First, it performs average pooling and max pooling on the output of each channel to aggregate spatial information and form two different contextual features. Second, it concatenates these two contextual features along the channel dimension and performs a convolution operation using a first-layer one-dimensional convolution, activated by the ReLU function. This is followed by a second-layer one-dimensional convolution, and finally, the Sigmoid function is used to scale the weights to [0,1]. The attention map is then element-wise multiplied with the fingerprint features to enhance important parts and weaken unimportant parts. The resulting features are then 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 along the channel dimension to form a fused feature, as shown below:

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

[0044] Finally, the fused features are input into the device identification module to perform device fingerprint recognition.

[0045] Preferably, the device identification module consists of two fully connected layers. First, the fingerprint features, which have been enhanced by feature extraction and attention mechanism, are input into the first fully connected layer to reduce the dimensionality of the input features and perform nonlinear transformation through the ReLU activation function. Then, the second fully connected layer projects the mapped vector onto the device category and outputs the final identification result.

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

[0047] Step S1: Use USRP to acquire signals represented in in-phase and quadrature components, i.e., in IQ symbol form;

[0048] Step S2: In the signal processing module, the acquired signal data is preprocessed. First, IQ is split into three time-domain features: I, Q and IQ. Then, using the corresponding FFT and STFT formulas, IQ is transformed to the frequency domain and time-frequency domain, and the corresponding features are extracted in these two domains.

[0049] Step S3: Input the features in the time domain, frequency domain, and time-frequency domain into the feature extraction module for feature extraction;

[0050] Step S4: Send the features extracted by the feature extraction module into the feature fusion module to generate the final fused features;

[0051] Step S5: Finally, the fused features are passed to the classification model to complete the fingerprint matching and recognition of the device, thereby realizing the identification of the physical layer of the wireless device.

[0052] Therefore, the present invention employs the above-mentioned enhanced device fingerprint recognition system and method based on time-frequency domain feature fusion, which improves the security of IoT devices and can be deployed on IoT devices with limited power consumption or older devices that cannot transmit encrypted data; it improves the accuracy and robustness of device recognition, and by supplementing frequency domain features, the system's sensitivity to noise interference is significantly reduced, maintaining high recognition accuracy even under low signal-to-noise ratio conditions; it reduces the cost of device fingerprint extraction, and can achieve high recognition accuracy even for single physical layer fingerprints (such as IQ signals), avoiding excessive feature extraction overhead.

[0053] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to the present invention;

[0055] Figure 2 This is a schematic diagram of the model framework of the enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to the present invention;

[0056] Figure 3 This is a schematic diagram of the structure of the one-dimensional spatial channel attention convolutional layer of the present invention;

[0057] Figure 4 This is a schematic diagram of the structure of the two-dimensional spatial channel attention convolutional layer of the present invention;

[0058] Figure 5 This is a schematic diagram of the shared attention strategy of the present invention;

[0059] Figure 6 This is a flowchart of the enhanced device fingerprint recognition method based on time-frequency domain feature fusion according to the present invention. Detailed Implementation

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

[0061] like Figure 1 and Figure 2 As shown, 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; wherein, the feature extraction module is constructed by a spatial channel attention convolution (SCAConv) layer.

[0062] Example

[0063] I. Signal Processing Module.

[0064] The signal processing module is used to preprocess the acquired raw IQ data, namely the in-phase component I and the quadrature component Q.

[0065] First, IQ is split into three temporal features: I, Q, and IQ. In order to meet the requirement of the convolutional neural network for the consistency of input tensor size, the signal processing module uses a moving window to divide the original IQ data into segments of a fixed size, and sets the moving step size to half the window size. This ensures the coverage of data samples and makes effective use of computing resources.

[0066] Then, using the corresponding Fast Fourier Transform (FFT) and Discrete Fourier Transform (STFT) formulas, the IQ is transformed to the frequency domain and time-frequency domain, and the corresponding features are extracted in these two domains.

[0067] 1. Temporal characteristics.

