Robust lightweight cross-domain distillation method for non-cooperative wireless signal modulation recognition

Abstract

A robust lightweight cross-domain distillation method for non-cooperative wireless signal modulation recognition, taking a time-frequency dual-flow robust model as a teacher model, and a lightweight model receiving only time-domain I / Q signals as a student model, through a cross-domain robust distillation mechanism, the stable soft label distribution of the teacher model on clean samples is migrated to the output distribution of the student model on perturbed samples, so that the student model can still inherit the robust discrimination ability of the time-frequency dual-flow teacher model under the condition of relying only on time-domain I / Q input in the inference stage, thereby balancing the modulation recognition accuracy, anti-perturbation robustness and lightweight inference efficiency.

Description

Technical Field

[0001] This invention relates to the fields of intelligent wireless signal identification, automatic modulation identification, model robustness enhancement, and lightweight inference technology, and particularly to a robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification. Background Technology

[0002] Automatic modulation identification (EMI) is a crucial link in the intelligent sensing link of wireless signals. It can determine the modulation type of a received signal based on the unknown modulation scheme or limited prior information, providing fundamental support for subsequent signal sorting, communication reconnaissance, jamming decision-making, and situational awareness. In application scenarios such as aerospace warfare, low-Earth orbit satellite terminal monitoring, UAV non-cooperative communication reconnaissance, airborne platforms, and software-defined radio, received signals are typically affected by factors such as high-speed maneuvering, multipath propagation, carrier frequency shift, phase perturbation, background noise, and human interference. This results in signal characteristics that are non-stationary, have low signal-to-noise ratios, and are easily disturbed, thus increasing the difficulty of modulation identification models.

[0003] In recent years, deep learning models have been widely used in automatic modulation recognition tasks. To improve the robustness of models in complex electromagnetic environments and under adversarial perturbation conditions, existing methods typically introduce time-frequency dual-stream structures, frequency domain feature generation, attention fusion, and cross-domain interaction modules to enhance the model's ability to express joint time-domain and frequency-domain features. While these methods have improved the model's defense against noise perturbations and adversarial examples to some extent, their inference phase usually requires retaining frequency domain transformation links and multi-branch network structures, resulting in high model parameter count, computational cost, storage overhead, and inference latency, making it difficult to meet the requirements of lightweight and real-time performance for constrained devices.

[0004] For airborne nodes, UAV payloads, low-Earth orbit satellite monitoring equipment, and software-defined radio platforms, their hardware resources are typically limited by factors such as payload, power consumption, computing power, storage, and real-time performance, making it difficult to directly deploy complex time-frequency dual-stream robust models. While pure time-domain lightweight models offer advantages such as simple structure, short inference links, and low deployment costs, their robustness against complex electromagnetic environments and disturbances is generally weaker than that of time-frequency dual-stream models under conventional training methods. Therefore, how to transfer the stable discrimination capability of time-frequency dual-stream robust models to pure time-domain lightweight models without significantly increasing the complexity of the inference link is a pressing technical problem to be solved in the automatic modulation and identification of non-cooperative wireless signals. Summary of the Invention

[0005] The technical problem this invention aims to solve is to address the issues of complex frequency domain operator links, large parameter scale, high inference delay, and insufficient robustness of lightweight models in existing automatic modulation recognition robust models. To address these problems, a robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation recognition is proposed. This method uses a time-frequency dual-stream robust model as the teacher model and a lightweight model that only receives time-domain I / Q signals as the student model. Through a cross-domain robust distillation mechanism, the stable soft-label distribution of the teacher model on clean samples is transferred to the output distribution of the student model on perturbed samples. This allows the student model to inherit the robust discrimination capability of the time-frequency dual-stream teacher model even when relying only on time-domain I / Q inputs during the inference stage, thus balancing modulation recognition accuracy, disturbance resistance robustness, and lightweight inference efficiency.

[0006] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] A robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation recognition is proposed. First, a robustly trained time-frequency dual-stream model is constructed as the teacher model, and a lightweight network receiving only time-domain I / Q signal inputs is constructed as the student model. Second, during offline training, the teacher model parameters are frozen, and adversarial examples are generated based on the current parameters of the student model. Then, through an adversarial example-driven cross-domain distillation mechanism, the stable soft-label distribution of the teacher model on clean samples is transferred to the output distribution of the student model on adversarial samples, enabling the student model to learn the robust discriminative ability of the teacher model. Finally, during the inference application stage, only the pure time-domain lightweight student model is retained, thereby avoiding the frequency domain branch, FFT preprocessing link, and cross-domain interaction module from entering the online inference process, reducing the number of model parameters, computational complexity, and inference latency.

[0008] A robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification includes the following steps:

[0009] S1: Collect non-cooperative wireless communication signals, and process them through radio frequency reception, down-conversion, analog-to-digital sampling, frame segmentation and energy normalization to obtain a time-domain I / Q signal sample library. Then, pair the time-domain I / Q signal samples with modulation category labels to construct an automatic modulation identification sample library.

[0010] S2: Construct a time-frequency dual-stream robust teacher model for offline training. The teacher model takes time-domain I / Q signal samples as input, and sequentially completes time-domain feature learning, frequency-domain feature learning, and attention multi-domain feature fusion. It outputs the teacher logits vector corresponding to the modulation category and trains and freezes the teacher model parameters.

[0011] S3: Construct a pure time-domain lightweight student model for the lightweight inference stage. The student model only accepts time-domain I / Q signal samples as input, without introducing frequency-domain input branches, spectrum generation modules, FFT preprocessing links, and cross-domain interaction modules, and outputs student logits vectors corresponding to the modulation category.

