A generalization defense method for electromagnetic signal recognition model in adversarial domain
Through the dual-stream recognition model and adversarial training technology, the adversarial domain generalization defense method solves the adversarial perturbation problem caused by the uninterpretation of deep learning models in modulation classification, and improves the robustness of the model and adversarial sample recognition ability.
Patent Information
- Application Number
- CN202310477430.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-28
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2043-04-28
AI Technical Summary
The uninterpretation of deep learning models in modulation classification makes their results susceptible to imperceptible perturbations added to the input data, and existing defense methods are inadequate in their anti-rosability.
Adopting a dual-stream recognition model architecture, including low-level feature extractors, advanced feature extractors, domain generalized feature extractors and classifiers, data augmentation through popular hybrid algorithms, generate adversarial samples, and aligning tasks through adversarial training and maximum mean square deviation distance. Related features enhance the adversarial robustness of the model.
It improves the adversarial robustness of the model, enhances the recognition ability of the adversarial sample, and ensures the domain independence of the feature extractor and the perturbation resistance of the overall model.
Smart Images

Figure CN117113073B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to an adversarial domain generalization defense method, and belongs to the technical field of communications. Background Art
[0002] With the continuous development of communication technology, the sixth generation mobile system (6G) will have higher data transmission speed, lower latency and higher spectrum efficiency, playing an important role in the development of human civilization. Automatic modulation classification (AMC) technology is one of the indispensable key technologies for the development and application of 6G. Modulation classification helps to better understand the signal property and transmission environment, allowing the optimization of channel models and modulation methods, improving spectrum efficiency and data transmission speed. Modulation classification can also be used to detect and identify abnormal users in the network, thereby improving network security. In recent years, deep learning (DL) technology has been widely used in modulation classification and has achieved remarkable results, becoming one of the most promising technical directions. However, the unexplainability of DL models makes their results easily affected by adding carefully designed imperceptible perturbations to the input data.
[0003] In order to improve the robustness of the AMC model, various defense methods have been proposed in succession, trying to enhance the neighborhood robustness of the AMC model, thereby improving the adversarial robustness. The introduction of these methods effectively reduces the threat of adversarial samples to the AMC model. These methods either only consider the defense problem from a computational perspective, or treat adversarial samples as training data, or simply treat adversarial perturbations as ordinary noise. There are still many unresolved challenges in the research field of adversarial defense. Summary of the invention
[0004] In order to solve the problem that the unexplainability of DL models makes their results easily affected by adding carefully designed imperceptible disturbances to the input data, the present invention proposes a method for generalization defense of electromagnetic signal recognition models against adversarial domains.
[0005] The technical solution adopted by the present invention to solve the above problems is: the steps of the present invention are as follows:
[0006] Step 1: Build a dual-stream recognition model architecture. Each stream includes a feature extractor and a classifier. The feature extractor is shared by two streams, including the low-level feature extractor E l , Advanced Feature Extractor E h , domain generalization feature extractor E g , classifier C1 and classifier C2;
[0007] Step 2: Train the low-level feature extractor E l , Advanced Feature Extractor E hand classifier C1 classifies the original signal domain samples, during which a popular hybrid algorithm is used for data augmentation;
[0008] Step 3: Use E l 、E h And C1 generates adversarial samples to obtain adversarial domain signal samples, and trains E l 、E h And classifier C2 classifies the adversarial domain signal samples;
[0009] Step 4: Enable the domain generalization feature extractor E g , extract task-related features of the two classifiers and align them;
[0010] Step 5: Perform adversarial training on the feature extractor and C1;
[0011] Step 6: Repeat training steps 4 and 5 until convergence.
[0012] Furthermore, in step 2, first let the original sample x s By E l Extract low-level features and then use popular hybrid algorithms to augment the data:
[0013]
[0014]
[0015] In formulas (1) and (2), S represents the original signal domain, x s represents the data sample in S, y s represents the label sample in S, f l s =E l (x s ), f l s′ Yes l s The out-of-order copy of y s and s′ is the corresponding label in one-hot encoding form, is the data after data expansion, is the label after data expansion, λ~Beta(α,α) is the mixing coefficient, α=1;
[0016] f l s and Combine and pass to E h In order to extract high-level features, the prediction output of high-level features is obtained through C1, and the loss function is composed of the original signal classification loss and data augmentation loss Composition, the optimization process is:
[0017]
[0018] E in formula (3) l represents the low-level feature extractor, E h represents the high-level feature extractor, C1 represents the classifier, and S represents the original signal domain.
