A radar transmitting signal generation method for modulated identification network stealth

Abstract

This invention discloses a method for generating radar transmission signals that can stealth from modulation recognition networks. First, the time spectrum of the original signal is obtained through short-time Fourier transform. Then, based on the modulation recognition model and the auxiliary modulation recognition model, the modified DeepFool algorithm is iteratively executed to generate modulation stealth information. Finally, the improved inverse short-time Fourier transform is used to obtain the time-domain anti-recognition transmission signal. The signal waveform generated by this method can cause intelligent modulation recognition networks with time-frequency diagrams as input to make errors with high confidence, and can gain during pulse compression, thus possessing the ability to perform subsequent target detection processing.

Description

Technical Field

[0001] This invention belongs to the technical field of radar electronic reconnaissance and artificial intelligence, and specifically relates to a method for generating radar transmission signals that can be used to stealth modulation recognition networks. Background Technology

[0002] With the development of intelligent methods such as deep learning, passive detection systems such as electronic reconnaissance systems, radar warning receivers, and electronic countermeasures equipment are continuously improving their ability to intercept active radio frequency (RF) radiation sources, and their sorting and identification accuracy is constantly increasing. This makes radar jamming more precise, posing a serious threat to radar systems. RF stealth technology is an effective means to cope with the ever-evolving passive threats and improve radar's anti-reconnaissance and anti-jamming capabilities. This technology reduces target characteristics by controlling the signal parameters of the radiation source, preventing the radar from being intercepted, sorted, and identified by electronic reconnaissance systems, thus improving the radar's battlefield survivability. The corresponding radar RF stealth measures are anti-interception, anti-sorting, and anti-identification. Among them, anti-identification stealth technology aims to prevent electronic reconnaissance systems from obtaining the modulation type of the radiation source from the intercepted signal, avoiding providing information for subsequent parameter measurement, state reasoning, and jamming generation operations. It is a crucial part of countering electronic reconnaissance systems. However, existing anti-identification RF stealth signal generation methods are few and mainly designed for traditional feature extraction-based identification methods. Summary of the Invention

[0003] In view of this, the present invention provides a method for generating radar transmission signals with modulation stealth capabilities for modulation identification networks. The modulation stealth signal generated by this method can improve the modulation stealth anti-identification capability of the radar transmission signal while maintaining normal pulse compression.

[0004] A method for generating radar transmission signals to identify network stealth through pulse modulation includes the following steps:

[0005] Step 1: Perform time-frequency analysis on the original radar transmitted signal x(n) to generate a complex time-frequency spectrum X(m, ω); then decompose the real part X(m, ω) into its components. r (m, ω) and imaginary part X i The time spectrum H(m, ω) is composed of two channels (m, ω), which is abbreviated as H; then the time-frequency diagram I is drawn based on the time spectrum H;

[0006] Step 2: Build a modulation recognition model f and an auxiliary modulation recognition model g. The modulation recognition model f takes the time-frequency graph as input and outputs the classification result of the time-frequency graph. The auxiliary recognition model g is used to identify the modulation type of the time spectrum. It adopts the same network structure as model f, takes the time spectrum as input, and outputs the model's recognition result of the input sample and the gradient value in the process.

[0007] Step 3: Use modulation recognition model f to perform anti-recognition test on time-frequency graph I, specifically as follows:

[0008] Let the true label of I be 1, and the recognition test result output by the modulation recognition model f be t. If t = l, it indicates that the time-frequency map does not have anti-recognition capability, and proceed to step four; otherwise, if t ≠ l, it indicates that the time-frequency map has anti-recognition capability, and there is no need to continue generating subsequent stealth information, and proceed to step five.

[0009] Step 4: Use the auxiliary recognition model g to identify the time spectrum H of the two channels and use the modified DeepFoo1 algorithm to generate stealth information, specifically:

[0010] The two-channel time spectrum should have the same true label and classification result as the time-frequency plot of the same type. Therefore, the true label and recognition test result of the two-channel time spectrum H are also denoted as 1 and t.

