Multi-task signal recognition method based on auxiliary information extraction and joint training

Abstract

The application discloses a kind of multi-task signal identification methods based on auxiliary information extraction and joint training, it is related to communication signal processing method technical field.The method includes the following steps: auxiliary network structure design: design an SNR estimation network as auxiliary network, for extracting auxiliary information from wireless signal sample;Auxiliary network pre-training: based on wireless signal sample, the training of auxiliary network is carried out;Multi-task network structure design: design a multi-task network, including a shared backbone network and multiple task branch networks;Unfrozen prior collaborative optimization adversarial training: based on collaborative adversarial training strategy, the training of auxiliary network and multi-task network is completed;Multi-task network deployment: the auxiliary network and multi-task network obtained by training are deployed to hardware device, and wireless signal is identified.The method has the advantages of high recognition accuracy and good robustness.

Description

Technical Field

[0001] This invention relates to the field of communication signal processing methods, and in particular to a multi-task signal recognition method based on auxiliary information extraction and joint training. Background Technology

[0002] Deep learning technology has been applied to various sub-fields of wireless communication, enabling automatic modulation recognition, signal type identification, and transmission channel modeling. For a given task, firstly, the input and output data need to be determined based on the task's characteristics; secondly, a deep neural network model is designed to map the input to the output; then, the deep neural network model is trained and used for task inference. Generally, a separate deep neural network model is needed for each task. For multiple tasks, multiple deep neural network models need to be designed, trained, and deployed separately. This increases design complexity and significantly increases the total amount of hardware and software resources required for deployment. To reduce complexity, a multi-task learning approach can be used, designing and training a common feature extraction network for multiple tasks, and then designing a task branch for each task. Through task association, information sharing between tasks can be achieved, enabling a synergistic improvement in model performance, generalization ability, and engineering efficiency under limited resources and data constraints.

[0003] Employing a multi-task architecture for wireless signal recognition is a current development trend. The paper "Multi-task learning approach for automatic modulation and wireless signal classification" (IEEE International Conference on Communications. (ICC), 2021) proposes a multi-task learning framework (referred to as MTL-CNN in this invention), which uses a convolutional neural network and designs two branches to complete the relevant tasks. Existing technology 1, "MTL-SRN: Multi-task Learning-based Signal Recognition Network" (IEEE Global Communications Conference. (GLOBECOM), 2024), proposes a multi-task learning framework called MTL-SRN, which uses an improved ResNet-34 as a shared feature extraction network and introduces an attention mechanism to improve feature representation capabilities. Existing technology 2, "Multitask Vehicle Signal Recognition With Dual‐Speed ​​Adaptive Weighting" (Journal of Advanced Transportation, 2025), proposes a multi-task learning framework called DSR-CNN-LSTM, which uses a hybrid architecture of CNN and LSTM for feature extraction and designs a dynamic weight adjustment strategy. Existing technology 3, "Improving the performance of wireless signal recognition using dimensionality reduction–based classifiers for low SNR signals" (International Journal of Communication Systems, 2025), proposes a multi-task learning framework called DR-MLC, which combines dimensionality reduction techniques, noise-suppressed projection techniques, and traditional machine learning classifiers to complete relevant tasks. However, from the perspective of multi-task wireless signal recognition, current work still has two problems: 1) The multi-task architecture still faces a severe threat from adversarial examples, and even small perturbations can significantly reduce recognition accuracy; 2) Directly adopting an adversarial training architecture can recover recognition accuracy to some extent and mitigate the impact of adversarial examples, but its accuracy is still low. Summary of the Invention

[0004] The technical problem to be solved by this invention is how to provide a recognition method with high accuracy and good robustness.

