A channel fingerprint generation and recovery method based on a conditional latent diffusion model

Abstract

The application discloses a channel fingerprint generation and recovery method based on a conditional latent diffusion model, comprising the following steps: converting the channel fingerprint into an initial image through a preset mapping rule, performing gray scale verification on the initial image, performing multi-scale directional filtering processing on the gray scale image, taking the filtered image as the condition input of the conditional latent diffusion model and initializing the latent variable, expanding the pixel value of the filtered image into a feature vector, performing a reverse sampling process, starting from the last step latent variable of the forward diffusion process, converting the final latent variable obtained through the reverse sampling process into data consistent with the dimension of the initial channel fingerprint, and obtaining the generated and recovered channel fingerprint. Thus, the preprocessing quality can be optimized to reduce feature loss, the feature capture integrity can be improved to strengthen the fingerprint distinguishability, the generated fingerprint similarity can be enhanced to guarantee the feature authenticity, the complex environment robustness can be improved to guarantee the recovery stability, and the dimension consistency can be guaranteed to improve the engineering practicability, and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, in particular to a channel fingerprint generation and recovery method based on a conditional latent diffusion model. BACKGROUND

[0002] In the field of wireless communication technology, channel fingerprints, as inherent characteristics of wireless channels with device or location uniqueness, are widely used in device identity authentication, indoor positioning, communication security encryption and other key scenarios. The core value lies in capturing the unique features formed by the environment (such as terrain, obstruction, multipath effect) and device hardware differences in the propagation process of wireless signals, to achieve accurate identification and state awareness of communication nodes. Therefore, efficient generation and reliable recovery of channel fingerprints are key technical links to ensure the performance of related applications.

[0003] Currently, the industry has formed various technical solutions for channel fingerprint generation and recovery. Traditional methods mainly focus on statistical feature extraction and classic machine learning models, such as calculating statistical quantities such as mean, variance, Doppler shift features of channel impulse response (CIR) and channel frequency response (CFR), and combining support vector machine (SVM) and shallow neural network models to realize fingerprint generation and recovery. Such methods can meet the basic accuracy requirements in relatively stable channel environments (such as low noise, few obstructions, weak multipath effect), and have the characteristics of low implementation complexity and small computational overhead, and have been applied in some low-speed and low-reliability communication systems.

[0004] However, with the development of wireless communication towards high speed, high reliability and complex dynamic environment (such as dense urban areas, industrial Internet of Things, vehicle networking), the existing technology needs further optimization in the following aspects in practical application. First, channel fingerprints are essentially high-dimensional and nonlinear data, and traditional statistical feature extraction methods cannot fully capture their deep structure information, resulting in insufficient similarity between the generated fingerprints and the real fingerprints, which further affects the recovery accuracy. Then, in complex dynamic channel environments (such as sudden noise interference, rapid changes in obstructions, and dramatic fluctuations in multipath components), the robustness of traditional models to environmental changes is weak, and the recovered channel fingerprints are easily deviated from the real features, making it difficult to meet the needs of high-reliability applications. Finally, some methods lack targeted preprocessing means when converting channel fingerprints, which are non-image data, into structured data (such as images) that can be used for feature processing, which easily leads to data feature loss or redundancy, reducing the efficiency and effectiveness of the subsequent generation and recovery process. SUMMARY

[0005] The present application aims to at least partially solve the above technical problems in the art.

[0006] Therefore, this invention discloses a method for generating and recovering channel fingerprints based on a conditional latent diffusion model, comprising the following steps:

[0007] S1: Obtain the channel fingerprint and convert the channel fingerprint into an initial image using a preset mapping rule. Each pixel in the initial image has a corresponding grayscale value.

[0008] S2: Perform grayscale verification on the initial image. If the initial image is multi-channel data, convert it into a single-channel grayscale image and record it as a grayscale image.

[0009] S3: Perform multi-scale directional filtering on the grayscale image, set at least 3 fixed-scale templates, each of the scale templates contains at least 4 fixed directions, perform convolution operation on each scale template with the grayscale image, process the image edge pixels, and obtain the filtered channel fingerprint image based on all convolution results, which is denoted as the filtered image.

