Radio frequency fingerprinting method and system based on feature latent space diffusion model

Abstract

The present application belongs to the technical field of radio frequency fingerprint identification, and aims at the problem of decline of recognition rate of existing methods in low signal-to-noise ratio environment, and proposes a radio frequency fingerprint identification method and system based on feature latent space diffusion model. The method comprises the following steps: after receiving the radio frequency signal, firstly, the system is classified and pre-trained, and the parameters of the feature extraction module and the classification module are updated; then, latent space diffusion pre-training is carried out, the noise time step prediction module and the latent space diffusion noise prediction module are trained based on the feature extraction and classification module, and the parameters are optimized in combination with feature reconstruction, time step prediction and classification loss; finally, systematic fine-tuning is carried out, and all modules are optimized as a whole. Repeat the training until the neural network converges, and output the optimal network parameters for testing. Through the combination of feature extraction and latent space noise suppression, the robustness and accuracy of radio frequency equipment identification under low signal-to-noise ratio conditions are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of radio frequency fingerprint recognition technology, and particularly relates to a radio frequency fingerprint recognition method and system based on a feature latent space diffusion model. Technical Background

[0002] With the continued proliferation of the Internet of Things (IoT), related devices have permeated almost every aspect of daily life. This has been accompanied by increased security risks, posing new challenges to device authentication technologies. Radio Frequency Fingerprint Identification (RFFI), as a physical layer authentication approach, utilizes subtle characteristics in radio frequency signals created by differences in device hardware manufacturing to distinguish identities. This method does not introduce significant computational or storage burdens on the terminal side, and hardware defects are difficult to accurately counterfeit, thus possessing inherent anti-counterfeiting properties.

[0003] However, existing RFFI schemes often experience a decline in recognition rate under low signal-to-noise ratio (SNR) conditions. To address this, research typically follows two paths: first, introducing a dedicated denoising module to denoise the received signal before feeding it into the classification model, thus mitigating noise interference; second, enhancing feature learning capabilities by using additional feature extraction units to capture more complex internal correlations in noisy signals, thereby improving the model's noise resistance and classification performance. However, the gains from these two approaches in low SNR scenarios remain limited, and RFFI still requires more efficient denoising mechanisms or more discriminative feature modeling methods to mitigate performance degradation. Summary of the Invention

[0004] To address the aforementioned issues, this paper proposes a radio frequency fingerprinting method and system based on a feature latent space diffusion model. This method combines two modules: feature extraction and noise reduction. First, a feature extraction network is used to extract subtle representations relevant to classification from the radio frequency signal. Then, the obtained features are mapped to the latent space, and a diffusion-based noise prediction module is used to model and suppress noise, achieving targeted denoising of discriminative representations. This design reduces noise interference while focusing on useful features, thus significantly improving the robustness and accuracy of radio frequency fingerprinting under low signal-to-noise ratio conditions.

[0005] Firstly, a radio frequency fingerprinting method based on a feature latent spatial diffusion model is provided, the method comprising the following steps:

[0006] Receive radio frequency signals;

[0007] Training the received radio frequency signal includes:

[0008] The system undergoes classification pre-training: the parameters of the feature extraction module and the classification module are updated based on the received radio frequency signals and the transmitter serial numbers;

[0009] The system is pre-trained for latent space diffusion: the latent space diffusion module and the noise time step prediction module are pre-trained based on the received signal and the feature extraction module and classification module obtained from the classification pre-training, and the parameters of the latent space diffusion module and the noise time step prediction module are updated based on the feature reconstruction loss, time step prediction loss and classification loss.

[0010] System fine-tuning: Perform systematic fine-tuning of all modules based on the received signals;

[0011] The system is repeatedly trained based on the received radio frequency signals until the neural network converges.

[0012] The optimal network parameters are output for testing.

[0013] Preferably, the classification pre-training of the system specifically includes:

[0014] The received signal is normalized, and the maximum and minimum values ​​of the in-phase and quadrature component sequences are calculated using the following formula:

[0015]

[0016] Where j∈{i,q}.

[0017] The received in-phase and quadrature component sequences are normalized:

[0018]

[0019] The obtained normalized in-phase component sequence and quadrature component sequence are spliced ​​together to obtain the normalized radio frequency signal tensor.

[0020] x N [n] = [x] i,N [n] x q,N [n]], n=0,1,...,m-1,

[0021] For the normalized radio frequency signal tensor x N The input is projected through a linear layer to obtain the specified projection result.

