Signal modulation type identification method and device, and electronic device

Abstract

This application provides a signal modulation type identification method, apparatus, and electronic device, which can be applied to the field of adversarial defense technology. It includes: acquiring a modulation signal to be identified; using a modulation type identification model to identify the modulation type of the signal to be identified, thereby obtaining the modulation type; the modulation type identification model is obtained through the following steps: iteratively processing a sample modulation signal under predetermined conditions to obtain a first adversarial sample; obtaining a second adversarial sample based on the sample modulation signal and the first adversarial sample; adjusting the parameters of an initial deep learning model based on the adversarial loss value obtained from the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label to obtain a target deep learning model; and adjusting the parameters of the target deep learning model based on the contrastive learning loss value obtained from the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label to obtain the modulation type identification model.

Description

Technical Field

[0001] This application relates to the field of countermeasures and defense technology, and more specifically to a signal modulation type identification method, apparatus, and electronic device. Background Technology

[0002] Modulation type identification is a crucial step in electronic detection and electronic countermeasures. By identifying the modulation type of the intermediate frequency (IF) signal received by the receiver, the modulation type of the IF signal can be determined, thus supporting subsequent processing tasks such as demodulation, decoding, and signal identification. However, with increasingly congested electromagnetic environments, time-frequency transform-based signal modulation type identification algorithms struggle to adapt to harsh conditions such as low signal-to-noise ratios and signal attenuation in complex electromagnetic environments, resulting in low identification accuracy and poor robustness. Summary of the Invention

[0003] In view of the above problems, this application provides a signal modulation type identification method, apparatus, electronic device, medium and program product.

[0004] According to a first aspect of this application, a signal modulation type identification method is provided, comprising: acquiring a modulation signal to be identified; using a modulation type identification model to identify the modulation type of the modulation signal to be identified, thereby obtaining the modulation type of the modulation signal to be identified; wherein the modulation type identification model is trained by: performing iterative processing on sample modulation signals under predetermined conditions, and determining sample modulation signals that meet the predetermined conditions as first adversarial samples; calculating the mean of the sample modulation signals and the first adversarial samples to obtain second adversarial samples; adjusting the parameters of an initial deep learning model based on the adversarial loss value obtained according to the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label, thereby obtaining a target deep learning model; and adjusting the parameters of the target deep learning model based on the contrastive learning loss value obtained according to the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label, thereby obtaining a modulation type identification model.

[0005] According to an embodiment of this application, a target deep learning model is obtained by adjusting the parameters of an initial deep learning model based on the adversarial loss value obtained from the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label. This includes: extracting features from the sample modulation signal, the first adversarial sample, and the second adversarial sample to obtain modulation type features corresponding to each of the sample modulation signal, the first adversarial sample, and the second adversarial sample; obtaining the adversarial loss value using the adversarial loss function based on the modulation type features corresponding to each of the sample modulation signal, the first adversarial sample, and the second adversarial sample, and the real modulation type label; and updating the network parameters of the initial deep learning model based on the adversarial loss value to obtain the target deep learning model.

[0006] According to an embodiment of this application, the adversarial loss function includes a cross-entropy loss function and a regularization loss function. The adversarial loss function is used to obtain an adversarial loss value based on the modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample, and the true modulation type label, respectively. This includes: using the cross-entropy loss function to obtain a cross-entropy loss value based on the modulation type features corresponding to the first adversarial sample and the true modulation type label; using the regularization loss function to calculate the mean of the modulation type features corresponding to the second adversarial sample, the modulation type features corresponding to the sample modulation signal, and the modulation type features corresponding to the first adversarial sample, to obtain a regularization loss value; and summing the cross-entropy loss value and the regularization loss value to obtain the adversarial loss value.

[0007] According to an embodiment of this application, a modulation type recognition model is obtained by adjusting the parameters of a target deep learning model based on a contrastive learning loss value obtained from a sample modulation signal, a first adversarial sample, a second adversarial sample, and a real modulation type label. This includes: extracting features from the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample to obtain modulation type features corresponding to each of the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, respectively, wherein the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample are obtained by masking the sample modulation signal, the first adversarial sample, and the second adversarial sample; using a contrastive learning loss function, a contrastive learning loss value is obtained based on the modulation type features corresponding to each of the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, and the real modulation type label; and updating the network parameters of the target deep learning model based on the contrastive learning loss value to obtain the modulation type recognition model.

[0008] According to embodiments of this application, the contrastive learning loss function includes a mask regularization loss function, a contrastive loss function, and a cross-entropy loss function. The contrastive learning loss function is used to obtain the contrastive learning loss value based on the modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, and the true modulation type label, respectively. This includes: using the mask regularization loss function to obtain a mask regularization loss value based on the modulation type features corresponding to the second masked adversarial sample, the modulation type features corresponding to the masked sample modulation signal, and the modulation type features corresponding to the first masked adversarial sample; treating the first masked adversarial sample and the second mask (which have different types of true modulation type labels from the sample modulation signal) as negative sample pairs, and treating the sample modulation signal as positive samples, and processing the negative sample pairs and positive samples using the contrastive loss function to obtain a contrastive loss value; using the cross-entropy loss function to obtain a cross-entropy loss value based on the modulation type features corresponding to the first adversarial sample and the true modulation type label; and summing the mask regularization loss value, the contrastive loss value, and the cross-entropy loss value to obtain the contrastive learning loss value.

[0009] According to an embodiment of this application, the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample are obtained by masking the sample modulation signal, the first adversarial sample, and the second adversarial sample, respectively. The process includes: randomly generating a mask matrix associated with a preset loss rate based on a preset loss rate; and performing masking processing on the sample modulation signal, the first adversarial sample, and the second adversarial sample based on the mask matrix to obtain a masked sample modulation signal corresponding to the sample modulation signal, a first masked adversarial sample corresponding to the first adversarial sample, and a second masked adversarial sample corresponding to the second adversarial sample.

[0010] According to an embodiment of this application, iterative processing of a sample modulation signal under predetermined conditions is performed, and the sample modulation signal that meets the predetermined conditions is determined as a first adversarial sample. This includes: differentiating an objective function to obtain a differentiated function; obtaining a gradient value of the sample modulation signal corresponding to the current iteration based on the differentiated function and the feature vector of the sample modulation signal corresponding to the current iteration; the objective function is a function calculated by the feature vector of the sample modulation signal corresponding to each iteration; obtaining the sample modulation signal corresponding to the next iteration based on the gradient value of the sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration; and determining the sample modulation signal that meets the predetermined conditions as a first adversarial sample.