[0012] S4: Generate adversarial samples corresponding to clean time-domain I / Q signal samples based on the current parameters of the student model;

[0013] S5: Input the clean time-domain I / Q signal samples into the frozen teacher model to obtain the temperature soft label distribution of the teacher model;

[0014] S6: Input the adversarial examples generated in step S4 into the student model to obtain the temperature soft label distribution of the student model under perturbation conditions;

[0015] S7: Calculate the cross-domain robust distillation loss based on the temperature soft label distribution of the teacher model on clean samples and the temperature soft label distribution of the student model on adversarial samples.

[0016] S8: Calculate the supervised classification loss based on the student model's output on adversarial examples and the real modulated class labels;

[0017] S9: The cross-domain robust distillation loss and the supervised classification loss are weighted and fused to construct a joint loss function, and the student model parameters are updated according to the joint loss function;

[0018] S10: Input the time-domain I / Q signal sample to be predicted into the student model after the model parameters are updated in step S9 to obtain the modulation recognition result.

[0019] Furthermore, step S1, the process of constructing the automatic modulation recognition sample library, is as follows:

[0020] Non-cooperative wireless communication signals are collected and processed through radio frequency reception, down-conversion, analog-to-digital sampling, frame segmentation, and energy normalization to obtain time-domain I / Q signal samples. These time-domain I / Q signal samples are then paired with modulation category tags to construct an automatic modulation identification sample library. , ,in, Indicates the first A normalized time-domain I / Q signal sample, express The corresponding modulation category label, This represents the total number of samples.

[0021] Furthermore, step S2, which involves constructing a time-frequency dual-stream robust teacher model, is as follows:

[0022] S201. Construct a temporal feature extraction branch;

[0023] Time-domain I / Q signal samples Input time-domain encoder Extracting temporal features:

[0024]

[0025] And through time-domain classification head Obtain the time-domain branch output:

[0026]

[0027] in, Representing time-domain characteristics, Indicates the parameters of the time-domain encoder. Represents the time-domain branch logits vector. This represents the time-domain classification header parameters.

[0028] Let the disturbance budget be The step size for a single-step update is The number of iterations is The temporal adversarial examples are initialized as follows:

[0029]

[0030] No. The next iteration update is as follows:

[0031]

[0032] in, Indicates the first The temporal adversarial examples obtained in the second iteration , This represents the projection operator, used to project samples back to the original sample. Centered on, with radius Within the disturbance constraint range, Represents a symbolic function. This indicates that the gradient is calculated over the input sample. This represents the cross-entropy loss function.

[0033] go through After the second iteration, temporal adversarial examples are obtained. .

[0034] With clean time-domain I / Q signal samples Adversarial examples in the temporal domain The input is used for training until a preset number of training epochs are reached or the temporal branch training loss meets the convergence condition. Training is then complete, and the trained temporal encoder is obtained. The temporal feature extraction branch serves as the teacher model; the training loss for the temporal branch is as follows:

[0035]

[0036] in, This represents the training loss in the time domain branch. This represents the weighting coefficient between the clean sample loss and the adversarial sample loss.

[0037] S202. Construct a frequency domain representation;

[0038] Time-domain I / Q signal samples The in-phase and quadrature components in the equation are denoted as follows: and Construct complex baseband sequences:

[0039]

[0040] in Represents the imaginary unit;

[0041] Performing a Fast Fourier Transform and spectral centering on the complex baseband sequence yields its frequency domain representation:

[0042]

[0043] Then, by separating the real and imaginary parts of the frequency domain representation, we obtain a dual-channel frequency domain input:

[0044]

[0045] in, This represents a complex baseband sequence composed of I / Q components. Represented by the spectrum. This represents the mapping function from time-domain I / Q samples to frequency-domain two-channel samples. and These represent taking the real part and taking the imaginary part, respectively.

[0046] S203. Construct a frequency domain feature extraction branch;

[0047] Input frequency domain Input frequency domain encoder Extracting frequency domain features:

[0048]

[0049] And through frequency domain classification head The frequency domain branch output is obtained:

[0050]

[0051] in, Represents frequency domain characteristics, Indicates the frequency domain encoder parameters. This represents the frequency domain branch of the logits vector. This represents the frequency domain classification header parameters.

[0052] Let the frequency domain perturbation budget be The step size for a single-step update is The number of iterations is The frequency domain adversarial examples are initialized as follows:

[0053]

[0054] No. The next iteration update is as follows:

[0055]

[0056] in, Indicates the first The frequency domain adversarial sample obtained in the second iteration , This represents the projection operator, used to project samples back to the original sample. Centered on, with radius Within the disturbance constraint range, Represents a symbolic function. This indicates that the gradient is calculated over the input sample. This represents the cross-entropy loss function.

[0057] go through After the second iteration, frequency domain adversarial examples are obtained. .

[0058] During frequency domain branch pre-training, adversarial perturbations are still applied to the original time-domain I / Q samples. Above, and through the frequency domain mapping function Input frequency domain branch. With clean time-domain I / Q signal samples. Adversarial examples in the frequency domain The frequency domain encoder is then trained using the input data until a preset number of training rounds is reached or the frequency domain branch training loss meets the convergence condition. Once trained, the frequency domain encoder is obtained. The frequency domain feature extraction branch serves as the teacher model; the training loss for the frequency domain branch is as follows:

[0059]

[0060] in, This represents the frequency domain branch training loss. This represents the weighting coefficient between the clean sample loss and the adversarial sample loss in the frequency domain branch. This represents a temporal adversarial example generated for the frequency domain branch.

