A Wi-Fi radio frequency fingerprinting method, storage medium and device

Abstract

This invention discloses a Wi-Fi radio frequency fingerprinting method, storage medium, and device, belonging to the field of wireless communication security technology. The invention addresses the problems of incomplete signal features extracted by existing methods, sensitivity of features to channel changes, and loss of key phase correlation information, leading to low recognition accuracy. The network model of this invention integrates time-domain, frequency-domain, and time-frequency-domain features through a multi-channel architecture, combining the local feature extraction capabilities of two channels with the global relationship modeling capabilities of a Transformer encoder, forming a more comprehensive and discriminative feature representation system. Furthermore, the model can directly process complex signals, completely preserving the amplitude and phase information of Wi-Fi signals, avoiding the loss of phase correlation information caused by separate processing of real and imaginary components. By extracting segmented relative power spectrum features, hardware features and channel features can be effectively separated, significantly reducing the impact of time-varying channels on recognition accuracy. This invention can be applied to the field of wireless communication security.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication security technology, specifically relating to a Wi-Fi radio frequency fingerprint recognition method, storage medium, and device. Background Technology

[0002] With the popularization of IoT and wireless communication technologies, in scenarios such as smart manufacturing, smart healthcare, and smart parks, a large number of Wi-Fi terminals need to complete "network access authentication" in zero contact and milliseconds.

[0003] Traditional authentication schemes rely on MAC addresses or digital certificates, but:

[0004] 1) MAC addresses can be forged with a single click at the driver layer, allowing for on-site breaches;

[0005] 2) Digital certificates require pre-installed keys or SE security elements on each terminal, resulting in long deployment cycles and high costs. Furthermore, certificate revocation / renewal requires line shutdown, which cannot meet the 7×24 flexible production cycle.

[0006] 3) Once the certificate private key is leaked in batches, it will cause the entire network to be paralyzed, and the recovery time will be measured in days, directly causing production line shutdown losses.

[0007] Radio frequency fingerprinting, as a physical layer non-cryptographic authentication technology, achieves identity recognition by extracting unique features of wireless device hardware, possessing inherent anti-counterfeiting and high security potential. Existing radio frequency fingerprinting methods mainly include: Reference 1 (Riyaz S, Kokalj-Filipovic S, Reising D. Deep Learning Convolutional Neural Networks for Radio Identification[J]. IEEE Communications Magazine, 2018, 56(10): 11-15.) directly uses the time-domain sampling points of the received signal as feature input to the classifier, without considering the influence of channel multipath effects; its recognition performance drops sharply when channel conditions change. Reference 2 (Merchant K, Revay S, Stantchev Z, et al. Deep Learning for RF Device Fingerprinting in Cognitive Communication Networks[J]. IEEE Journal on Selected Areas in Communications,2018, 36(10): 2160-2172.," IEEE JSAC,

[2018] Extracting the power spectrum of the signal in the frequency domain as a device fingerprint. This method decomposes the complex signal into real and imaginary parts for separate processing, loses phase correlation information, and the single frequency domain feature is insufficient for hardware representation; Reference 3 (Tu Y, Zhang J, Ma X, et al. Deep Residual Learning for RF Fingerprinting[J]. IEEE Internet of Things Journal, 2022, 9(11): 8854-8864.) uses STFT or wavelet transform to obtain time-frequency joint features, but fails to effectively separate the channel response from the device hardware features, and traditional real neural networks have difficulty processing the complete information of complex signals.

[0008] In summary, the existing technology mainly has the following problems:

[0009] 1) Failure to effectively separate the channel transfer function from the device hardware fingerprint, resulting in the feature being sensitive to channel changes;

[0010] 2) Traditional neural networks are mostly designed for real signals. When processing complex Wi-Fi signals, the I and Q signals need to be separated, which may cause the features to lose key phase correlation information and reduce the discriminative power of the features.

[0011] 3) A single model architecture is difficult to capture the time domain, frequency domain, and global and local features of a signal simultaneously.

[0012] Therefore, the above problems directly lead to a decrease in the accuracy of existing radio frequency fingerprint recognition methods, making it essential to propose a new radio frequency fingerprint recognition method to solve these problems. Summary of the Invention

[0013] This invention addresses the problems of low recognition accuracy caused by incomplete signal features extracted by existing methods, sensitivity of features to channel changes, and loss of key phase correlation information. It proposes a Wi-Fi radio frequency fingerprint recognition method to improve the robustness and accuracy of radio frequency fingerprint recognition.

[0014] The technical solution adopted by the present invention to solve the above-mentioned technical problems is: a Wi-Fi radio frequency fingerprint recognition method, the method specifically including the following steps:

[0015] Step 1: After the device to be accessed sends a Wi-Fi radio frequency signal, the receiving end collects the Wi-Fi radio frequency signal.