[0022] x1 = Linear 2→64 (x N ),

[0023] Perform three convolution operations on x1 to obtain local features, with kernel dimensions of 3×3, 5×5, and 7×7, respectively, to obtain the output tensor after convolution.

[0024] x2=Conv 7×7 (Conv 5×5 (Conv 3×3 (x1))),

[0025] The convolution result x2 tensor is used to perform global dependency modeling using a multi-head self-attention module to obtain a global feature tensor. Then, global pooling is used in the time dimension to obtain the latent space feature vector aggregated in the time dimension.

[0026] r = AvgPool(MHA(x2,x2,x2)),

[0027] For ease of explanation, the above feature extraction module will be abbreviated as ε:

[0028]

[0029] For the obtained latent space eigenvectors The probability vector is calculated using a multilayer perceptron with three linear layers and the GELU activation function. The resulting simplex vector is:

[0030] p = Linear 128→k (GELU(Linear 256→128 (GELU(Linear 512→256 (r))))),

[0031] Where each component of p

[0032] The module for generating the above feature vectors can be abbreviated as:

[0033]

[0034] We choose cross-entropy as the loss function, and the formula for calculating cross-entropy is:

[0035]

[0036] in, This is a one-bit hot code for the serial number of the radio frequency signal transmitter. If the serial number of the transmitter transmitting the current radio frequency signal is i, then q i =1 and q j =0, j≠i.

[0037] Backpropagation is performed on the L1 loss function, and the parameters are updated using the AdamW optimizer. Once the model converges, the optimal network parameters ε for the feature extractor and classifier are saved. * and

[0038] Preferably, the pre-training of the system for latent space diffusion specifically includes:

[0039] The optimal parameters ε for classification pre-training of the system * and Load and freeze model ε * and Parameters. During the freezing process, the model's ε... * and It will not be optimized by the optimizer.

[0040] For the normalized radio frequency signal sequence x N Generate features using the loaded model parameters.

[0041] r = ε * (x N ),

[0042] Uniformly distributed random sampling is performed on {0,1,...,T-1} to generate random noise time steps t. Noisy feature vectors are obtained based on time step t.

[0043]

[0044] Where α t For cumulative noise scheduling parameters, This is a sampling of a standard normal distribution.

[0045] For normalized radio frequency signals Perform a linear mapping to obtain an m×512 dimensional projected input sequence.

[0046] x p =Linear 2→512 (x N )

[0047] The noisy feature vector r t The query vector Q of the attention module is used as the key vector K and value vector V of the attention module, thereby achieving cross-attention fusion and obtaining the fused value.

[0048] h = MHA(r) t ,x p ,x p )

[0049] MHA(·) is the multi-head attention module.

[0050] After mapping the obtained fusion value h using a single linear layer, a nonlinear mapping is performed using the GELU activation function. Subsequently, a linear layer and the Sigmoid function are used to map the output to the (0,1) interval, and the step size is multiplied to obtain the step size estimate.

[0051]

[0052] The above noise time step prediction process can be summarized as follows:

[0053]

[0054] The noise time step t of the actual sampling and the obtained noisy feature r t Noise prediction is performed using a diffusion-backward process, which includes two downsampling operations, an intermediate layer, and two upsampling operations, with jumpers connecting the upsampling and downsampling operations. The noise time step is then passed through a linear layer and a GELU activation function and superimposed on the noisy features to obtain the noise time step embedding of the features.

[0055] r t,e =r t +Linear 512→512 (GELU(Linear 1→512 (t)))

[0056] Noise time step embedding of features r t,e Feature latent space diffusion is performed, comprising three steps: downsampling, intermediate self-attention information fusion, and upsampling. Downsampling and upsampling are implemented using residual blocks, with the input and output dimensions of the residual blocks being identical, and jump lines connecting different residual blocks. A total of two downsampling operations and two upsampling operations are used, with a multi-head self-attention module used in between. The feature's noisy time-step embedding r... t,e The outputs of downsampling block 1, downsampling block 2, multi-head auto-attention block, upsampling block 1, and upsampling block 2 are h1, h2, and h, respectively. b u1 and u2 are represented as follows:

[0057] h1 = ResidualBlock 512 (r t,e ),

[0058] h2 = ResidualBlock 512 (h1),

[0059] h b =h2+Linear512→512 (Mid(h2)),

[0060] u1 = Linear 1024→512 (ResidualBlock 1024 ([h b h2])),

[0061] u2 = Linear 1024→512 (ResidualBlock 1024 ([u1 h1]))

[0062] The intermediate layer Mid consists of a linear layer, a normalization layer, a multi-head attention layer, and a mean layer, connected by jumpers. A linear mapping is applied to h2, and its shape is altered to obtain four word segments from one feature.