[0068] The time-domain characteristics of a signal not only reflect its energy distribution and periodicity, but also reveal many subtle characteristics stemming from differences in hardware manufacturing. Due to unavoidable manufacturing deviations in components such as the two local oscillators, digital-to-analog converters, or analog-to-digital converters of the transmitter and receiver, the in-phase component I(t) and the quadrature component Q(t) of the modulation and demodulation signals are not exactly the same, and their phases cannot precisely maintain the ideal 90-degree difference.

[0069] Therefore, time-domain features can effectively reflect the hardware characteristics of the transmitter. Furthermore, the signals generated by the independent hardware of I(t) and Q(t) embed different fingerprint information. In-depth analysis of the time-domain characteristics of these two signals can more accurately uncover the unique fingerprint of the device. Traditional RF fingerprint extraction methods often treat I and Q as a whole, ignoring the independent characteristics of each hardware component. This neglect reduces the efficiency of device fingerprint recognition. Therefore, comprehensively utilizing time-domain features can not only capture overall features but also uncover the subtle differences hidden in each signal, significantly improving the recognition effect and accuracy of RF fingerprints.

[0070] 2. Frequency domain characteristics.

[0071] Frequency domain characteristics mainly reflect the spectral distribution of a signal. Using the Fast Fourier Transform (FFT) to perform a Discrete Fourier Transform on the IQ signal yields richer spectral information, as shown below:

[0072]

[0073] Among them, Y FFT (f) represents the signal Features in the frequency domain; N s Represents the number of signal samples; This represents the IQ signal.

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

[0075] 3. Time-frequency domain characteristics.

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

[0077] Let w(·) be a window function with length L. w L s This represents the window's movement step size. After windowing, the IQ signal is transformed into the frequency domain via Discrete Fourier Transform (STFT), as shown below:

[0078]

[0079] Among them, Y STFT (m,f) represents the STFT characteristics under discrete time component m and frequency component f.

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

[0081] II. Feature Extraction Module.

[0082] The feature extraction module is constructed by a Spatial Channel Attention Convolution (SCAConv) layer, 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.

[0083] 1. Design of five parallel channels.

[0084] In the feature extraction module, five parallel channels were designed to extract the five types of device fingerprint features mentioned above: I channel, Q channel, IQ channel, FFT channel, and STFT channel, to avoid mutual interference between features within 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 constructed using one-dimensional spatial channel attention convolutional layers (SCAConv); the time-frequency domain channels use two-dimensional spatial channel attention convolutional layers. The one-dimensional and two-dimensional spatial channel attention convolutional layers are, for example... Figure 3 and Figure 4 As shown, each channel consists of four layers of SCAConv cascaded sequentially, fully mining and characterizing the fingerprint features of their respective domains, ensuring that the network can fully learn the device fingerprint features contained in each channel.

[0086] The fingerprint features after passing through four spatial channel attention convolutional layers are denoted as F. I F Q F IQ F FFT F STFT .

[0087] 2. Spatial channel attention convolutional layer.

[0088] In traditional deep learning methods, convolutional neural networks use the same kernel weights for all fingerprints. However, the importance of each fingerprint is not the same, and different regions of the same fingerprint may also have different weights. Therefore, channel attention branches and spatial attention branches are introduced into the spatial channel attention convolutional layer to determine the overall weight of each fingerprint and the weight of each region of the fingerprint, respectively.

[0089] Therefore, the spatial channel attention convolutional layer consists of a fingerprint feature convolutional branch, a channel attention branch, a spatial attention branch, and an attention mechanism. Specifically, the fingerprint feature convolutional branch extracts fingerprint features from the input signal; the channel attention branch expresses the importance of each fingerprint feature; and the spatial attention branch expresses the local importance of each fingerprint feature.