[0011] Let g k (H) represents the probability value that the auxiliary recognition model g predicts the two-channel time spectrum H as the k-th class, where k = 1, 2, ..., K, and K represents the total number of signal modulation classes;

[0012] For other categories k that do not belong to category l, the following iterative calculation is performed:

[0013] g′ k =g k (H i )-g l (H i );

[0014]

[0015] Where i is the iteration number; g k (H i ) indicates that in the i-th iteration, the auxiliary recognition model g will use the time spectrum H of the two channels. i The probability of being identified as class k; g l (H i ) indicates that in the i-th iteration, the auxiliary recognition model g will use the time spectrum H of the two channels. i The probability of being identified as class l; k = 1, 2, ..., K, k ≠ l; This indicates calculating the gradient;

[0016] Based on the iterative calculation results, traverse k = 1, 2, ..., K, k ≠ l, and calculate the distance from the sample to each decision boundary. The category number corresponding to the decision boundary closest to the sample is denoted as... The change in the time spectrum is

[0017] Spectrum during update: H i =H i +ΔHi According to H i Plot the time-frequency graph, denoted as I, and update the iteration number to i = i + 1; return to step three;

[0018] Step 5: Based on the current time spectrum H i The time-domain modulated stealth transmission signal was reconstructed.

[0019] Preferably, in step five, the time-domain modulated stealth transmission signal is reconstructed. The method is as follows:

[0020] Let the time spectrum obtained by iteration be H i If denoted as Y(mS,ω), then the reconstructed time-domain signal is represented as:

[0021]

[0022] in, This represents the reconstructed time-domain signal, where n is an integer; ω is the angular frequency; w(n) represents the window function.

[0023] Preferably, the modulation recognition model f and the auxiliary modulation recognition model g adopt one of the following models: VGG16, ResNet18, or CNN-base.

[0024] Preferably, the time-frequency analysis method in step one uses the short-time Fourier transform.

[0025] The present invention has the following beneficial effects:

[0026] This method addresses the anti-identification problem of intelligent modulation and identification networks in electronic reconnaissance systems by proposing a modulation stealth radar transmission signal generation method. This method mainly comprises three stages: original signal time-frequency spectrum generation, iterative generation of stealth information, and time-domain waveform reconstruction of the stealth signal. First, in the original signal time-frequency spectrum generation stage, STFT is used to perform time-frequency analysis on the signal to obtain the corresponding time-frequency spectrum. Then, in the stealth information iterative generation stage, a modified DeepFool algorithm is used to attack the reconnaissance party's intelligent modulation and identification network to generate stealth information. Finally, in the stealth signal time-domain waveform reconstruction stage, the time-frequency spectrum containing modulation stealth information is converted into a time-domain modulated stealth signal using the IMSTFT method, achieving modulation-based stealth while ensuring normal radar detection. Attached Figure Description

[0027] Figure 1 This is a flowchart of the algorithm of the present invention.

[0028] Figure 2 The results of the anti-identification experiment are shown in the embodiments of the present invention.

[0029] Figure 3 The following are the pulse compression experimental results of the present invention: a) is the signal pulse compression result under ideal signal conditions; b) is a magnified view of a part. Detailed Implementation

[0030] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] This invention provides a method for generating radar transmission signals to achieve stealth through modulation recognition networks, comprising the following steps:

[0032] Step 1: Perform time-frequency analysis on the original radar transmitted signal x(n) to generate a complex time-frequency spectrum X(m, ω); then decompose the real part X(m, ω) into its components. r (m, ω) and imaginary part X i The two channels are composed of time spectrum H(m, ω), which is abbreviated as H. Then, the time-frequency diagram I is drawn based on the time spectrum H.

[0033] Step 2: Construct modulation recognition model f and auxiliary modulation recognition model g. VGG16, ResNet18, and CNN-based models are used as modulation recognition model f to identify the modulation type of the time-frequency graph. All three models take the time-frequency graph as input, with 3 input channels, and output the classification result of the time-frequency graph. The auxiliary recognition model is used to identify the modulation type of the time-spectrum. It adopts the same network structure as model f, takes the time-spectrum as input, with 2 input channels, and outputs the model's recognition result of the input sample and the gradient values ​​during the process.

[0034] Step 3: Use modulation recognition model f to perform anti-recognition test on time-frequency graph I, specifically as follows:

[0035] Let the true label of I be 1, and the recognition test result output by the modulation recognition model f be t. If t = l, it indicates that the time-frequency map does not have anti-recognition capability, and proceed to step four; otherwise, if t ≠ l, it indicates that the time-frequency map has anti-recognition capability, and there is no need to continue generating subsequent stealth information, and proceed to step five.