[0063] z = reshape(Linear) 512→2048 (h2))

[0064] Among them, the word segmentation tensor after shape change The segmentation tensor z is subjected to layer normalization and multi-head self-attention feature fusion. The input and output are connected by a jumper to obtain the output z':

[0065] z N =LayerNorm(z),

[0066] z'=MHA(z N ,z N ,z N )+z,

[0067] Layer normalization and mean averaging are performed on z', and the input and output are connected via jumpers to obtain the output.

[0068] h s =Mean(LayerNorm(z'))+z'

[0069] For h s Perform a linear mapping and connect using jumpers to obtain the output h of the intermediate layer. b :

[0070] h b =h s +h b2

[0071] By linearly mapping the output u2 of upsampling block 2, the noise prediction vector ∈ is obtained. θ :

[0072] ∈ θ =Linear(u2)

[0073] The above noise prediction process can be simplified as follows:

[0074]

[0075] For noisy features r t Denoising is performed to obtain the reconstructed features.

[0076]

[0077] Features of reconstruction Using a classifier Classification is performed to obtain probability simplex vectors.

[0078]

[0079] For the estimated step size value Reconstructed features The latent space diffusion loss is calculated using the estimated probability vector p, including: the squared error loss L for step size estimation. t Feature reconstruction loss L r And classification cross-entropy loss L c The calculation method is as follows:

[0080]

[0081] The calculated losses are weighted and summed to obtain the total latent space diffusion loss L2:

[0082] L2=αL t +βL r +γL c

[0083] Where α, β and γ are loss weight hyperparameters.

[0084] Backpropagation is performed on the L2 loss to obtain the gradient, and then the AdamW optimizer is used to perform a noisy time step predictor. and latent space diffusion noise predictor Optimize the network parameters. After the model converges, save the optimal network parameters for the feature extractor and classifier. and

[0085] Preferably, the fine-tuning of the system specifically includes:

[0086] The optimal parameters ε for classification pre-training of the system * , and Load the system to obtain all module parameters.

[0087] For in-phase and quadrature component sequences x i [n], x q [n] Calculate the average power P of the signal. s :

[0088]

[0089] For interval [SNR] min SNR max Uniformly distributed random sampling is used to obtain the random signal-to-noise power ratio (SNR). dB Convert it into a linear ratio SNR lin :

[0090]

[0091] For linear ratio SNR lin Calculate noise power P n :

[0092]

[0093] For noise power P n Generate complex Gaussian white noise w[n]:

[0094] w[n]=w i [n]+jw q [n],

[0095] in,

[0096] For the in-phase and quadrature component sequences of radio frequency signals x i [n], x q [n] Adds a noise component of specified intensity to obtain a sequence of noisy in-phase and quadrature components x. i,n [n], x q,n [n]:

[0097]

[0098] For noisy component sequence x i,n x q,n Normalization yields normalized vectors x with different components. i,n,N x q,n,N And splice them together to obtain the noisy radio frequency signal tensor.

[0099] For the noisy radio frequency signal tensor x n Noisy features are obtained by performing feature extraction operations.

[0100] r'=ε(xn )

[0101] For the noisy feature r' and the noisy radio frequency signal tensor x n The time step prediction value is obtained using a noise time step predictor.

[0102]

[0103] For noisy feature r' and time step prediction value Using a noise predictor Obtain the predicted cumulative noise

[0104]

[0105] Denoising the noisy features yields the reconstructed features.

[0106]

[0107] Classify the reconstructed features using a classifier Classification is performed to obtain probability simplex vectors.

[0108]

[0109] The cross-entropy loss is calculated on the obtained simplex vectors, and the classification cross-entropy loss is L3:

[0110]

[0111] Backpropagation is performed on the loss L2 to obtain the gradient, and then the AdamW optimizer is used to optimize all modules ε. and Optimize the network parameters. Save the optimal network parameters after the model converges.

[0112] Secondly, a radio frequency fingerprint recognition system based on a feature latent spatial diffusion model is provided, the system comprising the following modules:

[0113] The receiver module is used to receive radio frequency signals;

[0114] The training module, used to train the received radio frequency signal, includes:

[0115] The classification pre-training module is used for training and updating the parameters of the feature extraction module and the device classification module.