[0019] Furthermore, step 3 inherits the training E in step 2. l and E h Only update the parameters of C2, and the optimization process is:
[0020]
[0021] In formula (4), C2 represents the classifier and a represents the adversarial domain.
[0022] Furthermore, in step 4, the task-related features of the two classifiers are extracted. Taking the original signal as an example, the features with the label y s The corresponding prediction output is relative to the domain generalization feature The gradient is used as the relevant information of the feature for classification. A positive gradient indicates that the corresponding feature element has a positive impact on the classification, thus obtaining the task-related features of the original sample.
[0023] In formula (5) express The gradient of express The gradient of and The distribution shift of , its maximum mean square error MMD loss is:
[0024]
[0025] In formula (6), k(·,·) represents the Gaussian kernel function. In order to ensure the domain independence of the extracted features, E is trained in this step. l 、E h and E g , aligning task-related features:
[0026]
[0027] In formula (7), β is the adjustment factor, which is a non-negative hyperparameter with an unconstrained range.
[0028] Furthermore, the loss of adversarial training in step 5 is:
[0029]
[0030] In formula (8), A represents the adversarial signal domain, x a represents the data sample in the adversarial signal domain A, y a represents the label sample in the adversarial signal domain A;
[0031] After final convergence, the classifier predicts the output result as follows:
[0032] y = argmixC1(E g (E h (E l (x)))).
[0033] The beneficial effects of the present invention are:
[0034] 1. This paper analyzes the original signal and adversarial samples from the perspective of the domain, and concludes that the difference between them is reflected in the marginal distribution, and gives a system model summarized from the perspective of the domain;
[0035] 2. The present invention proposes the ADGD algorithm, which extracts the task-related features of the original signal and the adversarial sample and aligns them, thereby improving the adversarial robustness of the feature extractor;
[0036] 3. The present invention further uses adversarial training to improve the adversarial robustness of the entire model including the classifier, achieving incremental improvement in the robustness of the model;
[0037] 4. The core of the present invention is to enhance the adversarial robustness of the AMC model by aligning the task-related features of the original signal and the adversarial sample in combination with adversarial training. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic diagram of the overall model architecture of the present invention;
[0039] Figure 2 a is a schematic diagram of baseline model feature visualization;
[0040] Figure 2 b is a visualization diagram of the ADGD algorithm model features. DETAILED DESCRIPTION
[0041] Specific implementation method 1: Combination Figure 1 and Figure 2 The steps of the electromagnetic signal recognition model adversarial domain generalization defense method described in this embodiment are as follows:
[0042] Step 1: Build a dual-stream recognition model architecture. Each stream includes a feature extractor and a classifier. The feature extractor is shared by two streams, including the low-level feature extractor El , Advanced Feature Extractor E h , domain generalization feature extractor E g , classifier C1 and classifier C2;
[0043] Step 2: Train the low-level feature extractor E l , Advanced Feature Extractor E h and classifier C1 classifies the original signal domain samples, during which a popular hybrid algorithm is used for data augmentation;
[0044] Step 3: Use E l 、E h And C1 generates adversarial samples to obtain adversarial domain signal samples, and trains E l 、E h And classifier C2 classifies the adversarial domain signal samples;
[0045] Step 4: Enable the domain generalization feature extractor E g , extract task-related features of the two classifiers and align them;
[0046] Step 5: Perform adversarial training on the feature extractor and C1;
[0047] Step 6: Repeat training steps 4 and 5 until convergence.
[0048] Specific implementation method 2: Combination Figure 1 and Figure 2 In this embodiment, the electromagnetic signal recognition model described in this embodiment is used to defend against domain generalization. In step 2, the original sample x is first set to s By E l Extract low-level features and then use popular hybrid algorithms to augment the data:
[0049]
[0050]
[0051] In formulas (1) and (2), S represents the original signal domain, x s represents the data sample in S, y s represents the label sample in S, f l s =E l (x s ), f l s′ Yes l s The out-of-order copy of y s and s′ is the corresponding label in one-hot encoding form, is the data after data expansion, is the label after data expansion, λ~Beta(α,α) is the mixing coefficient, α=1;
[0052] Will and Combine and pass to E h In order to extract high-level features, the prediction output of high-level features is obtained through C1, and the loss function is composed of the original signal classification loss and data augmentation loss Composition, the optimization process is:
[0053]
[0054] E in formula (3) l represents the low-level feature extractor, E h represents the high-level feature extractor, C1 represents the classifier, and S represents the original signal domain.
