A countermeasure enhancement method based on radio signal classification
By introducing adversarial training methods into radio signal classification, boundary samples and strong adversarial samples are generated, which solves the problems of insufficient sample quantity and adversarial sample attacks, improves the model's recognition accuracy and robustness, and enhances the model's defense performance.
Patent Information
- Application Number
- CN202011451030.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2020-12-09
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2040-12-09
AI Technical Summary
Existing technologies for radio signal identification suffer from insufficient sample quantity and adversarial sample attacks, resulting in poor training performance of deep learning models. Furthermore, existing data augmentation methods have limited application in the signal field and are costly in terms of manpower and resources.
By introducing the concept of adversarial training, the residual neural network 1D-RESNET is used for pre-training, and subtle perturbations designed by the algorithm are added to the samples to generate boundary samples and strong adversarial samples. Combined with the decision boundary surface of the classification model, the model is enhanced.
It improves the recognition accuracy and robustness of the radio signal classification model, enhances the model's defensive performance, and balances classification accuracy with time complexity.
Smart Images

Figure CN112418347B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of signal enhancement, and in particular to a countermeasure enhancement method based on radio signal classification. Background Art
[0002] At present, radio signal recognition is widely used in the field of communications, including interference identification, spectrum sensing, and electronic countermeasures. The modulation recognition problem of signals is essentially a typical pattern recognition problem, the purpose of which is to identify the modulation type of the detected signal. Early signal recognition methods mainly used artificially designed features to make signal judgments, and the accuracy could not be guaranteed. Deep learning technology has the ability to automatically learn "features" from data, which provides an effective method to solve the problem of signal recognition. More and more scholars use deep learning technology to study signal modulation pattern recognition. However, in the field of deep learning, most models require a large amount of data for training and learning. With the emergence of more application scenarios, we are increasingly facing the problem of insufficient sample quantity. In signal recognition, the typical challenge is that it is real in various environments, and it is difficult to label data, so the sample set is limited, and it is not possible to establish a reliable database well. The sample set will directly affect the training effect.
[0003] In addition, with the resurgence of the artificial intelligence wave in recent years, researchers have found that deep neural networks are vulnerable to adversarial samples. Attackers can generate adversarial samples by adding specific perturbations to benign data to make the model wrong. Although these forged samples have no effect on human judgment, they are a fatal misleading for deep learning models. Recently, a series of adversarial attacks successfully implemented in the image field have proved that this problem is a security risk for all deep learning-based systems. Image data is generally two-dimensional data, while signal data is generally one-dimensional time series data. Although their essential meanings are very different, their expressions are similar. Therefore, the attack methods in the image field can be applied to the radio signal field, and the research on adversarial defense technology has attracted more and more attention from scholars in the communication field. In the prior art, the patent application number 201910694936.1 discloses a small sample radio signal enhancement recognition method based on ACGAN. The method first classifies the radio signal through the Long Short Term Memory (LSTM) network, and further uses the auxiliary classifier enerative adversarial network (ACGAN) to generate data to expand the small sample data set, and then enhances the recognition model to complete the small sample radio signal enhancement recognition. This technology is applied to radio signal data enhancement technology, but the training of ACGAN is very unstable and often requires a lot of training skills. Secondly, for the work of data enhancement, there are relatively complete analysis methods in the image field, but there are very few applications in the signal field. Based on the prior knowledge of radio signals, Huang L et al. appropriately transform the original data while maintaining the known information to achieve the effect of expanding the data set. Specifically for the classification task, each sample in the training set is randomly rotated and flipped left and right to a certain extent while keeping the data category unchanged. However, these transformations correspond to the observation results of the same target at different angles, and the enhancement effect is limited. Subsequently, Tang et al. preprocessed the radio signal data set and converted it into a constellation diagram that can represent the density information of sample points. They further expanded the data set by generating labeled constellation diagram samples through a generative adversarial network and completed the enhanced recognition of radio signal modulation types. However, the training of generative adversarial networks is very unstable, and preprocessing the data into a special constellation diagram requires professional knowledge and consumes a lot of manpower and material resources. Summary of the invention
[0004] In order to solve the above technical problems, an adversarial enhancement method based on radio signal classification is provided. By introducing the idea of adversarial training and controlling the EPS and iteration parameters, subtle perturbations carefully designed by the algorithm are added to the samples to generate boundary samples. The boundary samples are mixed with the original training samples, thereby enhancing the model's defense capability and improving the model's classification accuracy.
