A signal modulation recognition method based on self-paced self-distillation adversarial training

Abstract

The present invention discloses a signal modulation recognition method based on self-paced self-distillation adversarial training. The method uses ResNet18 as the basic network of the modulation recognition model to construct the modulation recognition model. The method defines a signal-to-interference ratio (SIR) to quantify the adversarial attack strength of communication signal samples, divides the training steps of the self-paced learning, and sets the number of training rounds for each training step to initiate the modulation recognition model training process. At each training step, an adversarial sample under the current modulation recognition model is generated for each communication signal sample in the training set. A model feedback mechanism based on loss quantiles is used to evaluate the difficulty of the adversarial sample for the communication signal and assign different weights. A total loss function is constructed based on weighted cross-entropy loss and self-distillation loss. The trained signal modulation recognition model is fed with a communication signal under adversarial attack conditions to perform modulation recognition. The present invention improves the accuracy of signal modulation recognition under adversarial attacks.

Description

Technical Field

[0001] The present invention relates to the field of communication signal modulation mode recognition, and in particular to a signal modulation mode recognition method based on self-paced self-distillation adversarial training. Background Art

[0002] Signal modulation recognition is a core technology in wireless communications, enabling intelligent analysis and efficient utilization of communication signals. In recent years, breakthroughs in deep learning technology have revolutionized modulation recognition. Using models such as convolutional neural networks and recurrent neural networks, the system can directly learn the deep characteristics of the modulation pattern from the original signal or time-frequency features, significantly improving recognition accuracy and robustness. However, research has shown that intelligent recognition models based on deep learning are vulnerable to malicious spoofing scenarios. Adding carefully designed, tiny perturbations to communication signal samples can deceive the model, causing it to incorrectly identify the signal's modulation mode. These samples are called adversarial examples, and the techniques used are known as adversarial attacks. Under signal adversarial attacks, modulation recognition accuracy decreases significantly, posing a serious threat to communication system security.

[0003] In complex electromagnetic environments, adversarial attacks pose a serious threat to the stability of traditional recognition models. Adversarial training has become a mainstream solution for improving the adversarial defense capabilities of modulation recognition models. This approach injects adversarial examples into the training process, forcing the model to learn robust features, thereby enhancing its recognition performance in adversarial environments. For example, the paper (A Madry, A Makelov, L Schmidt, et al. Towards deep learning models resistant to adversarial attacks. arXiv preprint arXiv:1706.06083, 2017.) proposed the classic projected gradient descent-based adversarial training (PAT) method, which significantly improved the model's robust recognition performance. The paper (H Zhang, Y Yu, J Jiao, et al. Theoretically principled trade-off between robustness and accuracy. International conference on machine learning. PMLR, 2019: 7472-7482.) proposed TRADES, an adversarial training method based on Kullback-Leibler (KL) divergence, which effectively counters various adversarial attacks and further fueled the research boom in adversarial training. The paper (Z Chen, Z Wang, D Xu, et al. Learn to defend: adversarial multi-distillation for automatic modulation recognition models. IEEE Transactions on Information Forensics and Security, 2024, 19: 3690-3702.) proposed AMD, an adversarial training method that utilizes two powerful teacher models to transfer knowledge about normal signal classification and adversarial signal defense to the student model, respectively, thereby improving the robustness of signal modulation recognition. Although existing signal modulation recognition methods based on adversarial training can improve the robust recognition performance of the model under adversarial attacks to a certain extent, they still face the following two major challenges in more complex and dynamic adversarial environments: First, the generalization ability to attacks of different intensities and types is limited. Traditional adversarial training is usually optimized for attacks of fixed intensity and fixed type, and it is difficult to adapt to complex attacks with sudden changes in intensity in communication environments.Second, there is a lack of a stable feature representation mechanism. As the attack intensity increases, the feature space of the signal modulation recognition model is prone to shift. There is a lack of effective feature constraints during the training process, and robustness is difficult to maintain stably. Summary of the Invention

[0004] The purpose of the present invention is to propose a signal modulation mode recognition method based on self-paced self-distillation adversarial training.