[0016] Step 2: The receiving end synchronizes the time of the collected Wi-Fi radio frequency signal and extracts the preamble portion from the time-synchronized Wi-Fi radio frequency signal. The preamble portion includes a short preamble sequence and a long preamble sequence.

[0017] Step 3: Perform short-time Fourier transform on the short and long preamble sequences using a sliding window of length W and overlap S respectively to obtain the power spectral density of the signal within each time window;

[0018] Step 4: Construct the relative power spectrum matrix based on the power spectral density of the short preamble sequence at each frequency and time window, and the power spectral density of the long preamble sequence at each frequency and time window. ;

[0019] Then consider the relative power spectrum matrix Extreme value clipping is performed to obtain a two-dimensional relative power spectrum time-frequency diagram;

[0020] Step 5: Construct a multi-channel complex ResNet-Transformer network, which includes a one-dimensional complex ResNet branch and a two-dimensional complex ResNet branch; wherein:

[0021] The input to the one-dimensional complex ResNet branch is the result of splicing the FFT transform results of the time-synchronized Wi-Fi RF signal and the time-synchronized Wi-Fi RF signal along the channel dimension.

[0022] The input to the two-dimensional complex ResNet branch is a two-dimensional relative power spectrum time-frequency plot;

[0023] The identification results of the devices to be accessed are output through a multi-channel complex ResNet-Transformer network.

[0024] Furthermore, in step three, a short-time Fourier transform is performed on the short preamble sequence to obtain the power spectral density of the signal within each time window, specifically as follows:

[0025] In frequency Down:

[0026] A short-time Fourier transform is performed on the short preamble sequence using a sliding window of length W and overlap S to obtain the spectrum of the signal within each time window corresponding to the short preamble sequence. Then, the power spectral density of the signal within each time window corresponding to the short preamble sequence is obtained from the spectrum. The first short code sequence corresponding to the next short code sequence The power spectral density of the signal within each time window is denoted as . .

[0027] Furthermore, in step three, a short-time Fourier transform is performed on the long preamble sequence to obtain the power spectral density of the signal within each time window, specifically as follows:

[0028] In frequency Down:

[0029] A short-time Fourier transform is performed on the long preamble sequence using a sliding window of length W and overlap S to obtain the spectrum of the signal within each time window corresponding to the long preamble sequence. Then, the power spectral density of the signal within each time window corresponding to the long preamble sequence is obtained from the spectrum. The first corresponding to the lower long guide sequence The power spectral density of the signal within each time window is denoted as . .

[0030] Furthermore, the relative power spectrum matrix The construction method is as follows:

[0031] Divide the power spectral density of the signal within the corresponding time window of the short precode sequence by the power spectral density of the long precode sequence:

[0032]

[0033] in, Represents the relative power spectrum matrix The Middle Line number Column elements, It represents any small positive number.

[0034] Furthermore, the relative power spectrum matrix Extreme value clipping is performed to obtain the two-dimensional relative power spectrum time-frequency diagram, specifically:

[0035] For the relative power spectrum matrix Sort the elements in any row in ascending order, keeping the elements that are not in the first 5% and the elements that are not in the last 5%, and replace the elements in the first 5% with the minimum value to be retained, and replace the elements in the last 5% with the maximum value to be retained.

[0036] Similarly, for the relative power spectrum matrix Each row is processed separately, and the resulting relative power spectrum matrix is ​​the two-dimensional relative power spectrum time-frequency diagram.

[0037] Furthermore, the operation of the multi-channel complex ResNet-Transformer network is as follows:

[0038] The FFT transform results of the time-synchronized Wi-Fi RF signal and the time-synchronized Wi-Fi RF signal are concatenated along the channel dimension as the input of the first complex convolutional layer, and the output of the first complex convolutional layer is used as the input of the first complex normalization layer.

[0039] The output of the first complex normalization layer is used as the input of the first complex activation function layer, and then the output of the first complex activation function layer is used as the input of the first complex max pooling layer.

[0040] The output of the first complex max pooling layer is used as the input of the first residual block, and the output of the first residual block is used as the input of the second residual block.

[0041] The output of the second residual block is used as the input of the third residual block, and then the output of the third residual block is used as the input of the fourth residual block.

[0042] The output of the fourth residual block is used as the input of the second complex max pooling layer, and the output of the second complex max pooling layer is used as the input of the first complex fully connected layer.

[0043] The two-dimensional relative power spectrum time-frequency plot is used as the input of the second complex convolutional layer, and the output of the second complex convolutional layer is used as the input of the second complex normalization layer.

[0044] The output of the second complex normalization layer is used as the input of the second complex activation function layer, and then the output of the second complex activation function layer is used as the input of the third complex max pooling layer.

