A method for generating deceptive signals based on feature transfer

Abstract

The present invention provides a method for generating a deception signal based on feature migration, which belongs to the field of electronic countermeasure technology. The method first collects communication signals, divides the signals into training sets and verification sets, and uses these data to train a neural network model; then, Fourier transforms the deception signals as content input, Fourier transforms the real communication signals as style input, constructs a loss function for style migration, and optimizes the randomly generated data and content input data as deception signals; when the training loss change is less than a preset threshold in multiple consecutive iterations, it is considered that the deception signal optimization is completed; finally, a time domain deception signal is generated through an inverse Fourier transform. The present invention generates a more realistic deception signal by migrating the features in the real signal of the communication device. The signal generated by the present invention not only matches the target signal in terms of parameters, but also has the real features of the communication signal, which improves its camouflage and makes it difficult for the anti-interference system to identify.

Description

Technical Field

[0001] The present invention relates to the field of electronic countermeasure technology, and in particular to a deception signal generation method based on feature migration, which can generate more realistic deception signals to improve the countermeasure effect. Background Art

[0002] In modern communication systems, electronic countermeasures equipment is used to interfere with frequency-using equipment. One common method is to generate spoof signals.

[0003] In the article “Design and Implementation of UAV Deception Interference Source” published by Yang Zaixiu et al. in the Journal of Navigation and Positioning, digital signal processing chips and field programmable logic gate arrays were used to generate deception signals based on UAV signal parameters.

[0004] In the article “SAR-GMTI / InSAR Active Deception Jamming Method Based on Dual Jammer Collaboration” published by Liu Yemin et al. in Aerospace Electronic Countermeasures, digital signal processing methods such as frequency modulation, amplitude modulation, phase modulation and delay are used to generate a deception jamming signal that is consistent with the mathematical expression of the echo signal.

[0005] In the patent "A jamming method for adaptive allocation of fused sampling deception suppression resources" applied by Yangzhou Yu'an Electronic Technology Co., Ltd. with authorization announcement number CN117572356B, digital synthesis technology is used to generate deception jamming signals based on radar parameters.

[0006] In the patent "Method for generating 1-bit multi-false target deception interference for synthetic aperture radar" applied by Zhao Bo et al., with authorization announcement number CN113484829B, the radar signal is processed to generate deception interference signals based on single-frequency threshold frequency and single-frequency threshold phase. Although the similarity is low, it can effectively reduce the complexity of the interference equipment.

[0007] However, the above-mentioned spoof signal generation methods rely on digital synthesis. Although these signals are consistent with real signals in parameters, due to the lack of characteristics in device signals (such as thermal noise, propagation channel characteristics, filter edge effects, etc.), these forged signals are often easily identified as non-real signals by anti-interference systems. Therefore, the existing technology is insufficient in the realism of spoof signals, and a method that can enhance the authenticity of spoof signals is urgently needed. Summary of the invention

[0008] In view of this, the present invention proposes a deception signal generation method based on feature migration. This method generates a more realistic interference signal by migrating the features of the real signal (such as thermal noise, propagation channel response, filter edge characteristics, etc.) to the synthesized deception signal, thereby improving its camouflage and enhancing the confrontation effect.

[0009] The technical solution adopted by the present invention is:

[0010] A method for generating a deception signal based on feature migration comprises the following steps:

[0011] Step 1: Collect target device signals and build training and test data sets;

[0012] Step 2: construct a signal modulation style classification neural network model and train it to obtain a trained signal modulation style classification neural network model;

[0013] Step 3: Use the trained signal modulation style classification neural network model to transfer the signal features to the spoofing signal.

[0014] Furthermore, the specific method of step 1 is:

[0015] Step 1-1: Collect N target device signals, each of which has a length of K and is divided into M categories according to the modulation style; perform fast Fourier transform on each signal to form sample data , sample data The corresponding signal modulation style is label ; Build sample set and label sets , , , sample set and label sets Together they form a data set;

[0016] Step 1-2: Set the data set split ratio , divide the data set into a training data set and a test data set, where the training data set includes the training sample data set And training label dataset The amount of data included is , the test data set includes the test sample set and the test label set The amount of data included is .

