A covert communication waveform generation method based on environmental signal feature learning

Abstract

The application relates to a covert communication waveform generation method based on environmental signal feature learning, and belongs to the technical field of wireless communication and signal processing. The application is implemented by constructing a covert communication waveform generation network based on a multi-generator adversarial network and decoupled training, wherein the network comprises an encoder, a decoder, a discriminator and a classifier. By constructing a multi-generator structure and introducing classifier-guided adversarial training, multiple statistical modes of real environmental signals are learned through adversarial training, and a camouflage waveform with high consistency and diversity in the time domain, the frequency domain and high-order statistical characteristics is generated, the mode collapse problem is avoided, and the concealment of communication is significantly improved. A decoupled two-stage training strategy is adopted to optimize the camouflage ability and information reconstruction ability of the signal respectively, the contradiction between high concealment and high reliability in the traditional method is solved, and the non-cooperative party cannot detect and recover the hidden information in the signal even if the signal is intercepted.

Description

Technical Field

[0001] This invention relates to a method for generating highly secure communication waveforms based on environmental signal feature learning, belonging to the field of wireless communication and signal processing technology, and is particularly suitable for secure communication scenarios such as the Industrial Internet that require high concealment and high reliability in complex electromagnetic environments. Background Technology

[0002] Faced with the security challenges brought about by the rapid development of wireless communication technology, covert communication has emerged as an new strategy. Currently, the research focus of covert communication is on how to achieve the secret transmission of information at the physical level, ensuring that communication activities are not detected by adversaries. Although traditional cryptographic methods provide a certain degree of security for wireless communication, the security of information still faces severe challenges as the computing power of non-cooperative parties continues to improve. Traditional upper-layer security solutions, such as application-layer encryption, mainly focus on protecting the security of the information content; however, these solutions cannot effectively prevent the existence of communication from being detected by adversaries. In contrast, physical-layer covert communication aims to utilize the inherent characteristics of wireless channels, such as noise and interference, to conceal communication behavior, making it impossible for eavesdroppers to obtain or decrypt the transmitted information. The core of covert communication is not simply protecting the transmitted content, but rather concealing the existence of the communication itself.

[0003] With the evolution of wireless communication technologies beyond 5G, covert communication technology can not only complement existing security mechanisms but also serve as an important means to enhance security. Currently, researchers are dedicated to exploring the fundamental theories of physical layer covert communication, such as channel models. Meanwhile, covert communication for single-antenna, multi-antenna three-node systems, and complex application scenarios with interference sources or relay assistance has also become a research hotspot.

[0004] In the field of secure wireless communication, GANs also play a crucial role. Some literature proposes physical-layer covert communication schemes based on GANs that optimize covert transmission power and maximize average communication rates by constructing a generator and a discriminator and engaging in a zero-sum game between them. Furthermore, research has proposed a wireless steganography technique that uses GANs to hide secret information within noisy signals for covert transmission. By jointly optimizing the encoding and decoding processes while maintaining the statistical similarity between the covert signal and hardware noise, the hidden information in the signal is difficult for non-cooperative parties to detect.

[0005] Despite significant advancements in artificial intelligence, particularly GANs, in secure wireless communication, existing technologies still face numerous limitations. Current covert communication systems primarily focus on preventing signal interception or detection by adversaries; however, these technologies often fall short of the ideal requirements for secure communication. Ideal secure communication should ensure that even if a non-cooperative party successfully intercepts the signal, they cannot discover any hidden information. These intercepted signals should appear indistinguishable from normal communication signals, and even during signal analysis, the non-cooperative party should be unable to detect any anomalies, mistaking them for ordinary, irrelevant data transmissions. Therefore, further enhancing the security of wireless communication systems to ensure that signals exhibit high visual and statistical consistency with signals in a normal electromagnetic environment, thereby preventing non-cooperative parties from detecting or deciphering hidden information, has become a core technological challenge that urgently needs to be overcome in the field of secure wireless communication. Summary of the Invention

[0006] To address the shortcomings of existing covert wireless communication technologies in waveform camouflage, generation diversity, and training stability, this invention aims to provide a covert communication waveform generation method based on a multi-generator adversarial network and a decoupled training strategy. By constructing a multi-generator structure and introducing classifier-guided adversarial training, the system can learn various statistical patterns of real-world signals, generating camouflaged waveforms that are highly consistent with the background signal in the time domain, frequency domain, and higher-order statistical characteristics, while exhibiting diversity. This avoids pattern collapse problems and significantly improves communication covertness. Simultaneously, a decoupled two-stage training strategy is employed to optimize both signal camouflage and information reconstruction capabilities, resolving the contradiction between high covertness and high reliability that traditional methods often struggle with a single network. This makes it difficult for non-cooperative parties to detect and recover the hidden information even if they intercept the signal.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention discloses a method for generating covert communication waveforms based on environmental signal feature learning, comprising the following steps:

[0009] Step 1: Construct a covert communication waveform generation network based on multi-generator adversarial network and decoupled training. The network includes an encoder, a decoder, a discriminator and a classifier.