[0005] A multi-task signal recognition method based on auxiliary information extraction and joint training.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a multi-task signal recognition method based on auxiliary information extraction and joint training, comprising the following steps:

[0007] S101, Auxiliary Network Structure Design: Design an SNR estimation network as an auxiliary network to extract auxiliary information from wireless signal samples;

[0008] S102, Auxiliary Network Pre-training: The auxiliary network is trained based on wireless signal samples;

[0009] S103, Multi-task network architecture design: Design a multi-task network, including a shared backbone network and multiple task branch networks;

[0010] S104, Unfreezing Priors: Collaborative Optimization Adversarial Training: Based on the collaborative adversarial training strategy, complete the training of the auxiliary network and the multi-task network;

[0011] S105, Multi-task Network Deployment: Deploy the trained auxiliary network and multi-task network onto the hardware device to identify wireless signals.

[0012] The beneficial effects of the above technical solution are as follows: The method designs an SNR estimation network that extracts SNR information end-to-end from the IQ signal and encodes it into a conditional vector, which serves as prior information to guide the decision-making process of the two task branches: modulation classification and signal type identification. Then, a co-optimized adversarial training strategy is proposed. This strategy, based on a high-performance, lightweight multi-task network, uses parameter grouping and differential learning rate mechanisms during training to enable the SNR estimation network and the multi-task network to co-optimize, thereby enhancing the model's adaptability in adversarial environments. Experiments on the RadCom dataset show that, compared with existing models, the proposed method exhibits better robustness. Attached Figure Description

[0013] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0014] Figure 1 This is the main flowchart of the method described in the embodiments of the present invention;

[0015] Figure 2 This is a detailed flowchart of the method described in the embodiments of the present invention;

[0016] Figure 3 This is a schematic diagram of the SNR estimation network in the method described in the embodiments of the present invention;

[0017] Figure 4 This is a schematic diagram of the structure of the multi-task network in the method described in the embodiment of the present invention;

[0018] Figure 5 This is a performance trend curve of the modulation recognition task AT-PGD as a function of SNR in an embodiment of the present invention;

[0019] Figure 6 This is a performance trend curve of the signal classification task AT-PGD as SNR changes in an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0022] Example 1:

[0023] Overall, such as Figure 1 As shown in the figure, this invention discloses a multi-task signal recognition method based on auxiliary information extraction and joint training. The method includes the following steps:

[0024] S101, Auxiliary Network Structure Design: Design an SNR estimation network as an auxiliary network to extract auxiliary information from wireless signal samples;

[0025] S102, Auxiliary Network Pre-training: The auxiliary network is trained based on wireless signal samples;

[0026] S103, Multi-task network architecture design: Design a multi-task network, including a shared backbone network and multiple task branch networks;

[0027] S104, Unfreezing Priors: Collaborative Optimization Adversarial Training: Based on the collaborative adversarial training strategy, complete the training of the auxiliary network and the multi-task network;

[0028] S105, Multi-task Network Deployment: Deploy the trained auxiliary network and multi-task network onto the hardware device to identify wireless signals.

[0029] The above steps will be explained in detail below with specific methods:

[0030] The overall process of the SA-MTR (SNR-Aware Multi-Task Robust Signal Recognition) method described in this application is as follows: Figure 2 As shown, by explicitly modeling and utilizing SNR, a key channel state information, SA-MTR provides environmental awareness for modulation classification and signal identification tasks, thereby enhancing the model's robustness against adversarial attacks. From a network structure perspective, SA-MTR mainly consists of two parts: an SNR estimation network and a multi-task network. The SNR estimation network extracts prior SNR information from the IQ signal and encodes it into a conditional vector. The multi-task network uses shared features from a lightweight ResNet feature extraction network and fuses this conditional vector in the modulation classification head and signal type identification head to achieve joint decision-making adaptively to the channel state. Based on this, SA-MTR employs a collaborative optimization adversarial training strategy, enabling the SNR estimation network and the multi-task network to jointly optimize in an adversarial environment, thus achieving robust multi-task wireless signal identification under adversarial conditions.

[0031] SNR estimation network:

[0032] Introducing prior SNR information can help the model adapt to different channel conditions and improve its robustness in complex environments. The structure of the SNR estimation network is as follows: Figure 3 As shown, a progressive feature extraction design is adopted. First, three convolutional layers are used to gradually fuse the spatiotemporal features of the IQ signal. After mapping through a fully connected layer, a dual-task output head is used to output discrete classification results (20 SNR intervals) and continuous regression values ​​(SNR dB values) respectively. This design enables the model to learn clear class boundaries and accurate numerical mappings simultaneously.

