Small-sample active jammer identification method based on adversarial sample generation and ensemble learning

Abstract

The present invention discloses a small sample active interference identification method based on adversarial sample generation and ensemble learning. By extracting time-frequency features, an expanded data set is generated using adversarial samples, a GRU network and a residual neural network are used as base classifiers to respectively identify the signal time-domain and frequency-domain characteristics and time-frequency characteristics, and the Adaboost algorithm is used to adjust the sample weights and retrain the network, and the final results are weighted and fused to obtain the identification results. In view of the problem that the number of active interference identification samples is small in practical applications, the present invention proposes to first increase the number of disturbed samples by adversarial sample generation, and then improve the recognition accuracy by ensemble learning, thereby realizing active interference identification under small sample conditions. Increasing the number of samples, combined with the characteristics of different networks with different recognition rates for different active interference patterns, thereby improving the accuracy of recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of active interference identification, and specifically relates to a small sample active interference identification method based on adversarial sample generation and ensemble learning. Background Art

[0002] With the continuous development of electronic countermeasure technology, many radar active jamming methods with specific jamming effects have been continuously proposed. The complex and changeable radar active jamming causes the radar to face serious electromagnetic interference threats and cannot effectively detect the real target and its parameters. Therefore, finding a fast and efficient radar active jamming identification algorithm in a complex electromagnetic environment has become an important research direction in the field of radar countermeasures in recent years.

[0003] At present, although the radar interference recognition algorithms of traditional convolutional neural networks (CNNs) can make great progress on measured or simulated data sets, the research on these algorithms is still in its infancy. Although the recognition algorithms of traditional CNNs can predict interference test samples with high similarity in the training stage with high accuracy, the recognition accuracy of these algorithms will be greatly reduced when facing test samples with large differences in signal-to-noise ratio and other factors in the training stage. The successful application of traditional CNNs in the field of computer vision is mainly due to the fact that large labeled data sets are used to learn model parameters during the network training stage. Once sufficient labeled data cannot be obtained, the CNN-based model will have the risk of overfitting, resulting in a serious decline in network performance. However, in reality, it is difficult for us to obtain a large amount of labeled high-quality radar interference data that meets the model training requirements, which directly limits the performance of traditional CNN interference recognition algorithms in actual scenarios. Data enhancement and ensemble learning in a certain way can effectively solve the problem of low accuracy of neural network recognition of active interference types under small sample conditions.

[0004] In summary, the existing radar active interference recognition algorithm has poor characterization ability for interference with discrete features and weak features due to the extreme lack of samples, and cannot meet the requirements of complex and changeable electromagnetic environments in practical applications. Therefore, using a more complex neural network working method for feature learning and classification prediction can better complete the recognition of active interference signals in complex environments. Summary of the invention

[0005] In order to solve the problem that the existing technology is prone to overfitting when using small samples, resulting in low accuracy, the present invention provides a small sample active interference identification method based on adversarial sample generation and ensemble learning. Existing active interference identification methods often use a single network for classification and identification, and the recognition rate of a single network for different interferences is often different. Therefore, the present invention considers combining adversarial sample generation and ensemble learning for interference identification, while considering the frequency domain characteristics and time-frequency characteristics of the signal to identify different active interference patterns.

[0006] The technical solution of the present invention is:

[0007] The present invention proposes a small sample active interference recognition method based on adversarial sample generation and ensemble learning. By extracting time-frequency features, using adversarial samples to generate an expanded data set, using GRU network and residual neural network as base classifiers to respectively identify the signal time-domain frequency-domain characteristics and time-frequency characteristics, adjusting the sample weights and retraining the network through the Adaboost algorithm, and weighted fusion of the final results to obtain the recognition result. Specifically, it includes the following steps:

[0008] Step 1: Obtain samples of different active interference types with labels, generate augmented active interference samples through adversarial samples, and use the original samples and augmented samples to form an active interference dataset;

[0009] Step 2: Perform fast Fourier transform on the active interference data set to obtain the spectrum of the signal, and perform Cui-Williams distribution transform on the expanded signal to obtain the time-frequency diagram of the signal;

[0010] Step 3: Construct two base classifier models, where the first base classifier uses a GRU network and the second base classifier uses a residual neural network; both base classifier models are composed of N base classifiers.