[0055] Specific implementation method three: Combination Figure 1 and Figure 2 The electromagnetic signal recognition model described in this embodiment inherits the training of E in step 2 in step 3 of the adversarial domain generalization defense method. l and E h Only update the parameters of C2, and the optimization process is:
[0056]
[0057] In formula (4), C2 represents the classifier and a represents the adversarial domain.
[0058] Specific implementation method four: Combination Figure 1 and Figure 2 The present embodiment is described. In the present embodiment, the electromagnetic signal recognition model described in the present embodiment extracts the task-related features of the two classifiers in step 4 of the adversarial domain generalization defense method. Taking the original signal as an example, the task-related features of the two classifiers are extracted using the label y s The corresponding prediction output is relative to the domain generalization feature The gradient is used as the relevant information of the feature for classification. A positive gradient indicates that the corresponding feature element has a positive impact on the classification, thus obtaining the task-related features of the original sample.
[0059]
[0060] In formula (5) express The gradient of express The gradient of and The distribution shift of , its maximum mean square error MMD loss is:
[0061]
[0062] In formula (6), k(·,·) represents the Gaussian kernel function. In order to ensure the domain independence of the extracted features, E is trained in this step. l 、E h and E g , aligning task-related features:
[0063]
[0064] In formula (7), β is the adjustment factor, which is a non-negative hyperparameter with an unconstrained range.
[0065] Specific implementation method five: Combination Figure 1 and Figure 2 To illustrate this embodiment, the loss of adversarial training in step 5 of the adversarial domain generalization defense method for an electromagnetic signal recognition model described in this embodiment is:
[0066]
[0067] In formula (8), A represents the adversarial signal domain, x a represents the data sample in the adversarial signal domain A, y a represents the label sample in the adversarial signal domain A;
[0068] After final convergence, the classifier predicts the output result as follows:
[0069] y=arg mixC1(E g (E h (E l (x)))).
[0070] Example
[0071] S1. Build a dual-stream recognition model architecture. Each stream includes a feature extractor and a classifier. The feature extractor is shared by two streams, including the low-level feature extractor E l , Advanced Feature Extractor E h and domain generalization feature extractor E g , the two classifiers are C1 and C2, the original signal domain is S, x s ,y s is the data and label sample in S, and the adversarial signal domain is A, x a ,y a is the data and label sample in A;
[0072] S2. Training low-level feature extractor E l, Advanced Feature Extractor E h And classifier C1 classifies the original signal domain samples, the original sample x s By E l Extract low-level features and then use popular hybrid algorithms to augment the data:
[0073]
[0074] where f l s =E l (x s ), f l s′ Yes l s The out-of-order copy of y s and s′ is the corresponding label in one-hot encoding form, and are the data and labels after data expansion, λ~Beta(α,α) is the mixing coefficient, α=1;
[0075] Will and Combine and pass to E h In order to extract high-level features, the prediction output of high-level features is obtained through C1, and the loss function is composed of the original signal classification loss and data augmentation loss Composition, the optimization process is:
[0076]
[0077] After iterative training, a feature extractor and classifier C1 capable of classifying the original signal are obtained.
[0078] S3, use E l 、E h And C1 generates adversarial samples to obtain adversarial domain signal samples:
[0079] x a =At(x s ,E l ,E h ,C1)
[0080] Among them, At(·) represents a specific adversarial sample generation method, the adversarial perturbation is set to 0.1, and the training E l 、E h And classifier C2 classifies the adversarial domain signal samples, E l and E h Inherit the parameters trained by S2 and only update the parameters of C2. The optimization process is:
[0081]
[0082] After iterative training, a feature extractor and classifier C2 are obtained that can classify adversarial domain signals.
[0083] Main training part
[0084] S4, enable domain generalization feature extractor E g , extract the task-related features of the two classifiers. Taking the original signal as an example, use s The corresponding prediction output is relative to the domain generalization feature The gradient is used as the relevant information of the feature to classify:
[0085]
[0086] in represents the predicted output corresponding to label y, express Relative to The gradient of
[0087] The gradient is positive, indicating that the corresponding feature element has a positive impact on the classification, so the task-related features of the original sample are obtained.