[0010] S4: Use the filtered image as the conditional input of the conditional latent diffusion model, initialize the latent variables, the dimension of the latent variables is consistent with the total number of pixels in the filtered image, and the initial values ​​of the latent variables follow a standard normal distribution within 0 to 1;

[0011] S5: Perform the forward diffusion process, setting a fixed number of diffusion steps, with the latent variable z at step t. t Through formula Calculate, where,

[0012] β t Let ε be the noise coefficient at step t, and let ε be the noise vector following a standard normal distribution. The dimension of ε is related to the latent variable z. t Consistent, the value of t ranges from 1 to the fixed number of diffusion steps;

[0013] S6: Construct linear conditional fusion features by expanding the pixel values ​​of the filtered image into a feature vector f, and then using the linear fusion formula f... merge =α·f+(1-α)·z t Obtain the fusion feature f merge Where α is the weighting coefficient;

[0014] S7: Perform the reverse sampling process, starting from the last latent variable of the forward diffusion process, and perform the sampling of the fixed diffusion steps in reverse. Based on the predictor, the noise is predicted, and the latent variable of each reverse sampling step is calculated.

[0015] S8: Convert the final latent variables obtained from the reverse sampling process into data with the same dimension as the initial channel fingerprint to obtain the generated and recovered channel fingerprint.

[0016] The channel fingerprint generation and recovery method based on the conditional latent diffusion model disclosed in this invention has at least the following beneficial effects:

[0017] (1) It can optimize the preprocessing quality and reduce feature loss. This method converts the channel fingerprint into an initial image with gray values ​​by using a preset mapping rule, and combines grayscale verification to unify multi-channel data into a single-channel grayscale image, thus standardizing the conversion process of non-image type channel fingerprints into structured image data.

[0018] (2) It can improve the completeness of feature capture and enhance fingerprint discrimination. This method performs multi-scale directional filtering on grayscale images, sets at least 3 fixed scale templates and at least 4 fixed directions, and combines convolution operation and edge pixel processing to fully capture the deep structural information and detailed features of channel fingerprints at different spatial scales and in different directions.

[0019] (3) It can enhance the similarity of generated fingerprints and ensure the authenticity of features. This method uses the filtered image as the conditional input of the conditional latent diffusion model and fuses the feature vector of the filtered image with the latent variable through the linear conditional fusion formula, so that the diffusion process is always constrained by the effective features of the channel fingerprint.

[0020] (4) It can improve the robustness of complex environments and ensure the stability of recovery. In the forward diffusion process, by setting a fixed number of diffusion steps and a clear latent variable update formula, this method ensures the controllability and stability of latent variable changes. In the reverse sampling process, based on the accurate prediction of noise by the predictor, it can effectively suppress the impact of interference factors such as sudden noise and multipath fluctuations on the recovery results in complex channel environments.

[0021] (5) It can ensure dimensional consistency and improve engineering practicality. This method ensures that the final generated and recovered channel fingerprint is completely consistent with the initial channel fingerprint by using a dimensional transformation method that is the inverse of the initial mapping rule.

[0022] In addition, the channel fingerprint generation and recovery method based on the conditional latent diffusion model disclosed in this invention may also have the following additional technical features:

[0023] In one embodiment of the present invention, in step S3, the at least three fixed-scale templates are specifically three scale templates, namely a 3×3 scale template, a 5×5 scale template, and a 7×7 scale template.

[0024] In one embodiment of the present invention, in step S3, the at least four fixed directions included in each scale template are specifically four directions, namely, horizontal direction, vertical direction, 45-degree direction, and 135-degree direction, and the element value in the template corresponding to each direction is set according to the pixel association feature of that direction.

[0025] In one embodiment of the present invention, in step S3, the image edge pixels are processed in a zero-fill mode, and the fill value is set to 0.

[0026] In one embodiment of the present invention, the specific method for obtaining the filtered image based on all convolution results in step S3 is as follows:

[0027] For each pixel in the grayscale image, the convolution results of that pixel at all scales and in all directions are collected, the arithmetic mean of the convolution results is calculated, and the arithmetic mean is used as the grayscale value of the corresponding pixel in the filtered image.

[0028] In one embodiment of the present invention, in step S4, the dimension of the latent variable is consistent with the total number of pixels in the filtered image.

[0029] In one embodiment of the present invention, in step S5, the fixed diffusion step number is set to 200 steps, and the noise figure β t The value sequence is a linear sequence increasing from 0.0001 to 0.02, and the value of t ranges from 1 to 200.