[0061] S204. Construct an attention-gated fusion module;

[0062] Time domain features and frequency domain features Mapping each to a unified dimension yields:

[0063]

[0064] in, Represents the temporal feature mapping layer. Represents the frequency domain feature mapping layer. and These represent the aligned time-domain and frequency-domain features, respectively.

[0065] The aligned time-domain and frequency-domain features are concatenated and then input into the attention-gated unit.

[0066]

[0067]

[0068] in, This represents the concatenated two-domain features. Represents the time-domain gating weights. Represents the frequency domain gating weights. This represents the trainable parameters in the attention gating unit. Represents a non-linear activation function. Used to normalize time-domain and frequency-domain weights.

[0069] Based on the aforementioned gating weights, the time-domain features and frequency-domain features are weighted and fused:

[0070]

[0071] in, This represents the multi-domain features after fusion. This indicates element-wise multiplication.

[0072] S205. Training a multi-domain attention fusion teacher model;

[0073] Fusion features Enter teacher category header The teacher model output is obtained as follows:

[0074]

[0075] in, This represents the logits vector output by the teacher model. This represents the header parameters for teacher classification.

[0076] During the fusion training phase, the temporal encoder is first fixed. and frequency domain encoder The parameters were used to train the attention-gated fusion module and the teacher classification head; then, a small learning rate was used to jointly fine-tune the temporal encoder, frequency encoder, attention-gated fusion module and teacher classification head.

[0077] The output of the above teacher model can also be equivalently represented as:

[0078] Let the complete teacher model be Its output is:

[0079]

[0080] in, This represents all the parameters of the teacher model.

[0081] Let the disturbance budget be The step size for a single-step update is The number of iterations is The teacher model adversarial examples are initialized as follows:

[0082]

[0083] No. The next iteration update is as follows:

[0084]

[0085] in, Indicates the first Adversarial examples of the teacher model obtained in the second iteration , This represents the projection operator, used to project samples back to the original sample. Centered on, with radius Within the disturbance constraint range, Represents a symbolic function. This indicates that the gradient is calculated over the input sample. This represents the cross-entropy loss function.

[0086] go through After the second iteration, adversarial examples of the teacher model were obtained. ;

[0087] With clean time-domain I / Q signal samples Adversarial examples against teacher models As input, the model is trained until a preset number of training rounds is reached or the joint training loss of the teacher model meets the convergence condition, resulting in a trained time-frequency dual-stream robust teacher model. After training, the teacher model parameters are frozen. In the subsequent cross-domain robust distillation stage, the teacher model is used only as a fixed teacher model to output stable teacher logits vectors and temperature soft label distributions. After robust training, the teacher model's parameters remain frozen during the distillation stage, serving only to provide soft supervision information and not participating in student model parameter updates. The joint training loss for the teacher model is:

[0088]

[0089] in, This represents the joint training loss of the teacher-model system. This represents the weighting coefficient between clean sample loss and adversarial sample loss during teacher model training. This represents adversarial examples generated for a complete teacher model.

[0090] Furthermore, step S3, which involves constructing a purely time-domain lightweight student model, is as follows:

[0091] Student model on input samples The output is represented as:

[0092]

[0093] in, Representing the student model, Represents the parameters of the student model. This represents the logits vector output by the student model. R represents the total number of modulation categories, and R represents the set of real numbers.

[0094] Furthermore, the process of generating adversarial examples in step S4 is as follows:

[0095] Let the original sample be Its adversarial examples are:

[0096]

[0097] in, This represents an adversarial perturbation applied to the original time-domain I / Q samples. To ensure the perturbation amplitude is limited, the adversarial perturbation satisfies:

[0098]

[0099] in, Represents the infinite norm, This represents the disturbance budget, used to limit the maximum range of variation of the counter-disturbance in the I / Q signal amplitude space.

[0100] Preferably, projective gradient descent is used to generate adversarial examples. Let the first... The adversarial sample obtained in the second iteration is Initialized as:

[0101]

[0102] Its iterative update process is as follows:

[0103]

[0104] in, Indicates the number of iterations. Indicates the first Adversarial examples obtained in the second iteration This indicates a single-step update step size. This indicates the budget for disturbances. This represents a projection operation used to restrict the updated samples to a subset of the specified values. Centered on, with radius Within the perturbation constraint neighborhood; Represents the cross-entropy classification loss. This indicates that the gradient is calculated over the input sample. Represents a symbolic function.

[0105] After a preset number of iterations Then, the final adversarial sample is obtained. .

[0106] Furthermore, the generation process of the temperature soft label distribution in step S5 of the teacher model is as follows:

[0107] Clean sample Teacher model frozen after inputting S2 training The teacher's logits vector is obtained. :

[0108]

[0109] in, This represents a time-frequency dual-stream robust teacher model with its training complete and parameters frozen. Indicates the teacher model parameters, This represents the logits vector output by the teacher model.

[0110] Introducing temperature coefficient The temperature soft-label distribution of the teacher model was obtained. :

[0111] ;

[0112] The teacher model's temperature soft-label probability in the c-th modulation category :

[0113]

[0114] in, This represents the temperature coefficient, used to smooth the output distribution of the teacher model; Indicates the teacher model on the sample The output value of logits for the c-th modulation category. Indicates the teacher model on the sample The output value of logits for the r-th modulation category.

[0115] Furthermore, step S6 uses the adversarial examples obtained in step S4. Input the student model and obtain the logits output of the student model under perturbation conditions:

[0116]

[0117] Temperature soft-label distribution in the student model:

[0118] ;

[0119] in, Indicates student model against adversarial examples Temperature soft label distribution.