[0011] Step 4: Initialize the sample weight to 1 / M, where M is the total number of samples. Input the original signal and spectrum into the GRU network, input the time-frequency graph into the residual neural network, and train the first base classifier in the two base classifier models;

[0012] Step 5: After the first base classifier is trained, save the trained network, assign weights to each base classifier in the two base classifier models according to the pseudo loss, change the sample weights according to the Adaboost algorithm, and input them into the network for training again. The classifier weights of the two base classifiers are updated in the same way as the sample weights;

[0013] Step 6: Repeat step 5 for the two base classifiers until the pseudo loss reaches the requirement, and then weight the final results to form a strong classifier.

[0014] Furthermore, in the above step 1, the adversarial sample generation method is selected as follows:

[0015] Use gradient-based adversarial samples (FGSM). FGSM can be understood as a gradient ascent algorithm, that is, using the plus sign to maximize the loss function.

[0016] The disturbance data is calculated by the following formula:

[0017]

[0018] Among them, x is the original sample, θ is the weight parameter of the model, which is 1, and real is the actual category of x. Input the original sample, weight parameter and real category, and use the J loss function to calculate the loss value of the neural network. The loss function uses SoftMax. It means to find the partial derivative of x, that is, the loss function J finds the partial derivative of x sample. sign is the sign function. ε is the perturbation ratio, and 8, 16, and 24 are taken to generate three sets of perturbed data respectively.

[0019] Furthermore, in the above step 2, the signal spectrum and time-frequency diagram are calculated as follows:

[0020] Perform fast Fourier transform (FFT) on the signal s(a) in the active interference data set to obtain the spectrum. Perform FFT on the signal to obtain the discrete spectrum S(k) of the signal:

[0021]

[0022] Among them, a is a discrete time point, A is the number of sample points, and k is a discrete frequency point.

[0023] Then, the signal s(a) in the active interference data set is transformed by Cui-Williams distribution (CWD) to obtain the time-frequency image of the signal and perform preprocessing, where the preprocessing includes cropping, graying, and median filtering. The time-frequency image CWD(t,f) of the signal obtained by performing CWD transformation on each frame of the signal is:

[0024]

[0025] Wherein, CWD(t,f) represents the CWD transformed time-frequency diagram of the radar active jammer signal at time t and frequency f, π represents pi, σ represents a positive scaling factor, σ represents the time delay of the radar active jammer signal, e(·) represents an exponential operation with the natural constant e as the base, u represents the local time of the radar active jammer signal, j represents the symbol of the imaginary unit, s(·) represents the time series of the radar active jammer signal, and * represents a conjugate operation.

[0026] Furthermore, the classifier weights of the first base classifier in step 5 are updated as follows:

[0027] Calculate the classification error rate of the nth base classifier:

[0028]

[0029] Where M is the total number of samples, D n (i) is the sample weight of the nth base classifier, G n (x i ) is the network prediction category after inputting the i-th sample, real i is the actual sample category of the i-th sample. The error rate is the sum of the weights of the misclassified samples.

[0030] Then calculate the weight of the nth base classifier:

[0031]

[0032] The weight update method of the input base classifier sample in step 5 is as follows:

[0033]

[0034] Where K is the number of sample categories, y i is the sample category label, p n (x i ) is the probability distribution of the output after the i-th data is input into the base classifier.

[0035] The base classifiers all use the Adam optimizer. Considering the different weights of samples, the loss function of the base classifier is:

[0036]

[0037] in, Output the results for the network.

[0038] Furthermore, the weighted fusion method in step 6 is as follows:

[0039] First, combine the base classifiers:

[0040]

[0041] Where G n (x) is the output of the nth base classifier in the first base classifier, α n is the weight of the nth base classifier in a base classifier, G n '(x') is the output of the nth base classifier in the second base classifier, β n is the weight of the nth base classifier in the second base classifier, and N is the total number of base classifiers in each base classifier.