[0011] According to an embodiment of this application, obtaining the sample modulation signal corresponding to the next iteration based on the gradient value of the sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration includes: taking the sign of the gradient value of the sample modulation signal corresponding to the current iteration to obtain the perturbation direction; perturbing the sample modulation signal corresponding to the current iteration along the perturbation direction based on a preset perturbation step size to obtain the intermediate sample modulation signal corresponding to the current iteration; summing the intermediate sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration to obtain the target sample modulation signal corresponding to the current iteration; and projecting the target sample modulation signal corresponding to the current iteration based on a preset perturbation radius to obtain the sample modulation signal corresponding to the next iteration.

[0012] A second aspect of this application provides a signal modulation type identification device, comprising: an acquisition module for acquiring a modulation signal to be identified; and an identification module for identifying the modulation type of the modulation signal to be identified using a modulation type identification model to obtain the modulation type of the modulation signal to be identified; wherein the modulation type identification model is trained by: performing iterative processing on sample modulation signals under predetermined conditions, and determining sample modulation signals that satisfy the predetermined conditions as first adversarial samples; calculating the mean of the sample modulation signals and the first adversarial samples to obtain second adversarial samples; adjusting the parameters of an initial deep learning model based on an adversarial loss value obtained according to the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label to obtain a target deep learning model; and adjusting the parameters of the target deep learning model based on a contrastive learning loss value obtained according to the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label to obtain the modulation type identification model.

[0013] A third aspect of this application provides an electronic device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0014] A fourth aspect of this application also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the steps of the above-described method.

[0015] The fifth aspect of this application also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0016] According to the signal modulation type identification method provided in this application, a target deep learning model is obtained by adjusting the parameters of an initial deep learning model based on adversarial loss. Then, a modulation type identification model is obtained by adjusting the parameters of the target deep learning model based on contrastive learning loss to identify the modulation type of the signal to be identified. The modulation type of the signal to be identified can be obtained. By introducing an adversarial loss function during the parameter adjustment of the initial deep learning model, the robustness accuracy of the model can be rapidly improved during the training phase of the initial deep learning model. On this basis, by introducing a contrastive learning loss function, robust overfitting can be suppressed during the training phase of the target deep learning model, thereby continuously improving the robustness accuracy of the target deep learning model and thus improving the type identification accuracy of the modulation type identification model. Attached Figure Description

[0017] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0018] Figure 1 This diagram illustrates an application scenario of the signal modulation type identification method according to an embodiment of this application.

[0019] Figure 2 A flowchart illustrating a signal modulation type identification method according to an embodiment of this application is shown schematically.

[0020] Figure 3 A flowchart illustrating a training method for a modulation type recognition model according to an embodiment of this application is shown in the schematic diagram.

[0021] Figure 4 A flowchart illustrating a training method for a modulation type recognition model according to yet another embodiment of this application is shown.

[0022] Figure 5 A schematic diagram illustrating the structure of a signal modulation type identification device according to an embodiment of this application is shown; and

[0023] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a signal modulation type identification method according to an embodiment of this application. Detailed Implementation

[0024] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0025] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0026] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0027] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0028] In related technologies, in order to meet the practical application needs of robust identification of signal modulation types in complex electromagnetic environments, deep neural network models, represented by recurrent neural networks, convolutional networks, and attention mechanisms, have been more widely used.

[0029] For example, signal modulation type recognition methods based on deep neural networks have advantages such as strong feature representation capabilities and noise adaptability, and have been applied to various signal detection systems. However, deep neural networks are vulnerable to adversarial attacks. For well-designed adversarial examples, such as those obtained by adding small perturbations to clean samples, the model's recognition accuracy can significantly degrade, thus affecting the security of the model's application. Adversarial training methods are an effective way to address adversarial attacks on neural networks. By generating adversarial training samples to assist the model in adversarial training, higher robust recognition accuracy can be obtained, thereby improving the model's defense against adversarial attacks. However, in the later stages of adversarial training, a special phenomenon called robust overfitting often occurs. That is, during continuous training, the model's loss function on the training set continuously decreases, but the loss function on the validation set continuously increases. Robust overfitting prevents the model's robust accuracy from improving further, and may even cause it to degrade drastically with continued training, until the model diverges, affecting the safe application of the model.

[0030] Because signal modulation type recognition models based on deep neural networks are vulnerable to adversarial examples, adversarial defense techniques employ adversarial training. This involves generating adversarial examples to assist the model in adversarial training, thereby enhancing its robustness and enabling it to adapt to such attacks. With advancements in adversarial training techniques, models have achieved higher robust accuracy and stronger overfitting suppression capabilities. However, overfitting still exists, and the gap between the model's natural accuracy and robust accuracy remains significant.

[0031] Furthermore, the adversarial training methods in related technologies mainly adopt single-stage adversarial training. In the initial stage of training, the robust accuracy of the model improves rapidly, and robust overfitting has not yet occurred in this stage. As training continues, the robust accuracy of the model deteriorates, and overfitting occurs, which leads to a decrease in model training accuracy, or even a sharp decline in instability, resulting in model training divergence.

[0032] In view of this, embodiments of this application provide a signal modulation type identification method, comprising: acquiring a modulation signal to be identified; using a modulation type identification model to identify the modulation signal to be identified, thereby obtaining the modulation type of the modulation signal to be identified; wherein the modulation type identification model is trained by the following method: performing iterative processing on sample modulation signals under predetermined conditions, and determining sample modulation signals that meet the predetermined conditions as first adversarial samples; calculating the mean of the sample modulation signals and the first adversarial samples to obtain second adversarial samples; adjusting the parameters of an initial deep learning model based on the adversarial loss value obtained according to the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label to obtain a target deep learning model; and adjusting the parameters of the target deep learning model based on the contrastive learning loss value obtained according to the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label to obtain a modulation type identification model.

[0033] Figure 1 The diagram illustrates an application scenario of the signal modulation type identification method according to an embodiment of this application.

[0034] like Figure 1 As shown, application scenario 100 according to this embodiment may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0035] Users can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as shopping applications, web browser applications, search applications, instant messaging tools, email clients, social media platform software, etc. (for example only).

[0036] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, laptops, and desktop computers.

[0037] Server 105 can be a server that provides various services, such as a backend management server that supports websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process data such as received user requests, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices.