[0120] Furthermore, the cross-domain robust distillation loss described in step S7 is:

[0121]

[0122] Expanding further:

[0123]

[0124] in, Indicates cross-domain robust distillation loss, Indicates the number of samples in a mini-batch. Denotes KL divergence, The teacher model indicates that the sample Belongs to the Soft probabilities of modulation-like methods This represents the first output of the student model on adversarial examples. Temperature-based soft probability.

[0125] Furthermore, the supervised classification loss in step S8 is:

[0126]

[0127]

[0128] in, Indicates the loss of the supervised classification. Indicates the number of samples in a mini-batch. This represents the category indicator function, when the sample When the true category is class c, Otherwise, it is 0; This indicates that the student model will use adversarial examples. The probability of classifying it as the c-th modulation scheme. Indicates student model against adversarial examples The output value of logits for the c-th modulation category. Indicates student model against adversarial examples The output value of logits for the r-th modulation class.

[0129] Furthermore, the process of step S9 is as follows:

[0130] By weighted fusion of the cross-domain robust distillation loss and the supervised classification loss, a joint optimization objective for the student model is obtained:

[0131]

[0132] in, This represents the joint loss function for training the student model. This represents the loss weighting coefficient, used to balance the guidance of teachers' soft labels and the supervision of real labels.

[0133] Adversarial Samples As input, the student model is trained until a preset number of training epochs or joint loss function is reached. The convergence condition is met, and the robust lightweight student model that has been trained is obtained. ,in, Represents the parameters of the student model after training, and represents the joint loss function. Regarding student model parameters The gradient.

[0134] Furthermore, in step S10, the time-domain I / Q signal samples to be predicted are... Input the student model after the model parameters are updated in step S9 (training complete), and obtain the modulation recognition result:

[0135]

[0136] in, Indicates the modulation recognition result. This indicates that the student model responds to the input samples. The prediction score for modulation scheme of type c.

[0137] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0138] (1) This invention transfers the robust discrimination capability of the time-frequency dual-stream teacher model to the pure time-domain lightweight student model through cross-domain adversarial knowledge distillation, so that the student model can still maintain good adversarial robustness without retaining the frequency domain branch.

[0139] (2) The present invention introduces an adversarial sample driving mechanism in the distillation process, so that the student model approximates the stable output distribution of the teacher model under clean input conditions under perturbation input conditions, thereby improving the discrimination stability of the lightweight model under complex electromagnetic interference and adversarial perturbation conditions.

[0140] (3) In the lightweight inference stage, this invention retains only the pure time-domain student model and does not perform FFT, spectrum generation and cross-domain interaction operations, thereby reducing the number of model parameters, storage overhead, computational complexity and inference latency.

[0141] (4) The present invention adopts the method of separating offline training and online deployment. The teacher model is only used to provide soft supervision information during the training stage and does not participate in the calculation during the online inference stage. It is suitable for resource-constrained airborne nodes, UAV payloads, low-orbit satellite monitoring equipment and software-defined radio platforms. Attached Figure Description

[0142] Figure 1 A schematic diagram of the automatic modulation and identification system for non-cooperative wireless signals;

[0143] Figure 2 This is a flowchart of the robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to the present invention. Detailed Implementation

[0144] The technical solution of the present invention will be described in further detail below, but the scope of protection of the present invention is not limited to the following description.

[0145] like Figure 1 As shown, in scenarios such as air-space warfare, UAV non-cooperative communication reconnaissance, low-orbit satellite terminal monitoring, and software-defined radio sensing, the front-end sensing node is responsible for receiving target communication signals via radio frequency, down-converting, analog-to-digital sampling, frame segmentation, and energy normalization to obtain standardized time-domain I / Q signal samples. Subsequently, these time-domain I / Q samples are input into a lightweight recognition model to perform automatic modulation recognition and output the corresponding modulation category results. To ensure good discrimination stability of the lightweight model under complex electromagnetic environments and disturbance conditions, this embodiment adopts a technical approach combining offline cross-domain distillation training and online lightweight inference.

[0146] S1. Construct an automatic modulation recognition sample library;

[0147] Non-cooperative wireless communication signals are collected, and time-domain I / Q samples are obtained after preprocessing at the receiver to construct an automatic modulation recognition sample library. :

[0148]

[0149] in, Indicates the first A normalized time-domain I / Q signal sample, express The corresponding modulation category label, This represents the total number of samples. In practical applications, time-domain I / Q samples can be obtained by the front-end sensing node through a sliding window truncation method. Each sample contains two channels: an in-phase component and a quadrature component. After energy normalization, different samples can participate in training and testing at a uniform amplitude scale, thereby ensuring the consistency of subsequent adversarial perturbation budget settings.

[0150] S2. Construct a robust time-frequency dual-stream teacher model ;

[0151] The time-frequency dual-stream robust teacher model includes a time-domain feature extraction branch, a frequency-domain feature extraction branch, and an attention-gated fusion module. Taking time-domain I / Q samples as input, it sequentially completes time-domain feature learning, frequency-domain feature learning, and attention-based multi-domain feature fusion, outputting a teacher logits vector corresponding to the modulation category.

[0152] S201. Construct a temporal feature extraction branch;

[0153] Time-domain I / Q signal samples Input time-domain encoder Extracting temporal features:

[0154]

[0155] And through time-domain classification head Obtain the time-domain branch output:

[0156]

[0157] in, Representing time-domain characteristics, Indicates the parameters of the time-domain encoder. Represents the time-domain branch logits vector. This represents the time-domain classification header parameters.