[0042] Then the combined base classifiers are processed by the symbolic function to obtain the final strong classifier:

[0043]

[0044] Compared with the prior art, the present invention has the following advantages:

[0045] In view of the problem that the number of active interference recognition samples is small in practical applications, the present invention proposes to first increase the number of disturbed samples by generating adversarial samples, and then improve the recognition accuracy by ensemble learning, thereby realizing active interference recognition under small sample conditions. Increasing the number of samples and combining the characteristics of different networks with different recognition rates for different active interference patterns can further improve the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the method proposed by the present invention;

[0047] Figure 2 It is a schematic diagram of the overall framework of the network designed in the embodiment of the present invention;

[0048] Figure 3 is an integrated network structure diagram of an embodiment of the present invention;

[0049] Figure 4 This is the basic structure diagram of the GRU network;

[0050] Figure 5 It is the GRU unit structure diagram;

[0051] Figure 6 It is the basic structure diagram of the residual neural network;

[0052] Figure 7 It is the structural diagram of the residual block. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0054] The present invention proposes a small sample active interference identification method based on ensemble learning and adversarial sample generation to solve the problem of low accuracy caused by overfitting when using small samples in the prior art. Existing active interference identification methods often use a single network for classification and identification, and the recognition rate of a single network for different interferences is often different. Therefore, the present invention considers combining adversarial sample generation and ensemble learning for interference identification, while considering the time-frequency characteristics and time domain characteristics of the signal to identify different active interference patterns.

[0055] like Figure 1-7 As shown in FIG. 1 , a small sample active interference identification method based on adversarial sample generation and ensemble learning specifically includes the following steps:

[0056] Step 1: Obtain samples of different active interference types with labels, generate augmented active interference samples through adversarial samples, and use the original samples and augmented samples to form an active interference dataset;

[0057] Generating augmented data through adversarial examples includes the following steps:

[0058] The disturbance data is calculated by FGSM:

[0059]

[0060] Among them, x is the original sample, θ is the weight parameter of the model, which is 1, and real is the actual category of x. Input the original sample, weight parameter and real category, and use the J loss function to calculate the loss value of the neural network. The loss function uses SoftMax. It means to find the partial derivative of x, that is, the loss function J finds the partial derivative of x sample. sign is the sign function. ε is the perturbation ratio, and 8, 16, and 24 are taken to generate three sets of perturbed data respectively.

[0061] Step 2: Perform fast Fourier transform on the active interference data set to obtain the spectrum of the signal, and perform Cui-Williams distribution transform on the expanded signal to obtain the time-frequency diagram of the signal;

[0062] The signal spectrum and time-frequency image are obtained as follows:

[0063] Step 2.1: Perform FFT on the signal in the active interference data set to obtain the spectrum. Perform FFT on each frame of the signal to obtain the discrete spectrum S(k) of the signal:

[0064]

[0065] Among them, a is a discrete time point, A is the number of sample points, and k is a discrete frequency point.

[0066] Step 2.2: Perform CWD transformation on the signal s(a) in the active interference data set to obtain the time-frequency image of the signal and perform preprocessing, where the preprocessing includes cropping, grayscale, and median filtering. Perform CWD transformation on each frame of the signal to obtain the time-frequency information CWD(t,f):

[0067]

[0068] Among them, CWD(t,f) represents the CWD transform time-frequency distribution power spectrum density of the radar active jamming signal at time t and frequency f, π represents pi, σ represents a positive scaling factor, τ represents the time delay of the radar active jamming signal, e(·) represents an exponential operation with the natural constant e as the base, u represents the local time of the radar active jamming signal, j represents the imaginary unit symbol, J1(·) represents the time series of the radar active jamming signal, and * represents a conjugate operation. After the above calculation formula, the signal spectrum and time-frequency diagram are obtained.

[0069] The CWD transformed time-frequency image of each interference signal in the active interference data set is cropped, grayed, and median filtered in turn to obtain a radar active interference signal time-frequency image training set; the cropping refers to cropping the radar active interference signal time-frequency image into a time-frequency image with a size of 224*224; the graying refers to graying the cropped time-frequency image to obtain a grayscale image; the median filtering refers to median filtering the grayscale image to obtain a time-frequency image after eliminating image noise.