[0045] The output of the third complex max pooling layer is used as the input of the fifth residual block, and the output of the fifth residual block is used as the input of the sixth residual block.

[0046] Use the output of the sixth residual block as the input of the seventh residual block, and then use the output of the seventh residual block as the input of the eighth residual block.

[0047] The output of the eighth residual block is used as the input of the fourth complex max pooling layer, and the output of the fourth complex max pooling layer is used as the input of the second complex fully connected layer.

[0048] The outputs of the first complex fully connected layer and the second complex fully connected layer are concatenated along the channel dimension to obtain the concatenated result a;

[0049] Position encoding is performed on the concatenated result a. The position encoding result is used as the input of the Transformer encoder. The output of the Transformer encoder is then concatenated with a along the channel dimension to obtain the concatenated result b.

[0050] The splicing result b is used as the input to the classification layer, and the classification layer outputs the identification result of the device to be accessed.

[0051] Furthermore, the working process of the first residual block is as follows:

[0052] Within the first residual block, the output of the first complex max pooling layer is used as the input of the third and fourth complex convolutional layers, and the output of the third complex convolutional layer is used as the input of the third complex normalization layer.

[0053] The output of the third complex normalization layer is used as the input of the third complex activation function layer, and then the output of the third complex activation function layer is used as the input of the fifth complex convolutional layer.

[0054] The output of the fifth complex convolutional layer is used as the input of the fourth complex normalization layer. Then, the outputs of the fourth complex convolutional layer and the fourth complex normalization layer are concatenated along the channel dimension to obtain the concatenated result c.

[0055] The concatenated result c is used as the input of the sixth complex convolutional layer, the output of the sixth complex convolutional layer is used as the input of the fifth complex normalization layer, and the output of the fifth complex normalization layer is used as the input of the fourth complex activation function layer.

[0056] The output of the fourth complex activation function layer is used as the input of the seventh complex convolutional layer, and the output of the seventh complex convolutional layer is used as the input of the sixth complex normalization layer.

[0057] c is concatenated with the output of the sixth complex normalization layer, and the concatenation result is used as the output of the first residual block.

[0058] A computer storage medium storing at least one instruction, which is loaded and executed by a processor to implement the Wi-Fi radio frequency fingerprint recognition method.

[0059] A Wi-Fi radio frequency fingerprint recognition device includes a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the Wi-Fi radio frequency fingerprint recognition method.

[0060] The beneficial effects of this invention are:

[0061] 1. Information integrity: The network model constructed in this invention can directly process complex signals, completely preserving the amplitude and phase information of Wi-Fi signals, and avoiding the loss of phase correlation information caused by separate processing of real and virtual parts.

[0062] 2. Feature complementarity: The model of this invention integrates time-domain, frequency-domain and time-frequency-domain features through a multi-channel architecture, and combines the local feature extraction capabilities of the two channels with the global relationship modeling capabilities of the Transformer encoder to form a more comprehensive and discriminative feature representation system.

[0063] 3. High robustness: The segmented relative power spectrum feature proposed in this invention, through theoretical derivation and experimental verification, can effectively separate hardware features and channel features, significantly reducing the impact of time-varying channels on recognition accuracy, and exhibiting excellent cross-day stability in tests with data from different dates.

[0064] 4. High accuracy: Experiments on public datasets show that the method of this invention has a significantly improved recognition accuracy compared to other comparative models.

[0065] 5. Easy to use: No need to modify existing Wi-Fi terminals, nor to use digital certificates or security elements. Attached Figure Description

[0066] Figure 1 This is a flowchart of a Wi-Fi radio frequency fingerprint recognition method according to the present invention;

[0067] Figure 2 This is a schematic diagram of the Wi-Fi signal physical layer preamble frame format;

[0068] Figure 3 This is a schematic diagram of the ResNet residual block structure;

[0069] Figure 4 This is a diagram of the overall architecture of the multi-channel complex ResNet-Transformer network;

[0070] Figure 5This is a visualization of the original IQ signal features under t-SNE dimensionality reduction;

[0071] Figure 6 It is a visualization of the frequency domain signal characteristics of the original time-domain signal after Fourier transform under t-SNE dimensionality reduction;

[0072] Figure 7 It is a visualization of the spectral characteristics of the original time-domain signal after short-time Fourier transform in t-SNE dimensionality reduction;

[0073] Figure 8 This is a visualization of the power spectral density characteristics of the original time-domain signal under t-SNE dimensionality reduction;

[0074] Figure 9 This is a visualization of the piecewise relative time-frequency features of the present invention under t-SNE dimensionality reduction;

[0075] Figure 10 This is a comparison chart of the recognition accuracy of the method of this invention and the comparative model. Detailed Implementation