[0033] The estimation network was trained on a clean RadCom dataset using the following composite loss function:

[0034]

[0035] in, For sparse classification cross-entropy loss, The mean squared error loss is represented by a coefficient of 0.35 to balance the magnitudes of the two losses. The SNR classification probability is then used. and regression value Feature fusion is performed, concatenating the two features into a 21-dimensional original feature vector, which is then projected onto a 64-dimensional space through a fully connected layer and normalized.

[0036]

[0037] Where W and b are the weights and biases of the fully connected layer, ReLU introduces nonlinearity, and LN represents layer normalization; / 20 is the scaling factor for the regression values ​​(for numerical stability), resulting in a 64-dimensional SNR conditional vector. This conditional vector effectively encodes the channel state information of the current signal and is injected as prior information into the multi-task classification head, guiding the model to adaptively adjust its decisions based on the SNR condition.

[0038] Multitasking Networks:

[0039] Multi-task learning (MTL) learns multiple related tasks simultaneously by sharing representations, mainly including two paradigms: hard parameter sharing and soft parameter sharing. This application adopts a hard parameter sharing architecture, which allows multiple tasks to jointly learn common feature representations by sharing underlying parameters.

[0040] Multi-task network architecture such as Figure 4 As shown, a ResNet-based feature extraction network and a dual-task classification head design are employed. The input is a 16×16 image reconstructed from I / Q signals. The feature extraction network uses a three-stage residual structure: the first stage uses 16 3×3 filters to maintain the 16×16 input resolution; the second stage uses 32 filters, downsampling to 8×8 resolution with a stride of 2; and the third stage uses 64 filters, further downsampling to 4×4 resolution. Each stage includes two basic residual blocks equipped with batch normalization and ReLU activation functions, and dimensionality matching is achieved at the residual connections using 1×1 convolutions. Global average pooling is used at the end of the network to obtain a 64-dimensional shared feature vector.

[0041] The dual-task classification heads are responsible for modulation classification and signal type identification tasks, respectively. Each classification head concatenates the 64-dimensional backbone features with the 64-dimensional SNR conditional vector to form a 128-dimensional fusion feature, which is then transformed through a 256-dimensional fully connected layer. ReLU activation and 0.5 Dropout are used to prevent overfitting, and finally, the task-specific classification result is output.

[0042] To effectively balance the training process of modulation classification and signal recognition tasks, a multi-task loss function based on homoscedastic uncertainty theory is adopted, which is defined as:

[0043]

[0044] in and These are the cross-entropy losses for modulation classification and signal recognition, respectively. and This is a trainable homoscedastic uncertainty parameter. The loss function automatically adjusts the loss weights for each task based on this uncertainty parameter. When the task training difficulty is high, the corresponding... As the value increases, the weighting coefficient The gradient is reduced accordingly, thereby balancing the learning progress of different tasks and effectively alleviating the gradient conflict problem in multi-task learning.

[0045] Collaborative optimization adversarial training strategy:

[0046] To improve the robustness of the model under adversarial perturbations, this application proposes an innovative adversarial training strategy. In deep learning models, a common practice is to optimize the auxiliary network separately from the main network or fix its parameters. Through system-wide comparative experiments, this application adopts a more effective strategy: allowing the SNR estimation network and the multi-task network to be co-optimized during adversarial training.

[0047] Basics of combat training:

[0048] This application employs the Projected Gradient Descent (PGD) method to generate adversarial examples. Starting from random initialization... Initially, the disturbance passed through Iterative updates are performed step by step, and the update formula is as follows:

[0049]

[0050] The final adversarial example is constructed as follows ,in To disrupt the budget, Step size, This represents the total number of PGD iterations. The update process aims to maximize the multi-task loss. The objective is to apply an adversarial perturbation to the image reconstructed from the IQ signal, and this perturbed representation serves as the input to a multi-task network.