[0116] The feature latent space diffusion pre-training module is used to train and update the parameters of the noise time step prediction module and the latent space diffusion noise prediction module.

[0117] The fine-tuning module is used to further train and update the parameters of the feature extraction module, noise time step prediction module, latent space diffusion noise prediction module, and device classification module obtained from the classification pre-training and feature latent space diffusion pre-training modules.

[0118] The update module is used to update the parameters of each training module based on the loss function of that module.

[0119] The iterative training module is used to repeatedly train the received radio frequency signals until the network converges.

[0120] The output module is used to output the optimal network parameters for the testing process.

[0121] Thirdly, a computer-readable storage medium is provided for storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform any of the methods described.

[0122] Fourthly, a computing device is provided, comprising:

[0123] One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described.

[0124] The beneficial effects of this invention are as follows:

[0125] 1. By pre-training the feature extraction module, priority is given to capturing radio frequency characteristics that are highly related to device number identification, enabling the model to focus on discriminative features and suppress interfering (or redundant) features, thereby improving the discriminative power and robustness of feature representation.

[0126] 2. A noise time step predictor based on cross-attention is introduced to jointly model the long-range dependency between the original signal and the extracted features, accurately estimating the noise intensity and temporal position corresponding to the target features, thus providing a more reliable prior for subsequent noise prediction and suppression.

[0127] 3. Based on the predicted noise time step, feature domain denoising is performed to effectively recover key representations related to equipment number identification, reduce noise disturbance to the decision boundary, and thus improve the accuracy and consistency of equipment number identification.

[0128] 4. The one-step prediction strategy using the latent space diffusion model avoids the computational burden and latency caused by traditional multi-step sampling, reduces energy consumption and computing power consumption on the system side, especially on the IoT side, and is more in line with the engineering requirements of low power consumption and low carbonization. Attached Figure Description

[0129] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are merely some examples of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0130] Appendix Figure 1 This is a flowchart illustrating the overall training process of the radio frequency fingerprint recognition algorithm based on the feature latent spatial diffusion model of the present invention.

[0131] Appendix Figure 2 This is a flowchart of the classification pre-training algorithm for the radio frequency fingerprint recognition algorithm based on the feature latent space diffusion model in this invention;

[0132] Appendix Figure 3 This is a flowchart of the latent space diffusion pre-training algorithm for the radio frequency fingerprint recognition algorithm based on the feature latent space diffusion model in this invention.

[0133] Appendix Figure 4 This is a flowchart of the systematic fine-tuning algorithm of the radio frequency fingerprint recognition algorithm based on the feature latent space diffusion model in this invention.

[0134] Appendix Figure 5 This is a network structure diagram of the feature extraction module in this invention.

[0135] Appendix Figure 6 This is a network structure diagram of the noise time step prediction module in this invention.

[0136] Appendix Figure 7 This is a network structure diagram of the latent space diffusion noise predictor module in this invention. Detailed Implementation

[0137] To make the objectives and technical solutions of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be described more clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0138] Since this invention is based on deep neural networks, the description of specific implementation methods will be separated into two aspects: the training process and the testing process.

[0139] like Figure 1The diagram shown is an overall flowchart of the training process of this invention. This invention provides a radio frequency fingerprinting method based on a feature latent space diffusion model, comprising the following steps:

[0140] Receive radio frequency signals;

[0141] Training the received radio frequency signal includes:

[0142] The system undergoes classification pre-training: the parameters of the feature extraction module and the classification module are updated based on the received radio frequency signals and the transmitter serial numbers;

[0143] The system is pre-trained for latent space diffusion: the latent space diffusion module and the noise time step prediction module are pre-trained based on the received signal and the feature extraction module and classification module obtained from the classification pre-training, and the parameters of the latent space diffusion module and the noise time step prediction module are updated based on the feature reconstruction loss, time step prediction loss and classification loss.

[0144] System fine-tuning: Perform systematic fine-tuning of all modules based on the received signals;

[0145] The system is repeatedly trained based on the received radio frequency signals until the neural network converges.

[0146] The optimal network parameters are output for testing.

[0147] Specifically, the system performs classification pre-training; the process can be found in [link to documentation]. Figure 2 This includes the following steps:

[0148] (5.1) Calculate the maximum and minimum values ​​of the in-phase and quadrature component sequences using the following formula:

[0149]

[0150] Where j∈{i,q}.

[0151] (5.2) Normalize the received in-phase and quadrature component sequences:

[0152]

[0153] (5.3) The obtained normalized in-phase component sequence and quadrature component sequence are spliced ​​together to obtain the normalized radio frequency signal tensor.