[0005] The technical solution to achieve the purpose of the present invention is: a signal modulation mode recognition method based on self-paced self-distillation adversarial training, the specific steps are as follows:

[0006] Step 1: Obtain a communication signal modulation mode dataset, including modulation mode category labels and communication signal samples, and divide the modulation signal into a training set and a validation set;

[0007] Step 2: Use ResNet18 as the basic network of the modulation recognition model, take the communication signal sample as input and the modulation category prediction as output to build the modulation recognition model;

[0008] Step 3: Define the signal-to-interference ratio to quantify the adversarial attack strength of the communication signal sample, divide the training steps of the self-paced learning, set the training rounds for each training step, and start the modulation mode recognition model training process;

[0009] Step 4: At each training step, an adversarial sample is generated for each communication signal sample in the training set under the current modulation recognition model. Based on the model feedback mechanism of the loss quantile, the difficulty of the adversarial sample of the communication signal is evaluated and assigned different weights for calculating the weighted cross entropy loss.

[0010] Step 5: Based on the weighted cross entropy loss and self-distillation loss, a total loss function is constructed for iterative training. In each training round, the adversarial samples of the communication signals in the training set are used to update the model parameters. The recognition accuracy of the original samples in the validation set is simultaneously calculated and the optimal model is saved. After completing all rounds of the current training step, the training step update is promoted, the proportion of easy samples and the minimum weight of difficult samples are updated, and the training is continued in step 4 until all training steps are completed.

[0011] Step 6: Input the communication signal under the anti-attack condition into the trained signal modulation mode recognition model to perform modulation mode recognition.

[0012] Furthermore, in step 1, a communication signal modulation mode dataset is obtained, including modulation mode category labels and various time domain IQ signal samples, and the modulation signal training set and validation set are divided. The specific method is as follows:

[0013] Set the category labels to , K represents the number of categories. For each communication signal sample, its I and Q signals are spliced ​​into The first row represents the I signal and the second row represents the Q signal. Indicates the number of sampling points of the communication signal samples, and divides the communication signal modulation mode training set and validation set into a ratio of 9:1.

[0014] Furthermore, in step 2, ResNet18 is used as the basic network of the modulation mode recognition model, with communication signal samples as input and modulation mode category prediction as output, to build a modulation mode recognition model. The specific method is as follows:

[0015] ResNet18 is used as the basic network of the modulation recognition model. For the training set, validation set, and subsequent test samples, in order to adapt to the input requirements of the ResNet18 neural network, a channel dimension with a value of 1 is added to reduce the size to The communication signal samples are reconstructed into three-dimensional , according to the batch size , stack multiple communication signal samples into a size of The network input data is , and the network output is the predicted probability of belonging to each category. The output size is , take the category with the highest probability as the signal category predicted by the modulation recognition model.

[0016] Furthermore, in step 3, the signal-to-interference ratio is defined to quantify the adversarial attack strength of the communication signal sample and to divide the training steps of the self-paced learning. The specific method is as follows:

[0017] The signal-to-interference ratio SPR is defined as:

[0018] (1)

[0019] (2)

[0020] (3)

[0021] in, and Respectively represent the communication signal I signal and Q signal sampling points; The smaller the signal-to-perturbation ratio (SPR), the greater the attack strength and the higher the difficulty of the adversarial sample.

[0022] Pre-set minimum SPR value and maximum SPR value To determine the range of SPR values, divide it into equal intervals. Discretized into training steps, each training step Corresponding to a fixed attack strength , which is used to gradually improve the adversarial attack strength of adversarial samples of communication signals and guide the modulation recognition model to gradually adapt from weak attacks to strong attacks.

[0023] Furthermore, in step 4, at each training step, an adversarial sample under the current modulation recognition model is generated for each communication signal sample in the training set. Based on the model feedback mechanism of the loss quantile, the difficulty of the adversarial sample of the communication signal is evaluated and assigned different weights. The specific method is as follows:

[0024] Step 4.1: At each training step, generate an adversarial sample under the current modulation recognition model for each communication signal sample in the training set;

[0025] In the training step , for any index number in the training set Communication signal samples , applying an attack strength of PGD ​​adversarial attack, according to The adversarial perturbation size for each training step , generate the number of iterations n PGD Corresponding adversarial examples The calculation formula is shown in formula (4):