[0051] Collaborative optimization mechanism:

[0052] Simple end-to-end collaborative training faces three challenges: (1) Gradient scale imbalance: Due to the heterogeneity of network structure, the gradient magnitude of the SNR estimation network is much larger than that of the backbone multi-task network, resulting in drastic parameter drift; (2) Optimization objective conflict: Discriminative learning of the multi-task network and metric learning of the SNR subnet antagonize each other at the feature level; (3) Unstable training convergence: The high variance gradient of PGD superimposes with the first two problems, forming a positive feedback that leads to training oscillation. To address these challenges, this application designs a refined grouping optimization mechanism:

[0053] Parameter grouping and independent optimization: dividing parameters into multi-task network parameters And SNR to estimate network parameters And configure independent Adam optimizers for the two sets of parameters. (2) Differentiated learning rate strategy: The learning rate for updating the multi-task network parameters is SNR estimation of network parameters uses a small learning rate. (3) Collaborative update process: In each training step, based on adversarial examples... Calculate the gradient of the loss with respect to the two sets of parameters, and after range-checking the gradient, update the parameters independently using their respective optimizers.

[0054] Through the collaborative optimization of the SNR estimation network and the multi-task network, the model can achieve a joint enhancement of channel perception and classification in adversarial environments. The SNR estimation network transforms from a fixed feature extractor into an adaptive environment perceiver, learning to provide more accurate channel state estimates under adversarial perturbations, ultimately significantly improving the overall model's adversarial robustness.

[0055] The complete core process of collaborative optimization adversarial training can be summarized as the algorithm:

[0056] enter Represents the dataset Each sample in the sample has a modulation type label. Signal type label And the actual SNR value label. Step 1: Randomly initialize the multi-task network parameters. And SNR to estimate network parameters Step 2: Initialize the Adam optimizer for the multi-task network and the SNR estimation network. and Set the learning rates for the two as follows: , In step 3, E represents the set number of training epochs. In step 4, the dataset is... Take a batch of data. In each training cycle, steps 5 and 6 respectively reconstruct the I / Q signal x into an image representation. (16×16 dimensions) and IQ representation (Dimensions are 2×128). Step 7: From IQ representation Extracting the SNR condition vector Step 8 generates adversarial examples by performing a PGD attack on the image representation. Steps 9 and 10 respectively clean the sample Adversarial Examples Combining SNR condition vector Inputting the data into a multi-task network for forward propagation yields modulation classification predictions for clean samples. Signal type classification prediction Modulation-based classification prediction of adversarial examples Signal type classification prediction Step 11: Calculate the clean sample loss. Step 12: Calculate the adversarial example loss. Step 13 involves weighted combination of the multi-task losses from clean samples and adversarial samples to calculate the total adversarial training loss. ,in These are weighting coefficients with values ​​between 0 and 1. Step 14 involves independently updating the multi-task network parameters using their respective Adam optimizers. And SNR to estimate network parameters Step 17 returns the final parameters obtained from training.

[0057] The SA-MTR collaborative optimization adversarial training algorithm in this application includes the following steps:

[0058] Input: Dataset ,in Raw I / Q signals, multi-task network SNR estimation network Clean sample weights PGD ​​parameters: , ,

[0059] Output: Robust multi-task network parameters SNR estimates network parameters ;

[0060] step:

[0061] 1) Random initialization , ;

[0062] 2) Initialize the optimizer , The learning rates are respectively , ;

[0063] 3) for epoch = 1 toE do;

[0064] 4) for batch from do;

[0065] 5) ;

[0066] 6) ;

[0067] 7) ;

[0068] 8) ;

[0069] 9) 10) ;

[0070] 11) ;

[0071] 12) ;

[0072] 13) ;

[0073] 14) ;

[0074] 15) endfor;

[0075] 16) end for;

[0076] 17) Return the optimal parameters ;

[0077] Experimental verification and result analysis:

[0078] Dataset:

[0079] (1) The RadComAWGN dataset is a public dataset for multi-task learning of wireless signals. The dataset includes 9 modulation formats and 11 signal classes (the first 5 classes are radar waveforms and the last 6 classes are communication waveforms). The last 3 classes of communication signals (Bluetooth, IEEE802.11bg, IEEE802.15.4) are derived from public interference datasets. The remaining waveforms are generated by GNU Radio software and superimposed with additive white Gaussian noise. The signal-to-noise ratio covers -20 dB to 18 dB (in 2 dB steps). This application mainly performs hyperparameter evaluation and model comparison on this dataset.