[0154] x N [n] = [x] i,N [n] x q,N [n]], n=0,1,…,m-1,

[0155] (5.4) For the normalized radio frequency signal tensor xN The input is projected through a linear layer to obtain the specified projection result.

[0156] x1 = Linear 2→64 (x N ),

[0157] (5.5) Perform three convolution operations on x1 to obtain local features. The dimensions of the convolution kernels are 3×3, 5×5, and 7×7, respectively, to obtain the output tensor after convolution.

[0158] x2=Conv 7×7 (Conv 5×5 (Conv 3×3 (x1))),

[0159] (5.6) Global dependency modeling is performed on the convolution result x2 tensor using a multi-head self-attention module to obtain a global feature tensor. Then, global pooling is used in the time dimension to obtain the latent space feature vector aggregated in the time dimension.

[0160] r = AvgPool(MHA(x2,x2,x2)),

[0161] (5.7) For ease of description, the above feature extraction module will be abbreviated as ε:

[0162]

[0163] The network structure of this feature extraction module can be found in [reference]. Figure 5 .like Figure 5 As shown, this module captures local multi-scale features of radio frequency signals through parallel convolutional kernels of different scales (3×3, 5×5, 7×7), then models the long-range dependencies between these features through a self-attention mechanism, and finally aggregates them into a fixed-length feature vector r through a pooling layer.

[0164] (5.8) The obtained latent space eigenvectors The probability vector is calculated using a multilayer perceptron with three linear layers and the GELU activation function. The resulting simplex vector is:

[0165] p = Linear 128→k (GELU(Linear 256→128 (GELU(Linear 512→256 (r))))),

[0166] Where each component of p

[0167] (5.9) The above feature vector generation module can be abbreviated as:

[0168]

[0169] (5.10) Cross-entropy is chosen as the loss function. The formula for calculating cross-entropy is:

[0170]

[0171] in, This is a one-bit hot code for the serial number of the radio frequency signal transmitter. If the serial number of the transmitter transmitting the current radio frequency signal is i, then q i =1 and q j =0, j≠i.

[0172] Backpropagation is performed on the L1 loss function, and the parameters are updated using the AdamW optimizer. Once the model converges, the optimal network parameters ε for the feature extractor and classifier are saved. * and

[0173] Specifically, the system undergoes latent space diffusion pre-training; the process can be found in [link to documentation]. Figure 3 This includes the following steps:

[0174] The optimal parameters ε for classification pre-training of the system * and Load and freeze model ε * and Parameters. During the freezing process, the model's ε... * and It will not be optimized by the optimizer.

[0175] (6.1) For the normalized radio frequency signal sequence x N Generate features using the loaded model parameters.

[0176] r = ε * (x N ),

[0177] (6.2) Perform uniformly distributed random sampling on {0,1,...,T-1} to generate a random noise time step t, and obtain the noisy feature vector based on the time step t.

[0178]

[0179] Where α t This is the cumulative noise scheduling parameter, whose value monotonically decreases as time step t increases, defining the progress of noise addition. This is a sampling of a standard normal distribution.

[0180] (6.3) For normalized radio frequency signals Perform a linear mapping to obtain an m×512 dimensional projected input sequence.

[0181] x p =Linear 2→512 (x N (6.4) The noisy feature vector r t The query vector Q of the attention module is used as the key vector K and value vector V of the attention module, thereby achieving cross-attention fusion and obtaining the fused value.

[0182] h = MHA(r) t ,x p ,x p )

[0183] MHA(·) is the multi-head attention module.

[0184] (6.5) After mapping the obtained fusion value h to a single linear layer, a nonlinear mapping is performed using the GELU activation function. Then, a linear layer and the Sigmoid function are used to map the output to the (0,1) interval, and the step size is multiplied to obtain the step size estimate.

[0185]

[0186] (6.6) The above noise time step prediction process can be summarized as follows:

[0187]

[0188] The network structure of the noise time step prediction module can be found in [reference]. Figure 6 .like Figure 6 As shown, this module uses a cross-attention mechanism to process noisy features r. t As a query, the projection x of the original signal p Using these as keys and values, global contextual information is fused to predict the noise intensity (time step) corresponding to the current feature.

[0189] (6.7) The noise time step t of the actual sampling and the obtained noisy feature r tNoise prediction is performed using a diffusion-backward process, which includes two downsampling operations, an intermediate layer, and two upsampling operations, with jumpers connecting the upsampling and downsampling operations. The noise time step is then passed through a linear layer and a GELU activation function and superimposed on the noisy features to obtain the noise time step embedding of the features.