[0010] The encoder network structure consists of multiple parallel sub-generators. Each sub-generator contains an input mapping layer and multiple transposed convolutional layers with shared parameters. The input mapping layer is a fully connected layer. The multiple transposed convolutional layers are used to progressively upsample high-dimensional features to the target signal length, and all sub-generators share the parameters of these transposed convolutional layers.

[0011] The encoder employs a multi-generator parallel structure to capture various patterns in the distribution of real-world signals, enhancing the diversity of generated waveforms and preventing all generators from converging to the same pattern (i.e., pattern collapse). This ensures that the generated camouflaged waveforms can cover different statistical characteristics of the real signal, reducing the risk of being identified by non-cooperative parties through pattern detection. The decoder is used to recover the original information from the received camouflaged waveforms. The discriminator is used to distinguish whether the input signal comes from a camouflaged waveform generated by the encoder or a real-world signal. The classifier is used to identify which sub-generator in the encoder produced the input camouflaged waveform. By maximizing the classifier's recognition loss, it forces each sub-generator to generate differentiated signals, preventing generators from falling into a single pattern and further improving the diversity of generated waveforms.

[0012] Step 2: Design the network structure of the encoder, decoder, discriminator, and classifier;

[0013] Step 2.1: The encoder network structure consists of multiple parallel sub-generators. Each sub-generator includes an input mapping layer and multiple transposed convolutional layers with shared parameters. The input mapping layer is a fully connected layer. The multiple transposed convolutional layers are used to progressively upsample high-dimensional features to the target signal length, and all sub-generators share the parameters of these transposed convolutional layers. The overall structure and time-frequency characteristics of the generated signal are determined by these upsampling layers, and the differences between different generators should only be reflected in the initial mapping of the input features. Through parameter sharing, the number of model parameters is reduced by approximately 66%, effectively reducing the risk of overfitting. At the same time, it forces all generators to learn their respective patterns in a unified feature space, significantly improving the stability and convergence speed of multi-generator collaborative training.

[0014] Step 2.2: The decoder network structure consists of multiple convolutional layers, an equal number of phase shuffling layers, an equal number of modulation layers, and fully connected layers. The phase shuffling layers are located after the convolutional layers. Phase shuffling layers with a shift factor of n1 are deployed after the first three convolutional layers, phase shuffling layers with a shift factor of n2 are deployed after the fourth convolutional layer, and no phase shuffling layers are deployed after the fifth convolutional layer; n1 > n2; the convolutional layer parameters are symmetrical to the transposed convolutional layers in the encoder.

[0015] The modulation layer contains multiple linear transformation layers, which combine the output of the phase shuffle layer with the waveform from the encoder output. That is, after each phase shuffle layer, the waveform from the encoder output is used to generate a modulation signal through an independent linear layer, and then added element by element to the output of the phase shuffle layer. This allows the decoder to adjust the feature extraction process according to different input waveforms, greatly enhancing its information reconstruction capability and adaptability to complex environments.

[0016] The function of a fully connected layer is to map the feature vector of the input itself into a reconstructed sequence of the same length as the encoder input.

[0017] The working method of the phase shuffling layer is as follows:

[0018] Each phase shuffling layer first generates a random integer shift amount Δ, with a value ranging from [-n, n], where n is a preset shift factor. Then, the entire feature map is cyclically shifted left or right by |Δ| time steps. Data shifted out of the boundary is discarded, and missing positions are filled using reflection filling. The phase shuffling module of this invention operates on the internal features of the neural network, introducing random temporal cyclic shifts during the training phase. Through hierarchical and differentiated deployment, it achieves progressive robust learning from local to global, significantly improving the reliability of the decoder in real wireless channel environments.