[0005] The specific steps include:
[0006] S1: preprocessing the radio signal as a sample data set, and dividing the sample data set;
[0007] S2: Pre-train the classification model using the divided sample data set;
[0008] S3: Attacking and calculating the pre-trained sample data set to obtain adversarial samples; based on the decision boundary surface of the classification model, introducing slight perturbations on the adversarial samples to obtain boundary samples and strong adversarial samples;
[0009] S4: Screening the boundary samples and strong adversarial samples;
[0010] S5: Fuse the filtered boundary samples and strong adversarial samples, retrain the classification model, and obtain an enhanced classification model;
[0011] S6: Use the enhanced recognition model to recognize radio signals, and complete the debugging recognition.
[0012] Preferably, the step S1 specifically comprises:
[0013] S11: taking a number of radio signals as sample data to obtain an electrical signal sample data set;
[0014] S12: performing maximum and minimum normalization processing on the electrical signal sample data set;
[0015] S13: Divide the preprocessed data into a training set and a test set in a ratio of 1:2.
[0016] Preferably, the classification model uses a residual neural network 1D-RESNET to classify the electrical signal modulation type.
[0017] Preferably, the step 3 is specifically:
[0018] S31: Performing a targeted attack on the sample data set and specifying a target label;
[0019] S32: Perform multiple iterative calculations on the sample data set after the attack to obtain adversarial samples;
[0020] S33: introducing a smaller coefficient and controlling the number of iterations before iteration, and adding perturbations to the adversarial sample;
[0021] S34: Based on the decision boundary surface of the classification model, the sample that successfully attacks the target label is called a weak adversarial sample, and the sample of the previous iteration of the adversarial sample is called a boundary sample.
[0022] Preferably, the target attack is performed using a BIM attack algorithm.
[0023] Preferably, the iterative method is calculated by a clip function, and the specific iterative method is as follows:
[0024] X * 0=X (1)
[0025]
[0026] Among them, X is the original sample, and the adversarial sample X * It is a synthesis of the original sample and noise, Y is the label corresponding to the original sample, sign() is the sign function, and e is the introduced coefficient eps.
[0027] Preferably, the smaller coefficient is an eps parameter, and the number of iterations is controlled by introducing a parameter Iteration for controlling the number of iterations.
[0028] Preferably, S4 uses the distance L2 between each type of samples to perform screening. Specifically, L2 is used as the natural metric difference between the original sample prediction vector and the generated sample prediction vector to perform screening. The expression is:
[0029]
[0030] Among them, g(v) is the output vector of the classification model generated by the softmax layer, V oi is the value of the i-th data of the original sample, V ai is the value of the i-th data generated by the sample.
[0031] Compared with the prior art, the present invention has the following beneficial effects:
[0032] (1) The present invention compares the signal classification models based on deep learning in recent years, balances the classification accuracy and time complexity, uses the 1D-ResNet signal classification model, introduces the adversarial training method into the signal field, introduces the eps and iteration parameters, adds the algorithm-designed subtle perturbations to the samples to generate boundary samples, realizes data enhancement, and improves the recognition accuracy of the model;
[0033] (2) The present invention puts the filtered boundary samples that are close to but do not cross the decision boundary into the model for training together with the original samples to enhance the robustness of the model. The gradient attack method is used to generate strong adversarial samples to attack the models before and after enhancement to verify the defensive performance of the adversarial enhancement method. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0035] Figure 1 The flowchart of the countermeasure enhancement method based on radio signal classification of the present invention;
[0036] Figure 2 The overall architecture diagram of the countermeasure enhancement method based on radio signal classification of the present invention;
[0037] Figure 3 This is a residual stack unit framework diagram of the present invention;
[0038] Figure 4 It is the network framework of the 1D-ResNet classification model of the present invention;
[0039] Figure 5 This is a waveform comparison diagram of the original sample and the generated sample of the present invention. DETAILED DESCRIPTION
[0040] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0041] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0042] Reference Figures 1-2 As shown, the present invention proposes a countermeasure enhancement method based on radio signal classification, which specifically includes the following steps:
[0043] S1: Use radio signals as sample data, preprocess the data set, and divide the training set and test set into a ratio of 2:1;
[0044] First, the present invention uses two radio signal data sets for evaluation. The data set uses the modulation signal data set RML2016.10a published by the University of Edinburgh. It uses GNU Radio to synthesize signal samples. It contains 11 modulation types, and the signal-to-noise ratio (SNR) ranges from -20dB to 18dB, and is evenly distributed at intervals of 2dB. The size of each radio signal sample is 128x2, the number of training set samples is 176,000, and the number of test set samples is 44,000. The data set is generated by the team simulation Sig2019-12, and some effects of a real communication system are considered during the simulation. It contains 12 modulation types: BPSK, QPSK, 8PSK, OQPSK, 2FSK, 4FSK, 8FSK, 16QAM, 32QAM, 64QAM, 4PAM and 8PAM. The radio signal data is generated in a random manner to ensure that the transmission bits have equal probability values. The pulse shaping filter uses a raised cosine filter, and the roll-off coefficient is randomly selected in the range of [0.2, 0.7]. The phase deviation is randomly selected in the range of [-π, π], and the normalized carrier frequency deviation (relative to the sampling frequency) is randomly selected in the range of [-0.1, 0.1]. The SNR range is from -20dB to 30dB, and is evenly distributed at intervals of 2dB. Each data sample contains 64 symbols, and the number of sampling points for each symbol is 8, so the number of sampling points for each sample is 512. The sizes of the training set and test set are 312,000 and 156,000 respectively, and the number of samples for each type of modulation signal is the same;
[0045] Then, the signal data set is preprocessed by maximum and minimum normalization, and the signal data is divided into a training set and a test set. Each modulation type with the same signal-to-noise ratio has only 1000 training set samples.