[0070] Step 3: Construct two base classifier models, where the first base classifier uses a GRU network and the second base classifier uses a residual neural network; both base classifier models are composed of N base classifiers.

[0071] Step 4: Initialize the sample weight to 1 / M, where M is the total number of samples. Input the original signal and spectrum into the GRU network, input the time-frequency graph into the residual neural network, and train the first base classifier in the two base classifier models;

[0072] The first base classifier, i.e., the GRU network, is trained using the expanded original signal and spectrum, and the second base classifier, i.e., the residual neural network, is trained using the time-frequency graph. The two networks are trained as the first base classifiers in their respective categories.

[0073] Step 4.1: The GRU network is used to identify the signal. The original signal and the spectrum are input into the GRU network for training.

[0074] The GRU network includes convolutional layer, pooling layer, fully connected layer 1, fully connected layer 2, GRU unit, fully connected layer 3 and SoftMax layer.

[0075] The input of the network is the active interference dataset x i , x i The feature dimension is 1×2048×2, where 1 represents a 1-dimensional signal, 2048 represents 2048 signal sampling points, and 2 represents the interference signal sample x i First, the interference signal sample x iThe data is input to the convolution layer to extract features. The convolution layer uses one-dimensional convolution, with a convolution kernel dimension of 1×9, a depth of 8, and a step size of 8. The convolution layer outputs the extracted features to the pooling layer. The pooling layer uses the maximum value pooling method, with the dimension also being one-dimensional, a window size of 1×4, and a step size of 4. That is, every 4 eigenvalues ​​of the input pooling layer data are taken to the maximum value, and the reduced-dimensional vector is input to the fully connected layer. The fully connected layer uses a two-layer hidden layer structure, including a fully connected layer 1 with 64 nodes and a fully connected layer 2 with 256 nodes. After receiving the output of the pooling layer and performing feature fusion, it outputs a feature vector of length 256 to the GRU unit.

[0076] The above fully connected layer also includes a fully connected layer 3 with K nodes, where K is the number of label categories, that is, the dimension of the final output confidence vector. The output of the GRU unit is subjected to feature fusion through the fully connected layer 3, and the output is a feature vector set U of K×1 dimensions (corresponding to K types of labels) i .

[0077] The GRU unit consists of an update gate and a reset gate. The update gate is used to control the influence of the output hidden layer at the previous moment on the output hidden layer at the current moment, and the reset gate is used to forget the unimportant parts of the hidden state.

[0078] The GRU unit formula is as follows:

[0079] Reset Gate R t The calculation formula is as follows:

[0080] R t =σ(W xr X t +W hr H t-1 +b r ) (13)

[0081] Where X t is the input, H t-1 is the hidden state at the previous moment, W hr is its corresponding weight, W xr is the updated weight, b r is the corresponding bias, σ is the sigmoid activation function;

[0082] Candidate hidden state The calculation formula is as follows:

[0083]

[0084] Where tanh is the activation function, ⊙ represents element-wise multiplication;

[0085] Update Gate Z t The calculation formula is as follows:

[0086] Z t =σ(W xz X t +W hz H t-1 +b z ) (15)

[0087] Where X t is the input, H t-1 is the hidden state at the previous moment, W hz is its corresponding weight, W xz is the updated weight, b z is the corresponding bias, σ is the sigmoid activation function;

[0088] Hidden state H t The calculation formula is as follows:

[0089]

[0090] Step 4.2: The residual neural network is used for image recognition. The time-frequency image is input into the residual neural network for training.

[0091] The residual neural network consists of a convolutional layer, four residual blocks, a global average pooling layer, and a fully connected layer.

[0092] The input of the network is a time-frequency image with a size of 224×224×1. First, the interference signal sample x i The input is sent to the convolution layer to extract features, the convolution kernel dimension is 7×7, and the step size is 2; the convolution layer outputs the extracted features to the residual block, which contains two basic residual blocks. The basic residual block structure is: input features, convolution layer, batch normalization, ReLU activation function, convolution layer, batch normalization, output features. The convolution kernel dimensions of the two convolution layers in the basic residual block are both 3×3, and the step size is 2; the final output of the residual block is input to the average pooling layer, which outputs a 1×1×512 feature map; the output of the pooling layer is input to the fully connected layer, which outputs the probability distribution of K categories.