[0026] (4)

[0027] Where, represents uniform distribution, The sample index number is The category label, = is the single-step perturbation step length, represents the symbolic function, is the cross entropy loss function, Crop the adversarial samples to Centered on The final adversarial sample is within the L2 norm ball with a radius of ;

[0028] Step 4.2: Based on the loss quantile model feedback mechanism, the difficulty of adversarial examples of communication signals is evaluated and assigned different weights;

[0029] Calculate the cross entropy loss value of all generated adversarial samples on the current modulation recognition model ,in is the number of training set samples;

[0030] use Indicates the The ratio of easy samples in the training step is determined The loss quantile threshold for training steps :

[0031] (5)

[0032] Where, Express The loss value set of adversarial examples Calculate the quantile threshold, sort the loss values ​​in ascending order, and take the value as the loss quantile threshold , so that at least The sample loss value of the proportion does not exceed , is the ceiling function;

[0033] Introducing the current The minimum weight of difficult samples in the training step , a linear weight descent strategy is used to weight samples: simple samples are given a weight of 1.0, and the weight of difficult samples is between 1.0 and The interpolation is uniform between them, and the loss decreases linearly in the order from small to large. The index number of the training step is The weight of the adversarial sample The calculation formula is expressed as:

[0034] (6).

[0035] Furthermore, in step 5, a total loss function is constructed based on the weighted cross entropy loss and the self-distillation loss for iterative training. In each training round, the adversarial samples of the communication signal in the training set are used to update the model parameters, and the recognition accuracy of the original samples in the validation set is simultaneously calculated and the optimal model is saved. After completing all rounds of the current training step, the training step update is promoted, the proportion of easy samples and the minimum weight of difficult samples are updated, and the training is continued by jumping to step 4 until all training steps are completed. The total loss function is constructed based on the weighted cross entropy loss and the self-distillation loss. The specific method is:

[0036] (1) Weighted cross entropy loss

[0037] In each training iteration, the adversarial examples of the communication signal under the current modulation recognition model are regenerated and the weighted cross entropy loss function is calculated;

[0038] In the In each training iteration of a training step, let the index set of the current batch training set samples be According to formulas (1)-(3), the attack intensity applied on the basis of the current modulation mode recognition model is PGD ​​counterattack, the index number is Adversarial examples ;

[0039] Based on the weight formula (6), the weighted cross entropy loss under the current modulation recognition model is calculated:

[0040] (7)

[0041] Where, is the batch size;

[0042] (2) Self-distillation loss

[0043] A teacher model is constructed based on the EMA strategy, and its parameters are obtained by smoothing and updating the historical state of the current modulation mode recognition model;

[0044] The parameters of the EMA teacher model are recorded as , which changes with the current model parameters at each training iteration Update according to the following rules:

[0045] (8)

[0046] Among them, the initial , is the attenuation coefficient, which is used to control the retention ratio of historical information;

[0047] In each training iteration, the current modulation recognition model is regarded as the student model, and the index number is The output logits of the adversarial sample are expressed as , the EMA teacher model output is , then the self-distillation loss is defined as the KL divergence between the two output distributions:

[0048] (9)

[0049] (3) Total loss function

[0050] Total loss By weighted cross entropy loss and self-distillation losses Joint composition:

[0051] (10)

[0052] in, is the self-distillation coefficient, which is used to balance the two losses.

[0053] Furthermore, in step 5, a total loss function is constructed based on weighted cross entropy loss and self-distillation loss for iterative training. In each training round, the adversarial samples of the communication signal of the training set are used to update the model parameters, and the recognition accuracy of the original samples of the validation set is simultaneously calculated and the optimal model is saved. After completing all rounds of the current training step, the training step update is promoted, the proportion of easy samples and the minimum weight of difficult samples are updated, and the training is continued in step 4 until all training steps are completed, where: the proportion of easy samples is updated. and the minimum weight of difficult samples , the specific method is:

[0054] (11)

[0055] (12)

[0056] in, is the maximum SPR value, Initial value of easy sample ratio Start with step length gradually decrease; Initialized by the lowest weight of the difficult sample Start with step length Gradually increase.