[0158] To enhance the stability of the time-domain branch under perturbation conditions, adversarial defense pre-training is performed on the time-domain branch. Let the perturbation budget be... The step size for a single-step update is The number of iterations is The temporal adversarial examples are initialized as follows:

[0159]

[0160] No. The next iteration update is as follows:

[0161]

[0162] in, Indicates the first The temporal adversarial examples obtained in the second iteration , This represents the projection operator, used to project samples back to the original sample. Centered on, with radius Within the disturbance constraint range, Represents a symbolic function. This indicates that the gradient is calculated over the input sample. This represents the cross-entropy loss function.

[0163] go through After the second iteration, temporal adversarial examples are obtained. .

[0164] With clean time-domain I / Q signal samples Adversarial examples in the temporal domain The input is used for training until a preset number of training epochs are reached or the temporal branch training loss meets the convergence condition. Training is then complete, and the trained temporal encoder is obtained. The temporal feature extraction branch serves as the teacher model; the training loss for the temporal branch is as follows:

[0165]

[0166] in, This represents the training loss in the time domain branch. This represents the weighting coefficient between the clean sample loss and the adversarial sample loss.

[0167] S202. Construct frequency domain representation;

[0168] Time-domain I / Q signal samples The in-phase and quadrature components in the equation are denoted as follows: and Construct complex baseband sequences:

[0169]

[0170] in Represents the imaginary unit;

[0171] Performing a Fast Fourier Transform and spectral centering on the complex baseband sequence yields its frequency domain representation:

[0172]

[0173] Then, by separating the real and imaginary parts of the frequency domain representation, we obtain a dual-channel frequency domain input:

[0174]

[0175] in, This represents a complex baseband sequence composed of I / Q components. Represented by the spectrum. This represents the mapping function from time-domain I / Q samples to frequency-domain two-channel samples. and These represent taking the real part and taking the imaginary part, respectively.

[0176] S203. Construct a frequency domain feature extraction branch;

[0177] Input frequency domain Input frequency domain encoder Extracting frequency domain features:

[0178]

[0179] And through frequency domain classification head The frequency domain branch output is obtained:

[0180]

[0181] in, Represents frequency domain characteristics, Indicates the parameters of the frequency domain encoder. This represents the frequency domain branch of the logits vector. This represents the frequency domain classification header parameters.

[0182] Let the frequency domain perturbation budget be The step size for a single-step update is The number of iterations is The frequency domain adversarial examples are initialized as follows:

[0183]

[0184] No. The next iteration update is as follows:

[0185]

[0186] in, Indicates the first The frequency domain adversarial sample obtained in the next iteration , This represents the projection operator, used to project samples back to the original sample. Centered on, with radius Within the disturbance constraint range, Represents a symbolic function. This indicates that the gradient is calculated over the input sample. This represents the cross-entropy loss function.

[0187] go through After the second iteration, frequency domain adversarial examples are obtained. .

[0188] During frequency domain branch pre-training, adversarial perturbations are still applied to the original time-domain I / Q samples. Above, and through the frequency domain mapping function Input frequency domain branch. With clean time-domain I / Q signal samples. Adversarial examples in the frequency domain The frequency domain encoder is then trained using the input data until a preset number of training rounds is reached or the frequency domain branch training loss meets the convergence condition. Once trained, the frequency domain encoder is obtained. The frequency domain feature extraction branch serves as the teacher model; the training loss for the frequency domain branch is as follows:

[0189]

[0190] in, This represents the frequency domain branch training loss. This represents the weighting coefficient between the clean sample loss and the adversarial sample loss in the frequency domain branch. This represents a temporal adversarial example generated for the frequency domain branch.

[0191] S204. Construct an attention-gated fusion module;

[0192] Time domain features and frequency domain features Mapping each to a unified dimension yields:

[0193]

[0194] in, Represents the temporal feature mapping layer. Represents the frequency domain feature mapping layer. and These represent the aligned time-domain and frequency-domain features, respectively.

[0195] The aligned time-domain and frequency-domain features are concatenated and then input into the attention-gated unit.

[0196]

[0197]

[0198] in, This represents the concatenated two-domain features. Represents the time-domain gating weights. Represents the frequency domain gating weights. This represents the trainable parameters in the attention gating unit. Represents a non-linear activation function. Used to normalize time-domain and frequency-domain weights.

[0199] Based on the aforementioned gating weights, the time-domain features and frequency-domain features are weighted and fused:

[0200]

[0201] in, This represents the multi-domain features after fusion. This represents element-wise multiplication. Through this attention-gating method, the teacher model can adaptively adjust the contribution of dual-domain features to the final recognition result based on the time-domain waveform features and frequency-domain spectral structure features of different samples.

[0202] S205. Training a multi-domain attention fusion teacher model;

[0203] Fusion features Enter teacher category header The teacher model output is obtained as follows:

[0204]

[0205] in, This represents the logits vector output by the teacher model. This represents the header parameters for teacher classification.

[0206] During the fusion training phase, the temporal encoder is first fixed. and frequency domain encoder The parameters were used to train the attention-gated fusion module and the teacher classification head; then, a small learning rate was used to jointly fine-tune the temporal encoder, frequency encoder, attention-gated fusion module and teacher classification head.

[0207] The output of the above teacher model can also be equivalently represented as:

[0208] Let the complete teacher model be Its output is:

[0209]

[0210] in, This represents all the parameters of the teacher model.

[0211] Let the disturbance budget be The step size for a single-step update is The number of iterations is The teacher model adversarial examples are initialized as follows:

[0212]

[0213] No. The next iteration update is as follows:

[0214]

[0215] in, Indicates the first Adversarial examples of the teacher model obtained in the second iteration , This represents the projection operator, used to project samples back to the original sample. Centered on, with radius Within the disturbance constraint range, Represents a symbolic function. This indicates that the gradient is calculated over the input sample. This represents the cross-entropy loss function.