[0093] Step 4.3: The two networks are trained separately. The cross entropy loss function (CrossEntropyLoss) is used to calculate the difference between the network output and the true label during the training process, and the Adam (adaptive moment estimation) algorithm is used for optimization. The learning rate of the GRU network is set to 10^-5, and the learning rate of the residual neural network is set to 0.001. The network parameters are updated through gradient back propagation to optimize the network prediction results. The training is stopped after the maximum number of iterations is reached.

[0094] Step 5: After the first base classifier is trained, save the trained network, assign weights to each base classifier in the two base classifier models according to the pseudo loss, change the sample weights according to the Adaboost algorithm, and input them into the network for training again. The classifier weights of the two base classifiers are updated in the same way as the sample weights.

[0095] The classifier weights of the first base classifier are updated as follows:

[0096] Calculate the classification error rate of the nth base classifier:

[0097]

[0098] Where M is the total number of samples, D n (i) is the sample weight of the nth base classifier, G n (x i ) is the network prediction category after inputting the i-th sample, real i is the actual sample category of the i-th sample. The error rate is the sum of the weights of the misclassified samples.

[0099] Then calculate the weight of the nth base classifier:

[0100]

[0101] The weight update method of the input base classifier sample in step 5 is as follows:

[0102]

[0103] Where K is the number of sample categories, y i is the sample category label, p n (x i ) is the probability distribution of the network output after the i-th data is input into the network.

[0104] Taking into account the different weights of samples, the loss function of the base classifier is:

[0105]

[0106] Where y i is the sample category label, Output the results for the network.

[0107] Step 6: Repeat step 5 for the two base classifiers until the pseudo loss reaches the requirement, and then weight the final results to form a strong classifier.

[0108] The weighted fusion method is as follows:

[0109] First, combine the base classifiers:

[0110]

[0111] Where G n (x) is the output of the nth base classifier in the first base classifier, α n is the weight of the nth base classifier in a base classifier, G n '(x') is the output of the nth base classifier in the second base classifier, β n is the weight of the nth base classifier in two base classifiers, and N is the total number of base classifiers in one base classifier.

[0112] Then the combined base classifiers are processed by the symbolic function to obtain the final strong classifier:

[0113]

[0114] The specific embodiments are:

[0115] 1: Generate interference data, specifically:

[0116] The required interference data is generated using an interference signal simulator for experimental verification. The radar active interference signal data set constructed by the simulation experiment of the present invention contains 8 types of interference signals: slicing interference, distance multi-false target interference, intermittent sampling forwarding interference, noise amplitude modulation interference, noise frequency modulation interference, sweep frequency interference, smart noise interference and comb spectrum interference. The pulse repetition period is 40μs, the sampling rate is 80MHz, the pulse width of each interference signal except the smart noise interference signal is fixed at 20μs, and the carrier frequency and bandwidth range of each interference signal are [6, 32] MHz and [1, 20] MHz respectively. Other variable parameters are set as follows: the frequency modulation slope range of the noise frequency modulation interference signal is [11, 14] MHz / s; the frequency sweep period range of the swept frequency interference signal is [27, 32] μs; the number of sub-bands of the comb spectrum interference signal is [2, 4]; the number of sub-pulse segments of the sliced ​​interference signal is [4, 5], and the number of time slots is [4, 5]; the real target distance range of the smart noise interference signal is [1000, 2000] m, the pulse width range is [10, 17] μs, and the frequency modulation slope range is [6, 50] MHz / s; the real target distance range of the intermittent sampling forwarding interference signal is [1000, 2000] m; the number of false targets of the distance multiple false target interference signal is [3, 4], and the delay range is [7, 10] μs.

[0117] 2: Data preprocessing, specifically:

[0118] The original signal is cropped and the size is set to 2048. Each class is kept to contain 50 samples and the total number of samples is 400.