[0038] It should be noted that the signal modulation type identification method provided in this application embodiment can generally be executed by server 105. Correspondingly, the signal modulation type identification device provided in this application embodiment can generally be located in server 105. The signal modulation type identification method provided in this application embodiment can also be executed by a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105. Correspondingly, the signal modulation type identification device provided in this application embodiment can also be located in a server or server cluster that is different from server 105 and capable of communicating with the first terminal device 101, the second terminal device 102, the third terminal device 103, and / or server 105.

[0039] It should be understood that Figure 1 The number of first terminal devices, second terminal devices, third terminal devices, networks, and servers shown in the diagram is merely illustrative. Depending on implementation needs, any number of first terminal devices, second terminal devices, third terminal devices, networks, and servers can be included.

[0040] Figure 2 A flowchart illustrating a signal modulation type identification method according to an embodiment of this application is shown schematically.

[0041] like Figure 2 As shown, the signal modulation type identification method 200 of this embodiment includes: operation S210 to operation S220.

[0042] In operation S210, the modulation signal to be identified is acquired.

[0043] In operation S220, the modulation type identification model is used to identify the type of the modulation signal to be identified, and the modulation type of the modulation signal to be identified is obtained.

[0044] The modulated signal to be identified can be an unknown communication signal whose modulation type needs to be identified. This unknown communication signal can be a signal obtained after modulation processing using modulation techniques. In one implementation, the modulated signal to be identified can be an intermediate frequency (IF) signal.

[0045] By using a modulation type identification model to identify the modulation signal to be identified, the modulation type of the signal can be obtained. For example, the modulation type of the signal to be identified can be Quadrature Phase Shift Keying (QPSK), 16-Quadrature Amplitude Modulation (16QAM), 32-Quadrature Amplitude Modulation (32QAM), or Binary Phase Shift Keying (BPSK), among other modulation types.

[0046] Figure 3 A flowchart illustrating a training method for a modulation type recognition model according to an embodiment of this application is shown.

[0047] like Figure 3 As shown, the training method for the modulation type recognition model includes operations S310 to S340.

[0048] In operation S310, the sample modulation signal is subjected to iterative processing under predetermined conditions, and the sample modulation signal that meets the predetermined conditions is determined as the first adversarial sample.

[0049] In operation S320, the mean of the sample modulation signal and the first adversarial sample is calculated to obtain the second adversarial sample.

[0050] In operation S330, based on the adversarial loss value obtained from the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label, the parameters of the initial deep learning model are adjusted to obtain the target deep learning model.

[0051] In operation S340, based on the contrastive learning loss value obtained from the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label, the parameters of the target deep learning model are adjusted to obtain the modulation type recognition model.

[0052] The true modulation type label can characterize the known modulation type of the sample modulation signal. For example, the true modulation type label can be other modulation types such as QPSK, 16QAM, 32QAM, or BPSK.

[0053] In one implementation, the sample modulation signal can represent a known communication signal for which modulation type identification is to be performed during the training of the modulation type identification model; that is, the modulation type of the sample modulation signal is known. In another implementation, there can be multiple sample modulation signals, and these multiple sample modulation signals can constitute a sample modulation signal dataset. ,in, This represents a dataset consisting of multiple sample modulated signals. This represents the actual modulation type label corresponding to each of the multiple sample modulation signals.

[0054] By iteratively processing the sample modulation signal under predetermined conditions, the sample modulation signal that meets the predetermined conditions can be identified as the first adversarial sample. In one implementation, the predetermined conditions can be a predetermined number of iteration rounds.

[0055] In one implementation, the second adversarial sample can be obtained by averaging the sample modulation signal and the first adversarial sample using the following formula (1).

[0056] (1);

[0057] in, This represents the second adversarial sample. This represents the first adversarial example. This represents the sample modulation signal.

[0058] In another implementation, the second adversarial sample can be obtained by multiplying the sample modulation signal and the first adversarial sample using the following formula (2).

[0059] (2);

[0060] By using the adversarial loss function, based on the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label, the adversarial loss value can be obtained. Based on the adversarial loss value, the parameters of the initial deep learning model can be adjusted to obtain the target deep learning model.

[0061] By using the contrastive learning loss function, based on the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label, the contrastive learning loss value can be obtained. Based on the contrastive learning loss value, the parameters of the target deep learning model are adjusted to obtain the modulation type recognition model.

[0062] By adjusting the parameters of an initial deep learning model using adversarial loss, a target deep learning model is obtained. Then, by adjusting the parameters of the target deep learning model again using contrastive learning loss, a modulation type recognition model is obtained to identify the modulation type of the modulation signal to be identified. In the process of adjusting the parameters of the initial deep learning model, the introduction of an adversarial loss function can quickly improve the robustness accuracy of the model during the training phase. On this basis, by introducing a contrastive learning loss function, robust overfitting can be suppressed during the training phase of the target deep learning model, thereby continuously improving the robustness accuracy of the target deep learning model and thus improving the type recognition accuracy of the modulation type recognition model.

[0063] The sample modulation signal is subjected to iterative processing under predetermined conditions, and the sample modulation signal that meets the predetermined conditions is determined as the first adversarial sample. This includes: taking the derivative of the objective function to obtain the differentiated function; based on the differentiated function, obtaining the gradient value of the sample modulation signal corresponding to the current iteration according to the feature vector of the sample modulation signal corresponding to the current iteration; obtaining the sample modulation signal corresponding to the next iteration according to the gradient value of the sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration; and determining the sample modulation signal that meets the predetermined conditions as the first adversarial sample.

[0064] The objective function can be a function calculated based on a preset loss function and using the feature vector of the sample modulation signal corresponding to each iteration. The feature vector of the sample modulation signal corresponding to each iteration can represent the perturbed intermediate adversarial sample obtained at the end of the previous iteration.

[0065] By differentiating the objective function, we can obtain the differentiated function. Substituting the feature vector of the sample modulation signal corresponding to the current iteration into the differentiated function, we can obtain the gradient value of the sample modulation signal corresponding to the current iteration.

[0066] In one implementation, obtaining the sample modulation signal corresponding to the next iteration based on the gradient value of the sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration may include: taking the sign of the gradient value of the sample modulation signal corresponding to the current iteration to obtain the perturbation direction; perturbing the sample modulation signal corresponding to the current iteration along the perturbation direction based on a preset perturbation step size to obtain the intermediate sample modulation signal corresponding to the current iteration; summing the intermediate sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration to obtain the target sample modulation signal corresponding to the current iteration; and projecting the target sample modulation signal corresponding to the current iteration based on a preset perturbation radius to obtain the sample modulation signal corresponding to the next iteration.

