Wireless communication signal confrontation sample generation method based on diffusion model

Through the wireless communication signal adversarial sample generation method based on the diffusion model, the security risks of deep learning algorithms against attacks in wireless communication are solved, efficient and diverse adversarial sample generation and targeted attacks are achieved, and the electromagnetic signal perception model is improved robustness.

CN119067193AActive Publication Date: 2024-12-03BEIHANG UNIV

Patent Information

Application Number
CN202410520512.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-04-28
Publication Date
2024-12-03
Estimated Expiration
2044-04-28

AI Technical Summary

Technical Problem

Existing deep learning algorithms have security risks of adversarial attacks in wireless communication signal modulation recognition. Traditional adversarial sample generation methods lack diversity and efficiency, and it is difficult to achieve targeted attacks.

Method used

Adopting the wireless communication signal adversarial sample generation method based on the diffusion model, the Markov chain diffusion and U-Net network training of the key samples is generated, and the original signal is restored through the reverse process denoising, realizing the Against the diversity of samples and targeted attacks.

Benefits of technology

Significantly improves the diversity and generation efficiency of adversarial samples, provides the same effect as classifier guidance without the need for a classifier, and can generate samples into specified categories for attack, improving the robustness of the electromagnetic signal perception model .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067193A_ABST
    Figure CN119067193A_ABST
Patent Text Reader

Abstract

The invention discloses a wireless communication signal confrontation sample generation method based on a diffusion model, and belongs to the field of wireless communication signal confrontation samples. The method specifically comprises the steps that firstly, stratified sampling is conducted on an existing data set, an AMR model is trained, the confidence degree of each sample under each modulation type is calculated, and the type corresponding to the maximum confidence degree value is selected as a prediction result corresponding to each sample; obtaining a sample of the classification boundary; selecting a key sample according to an obtained classification boundary sample result, modifying a label of the key sample, and gradually adding Gaussian noise as a new training set until sample distribution completely meets Gaussian distribution; and training the U-Net network to predict the noise added in each step by using the sample characteristic attributes as gradient guidance, and gradually removing Gaussian noise from the sample until the original signal is recovered to obtain a wireless communication signal adversarial sample generation model. According to the method, the adversarial sample shows relatively strong aggressiveness to the model, so that the classification accuracy of the model is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of wireless communication signal adversarial samples, and in particular, relates to a method for generating adversarial samples of wireless communication signals based on a diffusion model. Background Art

[0002] Automatic Modulation Recognition (AMR) technology for wireless communication signals can collect and extract signal features, such as signal spectrum, instantaneous amplitude and instantaneous phase, to provide support for efficient signal processing. It plays an important role in application scenarios such as network security, spectrum management, radio monitoring and intelligence collection. It can enhance system security, optimize spectrum resource utilization efficiency, improve monitoring accuracy, and provide decision makers with better decision-making basis.

[0003] With the rapid development of communication technology, artificial intelligence, software radio and other technologies, wireless communication signals are diverse and numerous, and the complexity of the wireless electromagnetic environment is increasing. Traditional likelihood-based and statistics-based perception and recognition methods are not effective, causing AMR technology to face great challenges. Against this background, AMR technology based on deep learning has become one of the mainstream development directions in the field of wireless communications.

[0004] Compared with traditional methods, deep learning-based AMR technology can significantly improve modulation recognition accuracy by utilizing large amounts of communication data. At the same time, automated feature extraction avoids the limitations of professional knowledge and manual experience, and can continuously benefit from the rapid iteration and evolution of deep learning tools, with the potential to solve complex modulation recognition problems. However, with the continuous deepening of research on deep learning technology, researchers have found that existing deep learning algorithms have serious security risks: attackers can easily deceive deep learning models by adding specific noise to benign samples, causing the normally trained model to output highly confident erroneous predictions without being discovered. Researchers call this phenomenon adversarial attack.