[0076] This invention relates to a method based on the IEEE 802.11 preamble of Wi-Fi radio frequency signals, specifically a method for extracting and identifying Wi-Fi radio frequency fingerprints using a multi-channel complex neural network. A multi-channel complex ResNet-Transformer network is used to classify the Wi-Fi radio frequency signals transmitted by the device to be accessed, in order to identify whether the device is legitimate. This invention is applicable to scenarios such as IoT device access authentication and wireless LAN security protection. The method is described in detail below with reference to the accompanying drawings:

[0077] Specific implementation method one: Combining Figure 1 This embodiment describes a Wi-Fi radio frequency fingerprint recognition method, which specifically includes the following steps:

[0078] Step 1: After the device to be accessed sends a Wi-Fi radio frequency signal, the receiving end uses the ZYNQ baseband board and AD9361 radio frequency board to collect the Wi-Fi radio frequency signal;

[0079] Step 2: The receiving end synchronizes the time of the acquired Wi-Fi radio frequency signal and extracts the preamble portion from the time-synchronized Wi-Fi radio frequency signal, such as... Figure 2 As shown, the preamble portion includes a short preamble sequence (STF) and a long preamble sequence (LTF).

[0080] Step 3: Perform short-time Fourier transform on the short and long preamble sequences using a sliding window of length W and overlap S respectively to obtain the power spectral density of the signal within each time window;

[0081] Specifically, in frequency Down:

[0082] A short-time Fourier transform is performed on the short preamble sequence using a sliding window of length W and overlap S to obtain the spectrum of the signal within each time window corresponding to the short preamble sequence. Then, the power spectral density of the signal within each time window corresponding to the short preamble sequence is obtained from the spectrum. The first short code sequence corresponding to the next short code sequence The power spectral density of the signal within each time window is denoted as . ;

[0083] A short-time Fourier transform is performed on the long preamble sequence using a sliding window of length W and overlap S to obtain the spectrum of the signal within each time window corresponding to the long preamble sequence. Then, the power spectral density of the signal within each time window corresponding to the long preamble sequence is obtained from the spectrum. The first corresponding to the lower long guide sequence The power spectral density of the signal within each time window is denoted as . ;

[0084] Step 4: Construct the relative power spectrum matrix based on the power spectral density of the short preamble sequence at each frequency and time window, and the power spectral density of the long preamble sequence at each frequency and time window. ;

[0085] Specifically, the power spectral density of the signal within the corresponding time window of the short preamble sequence is divided by the power spectral density of the long preamble sequence, and a local minimum is added to the denominator to enhance numerical stability. :

[0086]

[0087] in, Represents the relative power spectrum matrix The Middle Line number Column elements, Represents any small positive number;

[0088] Then consider the relative power spectrum matrix Extreme value clipping is performed to obtain a two-dimensional relative power spectrum time-frequency diagram; specifically:

[0089] For the relative power spectrum matrix Sort the elements in any row in ascending order, keeping the elements that are not in the first 5% and the elements that are not in the last 5%, and replace the elements in the first 5% with the minimum value to be retained, and replace the elements in the last 5% with the maximum value to be retained.

[0090] It should be noted that extreme value pruning only reassigns the smaller and larger elements in each row, and does not change the position of each element in the row.

[0091] Similarly, for the relative power spectrum matrix Each row is processed separately. Extreme value pruning can suppress residual outliers, resulting in a channel drift-resistant relative power spectrum matrix. The processed relative power spectrum matrix is ​​the two-dimensional relative power spectrum time-frequency diagram.

[0092] Step 5: Construct a multi-channel complex ResNet-Transformer network, which includes a one-dimensional complex ResNet branch and a two-dimensional complex ResNet branch; wherein:

[0093] The input to the one-dimensional complex ResNet branch is the result of splicing the FFT transformation results of the time-synchronized Wi-Fi RF signal and the time-synchronized Wi-Fi RF signal along the channel dimension. The one-dimensional complex ResNet branch can retain complete amplitude-phase joint information.

[0094] The input to the two-dimensional complex ResNet branch is a two-dimensional relative power spectrum time-frequency plot, and the two-dimensional complex ResNet branch can extract joint time-frequency features;

[0095] The features extracted by the one-dimensional complex ResNet branch and the two-dimensional complex ResNet branch are concatenated along the channel dimension. Then, the concatenated features are modeled by the Transformer encoder to model the long-distance dependencies between features, thereby achieving deep fusion of local detailed features and global related features.

[0096] Specifically, such as Figure 4 As shown, the operation of the multi-channel complex ResNet-Transformer network is as follows:

[0097] The FFT transform results of the time-synchronized Wi-Fi RF signal and the time-synchronized Wi-Fi RF signal are concatenated along the channel dimension as the input of the first complex convolutional layer, and the output of the first complex convolutional layer is used as the input of the first complex normalization layer.