[0046] S2: Use the training set to pre-train the 1D-RESNET classification model;
[0047] The modulation type of the signal is classified using the residual neural network 1D-RESNET. Figure 3 As shown in Figure 1, based on the residual block, a residual stack unit is constructed. Each residual stack unit includes a convolution layer with a convolution kernel size of 1×1 for calculation in the channel dimension, two residual blocks, and a maximum pooling layer. Figure 4As shown in the figure, the entire classification model contains 6 residual stack units. The dimension of the data is halved after each residual stack unit, and the number of convolution kernels in all convolutional layers in the model is 32. Self-normalized neural network, scaled exponential linear unit (SELU) activation function, mean response scaling initialization (MRSA) and Alpha Dropout are used in the fully connected layer of the network. First, the model is trained using the entire training set, and then the recognition accuracy is given on the test samples of each SNR.
[0048] S3: According to the existing model decision boundary, by introducing eps and iteration parameters, the BIM gradient attack method is used to attack the pre-trained model to generate boundary samples and strong adversarial samples;
[0049] The present invention performs a targeted attack on the sample, and specifies the target label as the second label in the probability distribution. The invention finally selects the BIM attack algorithm to generate boundary samples. BIM is a small-step iterative attack method that can generate adversarial samples of any infinite norm distance by controlling the hyperparameter ε. In each iteration, in order to avoid large changes, the clip function is used for calculation, and adversarial samples are obtained after multiple iterations.
[0050] Usually, when adversarial samples are used as an expanded data set for training, that is, adversarial training, while enhancing the model's anti-interference ability, it will reduce the model's classification accuracy. In order to make the generated samples closer to the decision boundary of the original model and obtain samples that are more effective for the enhanced model, the present invention adds a smaller coefficient (eps, 0.004~0.006) before the disturbance to reduce the disturbance, and then adds disturbances to the input samples in a directional manner through iteration. In order to avoid generating a sample for too long, a parameter (Iteration, 0~60) is introduced to control the number of iterations. The iterative method is as follows:
[0051] X * 0=X (1)
[0052]
[0053] Among them, X is the original sample, and adversarial sample X * It is a synthesis of the original sample and noise, Y is the label corresponding to the original sample, sign() is the sign function, and e is the introduced coefficient eps.
[0054] In combination with the decision boundary of the classification model, the present invention proposes the concepts of boundary samples, weak adversarial samples, and strong adversarial samples. The present invention refers to samples that successfully attack the target label as weak adversarial samples, and samples of the previous iteration of the adversarial samples as boundary samples. Boundary samples are very close to the decision boundary of the task but do not cross the decision boundary, which can effectively enhance the model while also increasing the generalization ability of the model.
[0055] Strong adversarial samples are generated in the same way as boundary samples. The BIM attack algorithm is used to generate samples, but the eps parameter is not used. Only the Iteration parameter is used to control the sample generation time. At the same time, in order to avoid adding too much perturbation, the perturbation is limited to 5%.
[0056] S4: Through the sample screening mechanism, the boundary sample set and the strong adversarial sample set are saved respectively;
[0057] The boundary samples need to be consistent with the original sample labels. In rare cases, the previous sample may be attacked to other boundaries, and then attacked to the target category boundary for the last time. Therefore, it is necessary to delete the samples whose predicted labels of boundary samples are different from the actual labels.
[0058] For strong adversarial samples, due to the multi-classification problem and the non-uniqueness of the boundaries, it is necessary to ensure that the weak adversarial samples and the strong adversarial samples have the same prediction label.