[0190] r t,e =r t +Linear 512→512 (GELU(Linear 1→512 (t)))(6.8) embedding the feature at the noisy time step r t,e Feature latent space diffusion is performed, comprising three steps: downsampling, intermediate self-attention information fusion, and upsampling. Downsampling and upsampling are implemented using residual blocks, with the input and output dimensions of the residual blocks being identical, and jump lines connecting different residual blocks. A total of two downsampling operations and two upsampling operations are used, with a multi-head self-attention module used in between. The feature's noisy time-step embedding r... t,e The outputs of downsampling block 1, downsampling block 2, multi-head auto-attention block, upsampling block 1, and upsampling block 2 are h1, h2, and h, respectively. b u1 and u2 are represented as follows:

[0191] h1 = ResidualBlock 512 (r t,e ),

[0192] h2 = ResidualBlock 512 (h1),

[0193] h b =h2+Linear 512→512 (Mid(h2)),

[0194] u1 = Linear 1024→512 (ResidualBlock 1024 ([h b h2])),

[0195] u2 = Linear 1024→512 (ResidualBlock 1024 ([u1 h1]))

[0196] The intermediate layer Mid consists of a linear layer, a normalization layer, a multi-head attention layer, and a mean layer, connected by jumpers. (6.9) A linear mapping is performed on h2, and the shape is changed to obtain four words from one feature:

[0197] z = reshape(Linear) 512→2048 (h2))

[0198] Among them, the word segmentation tensor after shape change (6.10) Perform layer normalization and multi-head self-attention feature fusion on the segmentation tensor z, and connect the input and output through a jumper line to obtain the output z':

[0199] z N =LayerNorm(z),

[0200] z'=MHA(z N ,z N ,z N )+z,

[0201] (6.11) Perform layer normalization and mean averaging on z', and connect the input and output via jumpers to obtain the output.

[0202] h s =Mean(LayerNorm(z'))+z'

[0203] (6.12) Regarding h s Perform a linear mapping and connect using jumpers to obtain the output h of the intermediate layer. b :

[0204] h b =h s +h b2

[0205] (6.13) Perform a linear mapping on the output u2 of upsampling block 2 to obtain the noise prediction vector ∈ θ :

[0206] ∈ θ =Linear(u2)

[0207] (6.14) The above noise prediction process can be simplified as follows:

[0208]

[0209] The network structure of the latent space diffusion noise predictor module can be found in [reference]. Figure 7 .like Figure 7 As shown, this module employs a U-Net-like architecture, comprising downsampling, self-attention intermediate layers, and upsampling. Skip connections ensure the fusion of feature information at different scales. The time step t is incorporated into the network through embedding layers (time step embedding), conditionalizing the noise prediction process. The self-attention mechanism of the intermediate layers further enhances the global modeling capability of the features.

[0210] (6.15) For noisy features r tDenoising is performed to obtain the reconstructed features.

[0211]

[0212] (6.16) Features of reconstruction Using a classifier Classification is performed to obtain probability simplex vectors.

[0213]

[0214] (6.17) For the estimated step size value Reconstructed features The latent space diffusion loss is calculated using the estimated probability vector p, including: the squared error loss L for step size estimation. t Feature reconstruction loss L r And classification cross-entropy loss L c The calculation method is as follows:

[0215]

[0216] Where q is a single valid bit of the transmitter serial number corresponding to the current training sample.

[0217] (6.18) The calculated losses are weighted and summed to obtain the total loss L2 of latent space diffusion:

[0218] L2=αL t +βL r +γL c

[0219] Where α, β and γ are loss weight hyperparameters.

[0220] Backpropagation is performed on the L2 loss to obtain the gradient, and then the AdamW optimizer is used to perform a noisy time step predictor. and latent space diffusion noise predictor Optimize the network parameters. After the model converges, save the optimal network parameters for the feature extractor and classifier. and

[0221] Specifically, systematic fine-tuning is performed; the process can be found in [link to relevant documentation]. Figure 4 This includes the following steps:

[0222] The optimal parameters ε for classification pre-training of the system * , and Load the system to obtain all module parameters.