[0057] A signal modulation mode recognition system based on self-paced self-distillation adversarial training implements the signal modulation mode recognition method based on self-paced self-distillation adversarial training to achieve signal modulation mode recognition based on self-paced self-distillation adversarial training, and performs steps 1 to 6 respectively in six modules.

[0058] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for signal modulation mode recognition based on self-paced self-distillation adversarial training is implemented to achieve signal modulation mode recognition based on self-paced self-distillation adversarial training.

[0059] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the method for signal modulation mode recognition based on self-paced self-distillation adversarial training is implemented to achieve signal modulation mode recognition based on self-paced self-distillation adversarial training.

[0060] Compared with the existing technology, the present invention has the following significant advantages: 1) It has good generalization for adversarial samples of communication signals with different attack intensities and attack types, significantly improving the accuracy of signal modulation mode recognition under adversarial attacks; 2) It provides a new adversarial training paradigm under adversarial attacks, which is suitable for various modulation mode recognition models. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a flow chart of the signal modulation mode recognition method based on self-paced self-distillation adversarial training of the present invention.

[0062] Figure 2 This is a graph showing the recognition accuracy of the four adversarial training methods under PGD adversarial attack conditions with different SPRs when the signal-to-noise ratio is 10 dB.

[0063] Figure 3 This is a graph showing the recognition accuracy of the four adversarial training methods under FGSM adversarial attack conditions with different SPRs when the signal-to-noise ratio is 10 dB. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0065] like Figure 1 As shown, the present invention provides a signal modulation mode recognition method based on self-paced self-distillation adversarial training, and the specific steps are as follows:

[0066] Step 1: Obtain a communication signal modulation mode dataset, including modulation mode category labels and communication signal samples, and divide the modulation signal into a training set and a validation set;

[0067] Set the category labels to , K represents the number of categories. For each communication signal sample, its I and Q signals are spliced ​​into The first row represents the I signal and the second row represents the Q signal. Indicates the number of sampling points of the communication signal sample. The communication signal modulation mode training set and validation set are divided into a ratio of 9:1.

[0068] Step 2: Use ResNet18 as the basic network of the modulation recognition model, take the communication signal sample as input and the modulation category prediction as output to build the modulation recognition model;

[0069] For the training set, validation set, and subsequent test samples, in order to adapt to the input requirements of the ResNet18 neural network, a channel dimension (value 1) is added to the training set, validation set, and subsequent test samples. The communication signal samples are reconstructed into three-dimensional , according to the batch size , stack multiple communication signal samples into a size of The network input data is , and the network output is the predicted probability of belonging to each category. The output size is , take the category with the highest probability as the signal category predicted by the modulation recognition model.

[0070] Step 3: Define the signal-to-interference ratio to quantify the adversarial attack strength of the communication signal sample, divide the training steps (pace) of the self-paced learning, set the training rounds for each training step, and start the modulation recognition model training process;

[0071] To quantify the strength of adversarial attacks, the definition of Signal-to-Perturbation Ratio (SPR) is adopted, namely:

[0072] (1)

[0073] (2)

[0074] (3)

[0075] in, and Respectively represent the I-channel signal and the Q-channel signal of the communication signal sampling points; The smaller the signal-to-perturbation ratio (SPR), the greater the attack intensity and the higher the difficulty of the adversarial sample.

[0076] Pre-set minimum SPR value and maximum SPR value To determine the range of SPR values. Discretized into training steps, each training step Corresponding to a fixed attack strength , which is used to gradually improve the adversarial attack strength of adversarial samples of communication signals and guide the modulation recognition model to gradually adapt from weak attacks to strong attacks.

[0077] It should be noted that the training step here refers to each stage of self-paced learning, rather than the training round (epoch) or iteration (iteration) in the traditional machine learning field. Each training step training In the specific implementation of the present invention, SPR takes integer values ​​in the range of [10, 20] dB, starting from 20 dB and gradually decreasing to 10 dB with a step size of 1 dB, with a total of 11 training steps. The number of training rounds for each training step is Set it to 10. Therefore, a total of 110 epochs are trained. Start the modulation recognition model training process and set the following training parameters: learning rate is 0.001, optimizer is Adam, and batch size of training data and validation data is 128.