[0098] The output of the first complex normalization layer is used as the input of the first complex activation function layer, and then the output of the first complex activation function layer is used as the input of the first complex max pooling layer.

[0099] The output of the first complex max pooling layer is used as the input of the first residual block, and the output of the first residual block is used as the input of the second residual block.

[0100] The output of the second residual block is used as the input of the third residual block, and then the output of the third residual block is used as the input of the fourth residual block.

[0101] The output of the fourth residual block is used as the input of the second complex max pooling layer, and the output of the second complex max pooling layer is used as the input of the first complex fully connected layer.

[0102] The two-dimensional relative power spectrum time-frequency plot is used as the input of the second complex convolutional layer, and the output of the second complex convolutional layer is used as the input of the second complex normalization layer.

[0103] The output of the second complex normalization layer is used as the input of the second complex activation function layer, and then the output of the second complex activation function layer is used as the input of the third complex max pooling layer.

[0104] The output of the third complex max pooling layer is used as the input of the fifth residual block, and the output of the fifth residual block is used as the input of the sixth residual block.

[0105] Use the output of the sixth residual block as the input of the seventh residual block, and then use the output of the seventh residual block as the input of the eighth residual block.

[0106] The output of the eighth residual block is used as the input of the fourth complex max pooling layer, and the output of the fourth complex max pooling layer is used as the input of the second complex fully connected layer.

[0107] The outputs of the first complex fully connected layer and the second complex fully connected layer are concatenated along the channel dimension to obtain the concatenated result a;

[0108] Position encoding is performed on the concatenated result a. The position encoding result is used as the input of the Transformer encoder. The output of the Transformer encoder is then concatenated with a along the channel dimension to obtain the concatenated result b.

[0109] The splicing result b is used as the input of the classification layer, and the classification layer outputs the identification result of the device to be accessed. The identification result includes whether the device to be accessed is a known legitimate device or does not belong to a known legitimate device.

[0110] like Figure 3 As shown, each residual block in this invention has the same structure. Taking the first residual block as an example, the working process of the residual block will be described in detail below:

[0111] Within the first residual block, the output of the first complex max pooling layer is used as the input of the third and fourth complex convolutional layers, and the output of the third complex convolutional layer is used as the input of the third complex normalization layer.

[0112] The output of the third complex normalization layer is used as the input of the third complex activation function layer, and then the output of the third complex activation function layer is used as the input of the fifth complex convolutional layer.

[0113] The output of the fifth complex convolutional layer is used as the input of the fourth complex normalization layer. Then, the outputs of the fourth complex convolutional layer and the fourth complex normalization layer are concatenated along the channel dimension to obtain the concatenated result c.

[0114] The concatenated result c is used as the input of the sixth complex convolutional layer, the output of the sixth complex convolutional layer is used as the input of the fifth complex normalization layer, and the output of the fifth complex normalization layer is used as the input of the fourth complex activation function layer.

[0115] The output of the fourth complex activation function layer is used as the input of the seventh complex convolutional layer, and the output of the seventh complex convolutional layer is used as the input of the sixth complex normalization layer.

[0116] c is concatenated with the output of the sixth complex normalization layer, and the concatenation result is used as the output of the first residual block.

[0117] By employing complex convolution, complex normalization, and complex activation functions in each layer of the network, the complex form of the signal can be maintained throughout the forward propagation process, avoiding the loss of phase information caused by the separation of the real and imaginary parts. Specific Implementation Method Two:

[0119] This embodiment is a computer storage medium that stores at least one instruction, which is loaded and executed by a processor to implement the Wi-Fi radio frequency fingerprint recognition method.

[0120] It should be understood that the instructions include computer program products, software, or computerized methods corresponding to any method described in this invention; the instructions can be used to program computer systems or other electronic devices. Computer storage media may include readable media on which instructions are stored, and may include, but are not limited to, magnetic storage media, optical storage media; magneto-optical storage media include read-only memory (ROM), random access memory (RAM), erasable programmable memory (e.g., EPROM and EEPROM), and flash memory layers, or other types of media suitable for storing electronic instructions. Specific implementation method three:

[0122] This embodiment is a Wi-Fi radio frequency fingerprint recognition device. The device includes a processor and a memory. It should be understood that it includes any device including a processor and a memory described in this invention. The device may also include other units or modules that perform display, interaction, processing, control and other functions through signals or instructions.

[0123] The memory stores at least one instruction, which is loaded and executed by the processor to implement the Wi-Fi radio frequency fingerprint recognition method.