[0067] The perturbation direction can be obtained by signifying the gradient value of the sample modulation signal corresponding to the current iteration. In one implementation, the gradient value can include both positive and negative values. By signifying the gradient value of the sample modulation signal corresponding to the current iteration, the gradient value can be converted into a sign vector containing only positive and negative values. The positive and negative sign vectors can be used as the perturbation direction. For example, if the sign vector is positive, the perturbation direction is positive; if the sign vector is negative, the perturbation direction is negative.

[0068] Based on a preset perturbation step size, the sample modulation signal corresponding to the current iteration is perturbed along the perturbation direction to obtain the intermediate sample modulation signal corresponding to the current iteration. Summing the intermediate sample modulation signal and the sample modulation signal corresponding to the current iteration yields the target sample modulation signal. Based on a preset perturbation radius, the target sample modulation signal corresponding to the current iteration is projected to obtain the sample modulation signal corresponding to the next iteration. The preset perturbation step size and preset perturbation radius can be set based on experience or requirements.

[0069] A sample modulation signal that meets predetermined conditions is identified as a first adversarial sample. In one implementation, the sample modulation signal that meets the predetermined conditions can be a sample modulation signal with a predetermined number of iteration rounds. For example, in the iterative processing of the sample modulation signal with predetermined conditions, the predetermined conditions can be a predetermined number of iteration rounds. When the predetermined number of iteration rounds is reached, the iteration stops, and the sample modulation signal corresponding to the predetermined number of iteration rounds is identified as the first adversarial sample.

[0070] In one implementation, the sample modulation signal is subjected to iterative processing under predetermined conditions, and the sample modulation signal that satisfies the predetermined conditions is determined as the first adversarial sample. For example, the process of processing the sample modulation signal corresponding to the i-th iteration to obtain the sample modulation signal corresponding to the (i+1)-th iteration satisfies the following formula (3).

[0071] (3);

[0072] in, This represents the sample modulation signal corresponding to the (i+1)th iteration. This represents the sample modulation signal corresponding to the i-th iteration. This indicates the label representing the actual modulation type corresponding to the sample modulation signal. Indicates the preset perturbation step size. Indicates the predetermined number of iteration rounds. Indicates the preset disturbance radius. Represents a symbolic function. This represents the sample modulation signal corresponding to the i-th iteration. gradient value, Indicates the preset loss function. Describe the objective function. This represents the initial deep learning model. Represents the set of sample modulated signals. This represents the set of labels for the actual modulation type.

[0073] In one implementation, the sample modulation signal corresponds to the (i+1)th iteration. Under predetermined conditions, the sample modulation signal corresponding to the (i+1)th iteration is... The sample modulation signal corresponding to the first adversarial example in the (i+1)th iteration is identified. If the predetermined conditions are not met, the sample modulation signal corresponding to the (i+1)th iteration and its corresponding gradient value are used to calculate the sample modulation signal corresponding to the (i+2)th iteration until the predetermined iteration conditions are met, and the sample modulation signal that meets the predetermined iteration conditions is used as the first adversarial sample.

[0074] By processing the gradient value by taking its sign, the perturbation direction can be determined, avoiding the instability caused by fluctuations in the gradient value's amplitude. This results in a more uniform distribution of the generated first adversarial samples and greater generalization of the perturbations. By performing multiple rounds of iterative processing on the sample modulation signal, perturbations are gradually applied to the sample modulation signal. Furthermore, by using projection processing to constrain the first adversarial samples within a preset perturbation radius neighborhood centered on the sample modulation signal, the difference between the first adversarial samples and the sample modulation signal can be minimized, making it more closely resemble actual communication and radar signal scenarios.

[0075] The process of generating a first adversarial sample based on the sample modulation signal, and then generating a second adversarial sample based on the sample modulation signal and the first adversarial sample, can be considered as the first stage of training the modulation type recognition model, i.e., the adversarial sample generation stage. The sample modulation signal, the first adversarial sample, and the second adversarial sample can form a triplet sample. .

[0076] Based on the adversarial loss value obtained from the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label, the parameters of the initial deep learning model are adjusted to obtain the target deep learning model. This includes: extracting features from the sample modulation signal, the first adversarial sample, and the second adversarial sample to obtain the modulation type features corresponding to each of the sample modulation signal, the first adversarial sample, and the second adversarial sample; using the adversarial loss function, obtaining the adversarial loss value based on the modulation type features corresponding to each of the sample modulation signal, the first adversarial sample, and the second adversarial sample, and the real modulation type label; and updating the network parameters of the initial deep learning model based on the adversarial loss value to obtain the target deep learning model.

[0077] By using the initial deep learning model to extract features from the sample modulation signal, the modulation type features corresponding to the sample modulation signal can be obtained. Similarly, by using the initial deep learning model to extract features from the first adversarial example, the modulation type features corresponding to the first adversarial example can be obtained. Finally, by using the initial deep learning model to extract features from the second adversarial example, the modulation type features corresponding to the second adversarial example can be obtained.

[0078] By substituting the modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample, and the true modulation type label into the adversarial loss function, the adversarial loss value can be obtained. In one implementation, the adversarial loss function may include a cross-entropy loss function and a regularization loss function. Specifically, using the adversarial loss function to obtain the adversarial loss value based on the modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample, and the true modulation type label, may include: using the cross-entropy loss function to obtain the cross-entropy loss value based on the modulation type features corresponding to the first adversarial sample and the true modulation type label; using the regularization loss function to calculate the mean of the modulation type features corresponding to the second adversarial sample, the modulation type features corresponding to the sample modulation signal, and the modulation type features corresponding to the first adversarial sample, to obtain the regularization loss value; and summing the cross-entropy loss value and the regularization loss value to obtain the adversarial loss value.

[0079] In one implementation, the adversarial loss function satisfies the following formula (4).

[0080] (4);

[0081] in, Represents the adversarial loss function. This represents the regularization loss function. Represents the cross-entropy loss function. This represents the first regularization coefficient. In one implementation, the regularization coefficient is... It can be set to 0.1.

[0082] By inputting the sample modulation signal, the first adversarial sample, and the second adversarial sample into the initial deep learning model, the model outputs the modulation type features corresponding to the sample modulation signal, the modulation type features corresponding to the first adversarial sample, and the modulation type features corresponding to the second adversarial sample.