[0080] (2) RadComDynamic: This dataset is based on RadComAWGN, but excludes the three types of signals from the interference dataset mentioned above. In addition to AWGN, it further introduces real propagation effects and hardware impairments, including multipath, fading, Doppler shift, oscillator drift, and sampling clock offset. The propagation channel is a Ricean channel (K factor = 3). This dataset is used to evaluate the model's generalization ability in more complex and realistic scenarios.

[0081] Table 1: Classification of RadCom series datasets

[0082]

[0083] The sample was divided into 70% (training), 20% (validation), and 10% (testing), and stratified sampling by signal category was used to ensure distribution consistency.

[0084] Parameter settings and comparison indicators:

[0085] In the signal preprocessing stage, after loading the dataset, the original 128-dimensional complex IQ signal is first separated into real and imaginary parts to obtain a 256-dimensional real vector, which is then normalized by unit energy to eliminate residual artifacts caused by analog propagation effects. The normalized 256-dimensional real vector is further reshaped into a 16×16 image view (input to the multi-task network) and a 2×128 IQ time series view (input to the SNR estimation network).

[0086] The experiment included two training modes: Standard Training (ST) and Adversarial Training (AT). Standard Training optimized the model based on clean samples and monitored the accuracy on the clean validation set. Adversarial Training mixed clean samples with adversarial samples generated by PGD attacks in equal proportions in each training batch to optimize the model's adversarial robustness. The accuracy on the adversarial samples on the validation set was monitored during training. During the training phase, adversarial training employed... Generate adversarial examples for norm-constrained PGD attacks, setting perturbation budgets. =0.10, attack step size = / 10 = 0.01, iteration count T = 10, and a random starting point is enabled. During the testing phase, clean samples and adversarial examples are used to evaluate both training modes; a stronger PGD attack is employed ( =0.20, T=20, with random starting point enabled) to evaluate the robustness of the model.

[0087] Training The warp-up strategy linearly increases the perturbation budget from 0 to the target value over the first 5 epochs, while gradient clipping (range [-1.0, 1.0]) controls the parameter update magnitude. The Adam optimizer is used, and the learning rate for the multi-task network parameters is... SNR estimation of network parameters uses a small learning rate. An early stopping mechanism was implemented, with monitoring based on validation set performance metrics for the corresponding training mode, and a patience value of 5 epochs. The learning rate scheduling employed the ReduceLROnPlateau strategy with a factor of 0.5 and a patience value of 2 epochs. The maximum number of training epochs was 30, and the batch size was 256. A fixed random seed was used to ensure reproducibility. All experiments were implemented using TensorFlow 2.13.1 on an NVIDIA RTX4090 GPU.

[0088] The evaluation metrics in this application include: ST-Clean (accuracy on clean samples after standard training), ST-PGD (accuracy on PGD adversarial samples after standard training), AT-Clean (accuracy on clean samples after adversarial training), and AT-PGD (accuracy on PGD adversarial samples after adversarial training). Furthermore, this application slices the test set according to SNR values ​​and plots SNR-accuracy curves to deeply analyze the model's performance under different channel qualities.

[0089] Analysis of experimental results:

[0090] 1) Comparative analysis with existing methods

[0091] To comprehensively evaluate the advancement of the SA-MTR framework, this application compares it with all existing state-of-the-art models for modulation recognition and signal classification tasks on the RadComAWGN and RadComDynamic datasets. To ensure consistency in task settings, all comparison models were uniformly adjusted to a dual-task architecture. The comparison covers a variety of representative solutions from traditional machine learning to deep learning. Due to its paradigm limitations, the traditional machine learning model DR-MLC cannot be adversarially trained; therefore, its performance metrics only provide accuracy on clean samples after standard training. The results are shown in Tables 2 and 3. Here, mod represents the modulation recognition task, and sig represents the signal type recognition task.