[0005] Adversarial attacks have been widely studied and applied in fields such as computer vision and natural language processing. A typical example is that when a carefully designed perturbation is added to a panda image, the deep learning classifier misclassifies it as a gibbon image. In the field of AMR, studies have shown that to construct adversarial samples in the signal field, it is only necessary to add carefully designed weak perturbations to clean signal samples. For example, adding a small adversarial perturbation to the original binary phase shift keying (BPSK) and quadrature phase shift keying (QPSK) modulation signals to form adversarial samples, and inputting the adversarial samples into the classifier may cause the pre-trained classifier to misclassify the signal as quadrature amplitude modulation (64QAM) or other modulation signals. In communication tasks, if the modulation recognition task is wrong, it will seriously threaten applications such as intelligent wireless communication systems, cognitive radio networks, electromagnetic reconnaissance and satellite navigation.

[0006] Adversarial sample generation methods were introduced into the field of electromagnetic signals in 2019. At present, traditional adversarial sample generation methods for AMR models are generally divided into two categories: gradient-based methods and optimization-based methods. The gradient-based method uses the back propagation of neural networks to generate adversarial samples by minimizing the difference between input samples and target samples, and adds perturbations in the direction of increasing gradients to make the model misjudge, but lacks specificity when adding adversarial perturbations; while the optimization-based method seeks input samples that maximize or minimize the objective function, and uses various variations to make the optimization problem easier to solve, so as to achieve the purpose of generating adversarial samples, but the generation process is time-consuming and prone to falling into local extremes. In addition, the purpose of adversarial sample attacks is to make the model make wrong decisions, but targeted attack methods that make the model identify samples as specified categories are rarely involved. Therefore, research on the generation of adversarial samples for electromagnetic signals is of practical significance.

[0007] Generative AI technology has received widespread attention in both industry and academia, among which the diffusion model is a representative model in the field of image generation. The diffusion model is a probability-based generative model that can generate new data samples by learning the potential distribution of data of given training samples. It surpasses the original SOTA: GAN model in image generation tasks and has outstanding performance in many application fields. Inspired by the image generation task, the diffusion model generation process is introduced into the electromagnetic field to conduct research on the generation of adversarial samples of wireless communication signals. Summary of the invention

[0008] In view of the above problems, the present invention provides a method for generating adversarial samples of wireless communication signals based on a diffusion model. By performing diffusion generation on key samples of a model training data set, adversarial samples that can significantly reduce the recognition accuracy of the model are obtained. Compared with traditional adversarial sample generation methods, the present invention can greatly improve the diversity of sample generation and can generate samples into specified categories for targeted attacks.

[0009] The method for generating adversarial samples of wireless communication signals based on a diffusion model comprises the following steps:

[0010] S1: Select a public wireless communication signal dataset D and divide it into several layers according to the feature attribute categories. Randomly extract signal samples from each layer; divide the dataset D into training set D in the form of stratified sampling. train and validation set D valid .

[0011] The elements in the initialization data set D are represented as: i (k), k = 1, 2, ..., K, K is the number of sampling points of a single signal, i = 1, 2, ..., N, N is the number of signal samples.

[0012] For R modulation types, each signal sample corresponds to a modulation type, and the modulation type label of the i-th signal sample is Represents, r∈{1,2,…,R};

[0013] S2: Input the training set into the AMR model with set network parameters for training, and verify each sample in the validation set. Calculate the confidence of each sample under each modulation type in turn, select the type corresponding to the maximum confidence value as the prediction result corresponding to each sample, and record the recognition accuracy P of the AMR model at this time. Accuracy ;

[0014] The confidence level corresponding to the i-th signal sample under R modulation types is Px i = {Pr1,Pr2,…,Pr R};

[0015] S3: for each signal sample of the stratified sampling, respectively calculate the variance array of the confidence between the prediction result of each sample and each modulation type, and when there is at least one variance value in the variance array that is less than the set threshold τ, save the signal sample corresponding to the variance array into the boundary sample;

[0016] The variance of the confidence between the prediction result A of the current i-th signal sample and each modulation type is expressed as an array as follows:

[0017]

[0018] Represents the variance of the prediction result A and the confidence under the r-th modulation type;

[0019] When the variance If the value is less than the threshold τ, the i-th signal sample is saved in the boundary sample and recorded as x j (k);

[0020] S4: Modify the modulation type label of each boundary sample to form a key sample set;