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

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

[0092] Then, the fingerprint features at these three levels are fed into the attention mechanism to further emphasize their key elements, thereby enhancing 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. It consists of a single convolutional layer containing multiple convolutional kernels. Under the control of a specified stride, the convolutional kernels perform element-wise multiplication and addition operations on each sub-region of the signal to generate feature values. After traversing all sub-regions, the complete fingerprint features of the input signal are obtained. In deep learning models, the convolutional kernel weights are automatically updated based on the data to extract differences in the signal related to the hardware fingerprint.

[0095] The calculation formulas for the one-dimensional fingerprint feature convolution branch and the 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 kernel size, stride, and padding; bias(C) j ) represents the bias of the j-th convolutional kernel; out(B i Cj (H,W) represents the feature map output by the two-dimensional fingerprint feature convolution branch; B i L represents the batch size of the training data; C represents the length of the signal sequence. j Indicates the number of output channels; C n The input signal has [number] channels; k represents the input channel index; * indicates 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 These represent the height and width of the convolution kernel, respectively.

[0099] The formulas for calculating L, H, and W are as follows:

[0100]

[0101] Among them, L out and L in W represents the length of the output feature map and the length of the input signal, respectively. out and W in H represents the time component lengths of the output signal feature map and the input signal feature map, respectively; out and H in These represent the frequency component lengths of the output signal feature map and the input signal feature map, respectively; K represents the size of the convolution kernel; S is the convolution stride; and P is the padding size.

[0102] (2) Channel attention branch.

[0103] The core objective of the channel attention branch is to assign appropriate weights to each fingerprint feature, thereby distinguishing the importance of different channels in the recognition task.

[0104] A. The one-dimensional channel attention branch consists of two fully connected (FC) layers, with the Tanh activation function used between the layers, and the weights are finally output through the Sigmoid function.

[0105] The first fully connected layer reduces the dimensionality of the fingerprint features in the channel dimension, transforming them into a one-dimensional vector. Subsequently, the Tanh activation function performs a non-linear transformation on the dimensionality-reduced data, with the Tanh function ranging from [-1, 1]. This distribution in neural networks helps avoid gradient vanishing and enriches the feature representation. The second fully connected layer then increases the dimensionality of the vector to match the original number of channels. Finally, the Sigmoid function is used to scale the weights to [0, 1], facilitating the multiplication of the channel attention branch with the fingerprint feature convolution branch output.

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

[0107] First, a 2D convolutional layer is used to adjust the number of channels in the channel attention branch to match the number of channels in the fingerprint feature convolutional branch. Then, average pooling and max pooling are used to highlight the overall fingerprint feature map and salient regions, respectively, and to some extent reduce the spatial dimensionality, thus mitigating the risk of overfitting. The first fully connected layer reduces the dimensionality of the pooled fingerprint feature map in the channel dimension, allowing the network to focus more on key channel information. Subsequently, the Tanh activation function performs a non-linear transformation on the data. The second fully connected layer increases the dimensionality of the fingerprint feature map in the channel dimension, similar to the first convolutional layer, again matching the number of channels in the fingerprint feature convolutional branch. Finally, the Sigmoid function is used to scale the weights to [0,1] so that they are multiplied by channel with the fingerprint feature convolutional branch.

[0108] (3) Spatial attention branch.

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

[0110] A. The one-dimensional attention branch consists of convolutional layers, max pooling layers, deconvolutional layers, and upsampling layers on specific channels.

[0111] Convolutional layers reduce the dimensionality of the data, thus minimizing spatial weight distortion in fingerprints and eliminating channel interference. Max pooling layers extract the strongest local responses, highlighting key features of the input tensor. Deconvolutional layers reconstruct the input fingerprint and enhance its salient features. Upsampling layers restore the fingerprint features to a size that matches the fingerprint feature convolutional branch. The Sigmoid function scales the weights to [0,1], facilitating the multiplication of the spatial attention branch with the fingerprint feature convolutional branch.

[0112] B. The two-dimensional spatial attention branch consists of max pooling layers, average pooling layers, and convolutional layers.