[0119] 3: Generate augmented data for adversarial samples, specifically:

[0120] The disturbance data is calculated by FGSM:

[0121]

[0122] Among them, x is the original sample, θ is the weight parameter of the model, which is 1, and real is the actual category of x. Input the original sample, weight parameter and real category, and use the J loss function to calculate the loss value of the neural network. The loss function uses SoftMax. It means to find the partial derivative of x, that is, the loss function J finds the partial derivative of x sample. sign is the sign function. ε is the perturbation ratio, and 8, 16, and 24 are taken to generate three sets of perturbed data respectively.

[0123] 4: Network parameter settings, specifically:

[0124] The input size of the GRU network is 1×2048×2. The convolution layer is a one-dimensional convolution with a convolution kernel size of 1×9×2×8. The four dimensions represent the length of the convolution kernel, the width of the convolution kernel, the number of input channels, and the number of output channels. The step length is 1×8, that is, eight units of length are moved each time in the second dimension of each sample, and the padding method is SAME. The pooling layer uses maximum pooling with a window size of 1×4, a step length of 1×4, and no padding. The output value of the pooling layer enters the fully connected layer. The fully connected layer has two hidden layers with 64 and 256 nodes respectively, and the output is a one-dimensional sequence. The output of the fully connected layer enters the GRU unit, and the output of the GRU unit passes through a fully connected layer to achieve feature fusion, and finally outputs the result through the SoftMax layer.

[0125] The input size of the residual neural network is 224×224×1. The convolution kernel size of the convolution layer is 7×7, and the step size is 2; the convolution layer outputs the extracted features to the residual block, which contains two basic residual blocks. The basic residual block structure is: input features, convolution layer, batch normalization, ReLU activation function, convolution layer, batch normalization, output features. The convolution kernel dimensions of the two convolution layers in the basic residual block are both 3×3, and the step size is 2; the final output of the residual block is input to the average pooling layer, and the output feature map of 1×1×512 is output; the output of the pooling layer is input to the fully connected layer, and finally the result is output through the SoftMax layer.

[0126] 5: Training parameter settings, specifically:

[0127] The loss function used in GRU network training is the cross entropy loss function, the learning rate is 10^-5, the training rounds are 100 times, and the training optimizer is Adam.

[0128] The loss function used in residual neural network training is the cross entropy loss function, the learning rate is 0.001, the training rounds are 100 times, and the training optimizer is Adam.

[0129] 6: Select the integration method, specifically:

[0130] Initialize the sample weight to 1 / M, where M is the total number of samples. Input the original signal and spectrum into the GRU network, and input the time-frequency graph into the residual neural network. Use the Adaboost algorithm for integration, change the sample weights of the input base classifier by calculation, and assign different weights to the base classifier according to the recognition accuracy, and finally perform weighted voting integration on the output results of the base classifier.

[0131] 7: In order to evaluate the model, the accuracy and F1-score are used as evaluation indicators for the designed model and the comparison model. The F1-score is obtained by weighting the precision and recall. The specific calculation formula is as follows:

[0132] Accuracy:

[0133]

[0134] Recall:

[0135]

[0136] Accuracy:

[0137]

[0138] F1-score:

[0139]

[0140] Among them, TP is correctly predicted as a positive example, TN is correctly predicted as a negative example, FP is incorrectly predicted as a positive example, and FN is incorrectly predicted as a negative example.

[0141] 8: Analysis of recognition results, specifically:

[0142] Compared with using residual neural network and GRU network alone, ensemble learning can improve recognition accuracy. After using adversarial sample generation to expand data, the recognition rate is significantly improved. Combining adversarial sample generation and ensemble learning can further improve the recognition rate.

[0143] Table 1 Accuracy comparison (unit: %)

[0144] Feature fusion solution Accuracy F1-score ResNet18 52.50 50.93 GRU 63.72 62.64 ResNet18+FGSM 75.42 74.24 Integrated + FGSM 92.15 91.08

[0145] In summary, the small sample active interference identification method based on adversarial sample generation and ensemble learning proposed in the present invention has higher recognition accuracy than other models both in terms of features and network design.