[0216] go through After the second iteration, adversarial examples of the teacher model were obtained. ;

[0217] With clean time-domain I / Q signal samples Adversarial examples against teacher models As input, the model is trained until a preset number of training rounds is reached or the joint training loss of the teacher model meets the convergence condition, resulting in a trained time-frequency dual-stream robust teacher model. After training, the teacher model parameters are frozen. In the subsequent cross-domain robust distillation stage, the teacher model is used only as a fixed teacher model to output stable teacher logits vectors and temperature soft label distributions. After robust training, the teacher model's parameters remain frozen during the distillation stage, serving only to provide soft supervision information and not participating in student model parameter updates. The joint training loss for the teacher model is:

[0218]

[0219] in, This represents the joint training loss of the teacher-model system. This represents the weighting coefficient between clean sample loss and adversarial sample loss during teacher model training. This represents adversarial examples generated for a complete teacher model.

[0220] S3. Construct a lightweight student model in the pure time domain. ;

[0221] The student model uses a pure time-domain lightweight network for the online lightweight inference stage. It only accepts time-domain I / Q samples as input and does not include frequency-domain input branches, FFT preprocessing, spectrum generation modules, or cross-domain interaction modules.

[0222] Student model on input samples The output is represented as:

[0223]

[0224] in, Representing the student model, Represents the parameters of the student model. This represents the logits vector output by the student model. R represents the total number of modulation categories, and R represents the set of real numbers.

[0225] In a preferred embodiment, the student model employs the lightweight EdgeResNet-8 network, comprising an input layer, a Stem layer, three lightweight one-dimensional residual stages, and a classification head. The Stem layer extracts shallow temporal features through one-dimensional convolution, batch normalization, and ReLU activation. The three residual stages sequentially extract hierarchical features. The classification head, composed of global average pooling and fully connected layers, outputs the modulation category recognition result. This structure can extract hierarchical discriminative features from temporal I / Q signals with low parameter count and computational complexity, making it suitable for lightweight inference on edge sides.

[0226] S4. Generate adversarial examples;

[0227] To improve the robustness of the student model under perturbed input conditions, this embodiment generates adversarial examples based on the current parameters of the student model.

[0228] Let the original sample be Its adversarial examples are:

[0229]

[0230] in, This represents an adversarial perturbation applied to the original time-domain I / Q samples. To ensure the perturbation amplitude is limited, the adversarial perturbation satisfies:

[0231]

[0232] in, Represents the infinite norm, This represents the disturbance budget, used to limit the maximum range of variation of the counter-disturbance in the I / Q signal amplitude space.

[0233] In a preferred embodiment, adversarial examples are generated using projective gradient descent. Let the first... The adversarial sample obtained in the second iteration is Initialized as:

[0234]

[0235] Its iterative update process is as follows:

[0236]

[0237] in, Indicates the number of iterations. Indicates the first Adversarial examples obtained in the second iteration This indicates a single-step update step size. This indicates the budget for disturbances. This represents a projection operation used to restrict the updated samples to a subset of the specified values. Centered on, with radius Within the perturbation constraint neighborhood; Represents the cross-entropy classification loss. This indicates that the gradient is calculated over the input sample. Represents a symbolic function.

[0238] After a preset number of iterations Then, the final adversarial sample is obtained. .

[0239] S5. Obtain the temperature soft label distribution of the teacher model;

[0240] Clean sample Teacher model frozen after inputting S2 training The teacher's logits vector is obtained. :

[0241]

[0242] in, This represents a time-frequency dual-stream robust teacher model with its training complete and parameters frozen. Indicates the teacher model parameters, This represents the logits vector output by the teacher model.

[0243] Introducing temperature coefficient The temperature soft-label distribution of the teacher model was obtained. :

[0244] ;

[0245] The teacher model's temperature soft-label probability in the c-th modulation category :

[0246]

[0247] in, This represents the temperature coefficient, used to smooth the output distribution of the teacher model; Indicates the teacher model on the sample The output value of logits for the c-th modulation category. Indicates the teacher model on the sample The output value of logits for the r-th modulation category;

[0248] By using temperature soft labels, the student model can not only learn the teacher model's judgment of the true category, but also learn the similarity relationship between different easily confused modulation categories.

[0249] S6. Obtain the perturbation domain output distribution of the student model;

[0250] The adversarial sample obtained in step S4 Input the student model and obtain the logits output of the student model under perturbation conditions:

[0251]

[0252] Further, the temperature soft-label distribution of the student model was obtained:

[0253] ;

[0254] in, Indicates student model against adversarial examples Temperature soft label distribution.

[0255] This establishes a cross-domain output alignment relationship between the teacher model's "clean sample output" and the student model's "adversarial sample output".

[0256] S7. Calculate the cross-domain robust distillation loss;

[0257] The KL divergence constraint is used to approximate the stable output distribution of the teacher model on clean samples by the student model's output distribution on adversarial examples. The cross-domain robust distillation loss is expressed as:

[0258]

[0259] Expanding further:

[0260]

[0261] in, Indicates cross-domain robust distillation loss, Indicates the number of samples in a mini-batch. Denotes KL divergence, The teacher model indicates that the sample Belongs to the Soft probability of modulation-like methods This represents the first output of the student model on adversarial examples. Temperature-based soft probability loss. This loss enables the student model to learn the stable discriminative behavior of the teacher model on clean samples under perturbation conditions, achieving robust knowledge transfer from a time-frequency dual-stream teacher model to a pure time-domain student model.