[0021] For the boundary sample x j (k) Select the modulation type corresponding to the minimum variance from the variance array of the corresponding i-th signal sample as the boundary sample x j (k) Boundary label

[0022] The boundary sample x j (k) and its corresponding boundary label As a set of data, save to the key sample set D key_d middle;

[0023] S5: Based on Markov chain, build a diffusion model for generating adversarial samples of wireless communication signals and use key sample sets for training;

[0024] Diffusion model training includes forward process and reverse process:

[0025] The forward process refers to: from the key sample set data D key_d The wireless communication signal x0 is loaded in the , and after T steps of forward iteration, Gaussian noise is gradually added to the wireless communication signal x0 until a sample x0 that completely satisfies the Gaussian distribution is obtained. T ;

[0026] First, the original signal x0 is transformed by adding noise to obtain the signal x t The expression is:

[0027]

[0028] It can be seen that the signal x at the current moment t is the signal x at the previous moment t-1 and Gaussian noise ε. Among them, β t is a pre-set hyperparameter, satisfying 0<β1<β2<…<β T <1, T is the number of diffusion steps, ε~Ν(0,I) obeys the standard normal distribution.

[0029] Then, set the intermediate variable 1-β t =α t , for the signal x tThe expression of is transformed to obtain the Markov recursive function:

[0030]

[0031] remember Rewrite the recursive function to get:

[0032]

[0033] Finally, according to the above formula, Gaussian noise is gradually added to the wireless communication signal x0, and Gaussian noise is gradually added to the data distribution of the wireless communication signal x0 through the Markov forward process until a sample data x0 that completely conforms to the Gaussian distribution is obtained. t .

[0034] The reverse process refers to: generating based on sampling of normal distribution, denoising the Gaussian noise data by training the U-Net network, and converting the sample data x after denoising t From a pure Gaussian data, gradually denoise and restore it to the original data x0;

[0035] Specifically:

[0036] First, according to the Bayesian formula, we get the posterior probability distribution of the forward process, that is, the reverse process q(x t-1 |x t ,x0):

[0037]

[0038] Then, combined with the rewritten recursive function (3), we get:

[0039]

[0040] Further simplification, we get the reverse process q(x t-1 |x t ,x0) are:

[0041]

[0042]

[0043] Next, the conditional sampling p is calculated using the U-Net network with parameters θ θ (x t-1 |x t ), so that its variance and posterior distribution are consistent with the reverse process q(x t-1 |x t ) to keep the variance consistent:

[0044] p θ (x t-1 |xt )=N(x t-1 ;μ θ (x t ,t),ε θ (x t ,t)) (8)

[0045] Where θ is the mean function μ of the Gaussian distribution θ (x t ,t) and standard deviation function ε θ (x t ,t) is the learnable parameter vector in .

[0046] Finally, an implicit classifier is designed to convert the boundary labels Merge into the training of the reverse process, that is, use Replacement conditional sampling p(x t-1 |x t ),get:

[0047]

[0048] By reverse solution, we get:

[0049]

[0050] in, The gradient of the implicit classifier is conditionally generated probability; ω is the weight, when ω is larger, the boundary label The greater the pulling strength, the more samples and boundary labels are generated. The higher the confidence level.

[0051] The training process is:

[0052] When the input of the U-Net neural network is given as (x t ,t), using the learned parameters θ and the obtained mean function μ θ (x t ,t) and standard deviation function ε θ (x t ,t), predict the noise ε at each step θ , noise ε θ Contains boundary labels The guidance can guide the noise towards the boundary label Direction denoising, narrowing the distribution and the reverse process q(x t-1 |x t ,x0), after T iterations, the noise ε θ The samples are removed until the original signal data is restored, and finally a diffusion model of the wireless communication signal against the samples is formed.

[0053] S6: Input a random sample x into the trained diffusion model T ~N(0,I), the number of iterations and the boundary labels of key samples, directly output the adversarial sample data with boundary labels, and form the adversarial sample dataset D adv .

[0054] S7: Generate adversarial sample dataset D adv Input it into the trained AMR model for classification prediction, and obtain the AMR model in the adversarial sample dataset D adv The recognition accuracy P′ Accuracy ; Compare P Accuracy and P′ Accuracy Verify the effect of adversarial sample generation.