[0146] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any simple modification, change and equivalent change made to the above embodiment according to the technical essence of the invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A small sample active interference identification method based on adversarial sample generation and ensemble learning, characterized in that: The steps include: Step 1: Obtain samples of different active interference types with labels, generate augmented active interference samples through adversarial samples, and use the original samples and augmented samples to form an active interference dataset; Step 2: Perform fast Fourier transform on the active interference data set to obtain the spectrum of the signal, and perform Cui-Williams distribution transform on the expanded signal to obtain the time-frequency diagram of the signal; Step 3: Construct two base classifier models, where the first base classifier uses a GRU network and the second base classifier uses a residual neural network; both base classifier models are composed of N base classifiers; Step 4: Initialize the sample weight to 1 / M, where M is the total number of samples; input the original signal and spectrum into the GRU network, input the time-frequency graph into the residual neural network, and train the first base classifier in the two base classifier models; Step 5: After the first base classifier is trained, save the trained network, assign weights to each base classifier in the two base classifier models according to the pseudo loss, change the sample weights according to the Adaboost algorithm, and input them into the network for training again; the classifier weights of the two base classifiers are updated in the same way as the sample weights; Step 6: Repeat step 5 for the two base classifiers until the pseudo loss reaches the requirement, and then weight the final results to form a strong classifier. In the above step 2, the signal spectrum and time-frequency diagram are calculated as follows: Perform fast Fourier transform FFT on the signal s(a) in the active interference data set to obtain the spectrum; perform FFT on the signal to obtain the discrete spectrum S(k) of the signal: Where a is a discrete time point, A is the number of sample points, k is a discrete frequency point, and j represents the symbol of the imaginary unit; Then, the signal s(a) in the active interference data set is transformed by Cui-Williams distribution CWD to obtain the time-frequency image of the signal, and preprocessing is performed, wherein the preprocessing includes cropping, graying, and median filtering; the time-frequency image CWD(t, f) of the signal obtained by CWD transformation of each frame of the signal is: Wherein, CWD(t, f) represents the CWD transformed time-frequency diagram of the radar active jammer signal at time t and frequency f, π represents pi, σ represents a positive scaling factor, τ represents the time delay of the radar active jammer signal, e(·) represents an exponential operation with the natural constant e as the base, u represents the local time of the radar active jammer signal, j represents the symbol of the imaginary unit, s(·) represents the time series of the radar active jammer signal, and * represents a conjugate operation; The classifier weights of the first base classifier in step 5 are updated as follows: Calculate the classification error rate of the nth base classifier: Where M is the total number of samples, D n (i) is the sample weight of the nth base classifier, G n (x i ) is the network prediction category after inputting the i-th sample, real i is the actual sample category of the i-th sample; the error rate is the sum of the weights of the misclassified samples; Then calculate the weight of the nth base classifier: The weight update method of the input base classifier sample in step 5 is as follows: Where K is the number of sample categories, y i is the sample category label, p n (x i ) is the probability distribution of the output after the i-th data is input into the base classifier; The base classifiers all use the Adam optimizer. Considering the different weights of samples, the loss function of the base classifier is: in, Output the results for the network.

2. The small sample active interference identification method based on adversarial sample generation and ensemble learning according to claim 1 is characterized in that: In step 1 above, the adversarial sample generation method is selected as follows: Use Gradient-Based Generative Adversarial Examples (FGSM); The disturbance data is calculated by the following formula: Among them, x is the original sample, θ is the weight parameter of the model, which is 1, and real is the actual category of x; input the original sample, weight parameter and real category, and use the J loss function to obtain the loss value of the neural network. The loss function uses SoftMax. It means to find the partial derivative of x, that is, the loss function J finds the partial derivative of the x sample; sign is the sign function; ε is the perturbation ratio, and 8, 16, and 24 are taken to generate three sets of perturbed data respectively.

3. The small sample active interference identification method based on adversarial sample generation and ensemble learning according to claim 2 is characterized in that: The weighted fusion method in step 6 is as follows: First, combine the base classifiers: Where G n (x) is the output of the nth base classifier in the first base classifier, α n is the weight of the nth base classifier in a base classifier, G n ′(x′) is the output of the nth base classifier in the second base classifier, β n is the weight of the nth base classifier in the second base classifier, and N is the total number of base classifiers in each base classifier; Then the combined base classifiers are processed by the symbolic function to obtain the final strong classifier:

Images