[0030] In one embodiment of the present invention, in step S6, the weighting coefficient α is fixedly set to 0.6.

[0031] In one embodiment of the present invention, in step S7, the predictor is a convolutional neural network predictor, which includes 3 convolutional layers and 2 fully connected layers, and the input of the predictor is the latent variable z of the current step. t The encoding vector of the diffusion step number t and the fusion feature f merge The output of the predictor is the prediction noise ε θ .

[0032] In one embodiment of the present invention, the method of converting the final latent variable into data consistent with the initial channel fingerprint dimension in step S8 is the inverse process of converting the channel fingerprint into the initial image using a preset mapping rule in step S1.

[0033] Additional features and advantages of this invention will be set forth in the description which follows, or may be learned by practicing the invention. Attached Figure Description

[0034] The technical solution and beneficial effects of the present invention will become apparent and readily understood from the following description in conjunction with the accompanying drawings, wherein:

[0035] Figure 1 This is a flowchart illustrating a channel fingerprint generation and recovery method based on a conditional latent diffusion model according to the present invention.

[0036] Figure 2 This is another flowchart of a channel fingerprint generation and recovery method based on a conditional latent diffusion model according to the present invention. Detailed Implementation

[0037] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0038] The following description, with reference to the accompanying drawings, outlines a method for generating and recovering channel fingerprints based on a conditional latent diffusion model.

[0039] like Figure 1 and Figure 2 As shown, a method for generating and recovering channel fingerprints based on a conditional latent diffusion model includes the following steps:

[0040] S1: Obtain the channel fingerprint and convert it into an initial image using a preset mapping rule. Each pixel in the initial image has a corresponding grayscale value.

[0041] S2: Perform grayscale verification on the initial image. If the initial image is multi-channel data, convert it into a single-channel grayscale image, denoted as grayscale image.

[0042] S3: Perform multi-scale directional filtering on the grayscale image, set at least 3 fixed-scale templates, each scale template contains at least 4 fixed directions, perform convolution operation on each scale template with the grayscale image, process the image edge pixels, and obtain the filtered channel fingerprint image based on all convolution results, which is denoted as the filtered image.

[0043] S4: Use the filtered image as the conditional input to the conditional latent diffusion model, initialize the latent variables, the dimension of the latent variables is the same as the total number of pixels in the filtered image, and the initial values ​​of the latent variables follow a standard normal distribution between 0 and 1.

[0044] S5: Perform the forward diffusion process, setting a fixed number of diffusion steps, with the latent variable z at step t. t Through formula Calculate, where,

[0045] β t Let ε be the noise coefficient at step t, and let ε be the noise vector following a standard normal distribution. The dimension of ε is related to the latent variable z. t Consistent, the value of t ranges from 1 to a fixed number of diffusion steps;

[0046] S6: Construct linear conditional fusion features by expanding the pixel values ​​of the filtered image into a feature vector f, and then using the linear fusion formula f... merge =α·f+(1-α)·z t Obtain the fusion feature f merge Where α is the weighting coefficient;

[0047] S7: Perform the reverse sampling process, starting from the last latent variable of the forward diffusion process, and perform sampling in reverse for a fixed number of diffusion steps. Based on the predictor, the noise is predicted, and the latent variable of each reverse sampling step is calculated.

[0048] S8: Convert the final latent variables obtained from the backsampling process into data with the same dimension as the initial channel fingerprint to obtain the generated and recovered channel fingerprint.

[0049] In step S3, the at least three fixed-scale templates are specifically three scale templates: a 3×3 scale template, a 5×5 scale template, and a 7×7 scale template.

[0050] In step S3, each scale template contains at least four fixed directions, specifically four directions: horizontal, vertical, 45-degree, and 135-degree. The element values ​​within the template corresponding to each direction are set according to the pixel association features of that direction.

[0051] In step S3, the image edge pixels are processed using a zero-fill method, with the fill value set to 0.

[0052] In step S3, the specific method for obtaining the filtered image based on all convolution results is as follows:

[0053] For each pixel in the grayscale image, the convolution results of that pixel at all scales and in all directions are collected, the arithmetic mean of the convolution results is calculated, and the arithmetic mean is used as the grayscale value of the corresponding pixel in the filtered image.