[0223] (7.1) For in-phase and quadrature component sequences x i [n], x q [n] Calculate the average power P of the signal. s :

[0224]

[0225] (7.2) For the interval [SNR min SNR max (e.g., [-10dB, 20dB]) Uniformly distributed random sampling is used to obtain the random signal-to-noise power ratio (SNR). dB Convert it into a linear ratio SNR lin :

[0226]

[0227] (7.3) For the linear ratio SNR lin Calculate noise power P n :

[0228]

[0229] (7.4) Regarding the noise power P n Generate complex Gaussian white noise w[n]:

[0230] w[n]=w i [n]+jw q [n],

[0231] in,

[0232] (7.5) For the in-phase and quadrature component sequences x of the radio frequency signal i [n], x q [n] Adds a noise component of specified intensity to obtain a sequence of noisy in-phase and quadrature components x. i,n [n], x q,n [n]:

[0233]

[0234] For noisy component sequence x i,n x q,n Normalization yields normalized vectors x with different components. i,n,N x q,n,N And splice them together to obtain the noisy radio frequency signal tensor.

[0235] (7.6) For the noisy radio frequency signal tensor x n Noisy features are obtained by performing feature extraction operations.

[0236] r'=ε(x n (7.7) For the noisy feature r' and the noisy radio frequency signal tensor x n The time step prediction value is obtained using a noise time step predictor.

[0237]

[0238] (7.8) Predicted values ​​of noisy feature r' and time step Using a noise predictor Obtain the predicted cumulative noise

[0239]

[0240] (7.9) Perform denoising operation on the noisy features to obtain the reconstructed features.

[0241]

[0242] (7.10) Classify the reconstructed features using a classifier Classification is performed to obtain probability simplex vectors.

[0243]

[0244] (7.11) Calculate the cross-entropy loss for the obtained simplex vectors. The classification cross-entropy loss is L3:

[0245]

[0246] Backpropagation is performed on the loss L3 to obtain the gradient, and then the AdamW optimizer is used to optimize all modules ε. and Optimize the network parameters. Save the optimal network parameters after the model converges.

[0247] The optimal network parameter set {ε} is obtained through the above training process. * D * ,T * U * After that, radio frequency fingerprint recognition can be performed. The testing process follows these steps in sequence:

[0248] Signal preprocessing: Receive the radio frequency signal to be identified, calculate the maximum and minimum values ​​of its in-phase and quadrature components in the same way as in the training phase, perform normalization processing, and concatenate the two normalized component sequences into a two-dimensional tensor as the input of the network.

[0249] Feature extraction: The preprocessed signal tensor is input into the trained feature extraction module to calculate the feature vector corresponding to the signal.

[0250] Noise level estimation: The feature vector obtained in the previous step is input together with the original input signal tensor into the trained noise time step prediction module. This module will analyze the information from both and output an estimated noise intensity value.

[0251] Feature denoising: The extracted feature vector and the estimated noise intensity value are input into the pre-trained latent space diffusion noise prediction module. This module predicts the noise component contained in the feature based on the noise intensity, and then subtracts the predicted noise from the original feature to obtain a denoised, cleaner feature vector.

[0252] Classification decision: The denoised feature vector is input into the pre-trained classifier module, which outputs a probability distribution vector, where the probability value at each position represents the likelihood that the input signal belongs to the corresponding transmitter device.

[0253] Output result: Select the device category corresponding to the largest probability value in the probability distribution vector as the final RFID fingerprint recognition result, that is, determine that the device sent the current signal.

[0254] This invention provides a radio frequency fingerprint recognition system based on a feature latent space diffusion model, the system comprising the following modules:

[0255] The receiver module is used to receive radio frequency signals;

[0256] The training module, used to train the received radio frequency signal, includes:

[0257] The classification pre-training module is used for training and updating the parameters of the feature extraction module and the device classification module.

[0258] The feature latent space diffusion pre-training module is used to train and update the parameters of the noise time step prediction module and the latent space diffusion noise prediction module.

[0259] The fine-tuning module is used to further train and update the parameters of the feature extraction module, noise time step prediction module, latent space diffusion noise prediction module, and device classification module obtained from the classification pre-training and feature latent space diffusion pre-training modules.

[0260] The update module is used to update the parameters of each training module based on the loss function of that module.

[0261] The iterative training module is used to repeatedly train the received radio frequency signals until the network converges.

[0262] The output module is used to output the optimal network parameters for the testing process.

[0263] Embodiments of this application may be provided as methods or computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may 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 may be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0264] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and 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.

[0265] 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.

[0266] These computer program instructions may 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.