[0055] The advantages of the present invention are:

[0056] The present invention discloses a method for generating adversarial samples of wireless communication signals based on a diffusion model. Compared with the traditional adversarial sample generation method, the present invention can greatly improve the diversity of sample generation, and provide the same effect as classifier guidance without the need for a classifier. It can generate samples into specified categories for targeted attacks. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of a method for generating adversarial samples of wireless communication signals based on a diffusion model. DETAILED DESCRIPTION

[0058] In order to facilitate those skilled in the art to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only partial embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.

[0059] The present invention proposes a method for generating adversarial samples of wireless communication signals based on a diffusion model. By performing conditional diffusion generation on key samples of a model training data set, adversarial samples of wireless communication signals that are difficult for the perception model to identify are generated. This method can greatly improve the efficiency of adversarial sample waveform design, generate effective new samples with strong specificity, ensure the optimization of generated samples, and enhance the robustness of the electromagnetic signal perception model.

[0060] The wireless communication signal adversarial sample generation method based on the diffusion model is as follows: Figure 1 As shown, the following steps are included:

[0061] S1: Select a public wireless communication signal dataset D and divide it into several layers according to the feature attribute categories. Randomly extract signal samples from each layer; divide the dataset D into training set D in the form of stratified sampling. train and validation set D valid .

[0062] For R modulation types, each signal sample corresponds to a modulation type, and the modulation type label of the i-th signal sample is Represents, r∈{1,2,…,R}; signal samples and labels together constitute the data set D;

[0063] In order to ensure that the number of samples of each characteristic attribute in the training set and the test set is relatively balanced, the data set is divided into several layers according to the overall characteristic attribute category, and then samples are randomly selected from each layer. The elements in the initialization data set D are represented as: i (k), k = 1, 2, ..., K, K is the number of sampling points of a single signal, i = 1, 2, ..., N, N is the number of signal samples.

[0064] S2: Input the training set into the AMR model with set network parameters for training, and verify each sample in the validation set. Calculate the confidence of each sample under each modulation type in turn, select the type corresponding to the maximum confidence value as the prediction result corresponding to each sample, and record the recognition accuracy P of the AMR model at this time. Accuracy ;

[0065] The confidence level corresponding to the i-th signal sample under R modulation types is Px i = {Pr1,Pr2,…,Pr R};

[0066] Initialize the deep convolutional neural network structure and parameter settings, call the existing network layer function, and build a deep network structure; the convolution module is designed based on the standard convolutional neural network, and is composed of convolutional layers and activation function layers arranged alternately and repeatedly; in the convolutional layer, the convolution kernel size, stride and inner margin padding are set to 3, 2, and 1 respectively, and the activation function layer uses the ReLU activation function; the ReLU activation function is applied to add nonlinear transformation to increase the network's expressive power, the pooling layer reduces the feature dimension, and the fully connected layer performs feature integration.

[0067] S3: for each signal sample of the stratified sampling, respectively calculate the variance array of the confidence between the prediction result of each sample and each modulation type, and when there is at least one variance value in the variance array that is less than the set threshold τ, save the signal sample corresponding to the variance array into the boundary sample;

[0068] The variance of the confidence between the prediction result A of the current i-th signal sample and each modulation type is expressed as an array as follows:

[0069]

[0070] Represents the variance of the prediction result A and the confidence under the r-th modulation type;

[0071] When the variance If the variance is less than the threshold τ, it means that the signal samples corresponding to the variance have small differences in confidence in each modulation type. Such samples are usually located between uncertain categories or close to the classification decision boundary. Then the i-th signal sample is saved in the boundary sample, denoted as x j (k);

[0072] S4: Modify the modulation type label of each boundary sample to form a key sample set;

[0073] For the boundary sample x j (k) Select the modulation type corresponding to the minimum variance from the variance array of the corresponding i-th signal sample as the boundary sample x j (k) Boundary label

[0074] The boundary sample x j (k) and its corresponding boundary label As a set of data, save to the key sample set D key_d middle;

[0075] S5: Based on Markov chain, build a diffusion model for generating adversarial samples of wireless communication signals and use key sample sets for training;

[0076] The diffusion model is a probability model based on the Markov chain. First, Gaussian noise is gradually added to the key samples until the sample distribution completely satisfies the Gaussian distribution. Then, the U-Net network is trained to predict the noise added at each step, and the Gaussian distribution samples are gradually restored to the original distribution state to complete the model training.