[0036] Step 4: Use the auxiliary recognition model g to identify the time spectrum H of the two channels and use the modified DeepFoo1 algorithm to generate stealth information, specifically:

[0037] The two-channel time spectrum should have the same true label and classification result as the time-frequency plot of the same type. Therefore, the true label and recognition test result of the two-channel time spectrum H are also denoted as 1 and t.

[0038] Let g k (H) represents the probability value that the auxiliary recognition model g predicts the two-channel time spectrum H as the k-th class, k = 1, 2, ..., K, where K represents the total number of signal modulation classes;

[0039] To shift the sample towards the incorrect category, the following iterative calculation is performed for other categories k that do not belong to category l:

[0040] g′ k =g k (H i )-g l (H i );

[0041]

[0042] Where i is the iteration number; g k (H i ) indicates that in the i-th iteration, the auxiliary recognition model g will use the time spectrum H of the two channels. i The probability of being identified as class k; g l (H i ) indicates that in the i-th iteration, the auxiliary recognition model g will use the time spectrum H of the two channels. i The probability of being identified as class l; k = 1, 2, ..., K, k ≠ l; This indicates calculating the gradient;

[0043] Based on the iterative calculation results, traverse k = 1, 2, ..., K, k ≠ l, and calculate the distance from the sample to each decision boundary. To ensure that the modification to the original sample is minimized, the category index corresponding to the decision boundary closest to the sample is denoted as... Right now The change in the time spectrum is

[0044] Spectrum during update: H i =H i +ΔH i This causes the time spectrum to move towards the nearest error category; according to H i Plot the time-frequency graph, denoted as I, and update the iteration number to i = i + 1; return to step three;

[0045] Step 5: Obtain the time-domain modulated stealth transmission signal using the improved inverse short-time Fourier transform.

[0046] Furthermore, the time-frequency analysis methods in step one include, but are not limited to, the Short-Time Fourier Transform (STFT). The STFT is calculated as follows:

[0047]

[0048] In the formula, x(n) is the time-domain signal, w(n) is the window function, and ω is the angular frequency.

[0049] Furthermore, the modulation recognition model in step two includes, but is not limited to, VGG, ResNet, and CNN-base.

[0050] Furthermore, the attack methods used in step four include, but are not limited to, the DeepFoo1 algorithm.

[0051] Furthermore, the improved inverse short-time Fourier transform method used in step five is IMSTFT (Inverse Modified STFT). This method reconstructs a time-domain signal based on the least squares approach, minimizing the difference between the time spectrum of the signal and the modified time spectrum, thus solving the problem that the standard inverse short-time Fourier transform cannot reconstruct the time-domain modulated stealth signal from the modulation stealth time spectrum. Let the iteratively obtained time spectrum H... i If denoted as Y(mS,ω), the reconstructed time-domain signal can be expressed as:

[0052]

[0053] in, This represents the reconstructed time-domain signal, where n is an integer; ω is the angular frequency; w(n) represents the window function.

[0054] Example:

[0055] (1) Experimental scenario setup

[0056] This experiment evaluates the proposed method using a radar adversarial attack scenario consisting of a radar employing an adversarial attack method to establish a transmitted waveform library and a reconnaissance receiver using an intelligent modulation and identification network (EMI). The radar side uses an adversarial attack method to generate a modulated stealth signal, constructing an EPI radar transmitted signal waveform library. The radar selects and transmits the modulated stealth signal from this library. After radiating and propagating, this signal is intercepted and received by the jamming system, causing the EPI network, which uses a time-frequency diagram as input, to make a high-confidence error prediction, thus reducing its threat to the radar. The echo signal scattered by the target can be pulse-compressed normally in the radar's receiving and processing branch for subsequent target echo detection processing.

[0057] (2) Evaluation indicator setting

[0058] The experiment used Attack Success Rate (ASR) and pulse compression results as evaluation metrics. ASR assesses the anti-identification performance of radar-modulated stealth signals. Pulse compression results assess the feasibility of pulse compression for radar-modulated stealth signals.

[0059] (3) Experimental Procedure

[0060] The algorithm flowchart of this invention is as follows: Figure 1As shown in the table. Specifically, the length of the original radar transmitted signal x(n) is 1024, and the signal-to-noise ratios are [-10dB: 4dB: 10dB] and 0dB, respectively. The signal type and parameters are shown in the table.