[0019] All training samples in the same batch are shifted by the same amount Δ and enter the phase shuffling module, passing through multiple phase shuffling layers in sequence to achieve phase shifting;

[0020] Step 2.3: The discriminator and classifier share a feature extraction module consisting of multiple convolutional layers. Both need to extract high-level semantic features from the original signal. Therefore, sharing this feature extraction module can significantly reduce the total number of network parameters, prevent model overfitting, and force the shared layer to learn the most representative expression of the essential features of the signal, thereby improving the learning efficiency and generalization ability of the two sub-tasks. The discriminator is connected to a fully connected layer after the shared module, outputting a scalar value to determine the authenticity of the input signal. The classifier is connected to an independent classification fully connected layer after the shared module, with an output dimension equal to the number of sub-generators. The output result is the probability distribution of the input signal belonging to each sub-generator.

[0021] Step 3: Define the objective function for each network; assume that the real environment signal follows a distribution. , recorded as The encoder input is a random noise vector. It follows a standard Gaussian distribution. ,remember encoder Depend on Child generator Composition, each sub-generator Input Mapped to generated signal Discriminator Output a scalar representing the authenticity score of the input signal; classifier Output a A probability vector of dimension represents the logarithmic probability distribution of the input signal belonging to each sub-generator.

[0022] Step 3.1, the optimization objective of the discriminator network is:

[0023]

[0024] in, The first term is the negative value of the true signal loss, which is minimized. Equivalent to maximizing The first term improves the discriminator's score for the real signal; the second term is a positive value of the generated signal loss, which is minimized. This is equivalent to reducing the discriminator's score on the generated signal; the first two terms together constitute the original form of the Wasserstein distance, enabling the discriminator to effectively distinguish between real and fake signals; the third term is the gradient penalty term, which forces the discriminator to satisfy the 1-Lipschitz continuity condition by constraining the gradient norm of the discriminator at the interpolation point to be close to 1, thus ensuring the accuracy of the Wasserstein distance calculation.

[0025] Step 3.2, the optimization objective of the classifier network is:

[0026]

[0027] in, This indicates that the classifier determines the generated signal. Belongs to the The probability of each sub-generator class; the loss function trains the classifier to accurately distinguish which sub-generator the generated signal comes from; as the classifier's ability improves, the encoder must generate more diverse waveforms in order to confuse the classifier, thus forming a game process that promotes diversity.

[0028] Step 3.3, the encoder network optimization objective is:

[0029]

[0030] The first term is the adversarial loss, which aims to train the encoder to generate realistic waveforms that can deceive the discriminator by minimizing the negative score of the discriminator on the generated signal, i.e., maximizing the discriminator's score on the generated signal. The second term is the classifier confusion loss, which aims to force the encoder to generate waveforms that can confuse the classifier by minimizing the negative log-likelihood of the classifier in identifying the source of the generated signal, i.e., maximizing the classifier's recognition loss. The two losses work together to enable the encoder to generate highly realistic waveforms while maintaining the diversity of generated waveforms, fundamentally avoiding the pattern collapse problem. The classifier confusion loss in the encoder optimization objective is the same in form but opposite in sign to the classification loss in the classifier optimization objective, reflecting an adversarial game relationship: the classifier tries to accurately identify the source of the generated signal, while the encoder tries to generate waveforms that are difficult for the classifier to identify. This game mechanism forces each sub-generator to learn different signal patterns.

[0031] Step 3.4, the optimization objective of the decoder network is:

[0032]

[0033] in, To balance the hyperparameters of reconstruction loss and cycle consistency loss, reconstruction loss Defined as the mean squared error between the original message and the message reconstructed by the decoder:

[0034]

[0035] In the formula, The bitstream input to the encoder is used to train the decoder. For the encoder to send messages Mapped to a masquerading waveform, Decoder This is the encoder input for the decoder to reconstruct from the masquerading waveform; this loss is designed to ensure the accuracy of information recovery.

[0036] Cycle consistency loss Defined as the mean square error between the original generated signal and the signal re-encoded from the decoder output:

[0037]

[0038] The loss will cause the decoder to reconstruct the message. Enter the encoder again Generate a cyclic signal And calculate its relationship with the original generated signal. The error is caused by directly imposing strong constraints on the decoder's output space, forcing the message decoded by the decoder to be identical to the encoder's original input. Highly consistent; while reconstruction loss alone can ensure that the decoder recovers the original message from the generated signal, this loss only constrains the output of the decoder and does not impose any restrictions on the intermediate signal between the encoder and decoder; cyclic consistency loss, by introducing a re-encoding stage, imposes an additional direct constraint on the output space of the decoder: requiring that the message output by the decoder can still generate a signal consistent with the original generated signal after passing through the encoder again; adopting a closed-loop structure of encoding-decoding-re-encoding is equivalent to adding a second-order consistency constraint on the basis of the original reconstruction loss, thereby further strengthening the bidirectional mapping relationship between the encoder and decoder and improving the reliability of information recovery.