[0083] By substituting the modulation type features corresponding to the first adversarial sample and the true modulation type label into the cross-entropy loss function, the cross-entropy loss value can be obtained.

[0084] In one implementation, the regularization loss function satisfies the following formula (5).

[0085] (5);

[0086] in, Represents the matrix norm. Represents the mean function, This represents the modulation type feature corresponding to the sample modulation signal. This represents the modulation type feature corresponding to the first adversarial example. This represents the modulation type feature corresponding to the second adversarial sample.

[0087] In one implementation, the mean function can reduce the influence of the matrix norm on the calculation of the adversarial loss function, preventing anomalies such as excessively large or small norms.

[0088] By substituting the modulation type features corresponding to the sample modulation signal, the modulation type features corresponding to the first adversarial sample, and the modulation type features corresponding to the second adversarial sample into the regularization loss function, and using the regularization loss function to calculate the mean of the modulation type features corresponding to the second adversarial sample, the modulation type features corresponding to the sample modulation signal, and the modulation type features corresponding to the first adversarial sample, the regularization loss value can be obtained.

[0089] The adversarial loss is calculated by summing the product of the regularization loss and the first regularization coefficient with the cross-entropy loss. Based on the adversarial loss, the network parameters of the initial deep learning model are updated to obtain the target deep learning model.

[0090] The process of updating the network parameters of the initial deep learning model based on the adversarial loss value to obtain the target deep learning model can be regarded as the second stage of training the modulation type recognition model, that is, the stability adversarial training stage.

[0091] By combining the modulation type characteristics of the sample modulation signal, the initial deep learning model can quickly learn the perturbation patterns of the first and second adversarial samples, effectively improving the model's adaptability to unknown perturbations. By introducing an adversarial loss function, the robustness accuracy of the model can be rapidly improved, reducing the risk of robust overfitting.

[0092] According to an embodiment of this application, a modulation type recognition model is obtained by adjusting the parameters of a target deep learning model based on a contrastive learning loss value obtained from a sample modulation signal, a first adversarial sample, a second adversarial sample, and a real modulation type label. This includes: extracting features from the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample to obtain modulation type features corresponding to each of the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample; obtaining a contrastive learning loss value using the contrastive learning loss function based on the modulation type features corresponding to each of the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, and the real modulation type label; and updating the network parameters of the target deep learning model based on the contrastive learning loss value to obtain the modulation type recognition model.

[0093] The masked sample modulated signal can be obtained by masking the sample modulated signal. The first masked adversarial sample can be obtained by masking the first adversarial sample. The second masked adversarial sample can be obtained by masking the second adversarial sample.

[0094] Specifically, the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample are obtained by masking the sample modulation signal, the first adversarial sample, and the second adversarial sample, respectively. This can include: randomly generating a mask matrix associated with a preset loss rate based on a preset loss rate; and performing masking processing on the sample modulation signal, the first adversarial sample, and the second adversarial sample based on the mask matrix to obtain the masked sample modulation signal corresponding to the sample modulation signal, the first masked adversarial sample corresponding to the first adversarial sample, and the second masked adversarial sample corresponding to the second adversarial sample.

[0095] A mask matrix associated with a preset loss rate can be randomly generated. Masking the sample modulation signal using the mask matrix yields a masked sample modulation signal. Masking the first adversarial sample using the mask matrix yields a first masked adversarial sample. Masking the second adversarial sample using the mask matrix yields a second masked adversarial sample. In one implementation, the preset loss rate can be set to 0.3. Masking can further mitigate robust overfitting.

[0096] By using a target deep learning model to extract features from the masked sample modulation signal, the modulation type features corresponding to the masked sample modulation signal can be obtained. Similarly, by using the target deep learning model to extract features from the first masked adversarial sample, the modulation type features corresponding to the first masked adversarial sample can be obtained. Finally, by using the target deep learning model to extract features from the second masked adversarial sample, the modulation type features corresponding to the second masked adversarial sample can be obtained.

[0097] In one implementation, the contrastive learning loss function may include a masking regularization loss function, a contrastive loss function, and a cross-entropy loss function.

[0098] Specifically, using the contrastive learning loss function, the contrastive learning loss value is obtained based on the modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, as well as the true modulation type label. This can include: using the mask regularization loss function, obtaining the mask regularization loss value based on the modulation type features corresponding to the second masked adversarial sample, the modulation type features corresponding to the masked sample modulation signal, and the modulation type features corresponding to the first masked adversarial sample; treating the first masked adversarial sample and the second mask, which have different types of true modulation type labels from the sample modulation signal, as negative sample pairs, and treating the sample modulation signal as positive samples, and processing the negative sample pairs and positive samples using the contrastive loss function to obtain the contrastive loss value; using the cross-entropy loss function, obtaining the cross-entropy loss value based on the modulation type features corresponding to the first adversarial sample and the true modulation type label; and summing the mask regularization loss value, the contrastive loss value, and the cross-entropy loss value to obtain the contrastive learning loss value.

[0099] In one implementation, the contrastive learning loss function satisfies the following formula (6).

[0100] (6);

[0101] in, This represents the contrastive learning loss function. This represents the masking regularization loss function. Represents the cross-entropy loss function. This represents the contrastive loss function.

[0102] By substituting the modulation type features corresponding to the second masked adversarial sample, the modulation type features corresponding to the masked sample modulation signal, and the modulation type features corresponding to the first masked adversarial sample into the mask regularization loss function, the mask regularization loss value can be obtained.

[0103] The first mask (different from the true modulation type label of the sample modulation signal) and the second mask (different from the true modulation type label of the sample modulation signal) can be used as negative sample pairs, while the sample modulation signal can be used as positive samples. The contrastive loss value can be obtained by substituting the negative sample pairs and the positive samples into the contrastive loss function.

[0104] In one implementation, the contrastive loss function satisfies the following formula (7);

[0105] (7);

[0106] in, Indicates chord distance, Indicates masking processing, This represents the modulated signal of the i-th sample. Let j be the j-th first adversarial sample. This represents the i-th first adversarial sample. Let j represent the j-th second adversarial sample. Indicates the temperature coefficient for comparative learning. Denotes the second regularization coefficient. Indicates the comparative learning coefficient. This represents the number of sample modulated signals. In one implementation, both the second regularization coefficient and the contrastive learning coefficient can be set to 0.1.

[0107] In one implementation, the chord distance satisfies the following formula (8), and the masking process satisfies the following formula (9).