[0113] Max pooling and average pooling are performed in parallel, which not only preserves the maximum local features of the feature map but also reflects the overall characteristics of the original feature map, achieving weighted summation for different spatial regions. Convolutional layers extract features within local regions, which can better capture contextual information in the spatial dimension, highlighting important regions and suppressing unimportant regions. Finally, the weights are scaled to [0,1] using the Sigmoid function, which facilitates the multiplication of the spatial attention branch with the fingerprint feature convolution branch.

[0114] III. Feature Fusion Module.

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

[0116] like Figure 5 As shown, the feature fusion module employs a shared attention strategy. First, it performs average pooling and max pooling on the output of each channel to aggregate spatial information and form two different contextual features. Second, these two contextual features are concatenated along the channel dimension and then convolved using a first-layer one-dimensional convolution operation activated by the ReLU function. This is followed by a second-layer one-dimensional convolution, and finally, the weights are scaled to [0,1] using the Sigmoid function. This facilitates element-wise multiplication of the attention map and fingerprint features, enhancing important parts and weakening unimportant parts. The obtained features are then denoted as F. r I F r Q F rIQ F r FFT and F r STFT These refined neural feature maps are then concatenated along the channel dimension to form a fused feature, as shown below:

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

[0118] Finally, the fused features are input into the device identification module to perform device fingerprint recognition.

[0119] IV. Device Identification Module.

[0120] The device identification module consists of two fully connected layers.

[0121] First, the fingerprint features, enhanced by feature extraction and attention mechanisms, are input into the first fully connected layer to reduce the dimensionality of the input features and perform a non-linear transformation using the ReLU activation function. This not only alleviates the vanishing gradient problem to some extent but also helps accelerate the model's training convergence. Then, the second fully connected layer projects the mapped vector onto the device category, outputting the final recognition result.

[0122] like Figure 6 As shown, based on the enhanced device fingerprint recognition system constructed above using time-frequency domain feature fusion, this invention also proposes an enhanced device fingerprint recognition method based on time-frequency domain feature fusion, applied to the above system, comprising the following steps:

[0123] Step S1: Use USRP (Universal Software Radio Peripheral) to acquire signals represented in in-phase and quadrature components, i.e., IQ symbol form.

[0124] Step S2: In the signal processing module, the acquired signal data is preprocessed. First, IQ is split into three time-domain features: I, Q and IQ. Then, using the corresponding FFT and STFT formulas, IQ is transformed to the frequency domain and time-frequency domain, and the corresponding features are extracted in these two domains.

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

[0126] Step S4: Send the features extracted by the feature extraction module into the feature fusion module to generate the final fused features.

[0127] Step S5: Finally, the fused features are passed to the classification model to complete the fingerprint matching and recognition of the device, realize the physical layer identity recognition of the wireless device, and improve the security of the wireless device.