[0061] Table 1 Original signal parameter settings

[0062]

[0063] Step 1: Generating the Time-Spectrum of the Original Signal. First, a Kaiser window with a window length of 161 and β = 9 is used to perform STFT on the three types of original signals x(n) to generate a complex time-spectrum X(m, ω) with a dimension of 1024 × 1024. Then, the real and imaginary parts of the complex time-spectrum are split to form a two-channel time-spectrum H with a dimension of 1024 × 1024 × 2.

[0064] Step 2: Model Construction. Construct modulation recognition model f and auxiliary modulation recognition model g. Build VGG16, ResNet18, and CNN-based models as modulation recognition networks. All three models take a 1024×1024 time-frequency image as input, with 3 input channels. The VGG16-based recognition network contains 13 convolutional layers and 3 fully connected layers. The first and second convolutional layers each have 64 3×3 kernels, the third and fourth convolutional layers each have 128 3×3 kernels, the fifth to seventh convolutional layers each have 256 3×3 kernels, and the eighth to thirteenth convolutional layers each have 512 3×3 kernels. There are 5 max-pooling layers (2×2 each) following the convolutional layers. The last pooling layer is followed by three fully connected layers. The ResNet18-based recognition network contains 17 convolutional layers and 1 fully connected layer. In the diagram, solid blue lines represent cross-layer connections where the feature dimension remains unchanged, while dashed blue lines represent cross-layer connections where the feature dimension changes. The first convolutional layer has 64 7×7 kernels, the second to fifth convolutional layers each have 64 3×3 kernels, the sixth to ninth convolutional layers each have 128 3×3 kernels, the tenth to thirteenth convolutional layers each have 256 3×3 kernels, and the fourteenth to seventeenth convolutional layers each have 512 3×3 kernels. Except for the first, sixth, tenth, and fourteenth convolutional layers, which have a stride of 2, the stride of the remaining convolutional layers is 1 by default. The 3×3 max pooling layer is located after the first convolutional layer, and the 1×1 average pooling layer is located after the last convolutional layer. The fully connected layer is located after the average pooling layer. The CNN-based recognition network consists of 5 layers: 2 convolutional layers and 3 fully connected layers. The first and second convolutional layers each have 16 3×3 convolutional kernels. Each convolutional layer is followed by a 2×2 max-pooling layer. The three fully connected layers are located after the last max-pooling layer. The auxiliary modulation recognition model uses the same network structure as the modulation recognition model, with 2 input channels.

[0065] Step 3: Anti-identification test. Draw an RGB time-frequency graph I with dimensions of 1024×1024×3 based on the magnitude values ​​of each element of the time-frequency spectrum H.

[0066] Let the true label of I be 1, and the recognition test result output by the modulation recognition model f be t. If t = l, it indicates that the time-frequency map does not have anti-recognition capability, and proceed to step four; otherwise, if t ≠ l, it indicates that the time-frequency map has anti-recognition capability, and there is no need to continue generating subsequent stealth information, and proceed to step five.

[0067] Step 4: Use the auxiliary recognition model g to identify the time spectrum H of the two channels and use the modified DeepFoo1 algorithm to generate stealth information, specifically:

[0068] The two-channel time spectrum should have the same true label and classification result as the time-frequency plot of the same type. Therefore, the true label and recognition test result of the two-channel time spectrum H are also denoted as 1 and t.

[0069] Let g k (H) represents the probability value that the auxiliary recognition model g predicts the two-channel time spectrum H as the k-th class, where k = 1, 2, ..., K, and K represents the total number of signal modulation classes;

[0070] For other categories k that do not belong to category l, the following iterative calculation is performed:

[0071] g′ k =g k (H i )-g l (H i );

[0072]

[0073] Where i is the iteration number; g k (H i ) indicates that in the i-th iteration, the auxiliary recognition model g will use the time spectrum H of the two channels. i The probability of being identified as class k; g l (H i ) indicates that in the i-th iteration, the auxiliary recognition model g will use the time spectrum H of the two channels. i The probability of being identified as class 1; k = 1, 2, ..., K, k ≠ l; This indicates calculating the gradient;

[0074] Based on the iterative calculation results, traverse k = 1, 2, ..., K, k ≠ l, and calculate the distance from the sample to each decision boundary. The category number corresponding to the decision boundary closest to the sample is denoted as... The change in the time spectrum is

[0075] Spectrum during update: H i =H i +ΔH i According to H i Plot the time-frequency graph, denoted as I, and update the iteration number to i = i + 1; return to step three;

[0076] Step 5: Based on the current time spectrum H i The time-domain modulated stealth transmission signal was reconstructed. The length of x(n) is the same.