[0267] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A radio frequency fingerprint recognition method based on a feature latent space diffusion model, characterized in that, Includes the following steps: Receive radio frequency signals; The received radio frequency signals are trained in stages, including: Classification pre-training phase: Using the received radio frequency signal and the corresponding transmitter serial number, the feature extraction module and classification module are trained and their parameters are updated to obtain the trained feature extraction module and classification module, and the optimal parameters of the feature extraction module and classification module are saved respectively. and ; Latent space diffusion pre-training stage: Based on the trained feature extraction module and classification module, the noise time step prediction module and latent space diffusion noise prediction module are trained and their parameters are updated to obtain the trained noise time step prediction module and latent space diffusion noise prediction module. The training process includes: extracting feature vectors of the radio frequency signal using the feature extraction module; adding noise of different intensities to the feature vectors; estimating the intensity information of the added noise using the noise time step prediction module; predicting the added noise using the latent space diffusion noise prediction module; and updating the parameters of the noise time step prediction module and latent space diffusion noise prediction module based on feature reconstruction loss, noise intensity information prediction loss, and classification loss. The specific steps are: loading and freezing the optimal parameters. and ;use Extracting normalized signals Features Random sampling noise time step And according to noise scheduling parameters Generate noisy feature vectors The noisy features are fused using a cross-attention mechanism. Estimating the noise time step using the original signal information. Embed time step information into the noisy feature. In the process, the embedding features of the noise time step are obtained. ;Through a diffusion network that includes downsampling, self-attention intermediate layers, and upsampling, for Processing is performed to predict the added noise. The noisy features are denoised to obtain the reconstructed features. and using a classifier Classify them; comprehensively calculate the step size to estimate the loss. Feature reconstruction loss and classification cross-entropy loss This constitutes the weighted total loss. The parameters of the noise time step prediction module and the latent space diffusion noise prediction module are updated through backpropagation and optimizer until convergence, and their optimal parameters are saved. and ; System fine-tuning stage: Based on the received signal, the feature extraction module, the classification module, the noise time step prediction module, and the latent space diffusion noise prediction module are fine-tuned and their parameters are updated. The received radio frequency signals are repeatedly trained in stages until the neural network converges, and the optimal network parameters are output for radio frequency fingerprint recognition.

2. The radio frequency fingerprinting method based on the feature latent space diffusion model according to claim 1, characterized in that, The classification pre-training phase specifically includes: The received radio frequency signal is preprocessed to obtain a normalized radio frequency signal tensor. ,in The signal sampling length is determined by the feature extraction module. The process involves sequentially performing linear projection, multi-scale convolution, multi-head self-attention global modeling, and global average pooling to obtain the latent space feature vector. The feature vector is processed by the classification module. Classification is performed to obtain the classification probability vector. ,in The total number of transmitter categories; based on the classification probability vector. Calculate the cross-entropy loss using the true label of the transmitter serial number. The parameters of the feature extraction module and classification module are updated through backpropagation and optimizer until the model converges, and the optimal parameters are saved. and .

3. The method according to claim 1, characterized in that, The systematic fine-tuning phase specifically includes: Optimal parameters for loading all modules , , and ; Based on a preset signal-to-noise ratio range, analog noise is added to the original radio frequency signal to construct a noisy radio frequency signal. The feature extraction module is used to extract noisy features. ; The noise time step prediction module is used to predict the noise time step corresponding to the noisy feature. ; Using the latent space diffusion noise prediction module, based on and Predicted noise ; Denoising the noisy features yields the reconstructed features. And classify using the classifier; Calculate the classification cross-entropy loss ; The parameters of all modules are jointly optimized through backpropagation and an optimizer until convergence, and the final network parameters are saved.

4. A radio frequency fingerprint recognition system based on a feature latent space diffusion model, characterized in that, include: The system for implementing the method of any one of claims 1 to 3 comprises: The receiver module is used to receive radio frequency signals; A training module, connected to the receiving module, is used to perform phased training on the received radio frequency signal. The training module includes: A classification pre-training unit is used to perform the classification pre-training phase; A latent space diffusion pre-training unit, connected to the classification pre-training unit, is used to perform the latent space diffusion pre-training phase; The fine-tuning unit, connected to the latent space diffusion pre-training unit, is used to perform the systematic fine-tuning phase; An iterative control module, connected to the training module, is used to control the repeated execution of the training process until the neural network converges; The output module, connected to the iterative control module, is used to output the optimal network parameters.

5. A computer-readable storage medium for storing one or more programs, characterized in that, The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform the method according to any one of claims 1 to 3.

6. A computing device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the one or more programs including instructions for performing the method according to any one of claims 1 to 3.

Images