[0108] (8);

[0109] (9);

[0110] in, Indicates the first parameter. Indicates the second parameter. This represents the third parameter. The first parameter can be... or The second parameter can be... or The first and second parameters are different. The third parameter can be... , or .

[0111] By substituting the modulation type features corresponding to the first adversarial sample and the true modulation type label into the cross-entropy loss function, the cross-entropy loss value can be obtained.

[0112] The contrastive learning loss value can be obtained by summing the product of the mask regularization loss value and the second regularization coefficient, the product of the contrastive loss value and the contrastive learning coefficient, and the cross-entropy loss value. Based on the contrastive learning loss value, the network parameters of the target deep learning model are updated to obtain the modulation type recognition model.

[0113] By introducing a contrastive learning loss function, robust overfitting can be suppressed, thereby further improving the robust accuracy of the model and enhancing the stability and continuous training capability of the model. Masking can further alleviate robust overfitting, ultimately resulting in a more robust and stable modulation type recognition model.

[0114] The process of updating the network parameters of the target deep learning model based on the contrastive learning loss value to obtain the modulation type recognition model can be regarded as the third stage of modulation type recognition model training, namely, the stability contrastive training stage.

[0115] Figure 4 A flowchart illustrating a training method for a modulation type recognition model according to yet another embodiment of this application is shown.

[0116] like Figure 4 As shown, the training method for the modulation type recognition model includes operations S401 to S411.

[0117] When operating S401, input the sample modulation signal.

[0118] In operation S402, the sample modulation signal is subjected to iterative processing under predetermined conditions, and the sample modulation signal that meets the predetermined conditions is determined as the first adversarial sample.

[0119] In operation S403, the mean of the sample modulation signal and the first adversarial sample is calculated to obtain the second adversarial sample.

[0120] In operation S404, feature extraction is performed on the sample modulation signal, the first adversarial sample, and the second adversarial sample to obtain the modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample respectively.

[0121] In operation S405, the cross-entropy loss function is used to obtain the cross-entropy loss value based on the feature vector associated with the sample modulation signal and the true modulation type label.

[0122] In operation S406, the regularization loss function is used to calculate the mean of the modulation type features corresponding to the second adversarial sample, the modulation type features corresponding to the sample modulation signal, and the modulation type features corresponding to the first adversarial sample, and obtain the regularization loss value.

[0123] In operation S407, the cross-entropy loss value and the regularization loss value are summed to obtain the adversarial loss value.

[0124] In operation S408, the network parameters of the initial deep learning model are updated based on the adversarial loss value to obtain the target deep learning model.

[0125] In operation S409, feature extraction is performed on the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, respectively, to obtain the modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample.

[0126] In operation S410, the contrastive learning loss function is used to obtain the contrastive learning loss value based on the modulation type features and real modulation type labels corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, respectively.

[0127] In operation S411, the network parameters of the target deep learning model are updated based on the contrastive learning loss value to obtain the modulation type recognition model.

[0128] In one implementation, operations S401 to S403 can be the adversarial example generation stage. Operations S404 to S408 can be the stability adversarial training stage. Operations S409 to S411 can be the stability comparison training stage.

[0129] Based on the above-described signal modulation type identification method, this application also provides a signal modulation type identification device. The following will be combined with... Figure 5 The device is described in detail.

[0130] Figure 5 A schematic block diagram of a signal modulation type identification device according to an embodiment of this application is shown.

[0131] like Figure 5 As shown, the signal modulation type identification device 500 of this embodiment includes an acquisition module 510 and an identification module 520.

[0132] The acquisition module 510 is used to acquire the modulation signal to be identified. In one embodiment, the acquisition module 510 can be used to perform the operation S210 described above, which will not be repeated here.

[0133] The identification module 520 is used to identify the modulation type of the modulation signal to be identified using a modulation type identification model, thereby obtaining the modulation type of the modulation signal to be identified. The modulation type identification model is trained using the following method: iterative processing of sample modulation signals under predetermined conditions, and identifying sample modulation signals that meet the predetermined conditions as first adversarial samples; calculating the mean between the sample modulation signals and the first adversarial samples to obtain second adversarial samples; adjusting the parameters of an initial deep learning model based on the adversarial loss value obtained from the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label to obtain a target deep learning model; and adjusting the parameters of the target deep learning model based on the contrastive learning loss value obtained from the sample modulation signals, the first adversarial samples, the second adversarial samples, and the real modulation type label to obtain the modulation type identification model. In one embodiment, the identification module 520 can be used to perform the operation S220 described above, which will not be repeated here.

[0134] According to an embodiment of this application, the identification module 520 includes a training submodule, wherein the training submodule includes an iteration unit for iteratively processing the sample modulation signal under predetermined conditions and determining the sample modulation signal that satisfies the predetermined conditions as a first adversarial sample; a first obtaining unit for calculating the mean of the sample modulation signal and the first adversarial sample to obtain a second adversarial sample; a second obtaining unit for adjusting the parameters of an initial deep learning model based on the adversarial loss value obtained according to the sample modulation signal, the first adversarial sample, the second adversarial sample and the real modulation type label to obtain a target deep learning model; and a third obtaining unit for adjusting the parameters of the target deep learning model based on the contrastive learning loss value obtained according to the sample modulation signal, the first adversarial sample, the second adversarial sample and the real modulation type label to obtain a modulation type identification model.

[0135] According to an embodiment of this application, the second obtaining unit includes: a first obtaining subunit, used to extract features from the sample modulation signal, the first adversarial sample, and the second adversarial sample respectively, to obtain modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample respectively; a second obtaining subunit, used to obtain an adversarial loss value using an adversarial loss function based on the modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample respectively and the real modulation type label; and a third obtaining subunit, used to update the network parameters of the initial deep learning model based on the adversarial loss value, to obtain the target deep learning model.

[0136] According to an embodiment of this application, the adversarial loss function includes a cross-entropy loss function and a regularization loss function. The second obtaining subunit is further configured to: use the cross-entropy loss function to obtain a cross-entropy loss value based on the feature vector associated with the sample modulation signal and the true modulation type label; use the regularization loss function to calculate the mean of the modulation type features corresponding to the second adversarial sample, the modulation type features corresponding to the sample modulation signal, and the modulation type features corresponding to the first adversarial sample to obtain a regularization loss value; and sum the cross-entropy loss value and the regularization loss value to obtain the adversarial loss value.