[0077] (4) Results Analysis

[0078] Figure 2 The results of this method in resisting identification are shown. It can be seen that the success rate of attacking various target models under different signal-to-noise ratios is above 70%, with most attacks achieving a success rate higher than 85%. Therefore, this method can effectively generate stealth signals that are resistant to the reconnaissance party's modulation and identification network.

[0079] Figure 3 The results of pulse compression of the stealth signal generated by this method under the condition that the original signal is an ideal signal are shown. It can be seen that the modulated stealth signal can gain during pulse compression and has a main-side ratio similar to that of the original ideal signal.

[0080] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating radar transmit signals that are stealthy to pulse modulation recognition networks, characterized by, The method comprises the following steps: Step one, original radar transmitted signal Time-frequency analysis to generate complex time-frequency spectrum ; then split Real part and imaginary part Two-channel time-frequency spectrum , abbreviated as H ; then draw the time-frequency graph H According to the time-frequency spectrum I ; Step two, build modulation identification model f and auxiliary modulation identification model ; modulation identification model f with time-frequency diagram as input, and the output is the classification result of the time-frequency diagram; auxiliary identification model for identifying time-frequency spectrum modulation type, using the same network structure as the model f with time-frequency spectrum as input, and the output is the identification result of the input sample by the model and the gradient value in this process; Step three, utilizing a modulation identification model f on a time-frequency plot I Anti-identification test was performed, specifically: Set I The real label of l , modulation identification model f The output of the identification test result is t If , it indicates that the time-frequency diagram does not have anti-identification ability, and enters step four; otherwise, if , it indicates that the time-frequency diagram has anti-identification ability, and does not need to continue to generate subsequent stealth information, and enters step five; Step four, utilizing the auxiliary identification model The two-channel time-frequency spectrum H Identification is performed and a modified DeepFool algorithm is used for stealth information generation, specifically: The real labels and classification results of the two-channel time-frequency spectrum should be the same as those of the same type of time-frequency diagram, so the real labels and recognition test results of the two-channel time-frequency spectrum H are also recorded as l and t ; Let denote an auxiliary recognition model two-channel time-frequency spectrum H predicted as the k probability value of the class, , denote the total number of signal modulation classes; For other categories not belonging to the category the following iterative calculation is performed: ​ ； ； in, i This represents the number of iteration rounds. Indicates the first i Auxiliary recognition model during round iteration Time spectrum of two channels Identified as the first k The probability of a class; Indicates the first i Auxiliary recognition model during round iteration Time spectrum of two channels Identified as the first l The probability of a class; ; This indicates calculating the gradient; Based on the iterative calculation results, traverse... Calculate the distance from the sample to each decision boundary. The category number corresponding to the decision boundary closest to the sample is denoted as... The change in the time spectrum is ; Update the time-frequency spectrum: , according to plot the time-frequency graph, denoted as I , and update the iteration round number as ; return to step three; Step five, based on the current time-frequency spectrum ， Reconstruction of time-domain modulation stealth transmission signal .

2. A method of generating radar transmit signals for network stealth against pulse modulation recognition as claimed in claim 1, characterized in that, In the fifth step, the time-domain modulation stealth transmitting signal is reconstructed The method is as follows: Let the time-frequency spectrum obtained by iteration be is expressed as The time-domain modulation stealth transmission signal reconstructed is is expressed as: where n is an integer; , is the angular frequency; denotes a window function.

3. The radar transmission signal generation method for pulse modulation identification network stealth as described in claim 1, characterized in that, Modulation identification model f and auxiliary modulation identification model One of VGG16, ResNet18, CNN-base model is adopted.

4. The radar transmission signal generation method for identifying network stealth based on pulse modulation as described in claim 1, characterized in that, The time-frequency analysis method in step one adopts a short-time Fourier transform.

Images