[0078] Step 4: At each training step, an adversarial sample under the current modulation recognition model is generated for each communication signal sample in the training set. A model feedback mechanism based on loss quantiles is proposed to evaluate the difficulty of the adversarial sample of the communication signal and assign different weights to them.

[0079] Step 4.1, in the training step , for any index number in the training set Communication signal samples , applying an attack strength of PGD ​​counterattack, according to formula (1)-(3), we can calculate the first The adversarial perturbation size for each training step , then the number of iterations is generated PGD ​​adversarial examples The calculation formula is shown in formula (4).

[0080] (4)

[0081] Where, represents uniform distribution, The sample index number is The category label, is the single-step perturbation step size (usually set to ), represents the symbolic function, is the cross entropy loss function, Crop the adversarial samples to Centered on The final adversarial sample is .

[0082] Note that at this time, the adversarial samples generated by the original communication signal samples are only used to calculate the weights and do not participate in the training of the subsequent modulation recognition model.

[0083] Step 4.2: Based on the loss quantile model feedback mechanism, the difficulty of adversarial examples of communication signals is evaluated and assigned different weights;

[0084] Calculate the cross entropy loss value of all generated adversarial samples of communication signals on the current modulation recognition model ,in is the number of samples in the training set. Note that the cross entropy loss calculated at this time is only used to calculate the weights and does not participate in the back propagation of the modulation recognition model gradient.

[0085] Step 4.3, use Indicates the The ratio of easy samples in the training step is determined The loss quantile threshold for training steps :

[0086] (5)

[0087] Where, Express The loss value set of adversarial examples Calculate the quantile threshold: Arrange the loss values ​​in ascending order and take the value as the loss quantile threshold , so that at least The proportion of sample loss values ​​does not exceed the threshold ( is the ceiling function);

[0088] According to the threshold, the adversarial samples of the communication signal are grouped and the current The minimum weight of difficult samples in the training step , a linear weight descent strategy is used to weight samples: simple samples are given a weight of 1.0, and the weight of difficult samples is between 1.0 and The interpolation is uniform between them, and the loss decreases linearly from small to large. Then the index number of the training step is The weight of the adversarial sample The calculation formula can be expressed as:

[0089] (6)

[0090] Step 5: Based on the weighted cross entropy loss and self-distillation loss, a total loss function is constructed for iterative training. In each training round, the adversarial samples of the communication signals in the training set are used to update the model parameters. The recognition accuracy of the original samples in the validation set is simultaneously calculated and the optimal model is saved. After completing all rounds of the current training step, the training step update is promoted, the proportion of easy samples and the minimum weight of difficult samples are updated, and the training is continued in step 4 until all training steps are completed.

[0091] In the In each training iteration of the training step, the weighted cross entropy loss function is calculated. At the same time, a self-distillation mechanism based on the exponential moving average (EMA) is constructed to calculate the self-distillation loss. Finally, the total loss is obtained, the modulation recognition model is trained, and the modulation recognition model with the best recognition accuracy on the validation set is saved.

[0092] (1) Weighted cross entropy loss

[0093] Regenerate adversarial examples under the current modulation recognition model in each training iteration and calculate the weighted cross entropy loss function;

[0094] In the In each training iteration of a training step, let the index set of the current batch training set samples be According to formulas (1)-(3), the attack intensity applied on the basis of the current modulation mode recognition model is PGD ​​counterattack, the index number is Adversarial examples of communication signals .

[0095] Based on the weight formula (6), the weighted cross entropy loss under the current modulation recognition model is calculated:

[0096] (7)

[0097] Where, is the batch size.

[0098] (2) Self-distillation loss

[0099] A teacher model is constructed based on the EMA strategy, and its parameters are obtained by smoothly updating the historical state of the current modulation mode recognition model.

[0100] The parameters of the EMA teacher model are recorded as , which changes with the current model parameters at each training iteration Update according to the following rules:

[0101] (8)

[0102] Among them, the initial , is a decay coefficient used to control the proportion of historical information retained. It is generally set to a value close to 1, and is set to 0.999 in the specific implementation of this invention. This strategy can effectively suppress gradient oscillation during training, making the teacher model output smoother and more robust, thereby providing a stable feature representation reference for the student model.