[0077] Diffusion model training includes forward process and reverse process:

[0078] The forward process refers to: from the key sample set data D key_d In the example, signal samples are selected as the initial wireless communication signal x0. After T steps of forward iteration, Gaussian noise is gradually added to the wireless communication signal x0 until a sample x0 that completely satisfies the Gaussian distribution is obtained. T ;

[0079] The forward process q(x t |x t-1 ) is the process of adding noise to the signal data at each step, and the formula is:

[0080]

[0081] Among them, N~(0,I) is a standard normal distribution, and the mean and variance are and β t ;

[0082] First, based on equation (1), the original signal x0 is transformed by adding noise, and the standard normal distribution ε~N(0,I) is sampled to obtain the signal x t The expression is:

[0083]

[0084] Formula (2) shows that the data x at the current moment t is the data x at the previous moment t-1 and Gaussian noise ε. Among them, β t is a pre-set hyperparameter, satisfying 0<β1<β2<…<β T <1, T is the number of diffusion steps, ε~Ν(0,I) obeys the standard normal distribution.

[0085] Set the intermediate variable 1-β t =α t ,get:

[0086]

[0087] x t is the noise data at time t, T is the number of diffusion steps, α t is the hyperparameter at time t, and ε is the Gaussian noise added at time t-1.

[0088] Formula (3) is a Markov recursive function, denoted by Formula (3) can be recursively rewritten as:

[0089]

[0090] According to the formula Gaussian noise is gradually added to the wireless communication signal x0, and the data distribution of the wireless communication signal x0 is gradually added with Gaussian noise through the Markov forward process until a sample data x0 that completely conforms to the Gaussian distribution is obtained. t .

[0091] The reverse process refers to: generating based on sampling of normal distribution, denoising the Gaussian noise data by training the U-Net network, and converting the sample data x after denoising t From a pure Gaussian data, gradually denoise and restore it to the original data x0;

[0092] The reverse process q(x t-1 |xt ,x0) is the forward process q(x t |x t-1 ), the posterior probability distribution, q(x t-1 |x t ,x0) is derived according to the Bayesian formula:

[0093]

[0094] According to formula (4), we can get:

[0095]

[0096] Substituting formula (6) into formula (5) we can obtain:

[0097]

[0098] Simplifying, we can get q(x t-1 |x t ,x0) are:

[0099]

[0100]

[0101] Substituting formula (4) into formula (8) yields:

[0102]

[0103] However, due to q(x t-1 |x t ,x0) cannot be obtained directly, so it is necessary to train a U-Net neural network for fitting. Use a UNet network with parameter θ to calculate the conditional sampling p θ (x t-1 |x t ), so that its variance and posterior distribution q(x t-1 |x t ) to keep the variance consistent:

[0104] p θ (x t-1 |x t )=N(x t-1 ;μ θ (x t ,t),ε θ (x t ,t)) (11)

[0105] Where θ is the mean function μ of the Gaussian distribution θ (x t ,t) and standard deviation function ε θ(x t ,t) in the learnable parameter vector. Equation (11) shows that the noise can be θ (x t ,t) is adjusted, and this value is predicted by the U-Net network. The input of the U-Net neural network is (x t ,t), the output is the noise ε that needs to be removed at the current moment θ .

[0106] In order to make the diffusion model generate boundary labels based on For samples, we design an implicit classifier to convert the boundary labels Merge into the reverse process and directly pass the boundary labels Add to the training process, that is, Replace p(x t-1 |x t ), the conditional sampling process is defined as:

[0107]

[0108] The gradient of the implicit classifier is expressed in terms of conditional and unconditional generation probabilities as:

[0109]

[0110] Transformation yields:

[0111]

[0112] Where ω is the weight. When ω is larger, the boundary label The greater the pulling strength, the more samples and boundary labels are generated. The higher the confidence level.