[0124] Those skilled in the art will understand that at least one stored instruction is a computer program product corresponding to a method or system. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0125] This application is described with reference to flowchart illustrations and / or block diagrams of methods, systems, and computer program products according to embodiments of this application, and can also be used with corresponding devices. It should be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0128] The theoretical derivation of the segmented relative power spectrum characteristics of this invention is based on the following three assumptions:

[0129] Assumption 1, transmitting signal To determine the signal or wide stationary random process;

[0130] Assumption 2: The channel h(n) remains unchanged during the STF and LTF durations (approximately 8μs + 8μs);

[0131] Assumption 3: The additive noise is zero-mean Gaussian white noise, which is uncorrelated with the signal;

[0132] For discrete input signals Performing the STFT (Short-Time Fourier Transform) yields:

[0133] (1)

[0134] in, Yes The complex matrix obtained after performing STFT is the first... Line number Column elements, complex matrix , Indicates the total number of frequencies. Indicates the total number of time windows. It is a window function with a window length of . The sliding distance is .

[0135] This invention uses a rectangular window, that is:

[0136] (2)

[0137] Substituting the received Wi-Fi short preamble into equation (1), the resulting complex matrix can be written as:

[0138] (3)

[0139] in, Indicates the received short preamble After being divided by sliding windows, the first The spectrum of a segment signal.

[0140] Similarly, the received Wi-Fi long preamble Substituting into equation (1), the resulting complex matrix can be written as:

[0141] (4)

[0142] According to Paswell's theorem, the square of the Fourier transform result of the signal is defined as the energy spectrum, and the time average of the energy spectral density is the power spectral density. Therefore, from equations (3) and (4), we can obtain the... Power spectral density of segment signal:

[0143] (5)

[0144] (6)

[0145] The Wiener-Khinchin theorem states that the Fourier transform of the autocorrelation function is the autopower spectral density, therefore:

[0146] (7)

[0147] (8)

[0148] According to ,make To obtain an ideal received signal with noise removed, then:

[0149] (9)

[0150] in, Indicates the first in the input signal The symbol, and the first The symbol in the first Within a time window;

[0151] Generally, the useful signal and noise can be considered uncorrelated, and their cross-correlation coefficient is 0, i.e.

[0152] (10)

[0153] Therefore, the autocorrelation function of equation (9) can be simplified to:

[0154] (11)

[0155] In practical applications, signals are always of finite length. The cross-correlation function for a finite-length discrete signal can also be expressed as:

[0156] (12)

[0157] Therefore, the correlation value between the two sequences within the short and long preambles after complex autocorrelation can partially eliminate additive noise. The effects are present, but cannot be completely eliminated.

[0158] Under high signal-to-noise ratio conditions, ignoring the influence of noise, equation (12) can be simplified to:

[0159] (13)

[0160] From equations (7), (8), and (12), we get:

[0161] (14)

[0162] make The power spectral density of the ideal received signal to remove noise. The power spectral density of the noise. It is a transmitted signal Given the power spectral density, according to Paswell's theorem, equation (14) can be expressed as:

[0163] (15)

[0164] Therefore, it is possible to... and Divide them to construct a piecewise relative power spectrum characteristic:

[0165] (16)

[0166] Under high signal-to-noise ratio conditions If we can ignore it, then equation (16) can be simplified to:

[0167] (17)

[0168] Therefore, under high signal-to-noise ratio conditions, the segmented relative power spectrum feature can effectively remove the influence of the channel and partially remove the influence of noise. Thus, this feature can be used as a fingerprint expression for wireless Wi-Fi signal transmitting devices, and this feature only retains the "device fingerprint" of transmitter hardware nonlinearity, I / Q imbalance, power amplifier memory effect, etc.

[0169] Experimental Section

[0170] The software environment selected for this invention is the Ubuntu system, using the PyTorch framework for model training and testing. The dataset used in the experiment is the ManySig subset of the public dataset WiSig. The WiSig dataset contains preambles of Wi-Fi radio frequency signals collected by multiple devices at multiple time periods. The ManySig subset contains data subsets collected by six devices on March 1, 8, 15, and 23, 2022. The dataset is divided into four subsets, D1, D2, D3, and D4, according to time. 300 data points are randomly selected from each subset, for a total of 1200 data points. The original IQ signal (i.e., the original preamble data), the frequency domain signal after Fourier transform of the original time domain signal, the spectrum of the original time domain signal after short-time Fourier transform, the power spectral density of the original time domain signal, and the segmented relative time-frequency characteristics of this invention are selected as contrast features. The t-SNE algorithm is used to visualize the contrast features. The visualization results are as follows: Figure 5 , Figure 6 , Figure 7 , Figure 8 and Figure 9 As shown.