[0017] Furthermore, the specific method of step 2 is:

[0018] Step 2-1: Construct a signal modulation style classification neural network model. The signal modulation style classification neural network model consists of a first feature extraction module, a second feature extraction module, a third feature extraction module and a classification model. , , and Indicates that the first feature extraction module, the second feature extraction module, and the third feature extraction module are all composed of convolutional layers, relu activation functions, pooling layers, and deconvolutional layers, and the classification model is composed of fully connected layers and relu activation functions; the input of the model is sample , the outputs of each module are:

[0019] ;

[0020] Step 2-2: Use Adam optimizer to optimize the network parameters of the signal modulation style classification neural network model For optimization, the loss function uses the cross entropy loss function and sets the training rate ;

[0021] Step 2-3: Set the signal modulation style classification neural network model to training mode and obtain the training sample data set And training label dataset Randomly extract B sample data and label data from the dataset to form a batch training sample data set and batch training labeled dataset ; Clear the optimizer's network parameters The cumulative gradient value of Obtained through the signal modulation style classification neural network model ,Will and Input the loss function and calculate the loss value loss; use the loss value loss and optimizer to classify the network parameters of the signal modulation style classification neural network model Optimize; repeat the process until the training sample data set And training label dataset Take it all;

[0022] Step 2-4: Set the classification deep neural network to evaluation mode and calculate the test sample set Obtained through the signal modulation style classification neural network model , get The type and test label set corresponding to the maximum value of the second dimension The number of the same type in the test sample set is divided by the total number of test samples to get the accuracy If the accuracy Less than threshold , then return to step 2-3 to retrain; otherwise, save the network parameters , and obtain the trained signal modulation style classification neural network model.

[0023] Furthermore, the specific method of step 3 is:

[0024] Step 3-1: Load network parameters ;

[0025] Step 3-2: Randomly extract a sample data from the dataset as style data ;

[0026] Step 3-3: Load the pre-spoofing signal, perform fast Fourier transform, and obtain the content data ;

[0027] Step 3-4: Construct random data with mean 0 and standard deviation 1 , set the weight , then the migration data is ;

[0028] Step 3-5: Use the Adam optimizer to migrate data Optimize and set the training rate ;

[0029] Step 3-6: Calculate style data The output values ​​of each module in the neural network model are classified by signal modulation style , , and ;

[0030] Step 3-7: Calculate content data The output values ​​of each module in the neural network model are classified by signal modulation style , , and ;

[0031] Step 3-8: Calculate migration data The output values ​​of each module in the neural network model are classified by signal modulation style , , and ;

[0032] Step 3-9: Clear the optimizer about migration data The accumulated gradient value is used to calculate the content loss value ,in Indicates the calculation of mean square error;

[0033] Calculate style loss ,in Indicates the calculation of the Gram matrix;

[0034] Setting weights , and the total migration loss is ;

[0035] Using the total transfer loss and optimizer, the transfer image Make updates;

[0036] Step 3-10: Repeat steps 3-5 to 3-9. If the change rate of the total migration loss obtained for M consecutive times is less than the preset threshold, the final migration data is obtained. ; For the final migration data Perform inverse Fourier transform to generate time domain spoofing signal.

[0037] The beneficial effects of the present invention are:

[0038] 1. The present invention generates a more realistic spoof signal by transferring the characteristics of the communication device signal (such as thermal noise, channel response, etc.) to the generated spoof signal.

[0039] 2. The deceptive signal generated by the present invention not only matches the target signal in terms of parameters, but also has similar characteristics to the real signal, thereby greatly improving the authenticity and disguise of the deceptive signal, making it difficult for the anti-interference system to identify it.

[0040] 3. The present invention is applicable to fields such as communication jamming and radar jamming, and can effectively improve the success rate of deceptive jamming. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Flowchart of a deception signal generation method based on feature migration.

[0042] Figure 2 Schematic diagram of the relationship between the number of training times and the loss value.

[0043] Figure 3 Schematic diagram of the relationship between the number of training times and accuracy.

[0044] Figure 4 Schematic diagram of the real part of the pre-spoofing signal.

[0045] Figure 5 Schematic diagram of the imaginary part of the pre-spoofing signal.

[0046] Figure 6 Schematic diagram of the power spectrum of the pre-spoofing signal.

[0047] Figure 7 Schematic diagram of the real part of the style signal.

[0048] Figure 8 Schematic diagram of the imaginary part of the style signal.

[0049] Fig. 9 Schematic diagram of the power spectrum of the style signal.

[0050] Fig.10 Schematic diagram of the relationship between total migration loss and number of iterations.

[0051] Fig.11 Schematic diagram of the real part of the pre-spoofing signal.

[0052] Fig.12 Schematic diagram of the imaginary part of the pre-spoofing signal.

[0053] Fig.13 Schematic diagram of the power spectrum of the pre-spoofing signal. DETAILED DESCRIPTION

[0054] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0055] A method for generating spoofing signals based on feature migration, which first collects the real signal of a communication device and converts the time domain signal into a frequency domain representation through Fourier transform. Then, the processed signal is divided into a training set and a validation set, and these data are used to train a neural network model so that it can accurately distinguish different types of communication signals. Next, the spoofing signal is Fourier transformed as the content input, the real communication signal is used as the style input, and the randomly generated data is used as the generation target. A loss function for style migration is constructed, and the randomly generated data and the content input data are used as spoofing signals for optimization training. When the loss value changes less than a preset threshold in multiple consecutive trainings, the model training is considered to be completed. Finally, after an inverse Fourier transform, the optimized frequency domain signal is converted back to the time domain signal to generate the final spoofing signal.