[0054] In step S4, the dimension of the latent variable is consistent with the total number of pixels in the filtered image.

[0055] In step S5, the fixed diffusion step number is set to 200 steps, and the noise figure β is... t The value sequence is a linear sequence increasing from 0.0001 to 0.02, and the value of t ranges from 1 to 200.

[0056] In step S6, the weighting coefficient α is fixed at 0.6.

[0057] In step S7, the predictor is a convolutional neural network predictor, which contains 3 convolutional layers and 2 fully connected layers. The input to the predictor is the latent variable z of the current step. tThe encoding vector of the diffusion step t and the fusion feature f merge The output of the predictor is the prediction noise ε θ .

[0058] In step S8, the method of converting the final latent variables into data consistent with the initial channel fingerprint dimension is the inverse process of converting the channel fingerprint into the initial image using the preset mapping rule in step S1.

[0059] Specifically:

[0060] The image obtained by converting the channel fingerprint through a preset mapping rule is set as the initial image I. The image size is fixed at H×W (H is the number of rows and W is the number of columns), and the gray value range of each pixel is [0,255].

[0061] It should be noted that, assuming there are 100 channel fingerprint feature values, they are arranged in 10 rows and 10 columns to convert into the initial image I with a size of 10×10. The gray value of pixel (1,1) is 89, the gray value of pixel (1,2) is 92, ..., the gray value of pixel (10,10) is 78, and so on.

[0062] Ensure that each pixel in the initial image I is a single-channel grayscale value. If multi-channel data exists (such as RGB format), convert it to a single-channel grayscale image using the formula G = 0.299R + 0.587G + 0.114B (where R, G, and B are the grayscale values ​​of the red, green, and blue channels, respectively), and denote it as grayscale image I. gray This ensures that subsequent filtering is based on uniform grayscale data.

[0063] It should be noted that if the initial image I is in RGB format, and the pixel (1,1) has R=85, G=90, and B=92, the calculated value is G=0.299×85+0.587×90+0.114×92≈89, which is consistent with the pixel grayscale value in step S1 and does not require adjustment. If the pixel grayscale value is consistent with the pixel grayscale value in step S1, then adjustment is required.

[0064] Three fixed-scale templates are defined: a 3×3 template, a 5×5 template, and a 7×7 template. Each template contains four fixed directions: horizontal, vertical, 45-degree, and 135-degree. The element values ​​within each template are set according to the corresponding directional characteristics, specifically as follows:

[0065] 3×3 scale horizontal template 3×3 scale vertical template 3×3 scale 45-degree orientation template 3×3 scale 135-degree orientation template

[0066] The 5×5 and 7×7 scale templates are arranged with logically expanded elements in the same direction.

[0067] For each scale, four directional templates are respectively compared with grayscale image I. gray The convolution operation is performed as follows:

[0068] Align the grayscale image I with the center pixel of the template gray For each pixel (image edge pixels are zero-padded with a padding value of 0), the template element is multiplied by the corresponding image pixel value, and the sum is obtained to get the convolution result of that pixel at the current scale and current direction, denoted as C. {s,d}{i,j} (s represents scale, d represents direction, i represents the row coordinate of the image, and j represents the column coordinate of the image).

[0069] It should be noted that, assuming we are taking a grayscale image I... gray The pixel (2,2) (row coordinate 2, column coordinate 2) in the image is convolved using a 3×3 horizontal template with the center of the template aligned to (2,2). The corresponding image pixels are (1,1)=89, (1,2)=92, (1,3)=90, (2,1)=91, (2,2)=93, (2,3)=94, (3,1)=88, (3,2)=90, (3,3)=87.

[0070] For grayscale image I gray For each pixel (i,j) in the dataset, collect its convolution results C in 12 directions across 3 scales (3 scales × 4 directions). {s,d}{i,j} Calculate the arithmetic mean of the 12 results, and use this average as the filtered image I. filtered The grayscale value of the middle pixel (i,j).

[0071] It should be noted that the sum of the template element multiplied by the pixel value is (0×89)+(0×92)+(0×90)+(1×91)+(1×93)+(1×94)+(0×88)+(0×90)+(0×87)=278, which is C. {3×3,水平}{2,2} =278. Similarly, calculate the convolution results of this pixel in other directions of 3×3, in all directions of 5×5, and in all directions of 7×7.