[0171] Following a 6:2:2 ratio and combining the rules in Tables 1 and 2, the entire ManyTx subset is divided into three parts: a training set, a validation set, and a test set. (Table 1 requires both the training and test sets to include four days of data. Table 2 includes two sets of experiments: the first requires the training set to include only data from day 1 to day 3, and the test set to include only data from day 4; the second requires the training set to include only data from day 2 to day 4, and the test set to include only data from day 1.) The network constructed in this invention is trained, validated, and tested using the training, validation, and test sets. The optimizer is Adam, and the learning rate is set to 1e-3. After training, the process of this invention is executed on new received signals from the device to be identified. The trained model is used for classification and decision-making to achieve device authentication.

[0172] The chosen comparison model is:

[0173] 1. Divide the complex signal into two parts: the real part and the imaginary part, and then use both parts as the input to a dual-channel ResNet-18 model;

[0174] 2. Divide the complex signal into real and imaginary parts, and then use both parts as input to the dual-channel ShuffleNet-V2 model;

[0175] 3. Divide the complex signal into real and imaginary parts, then truncate the real and imaginary parts respectively. Use the truncated real part subsequences to form the real part image, and use the truncated imaginary part subsequences to form the imaginary part image. Use the real part image and the imaginary part image as input to the DensnNet-121 model.

[0176] 4. The complex signal is directly used as the input of the Complex ResNet model (compared to the method of this invention, this model only includes one-dimensional channels, and the output of the complex fully connected layer in the one-dimensional channel is directly used as the input of the classification layer, and the output of the classification layer is directly used as the device identification result).

[0177] 5. The one-dimensional complex signal, the frequency domain signal, and the two-dimensional piecewise relative power spectrum are used together as inputs to the dual-channel ComplexResNet-Transformer model (i.e., the model of the method of this invention).

[0178] The comparison of the five methods is shown in Tables 1 and 2:

[0179] Table 1 Overall model recognition accuracy

[0180]

[0181] Experimental results show that the method of the present invention achieves a recognition accuracy of 91.49% on the overall test set. Figure 10 As shown, it can also be seen that the model proposed in this invention is superior to the other four models compared.

[0182] Table 2. Model stability verification across days

[0183]

[0184] In cross-day tests (e.g., training using data from days D1-D3 and testing using data from day D4), the recognition accuracy of the method of this invention was 71.11%, significantly higher than other features and models, demonstrating the robustness of the method of this invention to time-varying channels.

[0185] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A Wi-Fi radio frequency fingerprint recognition method, characterized in that, The method specifically includes the following steps: Step 1: After the device to be accessed sends a Wi-Fi radio frequency signal, the receiving end collects the Wi-Fi radio frequency signal. Step 2: The receiving end synchronizes the time of the collected Wi-Fi radio frequency signal and extracts the preamble portion from the time-synchronized Wi-Fi radio frequency signal. The preamble portion includes a short preamble sequence and a long preamble sequence. Step 3: Perform short-time Fourier transform on the short and long preamble sequences using a sliding window of length W and overlap S respectively to obtain the power spectral density of the signal within each time window; Step 4: Construct the relative power spectrum matrix based on the power spectral density of the short preamble sequence at each frequency and time window, and the power spectral density of the long preamble sequence at each frequency and time window. ; Then consider the relative power spectrum matrix Extreme value clipping is performed to obtain a two-dimensional relative power spectrum time-frequency diagram; Step 5: Construct a multi-channel complex ResNet-Transformer network, which includes a one-dimensional complex ResNet branch and a two-dimensional complex ResNet branch; wherein: The input to the one-dimensional complex ResNet branch is the result of splicing the FFT transform results of the time-synchronized Wi-Fi RF signal and the time-synchronized Wi-Fi RF signal along the channel dimension. The input to the two-dimensional complex ResNet branch is a two-dimensional relative power spectrum time-frequency plot; The identification results of the devices to be accessed are output through a multi-channel complex ResNet-Transformer network.

2. The Wi-Fi radio frequency fingerprint recognition method according to claim 1, characterized in that, In step three, a short-time Fourier transform is performed on the short preamble sequence to obtain the power spectral density of the signal within each time window, specifically: In frequency Down: A short-time Fourier transform is performed on the short preamble sequence using a sliding window of length W and overlap S to obtain the spectrum of the signal within each time window corresponding to the short preamble sequence. Then, the power spectral density of the signal within each time window corresponding to the short preamble sequence is obtained from the spectrum. The first short code sequence corresponding to the next short code sequence The power spectral density of the signal within each time window is denoted as . .

3. The Wi-Fi radio frequency fingerprint recognition method according to claim 2, characterized in that, In step three, a short-time Fourier transform is performed on the long preamble sequence to obtain the power spectral density of the signal within each time window, specifically: In frequency Down: A short-time Fourier transform is performed on the long preamble sequence using a sliding window of length W and overlap S to obtain the spectrum of the signal within each time window corresponding to the long preamble sequence. Then, the power spectral density of the signal within each time window corresponding to the long preamble sequence is obtained from the spectrum. The first corresponding to the lower long guide sequence The power spectral density of the signal within each time window is denoted as . .