[0113] The training process is:

[0114] When the input of the U-Net neural network is given as (x t ,t), using the learned parameters θ and the obtained mean function μ θ (x t ,t) and standard deviation function ε θ (x t ,t), predict the noise ε at each step θ , noise ε θ Contains boundary labels The guidance can guide the noise towards the boundary label Direction denoising, narrowing the distribution and the reverse process q(x t-1 |x t ,x0), after T iterations, the noise ε θThe samples are removed until the original signal data is restored, and finally a diffusion model of the wireless communication signal against the samples is formed.

[0115] The objective function of optimization is:

[0116]

[0117] The loss function used is: Lα=||ε(x t ,t)-ε θ (x t ,t,l)|| 2 ;

[0118] S6: Input a random sample x into the trained diffusion model T ~N(0,I), the number of iterations and the boundary labels to be generated, directly output the adversarial sample data with boundary labels, and form the adversarial sample dataset D adv .

[0119] Modulation type using samples l i The attribute, that is, the boundary label is used as a gradient guide to train the U-Net network to predict the noise added at each step, and gradually remove the Gaussian noise from the sample until the original signal is restored. Specifically: based on the noise data x t Compute a conditional distribution from which samples are drawn relative to x t A slightly cleaner data sample that is closer to the real signal data is used to train the U-Net neural network, starting from pure noise and gradually denoising until a real signal is obtained.

[0120] The random Gaussian noise signal and the feature labels of the samples to be generated are simultaneously input into the trained diffusion model, and a new time domain signal sample with the input features is output.

[0121] S7: Generate adversarial sample dataset D adv Input the trained AMR model for classification prediction and obtain the AMR model in the adversarial sample dataset D adv The recognition accuracy P′ Accuracy ; Compare P Accuracy and P′ Accuracy Verify the effect of adversarial sample generation.

[0122] It can be seen that adversarial samples are highly aggressive to the model, causing the model classification accuracy to decrease.

[0123] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope of the conclusion of the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for generating adversarial samples of wireless communication signals based on a diffusion model, characterized in that: The following steps are involved: S1: Select a public wireless communication signal dataset D and divide it into several layers according to the feature attribute categories. Randomly extract signal samples from each layer; divide the dataset D into training set D in the form of stratified sampling. train and validation set D valid ; S2: Input the training set into the AMR model with set network parameters for training, and verify each sample in the validation set. Calculate the confidence of each sample under each modulation type in turn, select the type corresponding to the maximum confidence value as the prediction result corresponding to each sample, and record the recognition accuracy P of the AMR model at this time. Accuracy ; The confidence level corresponding to the i-th signal sample under R modulation types is Px i = {Pr1,Pr2,…,Pr R }; S3: for each signal sample of the stratified sampling, respectively calculate the variance array of the confidence between the prediction result of each sample and each modulation type, and when there is at least one variance value in the variance array that is less than the set threshold τ, save the signal sample corresponding to the variance array into the boundary sample; S4: Modify the modulation type label of each boundary sample to form a key sample set; S5: Based on Markov chain, build a diffusion model for generating adversarial samples of wireless communication signals and use key sample sets for training; Diffusion model training includes forward process and reverse process: The forward process refers to: from the key sample set data D key_d The wireless communication signal x0 is loaded in the , and after T steps of forward iteration, Gaussian noise is gradually added to the wireless communication signal x0 until a sample x0 that completely satisfies the Gaussian distribution is obtained. T ; First, the original signal x0 is transformed by adding noise to obtain the signal x t The expression is: It can be seen that the signal x at the current moment t is the signal x at the previous moment t-1 and Gaussian noise ε weighted; where β t is a pre-set hyperparameter, satisfying 0<β1<β2<…<β T <1, T is the number of diffusion steps, εΝ(0,I) obeys the standard normal distribution; Then, set the intermediate variable 1-β t =α t , for the signal x t The expression of is transformed to obtain the Markov recursive function: remember Rewrite the recursive function to get: Finally, according to the above formula, Gaussian noise is gradually added to the wireless communication signal x0, and Gaussian noise is gradually added to the data distribution of the wireless communication signal x0 through the Markov forward process until a sample data x0 that completely conforms to the Gaussian distribution is obtained. t ; The reverse process refers to: generating based on sampling of normal distribution, denoising the Gaussian noise data by training the U-Net network, and converting the sample data x after denoising t From a pure Gaussian data, gradually denoise and restore it to the original data x0; Specifically: First, according to the Bayesian formula, we get the posterior probability distribution of the forward process, that is, the reverse process q(x t-1 |x t ,x0): Then, combined with the rewritten recursive function (3), we get: Further simplification, we get the reverse process q(x t-1 |x t ,x0) are: Next, the conditional sampling p is calculated using the U-Net network with parameters θ θ (x t-1 |x t ), so that its variance and posterior distribution are consistent with the reverse process q(x t-1 |x t ) to keep the variance consistent: Where θ is the mean function μ of the Gaussian distribution θ (x t ,t) and standard deviation function ε θ (x t ,t) in the learnable parameter vector; Finally, an implicit classifier is designed to convert the boundary labels Merge into the training of the reverse process, that is, use Replacement conditional sampling p(x t-1 |x t ),get: By reverse solution, we get: in, The gradient of the implicit classifier is conditionally generated probability; ω is the weight, when ω is larger, the boundary label The greater the pulling strength, the more samples and boundary labels are generated. The higher the confidence level; S6: Input a random sample x into the trained diffusion model T ~N(0,I), the number of iterations and the boundary labels of key samples, directly output the adversarial sample data with boundary labels, and form the adversarial sample dataset D adv ; S7: Generate adversarial sample dataset D adv Input it into the trained AMR model for classification prediction, and obtain the AMR model in the adversarial sample dataset D adv The recognition accuracy P A ' ccuracy ; Compare P Accuracy With P A ' ccuracy Verify the effect of adversarial sample generation.