[0092] Table 2 Performance Comparison of Different Multitasking Networks (RadComAWGN, Accuracy: %)

[0093]

[0094] Table 3 Performance Comparison of Different Multitasking Networks (RadComDynamic, Accuracy: %)

[0095]

[0096] Analysis shows that SA-MTR exhibits comprehensive advantages across all core metrics. On the RadComAWGN dataset, compared to the parameter-intensive MTL-SRN, SA-MTR, using only about 18.7% of its parameters, improves the adversarial example accuracy (AT-PGD) after adversarial training by approximately 6.5% and 7.63%, respectively, in automatic modulation classification and signal type recognition tasks. Compared to the ultra-lightweight MTL-CNN, SA-MTR improves the adversarial example accuracy after adversarial training by approximately 13.9% and 15.82%, respectively, and also maintains an advantage in clean sample accuracy. The overall performance of DSR-CNN-LSTM falls between MTL-CNN and MTL-SRN, showing no significant advantage in classification accuracy or adversarial robustness.

[0097] Figure 5 and Figure 6 This paper presents the performance trend of AT-PGD adversarial example accuracy as a function of SNR after adversarial training on the RadComAWGN dataset. SA-MTR exhibits a stable advantage over other models across all SNR ranges. Specifically, in the low SNR region (SNR < 0 dB), the performance of all models degrades, but SA-MTR's accuracy declines most gradually, significantly outperforming the comparison models. In the medium-to-high SNR region, SA-MTR reaches its performance saturation point first. These results demonstrate that SA-MTR possesses superior environmental adaptability, achieving a balance between high accuracy and high robustness.

[0098] In summary, the SA-MTR method described in this application achieves a good balance between model accuracy and adversarial robustness, providing an effective solution for the application of multi-task wireless signal recognition in complex scenarios.

[0099] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0100] This application uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. In summary, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-task signal recognition method based on auxiliary information extraction and joint training, characterized in that... Includes the following steps: S101, Auxiliary Network Structure Design: Design an SNR estimation network as an auxiliary network to extract auxiliary information from wireless signal samples; S102, Auxiliary Network Pre-training: The auxiliary network is trained based on wireless signal samples; S103, Multi-task network architecture design: Design a multi-task network, including a shared backbone network and multiple task branch networks; S104, Unfreezing Priors: Collaborative Optimization Adversarial Training: Based on the collaborative adversarial training strategy, complete the training of the auxiliary network and the multi-task network; S105, Multi-task Network Deployment: Deploy the trained auxiliary network and multi-task network onto the hardware device to identify wireless signals.

2. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 1, characterized in that, In step S101: The input to the auxiliary network is wireless signal samples, and the output is auxiliary information; The wireless signal sample refers to a labeled sample used for wireless signal identification. The number of labels is the same as the number of tasks. The wireless signal is represented in the form of an IQ sequence or a constellation diagram.

3. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 1, characterized in that, The SNR estimation network: A progressive feature extraction design is adopted, which uses three convolutional layers to gradually fuse the spatiotemporal features of wireless signals. After mapping through a fully connected layer, the back-end uses a dual-task output head to output discrete classification results and continuous regression values ​​respectively.

4. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 1, characterized in that: In step S102, the auxiliary network is trained on the RadCom dataset using the following composite loss function: ； in, For sparse classification cross-entropy loss, The mean squared error loss is represented by a coefficient of 0.35, which is used to balance the magnitudes of the two losses. Classification probability of SNR and regression value Feature fusion is performed, concatenating the two features into a 21-dimensional original feature vector, which is then projected onto a 64-dimensional space through a fully connected layer and normalized. ； Where W and b are the weights and biases of the fully connected layer, ReLU introduces nonlinearity, and LN represents layer normalization; / 20 is the scaling factor for the regression values, resulting in a 64-dimensional SNR condition vector. .

5. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 1, wherein the multi-task network in step S103 is: A ResNet-based feature extraction network and a dual-task classification head design are adopted. The input is a 16×16 image reconstructed from I / Q signals. The feature extraction network adopts a three-stage residual structure: the first stage uses 16 3×3 filters to maintain the 16×16 input resolution; the second stage uses 32 filters to downsample to 8×8 resolution with a stride of 2; and the third stage uses 64 filters to further downsample to 4×4 resolution. Each stage includes two basic residual blocks equipped with batch normalization and ReLU activation functions. Dimension matching is achieved at the residual connection points through 1×1 convolution. Global average pooling is used at the end of the network to obtain a 64-dimensional shared feature vector. The dual-task classification heads are responsible for modulation classification and signal type identification tasks, respectively. Each classification head concatenates the 64-dimensional backbone features with the 64-dimensional SNR conditional vector to form a 128-dimensional fusion feature. Then, the feature is transformed through a 256-dimensional fully connected layer. ReLU activation and 0.5 Dropout are used to prevent overfitting. Finally, the task-specific classification result is output.

6. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 5, characterized in that: A multi-task loss function based on homoscedasticity uncertainty theory is adopted to balance the two tasks of modulation classification and signal recognition. The multi-task loss function is defined as follows: ； in and These are the cross-entropy losses for modulation classification and signal recognition, respectively; and is a trainable homoscedasticity uncertainty parameter.

7. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 1, wherein the collaborative optimization adversarial training in step S104 includes: It allows the estimation network and the multi-task network to co-optimize during adversarial training, enabling the estimation network to dynamically adapt to channel changes under perturbation, thereby achieving the co-evolution of channel estimation and main task identification in an adversarial environment.

8. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 7, wherein the collaborative optimization adversarial training in step S104 specifically includes: 1) Basics of combat training: Adversarial examples are generated using the Projected Gradient Descent (PGD) method. From random initialization Initially, the disturbance passed through Iterative updates are performed step by step, and the update formula is as follows: ； The final adversarial example is constructed as follows ,in To disrupt the budget, Step size, This represents the total number of PGD iterations; the update process aims to maximize the multi-task loss. The objective is to apply an adversarial perturbation to an image reconstructed from the IQ signal, and this perturbed representation is used as input to a multi-task network. 2) Collaborative optimization mechanism: Parameter grouping and independent optimization: dividing parameters into multi-task network parameters And estimating network parameters And configure independent Adam optimizers for the two sets of parameters; Differential learning rate strategy: The learning rate for updating the parameters of a multi-task network is... The estimated learning rate of the network parameters is ; Collaborative update process: In each training step, based on adversarial examples... Calculate the gradient of the loss with respect to the two sets of parameters respectively, and update the parameters independently using their respective optimizers after range determination of the gradient; By optimizing the estimation network and the multi-task network together, the model can achieve a joint enhancement of channel perception and classification in adversarial environments.

9. The multi-task signal recognition method based on auxiliary information extraction and joint training as described in claim 8, characterized in that, Collaborative optimization adversarial training includes the following steps: Randomly initialize multi-task network parameters And SNR to estimate network parameters ; Adam optimizer for initializing multi-task networks and SNR estimation networks and Set the learning rates for the two as follows: , ; From the dataset Take a batch of data from the dataset, and in each training cycle, reconstruct the I / Q signal x into an image representation. and IQ ,in Represents the dataset Each sample in the dataset is labeled with a modulation type. Signal type label The labels are the actual SNR values, where i represents the index of the sample and N represents the number of samples. From IQ representation Extracting the SNR condition vector Adversarial examples are generated by performing PGD attacks on image representations. ; Clean samples respectively Adversarial Examples Combining SNR condition vector Inputting the data into a multi-task network for forward propagation yields modulation classification predictions for clean samples. Signal type classification prediction Modulation-based classification prediction of adversarial examples Signal type classification prediction ; Calculate clean sample loss and adversarial sample loss The total adversarial training loss is calculated by weighting and combining the multi-task losses of clean samples and adversarial samples. ; Update multi-task network parameters independently using their respective Adam optimizers. And SNR to estimate network parameters Then, return to the training to obtain the final parameters.

Images