[0088]
[0089] Similarly, we can get The maximum mean square deviation (MMD) distance is used to measure and The distribution shift of , its MMD loss is:
[0090]
[0091] where k(·,·) is the Gaussian kernel function. In order to ensure the domain independence of the extracted features, E is trained in this step. l 、E h and E g , align the task-related features, and the optimization process is:
[0092]
[0093] Where β is the adjustment factor, which is a non-negative hyperparameter with an unconstrained range, set β = 10;
[0094] S5, perform adversarial training on the feature extractor and C1, and the loss of adversarial training is:
[0095]
[0096] The optimization process is:
[0097]
[0098] S6, repeat training S4 and S5 until convergence. After final convergence, the classifier predicts the output result as follows:
[0099] y = argmixC1(E g (E h (E l (x)))).
[0100] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.
Claims
1. A method for defending electromagnetic signal recognition model against adversarial domain generalization, characterized by: The steps of the electromagnetic signal recognition model adversarial domain generalization defense method are as follows: Step 1: Build a dual-stream recognition model architecture. Each stream includes a feature extractor and a classifier. The feature extractor is shared by two streams, including the low-level feature extractor E l , Advanced Feature Extractor E h , domain generalization feature extractor E g , classifier C1 and classifier C2; Step 2: Train the low-level feature extractor E l , Advanced Feature Extractor E h and classifier C1 classifies the original signal domain samples, during which a popular hybrid algorithm is used for data augmentation; First, let the original sample x s By E l Extract low-level features and then use popular hybrid algorithms to augment the data: In formulas (1) and (2), S represents the original signal domain, x s represents the data sample in S, y s represents the label sample in S, f l s =E l (x s ), f l s′ Yes l s The out-of-order copy of y s and s′ is the corresponding label in one-hot encoding form, is the data after data expansion, is the label after data expansion, λ~Beta(α,α) is the mixing coefficient, α=1; f l s and Combine and pass to E h In order to extract high-level features, the prediction output of high-level features is obtained through C1, and the loss function is composed of the original signal classification loss and data augmentation loss Composition, the optimization process is: E in formula (3) l represents the low-level feature extractor, E h represents the high-level feature extractor, C1 represents the classifier, and S represents the original signal domain; Step 3: Use E l 、E h And C1 generates adversarial samples to obtain adversarial domain signal samples, and trains E l 、E h And classifier C2 classifies the adversarial domain signal samples; inherit the training E in step 2 l and E h Only update the parameters of C2, and the optimization process is: In formula (4), C2 represents the classifier, and a represents the adversarial domain; Step 4: Enable the domain generalization feature extractor E g , extract task-related features of the two classifiers and align them; Extract the task-related features of the two classifiers. Take the original signal as an example and use s The corresponding prediction output is relative to the domain generalization feature The gradient is used as the relevant information of the feature for classification. A positive gradient indicates that the corresponding feature element has a positive impact on the classification, thus obtaining the task-related features of the original sample. In formula (5) express The gradient of express The gradient of and The distribution shift of , its maximum mean square error MMD loss is: In formula (6), k(·,·) represents the Gaussian kernel function. In order to ensure the domain independence of the extracted features, E is trained in this step. l 、E h and E g , aligning task-related features: In formula (7), β is the adjustment factor, which is a non-negative hyperparameter with an unconstrained range; Step 5: Perform adversarial training on the feature extractor and C1; The loss of adversarial training is: In formula (8), A represents the adversarial signal domain, x a represents the data sample in the adversarial signal domain A, y a represents the label sample in the adversarial signal domain A; After final convergence, the classifier predicts the output result as follows: y=argmaxC1(E g (E h (E l (x)))); Step 6: Repeat training steps 4 and 5 until convergence.
Citation Information
Patent Citations
General countermeasure disturbance generation method based on electromagnetic signal modulation type identification of automatic encoder
CN112215078A
Supervised comparative learning-based electromagnetic signal open set identification method
CN115114958A
Unsupervised bearing fault diagnosis method based on adaptive residual adversarial network
CN115127814A
Metasurface unit structure inverse design method based on improved generative adversarial network
CN115169235A
Intelligent migration confrontation method and system for electromagnetic signal identification under incomplete information
CN115600083A