[0128] Therefore, the present invention employs the above-mentioned enhanced device fingerprint recognition system and method based on time-frequency domain feature fusion, which improves the security of IoT devices and can be deployed on IoT devices with limited power consumption or older devices that cannot transmit encrypted data; it improves the accuracy and robustness of device recognition, and by supplementing frequency domain features, the system's sensitivity to noise interference is significantly reduced, maintaining high recognition accuracy even under low signal-to-noise ratio conditions; it reduces the cost of device fingerprint extraction, and can achieve high recognition accuracy even for single physical layer fingerprints (such as IQ signals), avoiding excessive feature extraction overhead.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An enhanced device fingerprint recognition system based on time-frequency domain feature fusion, characterized in that, It consists of a signal processing module, a feature extraction module, a feature fusion module, and a device recognition module; The signal processing module is used to preprocess the raw IQ data, namely the in-phase component I and the quadrature component Q. First, the signal processing module uses a moving window to segment the original IQ data into three time-domain features: I, Q, and IQ. The window movement step size is set to half the window size. Then, the IQ data is transformed into the frequency domain and time-frequency domain using the Fast Fourier Transform (FFT) and Discrete Fourier Transform (STFT) formulas, and the corresponding features are extracted in these two domains. Frequency domain characteristics reflect the spectral distribution of a signal. The spectral information obtained by performing a Discrete Fourier Transform (FFT) on the IQ signal is shown below: ； in, Indicates signal Characteristics in the frequency domain; Represents the number of signal samples; Indicates the IQ signal; Time-frequency domain characteristics calculate the frequency change of a signal over a local time range and are suitable for signals whose frequency changes over time. set up It is a window function with a length of ; The window's movement step size; after windowing, the IQ signal is transformed into the frequency domain via Discrete Fourier Transform (STFT), as shown below: ； in, Represents the discrete-time components and frequency components The STFT features below; The feature extraction module is constructed by the Spatial Channel Attention Convolutional 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; Five parallel channels were designed for five signal characteristics: 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. In terms of network structure, both the time-domain and frequency-domain channels employ one-dimensional spatial channel attention convolutional layers; the time-frequency domain channels employ two-dimensional spatial channel attention convolutional layers; each channel consists of four cascaded spatial channel attention convolutional layers; the fingerprint features after passing through the four spatial channel attention convolutional layers are denoted as follows: , , , , ; The spatial channel attention convolutional layer consists of a fingerprint feature convolutional branch, a channel attention branch, a spatial attention branch, and an attention mechanism; The feature fusion module is used to fuse the features output from the I channel, Q channel, IQ channel, FFT channel, and STFT channel to generate the final enhanced fingerprint feature.

2. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1, characterized in that: The fingerprint feature convolution branch is used to extract fingerprint features from 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 convolutional layer is as follows: First, the output of the fingerprint feature convolution branch is multiplied by the outputs 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 to enhance the overall feature representation of the device fingerprint.

3. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1, characterized in that: The fingerprint feature convolution branch is used to extract fingerprint features from the input signal. It consists of a single convolutional layer containing multiple convolutional kernels. The convolutional kernels perform element-wise multiplication and addition operations on each sub-region of the signal under the control of a specified stride, thereby generating feature values. After traversing all sub-regions, the complete fingerprint features of the input signal are obtained. The calculation formulas for the one-dimensional fingerprint feature convolution branch and the two-dimensional fingerprint feature convolution branch are as follows: ； ； in, The feature map represents the output of the one-dimensional fingerprint feature convolution branch. The size of the feature map is determined by the convolution kernel size, stride, and padding. This represents the bias of the j-th convolutional kernel; This represents the feature map output by the convolution branch of the two-dimensional fingerprint feature; Indicates the batch size of the training data; Indicates the length of the signal sequence; Indicates the number of output channels; Indicates the number of channels for the input signal; Indicates the input channel index; Indicates cross-correlation operation; Represents the frequency components of the signal feature map; Represents the time component of the signal feature map; and These represent the height and width of the convolution kernel, respectively. in, , , The calculation formula is as follows: ； ； ； in, and These represent the length of the output feature map and the length of the input signal, respectively. and These represent the time component lengths of the output signal feature map and the input signal feature map, respectively. and These represent the frequency component lengths of the output signal feature map and the input signal feature map, respectively. Indicates the size of the convolution kernel; The convolution stride; This is the fill size.

4. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1, characterized in that: The channel attention branch assigns corresponding weights to each fingerprint feature, distinguishing the importance of different channels in the recognition task; A. The one-dimensional channel attention branch consists of two fully connected (FC) layers, with the Tanh activation function used between the layers, and the weights are finally output through the Sigmoid function. The first fully connected layer reduces the dimensionality of the fingerprint features in the channel dimension, converting them into a one-dimensional vector. Then, the Tanh activation function performs a non-linear transformation on the dimensionality-reduced data, and the range of the Tanh function is [-1,1]. The second fully connected layer then increases the dimensionality of the vector to match the original number of channels. Finally, the Sigmoid function is used to scale the weights to [0,1], and the output of the channel attention branch is multiplied by the output of the fingerprint feature convolution branch. B. The two-dimensional channel attention branch consists of one convolutional layer, one pooling layer, and two fully connected layers. First, the number of channels in the channel attention branch is adjusted to match the number of channels in the fingerprint feature convolution branch using a two-dimensional convolutional layer; Next, average pooling and max pooling are used to highlight the overall and salient regions 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 non-linear transformation on the data. The second fully connected layer increases the dimension of the fingerprint feature map in the channel dimension, the same as the first convolutional layer, and matches the number of channels of the fingerprint feature convolutional branch. Finally, the Sigmoid function is used to scale the weights to [0,1] and multiply them by the fingerprint feature convolutional branch by channel.

5. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1, characterized in that: The role of the spatial attention branch is to assign weights to each spatial location of the fingerprint feature, highlighting key regions and suppressing irrelevant and noisy regions; A. The one-dimensional attention branch consists of a convolutional layer, a max pooling layer, a deconvolutional layer, and an upsampling layer; The data is reduced in dimensionality through convolutional layers; max pooling layers are used to extract local responses and highlight key features of the input tensor; deconvolutional layers are used to recover the input fingerprint and enhance its salient features; upsampling layers are used to restore the fingerprint features to a size that matches the fingerprint feature convolutional branch; and the sigmoid function is used to scale the weights to [0,1] to facilitate the multiplication of the spatial attention branch with the fingerprint feature convolutional branch. B. The two-dimensional spatial attention branch consists of max pooling layers, average pooling layers, and convolutional layers; Max pooling and average pooling are performed in parallel to preserve the maximum local features of the feature map and the overall characteristics of the original feature map, thereby achieving weighting of different spatial regions. Convolutional layers extract features within local regions, capturing contextual information in the spatial dimension, highlighting important regions and suppressing unimportant regions. Finally, the weights are scaled to [0,1] using the Sigmoid function, which facilitates the multiplication of the spatial attention branch with the fingerprint feature convolutional branch.

6. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1, characterized in that: The feature fusion module employs a shared attention strategy. First, it performs average pooling and max pooling on the output of each channel to aggregate spatial information and form two different contextual features. Second, it concatenates these two contextual features along the channel dimension and performs a convolution operation using a first-layer one-dimensional convolution, activated by the ReLU function. This is followed by a second-layer one-dimensional convolution, and finally, the Sigmoid function is used to scale the weights to [0,1]. The attention map is then element-wise multiplied with the fingerprint features to enhance important parts and weaken unimportant parts. The resulting features are then denoted as follows: , , , and The obtained neural feature maps are concatenated along the channel dimension to form a fused feature, as shown below: ； Finally, the fused features are input into the device identification module to perform device fingerprint recognition.

7. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to claim 1, characterized in that: The device identification module consists of two fully connected layers. First, the fingerprint features, which have been enhanced by feature extraction and attention mechanism, are input into the first fully connected layer to reduce the dimensionality of the input features and perform nonlinear transformation through the ReLU activation function. Then, the second fully connected layer projects the mapped vectors onto the device category, outputting the final recognition result.

8. The enhanced device fingerprint recognition system based on time-frequency domain feature fusion according to any one of claims 1-7, proposes an enhanced device fingerprint recognition method based on time-frequency domain feature fusion, applied to the above system, characterized in that, Includes the following steps: Step S1: Use USRP to acquire signals represented in in-phase and quadrature components, i.e., in IQ symbol form; Step S2: In the signal processing module, the acquired signal data is preprocessed. First, IQ is split into three time-domain features: I, Q and IQ. Then, using the corresponding FFT and STFT formulas, IQ is transformed to the frequency domain and time-frequency domain, and the corresponding features are extracted in these two domains. Step S3: Input the features in the time domain, frequency domain, and time-frequency domain into the feature extraction module for feature extraction; Step S4: Send the features extracted by the feature extraction module into the feature fusion module to generate the final fused features; Step S5: Finally, the fused features are passed to the classification model to complete the fingerprint matching and recognition of the device, thereby realizing the identification of the physical layer of the wireless device.

Images