[0103] In each training iteration, the current modulation recognition model is regarded as the student model, and the index number is The output logits of the adversarial sample are expressed as , the EMA teacher model output is The self-distillation loss is then defined as the KL divergence between the two output distributions:

[0104] (9)

[0105] (3) Calculate total loss

[0106] Total loss By weighted cross entropy loss and self-distillation losses Joint composition:

[0107] (10)

[0108] in, is the self-distillation coefficient, which is used to balance the two losses. In the specific implementation of the present invention, .

[0109] After each training round, the accuracy of the current modulation recognition model is calculated using the validation set, and the initial highest historical accuracy is set to 0. If the current accuracy exceeds the highest historical accuracy, the model parameters are immediately saved and the highest historical accuracy is updated. After the entire training is completed, the modulation recognition model that performs best on the validation set is saved.

[0110] As the training step progresses, the proportion of easy samples is updated and the minimum weight of difficult samples , and jump to step 4 to continue training until all training steps are completed. Specifically, according to formulas (11) and (12), the proportion of easy samples is updated respectively. and the minimum weight of difficult samples :

[0111] (11)

[0112] (12)

[0113] in, Initialized by a larger easy sample ratio Start with step length Gradually (pace) decrease; Initialized by a smaller minimum weight of the difficult sample Start with step length This dynamic update strategy ensures that the modulation recognition model learns simple samples and weak attacks in the early stage of training, and then gradually transitions to difficult samples and strong attack scenarios, thereby effectively improving the overall robustness and generalization ability. In the specific implementation of the present invention, , , .

[0114] At this point, an adaptive difficulty assessment that does not depend on the absolute loss value is implemented within each training step, and self-paced learning of the modulation mode recognition model is achieved through weighted cross-entropy loss and dynamic update strategy.

[0115] Step 6: Input the communication signal sample under the adversarial attack condition to be tested into the trained signal modulation mode recognition model to perform robust recognition of the modulation mode.

[0116] Example

[0117] To verify the effectiveness of this invention, we conducted experiments using the Python language and the PyTorch framework using the public communication signal modulation dataset RML2016.10a. This dataset contains 11 modulation modes: 8PSK, AM-DSB, AM-SSB, BPSK, CPFSK, GFSK, 4-PAM, 16-QAM, 64-QAM, QPSK, and WBFM.

[0118] In this embodiment, the data set is first divided, and a network architecture based on ResNet18 is constructed as a signal modulation mode recognition model. Secondly, the strength of the adversarial attack is quantified, the training steps are divided according to the attack strength, the training parameters are initialized, and the training begins. Then, the difficulty of the signal adversarial samples is adaptively evaluated at each training step, and different weights are assigned. Next, in each training iteration, the weighted cross entropy loss function is calculated; at the same time, a self-distillation mechanism based on EMA is constructed, and the self-distillation loss is calculated; finally, the total loss is obtained and used to train the modulation mode recognition model, and the modulation mode recognition model with the best recognition accuracy on the validation set is saved. Then, as the training steps advance, the proportion of easy samples and the minimum weight of difficult samples are updated until the training is completed. Finally, the effectiveness of the method proposed in the present invention is verified on the communication signal data to be tested.

[0119] Figure 2 The effectiveness of the proposed method is verified by comparing the recognition accuracy of four adversarial training methods under PGD adversarial attack conditions with different SPRs when the signal-to-noise ratio is 10dB. Figure 3 The effectiveness of the proposed method was verified by comparing the recognition accuracy of four adversarial training methods under FGSM adversarial attack conditions with different SPRs at a signal-to-noise ratio of 10dB. These results demonstrate that the proposed method has good generalization under adversarial attacks of different types and intensities, improving the accuracy of signal modulation recognition under adversarial attacks.