[0056] like Figure 1 As shown, the method specifically comprises the following steps:

[0057] Step 1: Target device signal acquisition and data set construction;

[0058] Step 1-1: Collect target device signals and build a data set.

[0059] The number of target device signal data collected is N = 11000, which is divided into M = 11 categories according to the modulation style, namely 8PSK, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK and WBFM. Fast Fourier transform is performed on each signal to form sample data: , respectively save the real and imaginary parts of the data, the dimension is , K is 128. The corresponding modulation pattern is used as the category label , the dimension is 1. The sample set is constructed as , , the dimension is , the label set is , , the dimension is .

[0060] Step 1-2: Split the dataset into training dataset and test dataset.

[0061] The data set is divided into =0.9. The training data set includes the training sample data set And training label dataset , the dimensions are and The test data set includes the test sample set and the test label set , the dimensions are and .

[0062] Step 2: Train the signal modulation style classification neural network model;

[0063] Step 2-1: Construct a signal modulation style classification neural network model.

[0064] The signal modulation style classification neural network model consists of a first feature extraction module, a second feature extraction module, a third feature extraction module and a classification model. , , and The feature extraction module consists of a convolutional layer, a relu activation function, a pooling layer, and a deconvolution layer. The classification model consists of a fully connected layer and a relu activation function. The input of the model is a sample , the outputs of each module are:

[0065] .

[0066] Specifically, the signal modulation style classification neural network model structure is shown in the following table:

[0067] .

[0068] Step 2-2: Use Adam optimizer to classify the parameters of the neural network model for signal modulation style Optimization, the loss function uses the cross entropy loss function and sets the training rate =0.0003.

[0069] Step 2-3: Batch training of the signal modulation style classification neural network model:

[0070] Set the signal modulation style classification neural network model to training mode. From the training sample dataset And training label dataset Randomly extract B=64 sample data and label data to form a batch training sample data set and batch training labeled dataset . Clear optimizer about network parameters The cumulative gradient value of . Calculate Obtained through the network .Will and Input the loss function and calculate the loss value. Use the loss value and optimizer to classify the parameters of the signal modulation style neural network model. Optimize. Repeat this process until the training sample data set Set and training label dataset Take it all.

[0071] Step 2-4: Evaluate the signal modulation pattern classification neural network model:

[0072] Set the classification deep neural network to evaluation mode. Calculate the test sample set Obtained through the network Get The type and test label set corresponding to the maximum value of the second dimension The number of the same type in the test sample set is divided by the total number of test samples to get the accuracy If the accuracy Less than threshold =0.99, then return to step 2-3; otherwise, save the network parameters of the signal modulation style classification neural network model . Figure 2 Schematic diagram of the relationship between the number of training times and the loss value. Figure 3 Schematic diagram of the relationship between the number of training times and accuracy.

[0073] Step 3: Transfer signal features to spoof signal:

[0074] Step 3-1: Load the network parameters of the signal modulation style classification neural network model .

[0075] Step 3-2: Randomly extract a signal data sample from the dataset as style data , the dimension is ;

[0076] Step 3-3: Load the pre-spoofing signal and perform fast Fourier transform, the dimension is , as content data The real part of the pre-spoofing signal is Figure 4 As shown, the imaginary part of the pre-spoofing signal is Figure 5As shown, the power spectrum of the pre-spoofing signal is Figure 6 As shown;

[0077] Step 3-4: Initialize spoof data : Construct random data with a mean of 0 and a standard deviation of 1 , the dimension is . Set the weight value , then the migration data is ,in =0.1.

[0078] Step 3-5: Use the Adam optimizer to migrate data Optimize and set the training rate is 0.001.

[0079] Step 3-6: Calculate style data The output value of the neural network model after signal modulation pattern classification , , and The style data is the frequency domain signal obtained by Fourier transforming the time domain signal of the frequency device. The real part of the style signal is as follows: Figure 7 As shown in Figure 8 As shown in Fig. 9 As shown;

[0080] Step 3-7: Calculate content data The output value of the neural network model after signal modulation pattern classification , , and .

[0081] Step 3-8: Calculate the migration data The output value of the neural network model after signal modulation pattern classification , , and ;

[0082] Step 3-9: Optimize the migration data:

[0083] Clear optimizer about migration data The accumulated gradient value of . Calculate the content loss value ,in Indicates the calculation of mean square error.

[0084] Calculate style loss ,in Represents the calculation of the Gram matrix.

[0085] Setting weights , and the total migration loss is , where the weight is set to 5.