[0059] S4.3: Calculate the L2 distance between samples:
[0060] In order to ensure the similarity of features with the original data set, the waveforms of the original samples and the generated samples are almost indistinguishable to the naked eye. In order to verify that the distance of each type of sample has a certain degree of distinction, after preliminary screening, the L2 distance calculation is performed on the four types of samples to determine whether the generated samples meet the requirements. The present invention selects the L2 norm as a natural measure of the difference between the original sample prediction vector and the generated sample prediction vector:
[0061]
[0062] Where g(v) is the output vector of the classification model generated by the softmax layer. The higher the L2 score, the greater the difference between the original sample prediction and the generated sample. oi is the value of the i-th data of the original sample, V ai is the value of the i-th data generated by the sample.
[0063] S5: Mix boundary samples with training samples and retrain the classification model;
[0064] The boundary samples are mixed with the original data set, and the 1D-ResNet classification model is retrained to obtain the classification accuracy of enhanced recognition. In order to ensure fairness, the network structure, test set, learning rate, training times and other parameters of the classification model before and after enhancement will remain consistent.
[0065] Reference Figure 5As shown in the figure, after the model training is completed, the strong adversarial sample is used to conduct the adversarial sample attack experiment. The original model and the enhanced recognition model are tested respectively using the strong adversarial sample, and the recognition accuracy of the two models after the adversarial sample attacks is compared, which verifies the defense performance of the present invention.
[0066] S6: Using the enhanced recognition model to recognize radio signals, and completing debugging recognition.
[0067] Radio signals of different modulation types are input into the enhanced recognition model to complete signal recognition of different modulation types, and then complete tasks such as interference identification, spectrum perception and electronic countermeasures.
[0068] In summary, the present invention compares the signal classification models based on deep learning in recent years, balances the classification accuracy and time complexity, uses the 1D-ResNet signal classification model, and introduces the adversarial training method into the signal field, introduces the eps and iteration parameters, and adds the algorithm to the sample to generate boundary samples with subtle perturbations carefully designed to achieve data enhancement and improve the recognition accuracy of the model; the boundary samples that are close to but do not cross the decision boundary after screening are put into the model together with the original samples for training to enhance the robustness of the model. The gradient attack method is used to generate strong adversarial samples to attack the models before and after enhancement, and verify the defensive performance of the adversarial enhancement method.
[0069] The embodiments described above are only descriptions of the preferred modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.
Claims
1. An adversarial enhancement method based on radio signal classification, characterized in that: The specific steps include: S1: pre-processing the radio signal as a sample data set, and dividing the sample data set; specifically: S11: taking a number of radio signals as sample data to obtain an electrical signal sample data set; S12: performing maximum and minimum normalization processing on the electrical signal sample data set; S13: Divide the preprocessed data into a training set and a test set in a ratio of 1:2; S2: Pre-train the classification model using the divided sample data set; S3: Attack and calculate the pre-trained sample data set to obtain adversarial samples; based on the decision boundary surface of the classification model, introduce slight perturbations on the adversarial samples to obtain boundary samples and strong adversarial samples; specifically: S31: Performing a targeted attack on the sample data set and specifying a target label; S32: Perform multiple iterative calculations on the sample data set after the attack to obtain adversarial samples; S33: introducing the eps parameter and controlling the number of iterations before iteration, and adding perturbations to the adversarial sample; S34: Based on the decision boundary surface of the classification model, the sample whose attack is successfully targeted is called a weak adversarial sample, and the sample of the previous iteration of the adversarial sample is called a boundary sample; S4: Screening the boundary samples and strong adversarial samples; screening is performed by using the different distances L2 between each type of samples. Specifically, L2 is used as a natural measure of the difference between the original sample prediction vector and the generated sample prediction vector. The expression is: Among them, g(v) is the output vector of the classification model generated by the softmax layer, V oi is the value of the i-th data of the original sample, V ai To generate the value of the i-th data of the sample; S5: Fuse the filtered boundary samples and strong adversarial samples, retrain the classification model, and obtain an enhanced classification model; S6: Use the enhanced recognition model to recognize radio signals, and complete the debugging recognition.
2. The method for countermeasure enhancement based on radio signal classification according to claim 1, characterized in that: The classification model uses a residual neural network 1D-RESNET to classify the electrical signal modulation type.
3. The method for countermeasure enhancement based on radio signal classification according to claim 1, characterized in that: The target attack is carried out using a BIM attack algorithm.
4. The method for countermeasure enhancement based on radio signal classification according to claim 1, characterized in that: The iterative method is calculated by the clip function, and the specific iterative method is as follows: X * 0=X (1) Among them, X is the original sample, and adversarial sample X * It is a synthesis of the original sample and noise, Y is the label corresponding to the original sample, sign() is the sign function, and e is the introduced coefficient eps.
5. The method for countermeasure enhancement based on radio signal classification according to claim 1, characterized in that: The control of the number of iterations is achieved by introducing a parameter Iteration for controlling the number of iterations.