[0120] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0121] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A signal modulation recognition method based on self-paced self-distillation adversarial training, characterized in that: The specific steps are as follows: Step 1: Obtain a communication signal modulation mode dataset, including modulation mode category labels and communication signal samples, and divide the modulation signal into a training set and a validation set; Step 2: Use ResNet18 as the basic network of the modulation recognition model, take the communication signal sample as input and the modulation category prediction as output to build the modulation recognition model; Step 3: Define the signal-to-interference ratio to quantify the adversarial attack strength of the communication signal sample, divide the training steps of the self-paced learning, set the training rounds for each training step, and start the modulation mode recognition model training process; Step 4: At each training step, an adversarial sample is generated for each communication signal sample in the training set under the current modulation recognition model. Based on the model feedback mechanism of the loss quantile, the difficulty of the adversarial sample of the communication signal is evaluated and assigned different weights for calculating the weighted cross entropy loss. Step 5: Based on the weighted cross entropy loss and self-distillation loss, a total loss function is constructed for iterative training. In each training round, the adversarial samples of the communication signals in the training set are used to update the model parameters. The recognition accuracy of the original samples in the validation set is simultaneously calculated and the optimal model is saved. After completing all rounds of the current training step, the training step update is promoted, the proportion of easy samples and the minimum weight of difficult samples are updated, and the training is continued in step 4 until all training steps are completed. Step 6: Input the communication signal under the anti-attack condition into the trained signal modulation recognition model to perform modulation recognition; in, Step 3: Define the signal-to-interference ratio to quantify the adversarial attack strength of the communication signal sample and divide the training steps of the self-paced learning. The specific method is: The signal-to-interference ratio SPR is defined as: （1） （2） （3） in, and Respectively represent the communication signal I signal and Q signal sampling points; The smaller the signal-to-perturbation ratio (SPR), the greater the attack strength and the higher the difficulty of the adversarial sample. Pre-set minimum SPR value and maximum SPR value To determine the range of SPR values, divide it into equal intervals. Discretized into training steps, each training step Corresponding to a fixed attack strength , which is used to gradually improve the adversarial attack strength of adversarial samples of communication signals and guide the modulation recognition model to gradually adapt from weak attacks to strong attacks.

2. The signal modulation recognition method based on self-paced self-distillation adversarial training according to claim 1, characterized in that: Step 1: Obtain a communication signal modulation data set, including modulation type labels and various time-domain IQ signal samples, and divide the modulation signal into a training set and a validation set. The specific method is as follows: Set the category labels to , K represents the number of categories. For each communication signal sample, its I and Q signals are spliced ​​into The first row represents the I signal and the second row represents the Q signal. Indicates the number of sampling points of the communication signal samples, and divides the communication signal modulation mode training set and validation set into a ratio of 9:

1.

3. The signal modulation recognition method based on self-paced self-distillation adversarial training according to claim 1, characterized in that: Step 2: Use ResNet18 as the basic network of the modulation recognition model, take the communication signal sample as input and the modulation category prediction as output to build the modulation recognition model. The specific method is as follows: ResNet18 is used as the basic network of the modulation recognition model. For the training set, validation set, and subsequent test samples, in order to adapt to the input requirements of the ResNet18 neural network, a channel dimension with a value of 1 is added to reduce the size to The communication signal samples are reconstructed into three-dimensional , according to the batch size , stack multiple communication signal samples into a size of The network input data is , and the network output is the predicted probability of belonging to each category. The output size is , take the category with the highest probability as the signal category predicted by the modulation recognition model.