4. The Wi-Fi radio frequency fingerprint recognition method according to claim 3, characterized in that, The relative power spectrum matrix The construction method is as follows: Divide the power spectral density of the signal within the corresponding time window of the short precode sequence by the power spectral density of the long precode sequence: in, Represents the relative power spectrum matrix The Middle Line number Column elements, It represents any small positive number.

5. The Wi-Fi radio frequency fingerprint recognition method according to claim 4, characterized in that, The relative power spectrum matrix Extreme value clipping is performed to obtain the two-dimensional relative power spectrum time-frequency diagram, specifically: For the relative power spectrum matrix Sort the elements in any row in ascending order, keeping the elements that are not in the first 5% and the elements that are not in the last 5%, and replace the elements in the first 5% with the minimum value to be retained, and replace the elements in the last 5% with the maximum value to be retained. Similarly, for the relative power spectrum matrix Each row is processed separately, and the resulting relative power spectrum matrix is ​​the two-dimensional relative power spectrum time-frequency diagram.

6. The Wi-Fi radio frequency fingerprint recognition method according to claim 5, characterized in that, The working process of the multi-channel complex ResNet-Transformer network is as follows: The FFT transform results of the time-synchronized Wi-Fi RF signal and the time-synchronized Wi-Fi RF signal are concatenated along the channel dimension as the input of the first complex convolutional layer, and the output of the first complex convolutional layer is used as the input of the first complex normalization layer. The output of the first complex normalization layer is used as the input of the first complex activation function layer, and then the output of the first complex activation function layer is used as the input of the first complex max pooling layer. The output of the first complex max pooling layer is used as the input of the first residual block, and the output of the first residual block is used as the input of the second residual block. The output of the second residual block is used as the input of the third residual block, and then the output of the third residual block is used as the input of the fourth residual block. The output of the fourth residual block is used as the input of the second complex max pooling layer, and the output of the second complex max pooling layer is used as the input of the first complex fully connected layer. The two-dimensional relative power spectrum time-frequency plot is used as the input of the second complex convolutional layer, and the output of the second complex convolutional layer is used as the input of the second complex normalization layer. The output of the second complex normalization layer is used as the input of the second complex activation function layer, and then the output of the second complex activation function layer is used as the input of the third complex max pooling layer. The output of the third complex max pooling layer is used as the input of the fifth residual block, and the output of the fifth residual block is used as the input of the sixth residual block. Use the output of the sixth residual block as the input of the seventh residual block, and then use the output of the seventh residual block as the input of the eighth residual block. The output of the eighth residual block is used as the input of the fourth complex max pooling layer, and the output of the fourth complex max pooling layer is used as the input of the second complex fully connected layer. The outputs of the first complex fully connected layer and the second complex fully connected layer are concatenated along the channel dimension to obtain the concatenated result a; Position encoding is performed on the concatenated result a. The position encoding result is used as the input of the Transformer encoder. The output of the Transformer encoder is then concatenated with a along the channel dimension to obtain the concatenated result b. The splicing result b is used as the input to the classification layer, and the classification layer outputs the identification result of the device to be accessed.

7. The Wi-Fi radio frequency fingerprint recognition method according to claim 6, characterized in that, The working process of the first residual block is as follows: Within the first residual block, the output of the first complex max pooling layer is used as the input of the third and fourth complex convolutional layers, and the output of the third complex convolutional layer is used as the input of the third complex normalization layer. The output of the third complex normalization layer is used as the input of the third complex activation function layer, and then the output of the third complex activation function layer is used as the input of the fifth complex convolutional layer. The output of the fifth complex convolutional layer is used as the input of the fourth complex normalization layer. Then, the outputs of the fourth complex convolutional layer and the fourth complex normalization layer are concatenated along the channel dimension to obtain the concatenated result c. The concatenated result c is used as the input of the sixth complex convolutional layer, the output of the sixth complex convolutional layer is used as the input of the fifth complex normalization layer, and the output of the fifth complex normalization layer is used as the input of the fourth complex activation function layer. The output of the fourth complex activation function layer is used as the input of the seventh complex convolutional layer, and the output of the seventh complex convolutional layer is used as the input of the sixth complex normalization layer. c is concatenated with the output of the sixth complex normalization layer, and the concatenation result is used as the output of the first residual block.

8. A computer storage medium, characterized in that, The storage medium stores at least one instruction, which is loaded and executed by a processor to implement a Wi-Fi radio frequency fingerprint recognition method according to any one of claims 1 to 7.

9. A Wi-Fi radio frequency fingerprint recognition device, characterized in that, The device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement a Wi-Fi radio frequency fingerprint recognition method according to any one of claims 1 to 7.

Images