2. The method for generating adversarial samples of wireless communication signals based on a diffusion model according to claim 1, characterized in that: In S1, the elements in the initialization data set D are represented as: i (k), k = 1, 2, ..., K, K is the number of sampling points of a single signal, i = 1, 2, ..., N, N is the number of signal samples; For R modulation types, each signal sample corresponds to a modulation type, and the modulation type label of the i-th signal sample is l i r Represents, r∈{1,2,…,R}.

3. The method for generating adversarial samples of wireless communication signals based on a diffusion model according to claim 1, characterized in that: In S3, the variance of the confidence between the prediction result A of the current i-th signal sample and each modulation type is represented by an array as follows: Represents the variance of the prediction result A and the confidence under the r-th modulation type; When the variance If the value is less than the threshold τ, the i-th signal sample is saved in the boundary sample and recorded as x j (k).

4. The method for generating adversarial samples of wireless communication signals based on a diffusion model according to claim 1, characterized in that: The S4 is specifically as follows: for the boundary sample x j (k) Select the modulation type corresponding to the minimum variance from the variance array of the corresponding i-th signal sample as the boundary sample x j (k) Boundary label The boundary sample x j (k) and its corresponding boundary label As a set of data, save to the key sample set D key_d middle.

5. The method for generating adversarial samples of wireless communication signals based on a diffusion model according to claim 1, characterized in that: The training process in S5 is: When the input of the U-Net neural network is given as (x t ,t), using the learned parameters θ and the obtained mean function μ θ (x t ,t) and standard deviation function ε θ (x t ,t), predict the noise ε at each step θ , noise ε θ Contains boundary labels The guidance can guide the noise towards the boundary label Direction denoising, narrowing the distribution and the reverse process q(x t-1 |x t ,x0), after T iterations, the noise ε θ The samples are removed until the original signal data is restored, and finally a diffusion model of the wireless communication signal against the samples is formed.

Citation Information

Patent Citations

  • Unrestricted adversarial sample generation method and device, electronic equipment and storage medium

    CN116152087A

  • Defense method facing machine learning to confront attacks

    CN116543240A

  • Signal modulation identification countermeasure attack sample generation method based on thermodynamic diagram and characteristic difference

    CN117874581A

Cited By

  • Black box attack method and device based on few samples, equipment and medium

    CN120337987A