[0137] According to an embodiment of this application, the third obtaining unit includes: a fourth obtaining subunit, used to extract features from the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample respectively, to obtain modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample respectively, wherein the first masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample are obtained by masking the sample modulation signal, the first adversarial sample, and the second adversarial sample respectively; a fifth obtaining subunit, used to obtain a contrastive learning loss value based on the contrastive learning loss function, according to the modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample respectively, and the real modulation type label; and a sixth obtaining subunit, used to update the network parameters of the target deep learning model based on the contrastive learning loss value, to obtain a modulation type recognition model.

[0138] According to an embodiment of this application, the contrastive learning loss function includes a mask regularization loss function, a contrastive loss function, and a cross-entropy loss function. The fifth sub-unit is further configured to: use the mask regularization loss function to obtain a mask regularization loss value based on the modulation type features corresponding to the second mask adversarial sample, the modulation type features corresponding to the mask sample modulation signal, and the modulation type features corresponding to the first mask adversarial sample; treat the first mask adversarial sample and the second mask (which are of different types from the true modulation type label of the sample modulation signal) as negative sample pairs, treat the sample modulation signal as a positive sample, and process the negative sample pairs and positive samples using the contrastive loss function to obtain a contrastive loss value; use the cross-entropy loss function to obtain a cross-entropy loss value based on the modulation type features corresponding to the sample modulation signal and the true modulation type label; and sum the mask regularization loss value, the contrastive loss value, and the cross-entropy loss value to obtain the contrastive learning loss value.

[0139] According to an embodiment of this application, the fourth obtaining subunit is further configured to: randomly generate a mask matrix associated with a preset loss rate based on a preset loss rate; and perform masking processing on the sample modulation signal, the first adversarial sample, and the second adversarial sample based on the mask matrix to obtain a masked sample modulation signal corresponding to the sample modulation signal, a first masked adversarial sample corresponding to the first adversarial sample, and a second masked adversarial sample corresponding to the second adversarial sample.

[0140] According to an embodiment of this application, the iteration unit includes: a first iteration subunit, used to differentiate the objective function to obtain the differentiated function, and based on the differentiated function, to obtain the gradient value of the sample modulation signal corresponding to the current iteration according to the feature vector of the sample modulation signal corresponding to the current iteration, wherein the objective function is a function calculated by the feature vector of the sample modulation signal corresponding to each iteration; a second iteration subunit, used to obtain the sample modulation signal corresponding to the next iteration according to the gradient value of the sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration; and a third iteration subunit, used to determine the sample modulation signal that meets predetermined conditions as the first adversarial sample.

[0141] According to an embodiment of this application, the second iterative subunit is further configured to: perform sign processing on the gradient value of the sample modulation signal corresponding to the current iteration to obtain the perturbation direction; based on a preset perturbation step size, perturb the sample modulation signal corresponding to the current iteration along the perturbation direction to obtain the intermediate sample modulation signal corresponding to the current iteration; sum the intermediate sample modulation signal and the sample modulation signal corresponding to the current iteration to obtain the target sample modulation signal corresponding to the current iteration; and perform projection processing on the target sample modulation signal corresponding to the current iteration based on a preset perturbation radius to obtain the sample modulation signal corresponding to the next iteration.

[0142] According to embodiments of this application, any plurality of modules in the acquisition module 510 and the identification module 520 may be combined into one module, or any one of these modules may be split into multiple modules. Alternatively, at least a portion of the functionality of one or more of these modules may be combined with at least a portion of the functionality of other modules and implemented in one module. According to embodiments of this application, at least one of the acquisition module 510 and the identification module 520 may be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any appropriate combination of any of these three implementation methods. Alternatively, at least one of the acquisition module 510 and the identification module 520 may be at least partially implemented as a computer program module, which, when run, can perform corresponding functions.

[0143] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a signal modulation type identification method according to an embodiment of this application.

[0144] like Figure 6 As shown, an electronic device 600 according to an embodiment of this application includes a processor 601, which can perform various appropriate actions and processes according to a program stored in ROM 602 or a program loaded from storage portion 608 into RAM 603. The processor 601 may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor 601 may also include onboard memory for caching purposes. The processor 601 may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of this application.

[0145] RAM 603 stores various programs and data required for the operation of electronic device 600. Processor 601, ROM 602, and RAM 603 are interconnected via bus 604. Processor 601 executes various operations of the method flow according to embodiments of this application by executing programs in ROM 602 and / or RAM 603. It should be noted that the programs may also be stored in one or more memories other than ROM 602 and RAM 603. Processor 601 may also execute various operations of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0146] According to embodiments of this application, the electronic device 600 may further include an input / output (I / O) interface 605, which is also connected to a bus 604. The electronic device 600 may also include one or more of the following components connected to the input / output (I / O) 605: an input section 606 including a keyboard, mouse, etc.; an output section 607 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the input / output (I / O) 605 as needed. A removable medium 611, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 610 as needed so that computer programs read from it can be installed into the storage section 608 as needed.

[0147] This application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0148] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, RAM (Random Access Memory), ROM (Read Only Memory), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include ROM 602 and / or RAM 603 and / or one or more memories other than ROM 602 and RAM 603 described above.

[0149] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code enables the computer system to implement the signal modulation type identification method provided in the embodiments of this application.

[0150] When the computer program is executed by the processor 601, it performs the functions defined in the system / apparatus of this application embodiment. According to the embodiments of this application, the systems, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0151] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and downloaded and installed via the communication section 609, and / or installed from the removable medium 611. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0152] In such an embodiment, the computer program can be downloaded and installed from a network via the communication section 609, and / or installed from the removable medium 611. When the computer program is executed by the processor 601, it performs the functions defined in the system of this application embodiment. According to the embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0153] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, languages ​​such as Java, C++, Python, "C", or similar programming languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0154] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0155] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

[0156] The embodiments of this application have been described above. However, these embodiments are merely illustrative and not intended to limit the scope of this application. Although various embodiments have been described above, this does not mean that the measures in the various embodiments cannot be used advantageously in combination. Without departing from the scope of this application, those skilled in the art can make various substitutions and modifications, all of which should fall within the scope of this application.

Claims

1. A signal modulation type identification method, characterized by, The method includes: Acquire the modulation signal to be identified; The modulation type of the modulation signal to be identified is obtained by using a modulation type identification model. The modulation type identification model is trained using the following method: The sample modulation signal is subjected to iterative processing under predetermined conditions, and the sample modulation signal that satisfies the predetermined conditions is determined as the first adversarial sample; The second adversarial sample is obtained by averaging the modulated sample signal and the first adversarial sample. Based on the adversarial loss value obtained according to the sample modulation signal, the first adversarial sample, the second adversarial sample and the real modulation type label, the parameters of the initial deep learning model are adjusted to obtain the target deep learning model; Based on the contrastive learning loss value obtained according to the sample modulation signal, the first adversarial sample, the second adversarial sample and the real modulation type label, the parameters of the target deep learning model are adjusted to obtain the modulation type recognition model.