[0072] It should be noted that, assuming the convolution results of pixel (2,2) in the four directions of a 3×3 matrix are 278, 291, 279, 285, the convolution results in the four directions of a 5×5 matrix are 1389, 1380, 1377, 1392, and the convolution results in the four directions of a 7×7 matrix are 3051, 3045, 3039, 3044, the sum of the 12 results is 18850. The arithmetic mean is 18850 ÷ 12 ≈ 1570.83, which is rounded down to I. filteredThe grayscale value of the middle pixel (2,2) is 1571.

[0073] After calculating the mean of all pixels, the filtered channel fingerprint image I is obtained. filtered The image preserves the edge features of the channel fingerprint while filtering out noise interference from different directions.

[0074] It should be noted that, after processing steps S1 to S5, assuming a 10×10 grayscale image I... gray Convert to a 10×10 filtered image I filtered The grayscale value of pixel (1,1) is 1575, the grayscale value of pixel (1,2) is 1572, ..., the grayscale value of pixel (10,10) is 1569. The overall noise interference of the image is reduced and the fingerprint edge is clearer.

[0075] The output filtered image I filtered As a conditional input, the latent variable z0 is initialized simultaneously, and the dimension of the latent variable is the same as that of I. filtered The total number of pixels is consistent (i.e., H×W dimension), and the initial values ​​of the latent variables follow a standard normal distribution N~(0,1).

[0076] It should be noted that, assuming the filtered image I... filtered The value is 10×10, which means the total number of pixels is 100. The latent variable z0 is initialized as a 100-dimensional vector, and each element value is randomly drawn from the standard normal distribution. For example, z0(1)=0.19, z0(2)=0.22,...,z0(100)=0.29.

[0077] First, there's forward diffusion, with a fixed diffusion step count of 200 steps. The latent variable z at step t (t ranges from 1 to 200) is... t Through formula Calculate, where β t Let be the noise coefficient at step t (pre-set to increase linearly from 0.0001 to 0.02), and ε be the noise vector following a standard normal distribution, with dimensions equal to z. t Consistent;

[0078] Next is linear conditional fusion, which transforms the filtered image I... filtered The pixel values ​​are expanded into a 100-dimensional feature vector f, and then fused using the linear fusion formula f. merge =α·f+(1-α)·z t Construct fusion features (α is a weighting coefficient, fixed at 0.6), where z t Let be the latent variable at step t of the forward diffusion;

[0079] Then comes backsampling, from z 200 Initially, perform 200 sampling steps in reverse order. The latent variable z at step t (t ranges from 200 to 1) is...t-1 Through formula Calculate, where ε θ For a simple convolutional neural network (CNN) predictor (containing only 3 convolutional layers and 2 fully connected layers), the input is z. t The encoding vector of the diffusion step t, and the fusion feature f merge The output is the predicted noise;

[0080] Finally, the output results are obtained after backsampling. The final latent variable z'0 is then converted into data with the same dimension as the initial channel fingerprint, which is the generated and recovered channel fingerprint.

[0081] In summary, the channel fingerprint generation and recovery method based on the conditional latent diffusion model disclosed in this invention has at least the following beneficial effects:

[0082] (1) It can optimize the preprocessing quality and reduce feature loss. This method converts the channel fingerprint into an initial image with gray values ​​by using a preset mapping rule, and combines grayscale verification to unify multi-channel data into a single-channel grayscale image, thus standardizing the conversion process of non-image type channel fingerprints into structured image data.

[0083] (2) It can improve the completeness of feature capture and enhance fingerprint discrimination. This method performs multi-scale directional filtering on grayscale images, sets at least 3 fixed scale templates and at least 4 fixed directions, and combines convolution operation and edge pixel processing to fully capture the deep structural information and detailed features of channel fingerprints at different spatial scales and in different directions.

[0084] (3) It can enhance the similarity of generated fingerprints and ensure the authenticity of features. This method uses the filtered image as the conditional input of the conditional latent diffusion model and fuses the feature vector of the filtered image with the latent variable through the linear conditional fusion formula, so that the diffusion process is always constrained by the effective features of the channel fingerprint.

[0085] (4) It can improve the robustness of complex environments and ensure the stability of recovery. In the forward diffusion process, by setting a fixed number of diffusion steps and a clear latent variable update formula, this method ensures the controllability and stability of latent variable changes. In the reverse sampling process, based on the accurate prediction of noise by the predictor, it can effectively suppress the impact of interference factors such as sudden noise and multipath fluctuations on the recovery results in complex channel environments.