4. The signal modulation recognition method based on self-paced self-distillation adversarial training according to claim 1, characterized in that: Step 4: At each training step, generate an adversarial sample under the current modulation recognition model for each communication signal sample in the training set. Based on the model feedback mechanism of the loss quantile, evaluate the difficulty of the adversarial sample of the communication signal and assign different weights. The specific method is as follows: Step 4.1: At each training step, generate an adversarial sample under the current modulation recognition model for each communication signal sample in the training set; In the training step , for any index number in the training set Communication signal samples , applying an attack strength of PGD ​​adversarial attack, according to The adversarial perturbation size for each training step , generate the number of iterations n PGD Corresponding adversarial examples The calculation formula is shown in formula (4): （4） Where, represents uniform distribution, The sample index number is The category label, = is the single-step perturbation step length, represents the symbolic function, is the cross entropy loss function, Crop the adversarial samples to Centered on The final adversarial sample is within the L2 norm ball with a radius of ; Step 4.2: Based on the loss quantile model feedback mechanism, the difficulty of adversarial examples of communication signals is evaluated and assigned different weights; Calculate the cross entropy loss value of all generated adversarial samples on the current modulation recognition model ,in is the number of training set samples; use Indicates the The ratio of easy samples in the training step is determined The loss quantile threshold for training steps : （5） Where, Express The loss value set of adversarial examples Calculate the quantile threshold, sort the loss values ​​in ascending order, and take the value as the loss quantile threshold , so that at least The sample loss value of the proportion does not exceed , is the ceiling function; Introducing the current The minimum weight of difficult samples in the training step , a linear weight descent strategy is used to weight samples: simple samples are given a weight of 1.0, and the weight of difficult samples is between 1.0 and The interpolation is uniform between them, and the loss decreases linearly in the order from small to large. The index number of the training step is The weight of the adversarial sample The calculation formula is expressed as: （6）。 5. The signal modulation recognition method based on self-paced self-distillation adversarial training according to claim 1, characterized in that: Step 5: Based on the weighted cross entropy loss and self-distillation loss, a total loss function is constructed for iterative training. In each training round, the adversarial samples of the communication signal in the training set are used to update the model parameters, and the recognition accuracy of the original samples in the validation set is simultaneously calculated and the optimal model is saved. After completing all rounds of the current training step, the training step is updated, the proportion of easy samples and the minimum weight of difficult samples are updated, and the training is continued in step 4 until all training steps are completed. The total loss function is constructed based on the weighted cross entropy loss and self-distillation loss. The specific method is: (1) Weighted cross entropy loss In each training iteration, the adversarial examples of the communication signal under the current modulation recognition model are regenerated and the weighted cross entropy loss function is calculated; In the In each training iteration of a training step, let the index set of the current batch training set samples be , based on the current modulation recognition model, the attack strength is PGD ​​counterattack, the index number is Adversarial examples ; Based on the weight formula (6), the weighted cross entropy loss under the current modulation recognition model is calculated: （7） Where, is the batch size; (2) Self-distillation loss A teacher model is constructed based on the EMA strategy, and its parameters are obtained by smoothing and updating the historical state of the current modulation mode recognition model; The parameters of the EMA teacher model are recorded as , which changes with the current model parameters at each training iteration Update according to the following rules: （8） Among them, the initial , is the attenuation coefficient, which is used to control the retention ratio of historical information; In each training iteration, the current modulation recognition model is regarded as the student model, and the index number is The output logits of the adversarial sample are expressed as , the EMA teacher model output is , then the self-distillation loss is defined as the KL divergence between the two output distributions: （9） (3) Total loss function Total loss By weighted cross entropy loss and self-distillation losses Joint composition: （10） in, is the self-distillation coefficient, which is used to balance the two losses.

6. The signal modulation recognition method based on self-paced self-distillation adversarial training according to claim 1, characterized in that: Step 5: Based on the weighted cross entropy loss and self-distillation loss, the total loss function is constructed for iterative training. In each training round, the adversarial samples of the communication signal in the training set are used to update the model parameters, and the recognition accuracy of the original samples in the validation set is calculated synchronously and the optimal model is saved. After completing all rounds of the current training step, the training step update is promoted, the proportion of easy samples and the minimum weight of difficult samples are updated, and the training is continued in step 4 until all training steps are completed, where: the proportion of easy samples is updated. and the minimum weight of difficult samples , the specific method is: （11） （12） in, is the maximum SPR value, Initial value of easy sample ratio Start with step length gradually decrease; Initialized by the lowest weight of the difficult sample Start with step length Gradually increase.

7. A signal modulation mode recognition system based on self-paced self-distillation adversarial training, implementing the signal modulation mode recognition method based on self-paced self-distillation adversarial training according to any one of claims 1 to 6, realizing signal modulation mode recognition based on self-paced self-distillation adversarial training, and performing steps 1 to 6 in six modules.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for signal modulation mode recognition based on self-paced self-distillation adversarial training according to any one of claims 1 to 6 is implemented to realize signal modulation mode recognition based on self-paced self-distillation adversarial training.

9. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for signal modulation mode recognition based on self-paced self-distillation adversarial training according to any one of claims 1 to 6 is implemented to realize signal modulation mode recognition based on self-paced self-distillation adversarial training.

Images