[0086] Using the total transfer loss and optimizer, the transfer image to update.

[0087] Step 3-10: Repeat steps 3-5 to 3-9. If the change rate of the total migration loss obtained for M consecutive times is less than the preset threshold of 5%, the final migration data is obtained. , the relationship between the total migration loss and the number of iterations is shown in the figure Fig.10 As shown in the figure. Perform inverse Fourier transform to generate the time domain spoofing signal. The real part of the spoofing signal is as follows Fig.11 As shown in Figure 2, the imaginary part of the spoof signal is Fig.12 As shown, the power spectrum of the deceptive signal is Fig.13 shown.

[0088] In summary, the present invention extracts the characteristics of the frequency-using equipment and transfers the characteristics of the real communication signal to the synthetic deception signal to form a more realistic deception signal, thereby enhancing the authenticity and disguise of the deception signal and improving the deception ability of the deception signal to the electronic countermeasure equipment, making it difficult for the anti-interference system to accurately identify it.

Claims

1. A method for generating a spoofing signal based on feature migration, characterized in that: The following steps are involved: Step 1: Collect target device signals and construct training and test data sets; the specific method is as follows: Step 1-1: Collect N target device signals, each of which has a length of K and is divided into M categories according to the modulation style; perform fast Fourier transform on each signal to form sample data , sample data The corresponding signal modulation style is label ; Build sample set and label sets , , , sample set and label sets Together they form a data set; Step 1-2: Set the data set split ratio , divide the data set into a training data set and a test data set, where the training data set includes the training sample data set And training label dataset , the amount of data included is , the test data set includes the test sample set and the test label set , the amount of data included is ; Step 2: Construct a signal modulation style classification neural network model and train it to obtain a trained signal modulation style classification neural network model; the specific method is: Step 2-1: Construct a signal modulation style classification neural network model. The signal modulation style classification neural network model consists of a first feature extraction module, a second feature extraction module, a third feature extraction module and a classification model. , , and Indicates that the first feature extraction module, the second feature extraction module, and the third feature extraction module are all composed of convolutional layers, relu activation functions, pooling layers, and deconvolutional layers, and the classification model is composed of fully connected layers and relu activation functions; the input of the model is sample , the outputs of each module are: ； Step 2-2: Use Adam optimizer to optimize the network parameters of the signal modulation style classification neural network model For optimization, the loss function uses the cross entropy loss function and sets the training rate ; Step 2-3: Set the signal modulation style classification neural network model to training mode and obtain the training sample data set And training label dataset Randomly extract B sample data and label data from the dataset to form a batch training sample data set and batch training labeled dataset ; Clear the optimizer's network parameters The cumulative gradient value of Obtained through the signal modulation style classification neural network model ,Will and Input the loss function and calculate the loss value loss; use the loss value loss and optimizer to classify the network parameters of the signal modulation style classification neural network model Optimize; repeat the process until the training sample data set And training label dataset Take it all; Step 2-4: Set the classification deep neural network to evaluation mode and calculate the test sample set Obtained through the signal modulation style classification neural network model , get The type and test label set corresponding to the maximum value of the second dimension The number of the same type in the test sample set is divided by the total number of test samples to get the accuracy If the accuracy Less than threshold , then return to step 2-3 to retrain; otherwise, save the network parameters , obtain the trained signal modulation style classification neural network model; Step 3: Use the trained signal modulation pattern classification neural network model to transfer the signal features to the spoof signal; the specific method is as follows: Step 3-1: Load network parameters ; Step 3-2: Randomly extract a sample data from the dataset as style data ; Step 3-3: Load the pre-spoofing signal, perform fast Fourier transform, and obtain the content data ; Step 3-4: Construct random data with mean 0 and standard deviation 1 , set the weight , then the migration data is ; Step 3-5: Use the Adam optimizer to migrate data Optimize and set the training rate ; Step 3-6: Calculate style data The output values ​​of each module in the neural network model are classified by signal modulation style , , and ; Step 3-7: Calculate content data The output values ​​of each module in the neural network model are classified by signal modulation style , , and ; Step 3-8: Calculate migration data The output values ​​of each module in the neural network model are classified by signal modulation style , , and ; Step 3-9: Clear the optimizer about migration data The accumulated gradient value is used to calculate the content loss value ,in Indicates the calculation of mean square error; Calculate style loss ,in Indicates the calculation of the Gram matrix; Setting weights , and the total migration loss is ; Using the total transfer loss and optimizer, the transfer image Make updates; Step 3-10: Repeat steps 3-5 to 3-9. If the change rate of the total migration loss obtained for M consecutive times is less than the preset threshold, the final migration data is obtained. ; For the final migration data Perform inverse Fourier transform to generate time domain spoofing signal.

Images