[0086] (5) It can ensure dimensional consistency and improve engineering practicality. This method ensures that the final generated and recovered channel fingerprint is completely consistent with the initial channel fingerprint by using a dimensional transformation method that is the inverse of the initial mapping rule.

[0087] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for generating and recovering channel fingerprints based on a conditional latent diffusion model, characterized in that, Includes the following steps: S1: Obtain the channel fingerprint and convert the channel fingerprint into an initial image using a preset mapping rule. Each pixel in the initial image has a corresponding grayscale value. S2: Perform grayscale verification on the initial image. If the initial image is multi-channel data, convert it into a single-channel grayscale image and record it as a grayscale image. S3: Perform multi-scale directional filtering on the grayscale image, set at least 3 fixed-scale templates, each of the scale templates contains at least 4 fixed directions, perform convolution operation on each scale template with the grayscale image, process the image edge pixels, and obtain the filtered channel fingerprint image based on all convolution results, which is denoted as the filtered image. S4: Use the filtered image as the conditional input of the conditional latent diffusion model, initialize the latent variables, the dimension of the latent variables is consistent with the total number of pixels in the filtered image, and the initial values ​​of the latent variables follow a standard normal distribution within 0 to 1; S5: Perform the forward diffusion process, setting a fixed number of diffusion steps, with the latent variable z at step t. t Through formula Calculate, where, β t Let ε be the noise coefficient at step t, and let ε be the noise vector following a standard normal distribution. The dimension of ε is related to the latent variable z. t Consistent, the value of t ranges from 1 to the fixed number of diffusion steps; S6: Construct linear conditional fusion features by expanding the pixel values ​​of the filtered image into a feature vector f, and then using the linear fusion formula f... merge =α·f+(1-α)·z t Obtain the fusion feature f merge Where α is the weighting coefficient; S7: Perform the reverse sampling process, starting from the last latent variable of the forward diffusion process, and perform the sampling of the fixed diffusion steps in reverse. Based on the predictor, the noise is predicted, and the latent variable of each reverse sampling step is calculated. S8: Convert the final latent variables obtained from the reverse sampling process into data with the same dimension as the initial channel fingerprint to obtain the generated and recovered channel fingerprint.

2. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 1, characterized in that, In step S3, the at least three fixed-scale templates are specifically three scale templates: a 3×3 scale template, a 5×5 scale template, and a 7×7 scale template.

3. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 2, characterized in that, In step S3, each scale template includes at least four fixed directions, specifically four directions: horizontal, vertical, 45-degree, and 135-degree. The element values ​​within the template corresponding to each direction are set according to the pixel association features of that direction.

4. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 3, characterized in that, In step S3, the image edge pixels are processed using a zero-fill method, with the fill value set to 0.

5. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 4, characterized in that, In step S3, the specific method for obtaining the filtered image based on all convolution results is as follows: For each pixel in the grayscale image, the convolution results of that pixel at all scales and in all directions are collected, the arithmetic mean of the convolution results is calculated, and the arithmetic mean is used as the grayscale value of the corresponding pixel in the filtered image.

6. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 5, characterized in that, In step S4, the dimension of the latent variable is consistent with the total number of pixels in the filtered image.

7. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 6, characterized in that, In step S5, the fixed diffusion step number is set to 200 steps, and the noise figure β t The value sequence is a linear sequence increasing from 0.0001 to 0.02, and the value of t ranges from 1 to 200.

8. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 7, characterized in that, In step S6, the weighting coefficient α is fixed at 0.

6.

9. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 8, characterized in that, In step S7, the predictor is a convolutional neural network predictor, which includes 3 convolutional layers and 2 fully connected layers. The input of the predictor is the latent variable z of the current step. t The encoding vector of the diffusion step number t and the fusion feature f merge The output of the predictor is the prediction noise ε θ .

10. The channel fingerprint generation and recovery method based on the conditional latent diffusion model as described in claim 9, characterized in that, In step S8, the method of converting the final latent variables into data consistent with the initial channel fingerprint dimension is the inverse process of converting the channel fingerprint into the initial image using a preset mapping rule in step S1.

Images