[0262] S8. Calculate the supervised classification loss;

[0263] To ensure that the student model still has the ability to directly distinguish the real modulation category, a supervised classification loss is introduced into the student model's output on adversarial examples.

[0264] First, convert the student's logits into a normal probability distribution:

[0265]

[0266] The supervised classification loss is expressed as:

[0267]

[0268] in, Indicates the loss of the supervised classification. Indicates the number of samples in a mini-batch. This represents the category indicator function, when the sample When the true category is class c, Otherwise, it is 0; This indicates that the student model will use adversarial examples. The probability of classifying it as the c-th modulation scheme. Indicates student model against adversarial examples The output value of logits for the c-th modulation class. Indicates student model against adversarial examples The logits output value for the r-th modulation category. This loss is used to constrain the student model to still output the correct modulation category on perturbed samples.

[0269] S9. Construct a joint optimization objective and update the student model;

[0270] By weighted fusion of the cross-domain robust distillation loss and the supervised classification loss, a joint optimization objective for the student model is obtained:

[0271]

[0272] in, This represents the joint loss function for training the student model. This represents the loss weighting coefficient, used to balance the guidance of teachers' soft labels and the supervision of real labels.

[0273] Adversarial Samples As input, the student model is trained until a preset number of training epochs or joint loss function is reached. The convergence condition is met, and the robust lightweight student model that has been trained is obtained. ,in, Represents the parameters of the student model after training, and represents the joint loss function. Regarding student model parameters The gradient of the parameter is used to update the student model. Through this parameter update process, the student model learns the stable soft label distribution and the true modulated category label of the teacher model simultaneously under adversarial example input conditions, thereby improving its robustness in perturbation environments.

[0274] S10. Lightweight inference and modulation recognition;

[0275] After the student model is trained, only the trained, lightweight time-domain student model is retained during the online inference phase. ;

[0276] The teacher model is no longer used during the deployment phase, nor are frequency domain feature generation, FFT transformation, frequency domain branch inference, and cross-domain attention interaction performed.

[0277] Edge nodes receive temporal I / Q samples uploaded by front-end sensing nodes. Then, directly input the trained student model to obtain the modulation category prediction result:

[0278]

[0279] in, This indicates the final modulation recognition result. This indicates that the student model responds to the input samples. The prediction score for modulation scheme of type c.

[0280] This embodiment, through the aforementioned steps, maintains a pure time-domain, lightweight, and low-complexity inference link in the edge-side online model. Simultaneously, during the training phase, it leverages the robust soft labels and adversarial example-driven distillation mechanism of the time-frequency dual-stream teacher model to enable the student model to inherit the teacher model's stable discrimination capability within the perturbation neighborhood. Therefore, this method can improve robustness and recognition stability in non-cooperative wireless signal modulation recognition tasks while reducing the number of model parameters, computational complexity, and inference latency.

[0281] The above description represents preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technical or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification, characterized in that, Includes the following steps: S1: Collect non-cooperative wireless communication signals, preprocess them to obtain time-domain I / Q signal samples, and pair the time-domain I / Q signal samples with modulation category labels to build an automatic modulation recognition sample library; S2: Construct a robust teacher model for time-frequency dual-stream signals. The teacher model takes time-domain I / Q signal samples as input, and sequentially completes time-domain feature learning, frequency-domain feature learning, and attention multi-domain feature fusion. It outputs a teacher vector corresponding to the modulation category and trains and freezes the teacher model parameters. S3: Construct a lightweight student model in the pure time domain; S4: Generate adversarial samples corresponding to time-domain I / Q signal samples based on the current parameters of the student model; S5: Input the time-domain I / Q signal samples into the frozen teacher model to obtain the temperature soft label distribution of the teacher model; S6: Input the adversarial examples generated in step S4 into the student model to obtain the temperature soft label distribution of the student model; S7: Calculate the cross-domain robust distillation loss based on the temperature soft-label distribution of the teacher model on clean samples and the temperature soft-label distribution of the student model on adversarial samples. S8: Calculate the supervised classification loss based on the student model's output on adversarial examples and the real modulation category labels; S9: Weigh and fuse the cross-domain robust distillation loss and the supervised classification loss to construct a joint loss function, and update the student model parameters according to the joint loss function; S10: Input the time-domain I / Q signal sample to be predicted into the student model after the model parameters are updated in step S9 to obtain the modulation recognition result.

2. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, Automatic modulation recognition sample library in step S1 ,in, Indicates the first One time-domain I / Q signal sample, express The corresponding modulation category label, This represents the total number of samples.

3. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, Step S2 involves constructing a robust time-frequency dual-stream teacher model as follows: S201. Construct a temporal feature extraction branch; Time-domain I / Q signal samples Input time-domain encoder Extracting temporal features: And through time-domain classification head Obtain the time-domain branch output: in, Representing time-domain features, Indicates the parameters of the time-domain encoder. Represents the time-domain branch logits vector. Indicates the time-domain classification header parameters; Let the disturbance budget be The step size for a single-step update is The number of iterations is The temporal adversarial examples are initialized as follows: No. The next iteration update is as follows: in, Indicates the first The temporal adversarial examples obtained in the second iteration , Represents the projection operator. Represents a symbolic function. This indicates that the gradient is calculated over the input sample. Represents the cross-entropy loss function; go through After the second iteration, temporal adversarial examples are obtained. ; Time-domain I / Q signal samples Adversarial examples in the temporal domain The input is used for training until a preset number of training epochs are reached or the temporal branch training loss meets the convergence condition. Training is then complete, and the trained temporal encoder is obtained. The temporal feature extraction branch serves as the teacher model; the training loss for the temporal branch is as follows: in, This represents the training loss in the time domain branch. The weighting coefficients between clean sample loss and adversarial sample loss; S202. Construct a frequency domain representation; Time-domain I / Q signal samples The in-phase and quadrature components in the equation are denoted as follows: and Construct complex baseband sequences: in Represents the imaginary unit; Performing a Fast Fourier Transform and spectral centering on the complex baseband sequence yields its frequency domain representation: Then, by separating the real and imaginary parts of the frequency domain representation, we obtain a dual-channel frequency domain input: in, This represents the mapping function from time-domain I / Q samples to frequency-domain two-channel samples. and These represent taking the real part and taking the imaginary part, respectively. S203. Construct a frequency domain feature extraction branch; Input frequency domain Input frequency domain encoder Extracting frequency domain features: And through frequency domain classification head The frequency domain branch output is obtained: in, Represents frequency domain characteristics, Indicates the frequency domain encoder parameters. This represents the frequency domain branch of the logits vector. Indicates the frequency domain classification header parameters; Time-domain I / Q signal samples Adversarial examples in the frequency domain The frequency domain encoder is then trained using the input data until a preset number of training rounds is reached or the frequency domain branch training loss meets the convergence condition. Once trained, the frequency domain encoder is obtained. The frequency domain feature extraction branch serves as the teacher model; the training loss for the frequency domain branch is as follows: in, This represents the frequency domain branch training loss. This represents the weighting coefficient between the clean sample loss and the adversarial sample loss in the frequency domain branch; S204. Construct an attention-gated fusion module; Time-domain features and frequency domain features Mapping each to a unified dimension yields: in, Represents the temporal feature mapping layer. Represents the frequency domain feature mapping layer. and These represent the aligned time-domain and frequency-domain features, respectively. The gating weights are calculated by concatenating the aligned time-domain and frequency-domain features: in, This represents the concatenated two-domain features. Represents the time-domain gating weights. Represents the frequency domain gating weights. Represents training parameters, Represents a non-linear activation function. Used to normalize time-domain and frequency-domain weights; Based on the gating weights, the time-domain features and frequency-domain features are weighted and fused: in, This represents the multi-domain features after fusion. This represents element-wise multiplication; S205. Training a multi-domain attention fusion teacher model; Fusion features Enter teacher category header The teacher model output is obtained as follows: in, This represents the logits vector output by the teacher model. Indicates the teacher category header parameters; Let the complete teacher model be Its output is: in, This represents all the parameters of the teacher model; Time-domain I / Q signal samples Adversarial examples against teacher models As input, the model is trained until a preset number of training rounds is reached or the joint training loss of the teacher model meets the convergence condition, resulting in a trained time-frequency dual-stream robust teacher model. After training, the teacher model parameters are frozen. The joint training loss for the teacher model is: in, This represents the joint training loss of the teacher-model system. This represents the weighting coefficient between clean sample loss and adversarial sample loss during teacher model training.

4. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, Step S3: Pure Time-Domain Lightweight Student Model in, Representing the student model, Represents the parameters of the student model. This represents the logits vector output by the student model. R represents the total number of modulation categories, and R represents the set of real numbers.

5. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, The adversarial example for step S4 is: in, This indicates resistance to disturbances. , Represents the infinite norm, This indicates the budget for disturbances.

6. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 5, characterized in that, The process of generating adversarial examples using projective gradient descent is as follows: Let the first The adversarial sample obtained in the second iteration is Initialized as: Its iterative update process is as follows: in, Indicates the number of iterations. Indicates the first Adversarial examples obtained in the second iteration This indicates a single-step update step size. This indicates the budget for disturbances. Represents the projection operator. Represents the cross-entropy classification loss. This indicates that the gradient is calculated over the input sample. Represents a symbolic function; After a preset number of iterations Then, adversarial examples are obtained. .

7. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, Step S5, the process of generating the temperature soft label distribution for the teacher model, is as follows: Clean sample Teacher model frozen after inputting S2 training The teacher's logits vector is obtained. : in, This represents a time-frequency dual-stream robust teacher model with its training complete and parameters frozen. Indicates the teacher model parameters, This represents the logits vector output by the teacher model; Temperature soft label distribution in teacher model : ； The teacher model's temperature soft-label probability in the c-th modulation category : in, Indicates the temperature coefficient. Indicates the teacher model on the sample The output value of logits for the c-th modulation category. Indicates the teacher model on the sample The output value of logits for the r-th modulation class.

8. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, Step S6 involves using the adversarial examples obtained in step S4. Input the student model and obtain its output under perturbation conditions: Temperature soft-label distribution in the student model: ； in, Indicates student model against adversarial examples Temperature soft label distribution.

9. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, The cross-domain robust distillation loss mentioned in step S7 is: in, Indicates cross-domain robust distillation loss, Indicates the number of samples in a mini-batch. Denotes KL divergence, The teacher model indicates that the sample Belongs to the Soft probability of modulation-like methods This represents the first output of the student model on adversarial examples. Temperature-based soft probability.

10. The robust and lightweight cross-domain distillation method for non-cooperative wireless signal modulation identification according to claim 1, characterized in that, The supervised classification loss in step S8 is: in, Indicates the loss of the supervised classification. Indicates the number of samples in a mini-batch. This represents the category indicator function, when the sample When the true category is class c, Otherwise, it is 0; This indicates that the student model will use adversarial examples. The probability of classifying it as the c-th modulation scheme. Indicates student model against adversarial examples The output value of logits for the c-th modulation class. Indicates student model against adversarial examples The output value of logits for the r-th modulation class.

Images