2. The method of claim 1, wherein, The step of adjusting the parameters of the initial deep learning model based on the adversarial loss value obtained according to the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label to obtain the target deep learning model includes: Feature extraction is performed on the sample modulation signal, the first adversarial sample, and the second adversarial sample respectively to obtain the modulation type features corresponding to each of the sample modulation signal, the first adversarial sample, and the second adversarial sample. Using the adversarial loss function, the adversarial loss value is obtained based on the modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample, and the real modulation type label. Based on the adversarial loss value, the network parameters of the initial deep learning model are updated to obtain the target deep learning model.

3. The method of claim 2, wherein, The adversarial loss function includes the cross-entropy loss function and the regularization loss function. The step of using the adversarial loss function to obtain the adversarial loss value based on the modulation type features corresponding to the sample modulation signal, the first adversarial sample, and the second adversarial sample, and the real modulation type label, includes: Using the cross-entropy loss function, the cross-entropy loss value is obtained based on the modulation type feature corresponding to the first adversarial sample and the real modulation type label; Using the regularization loss function, the mean values ​​of the modulation type features corresponding to the second adversarial sample, the modulation type features corresponding to the sample modulation signal, and the modulation type features corresponding to the first adversarial sample are calculated to obtain the regularization loss value. The adversarial loss value is obtained by summing the cross-entropy loss value and the regularization loss value.

4. The method of claim 1, wherein, The step of obtaining the modulation type recognition model by adjusting the parameters of the target deep learning model based on the sample modulation signal, the first adversarial sample, the second adversarial sample, and the real modulation type label includes: Feature extraction is performed on the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample respectively to obtain the modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample respectively. The masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample are obtained by masking the sample modulation signal, the first adversarial sample, and the second adversarial sample respectively. The contrastive learning loss function is used to obtain the contrastive learning loss value based on the modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, and the real modulation type label. Based on the contrastive learning loss value, the network parameters of the target deep learning model are updated to obtain the modulation type recognition model.

5. The method of claim 4, wherein, Contrastive learning loss functions include mask regularization loss function, contrastive loss function, and cross-entropy loss function. The step of using a contrastive learning loss function to obtain the contrastive learning loss value based on the modulation type features corresponding to the masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample, and the real modulation type label, includes: Using the mask regularization loss function, the mask regularization loss value is obtained based on the modulation type features corresponding to the second mask adversarial sample, the modulation type features corresponding to the modulation signal of the mask sample, and the modulation type features corresponding to the first mask adversarial sample. The first mask adversarial sample and the second mask, which are of different types from the real modulation type label of the sample modulation signal, are used as negative sample pairs, and the sample modulation signal is used as a positive sample. The negative sample pairs and the positive samples are processed using the contrast loss function to obtain the contrast loss value. Using the cross-entropy loss function, the cross-entropy loss value is obtained based on the modulation type feature corresponding to the first adversarial sample and the real modulation type label; The contrastive learning loss value is obtained by summing the mask regularization loss value, the contrastive loss value, and the cross-entropy loss value.

6. The method of claim 4, wherein, The masked sample modulation signal, the first masked adversarial sample, and the second masked adversarial sample are obtained by masking the sample modulation signal, the first adversarial sample, and the second adversarial sample, respectively, including: Based on a preset loss rate, a mask matrix associated with the preset loss rate is randomly generated; Based on the mask matrix, the sample modulation signal, the first adversarial sample, and the second adversarial sample are masked respectively to obtain the masked sample modulation signal corresponding to the sample modulation signal, the first masked adversarial sample corresponding to the first adversarial sample, and the second masked adversarial sample corresponding to the second adversarial sample.

7. The method of any one of claims 1-5, wherein, The iterative processing of the sample modulation signal under predetermined conditions, and the determination of the sample modulation signal that satisfies the predetermined conditions as the first adversarial sample, includes: The objective function is differentiated to obtain the differentiated function. Based on the differentiated function, the gradient value of the sample modulation signal corresponding to the current iteration is obtained according to the feature vector of the sample modulation signal corresponding to the current iteration. The objective function is a function calculated by the feature vector of the sample modulation signal corresponding to each iteration. Based on the gradient value of the sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration, the sample modulation signal corresponding to the next iteration is obtained; The sample modulation signal that meets the predetermined conditions is determined as the first adversarial sample.

8. The method of claim 7, wherein, The step of obtaining the sample modulation signal corresponding to the next iteration based on the gradient value of the sample modulation signal corresponding to the current iteration and the sample modulation signal corresponding to the current iteration includes: The gradient value of the sample modulation signal corresponding to the current iteration is sign-processed to obtain the perturbation direction; Based on a preset perturbation step size, the sample modulation signal corresponding to the current iteration is perturbed along the perturbation direction to obtain the intermediate sample modulation signal corresponding to the current iteration. The target sample modulation signal corresponding to the current iteration is obtained by summing the intermediate sample modulation signal and the sample modulation signal corresponding to the current iteration. Based on a preset disturbance radius, the target sample modulation signal corresponding to the current iteration is projected to obtain the sample modulation signal corresponding to the next iteration.

9. A signal modulation type identifying apparatus characterized by comprising: The device includes: The acquisition module is used to acquire the modulation signal to be identified; The identification module is used to identify the type of the modulation signal to be identified using a modulation type identification model, so as to obtain the modulation type of the modulation signal to be identified; The modulation type identification model is trained using the following method: The sample modulation signal is subjected to iterative processing under predetermined conditions, and the sample modulation signal that satisfies the predetermined conditions is determined as the first adversarial sample; The second adversarial sample is obtained by averaging the modulated sample signal and the first adversarial sample. Based on the adversarial loss value obtained according to the sample modulation signal, the first adversarial sample, the second adversarial sample and the real modulation type label, the parameters of the initial deep learning model are adjusted to obtain the target deep learning model; Based on the contrastive learning loss value obtained according to the sample modulation signal, the first adversarial sample, the second adversarial sample and the real modulation type label, the parameters of the target deep learning model are adjusted to obtain the modulation type recognition model.

10. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 8.

Images