[0039] Step 4: Perform decoupled two-stage joint training of the network using the objective function from Step 3;

[0040] Phase 1: Optimizing signal camouflage capabilities; sampling latent vectors from a standard Gaussian distribution. As a simulated representation of secret information, it is input to the encoder. The sub-generator generates candidate camouflage waveforms. The generated camouflaged waveforms are input into the discriminator and classifier respectively; based on the discriminator loss... Update the discriminator parameters based on the classifier loss. Update the classifier parameters; then, fixing the discriminator and classifier parameters, calculate the joint loss based on the encoder. Update the encoder parameters and train the encoder to generate high-quality spoofed waveforms that can both deceive the discriminator and confuse the classifier; repeat the above process until the encoder model converges; by freeing the encoder from information recovery constraints in the first stage, it can fully learn the statistical distribution of the background signal and focus on improving the realism and diversity of the waveform.

[0041] Phase Two: Information Reconstruction Capability Optimization; A batch of binary bit streams m is generated as the secret message to be transmitted, and input into an encoder with frozen parameters to convert it into a disguised waveform. Input the masquerading waveform into the decoder and output the reconstructed message. Calculate reconstruction loss The message reconstructed by the decoder is input again into the same frozen encoder to generate a fake signal. Calculate the cycle consistency loss between the original generated signal and the forged signal. According to the decoder joint loss Update the decoder parameters until the decoder model converges. Since the encoder parameters are frozen, the decoder faces a fixed encoder with excellent masquerading capabilities, allowing it to focus on learning accurate inverse mappings and avoid interference from encoder parameter changes to decoder training.

[0042] Step 5: After training, the inference phase begins, enabling highly covert and reliable communication. At the transmitting end, the secret information to be transmitted is input into the trained encoder. The encoder generates a camouflaged waveform that closely matches the statistical characteristics of the background signal and transmits it to the wireless channel. At the legitimate receiving end, the received signal is input into the trained decoder. The decoder reconstructs the original transmission information, completing the covert and reliable communication. During the communication process, non-cooperative parties cannot distinguish between the generated camouflaged waveform and the real environmental signal, making it difficult for them to detect the communication. Even if they intercept the signal in the channel, they cannot recover the correct transmission information due to the lack of the decoder's mapping capability.

[0043] Beneficial effects:

[0044] 1. The present invention discloses a covert communication waveform generation method based on environmental signal feature learning. Through joint adversarial training of encoder, discriminator and classifier, the generated camouflage waveform is forced to closely approximate the real environmental signal in statistical distribution, which significantly improves the covertness and anti-reconnaissance capability of the signal.

[0045] 2. The present invention discloses a covert communication waveform generation method based on environmental signal feature learning. It adopts a multi-generator encoder structure, which can generate diverse camouflage signals, increasing the difficulty for non-cooperative parties to detect and classify signals through machine learning methods, and further enhancing the security of the system.

[0046] 3. The present invention discloses a covert communication waveform generation method based on environmental signal feature learning. It adopts a decoupled two-stage training strategy to separate the signal camouflage modeling and information reconstruction tasks, thereby avoiding training instability and performance degradation problems that may be caused by multi-task joint optimization, and ensuring the dual goals of high concealment and high reliability.

[0047] 4. The present invention discloses a covert communication waveform generation method based on environmental signal feature learning, which introduces cyclic consistency loss to constrain decoder training, enhances the system's robustness to channel noise and encoder output changes, and ensures reliable information recovery under complex channel conditions.

[0048] 5. The present invention discloses a method for generating covert communication waveforms based on environmental signal feature learning. The generated waveforms are consistent with the characteristics of conventional communication signals. Even if intercepted by a non-cooperative party, only the apparent information can be demodulated, and the hidden secret information cannot be recovered, thus achieving a deeper level of communication security. Attached Figure Description

[0049] Figure 1 This is a diagram of the covert signal generation network architecture based on autoencoders and generative adversarial networks in an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the encoder network structure in an embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the decoder network structure in an embodiment of the present invention.

[0052] Figure 4 This is a time-domain waveform diagram of the generated signal in an embodiment of the present invention.

[0053] Figure 5 This is a time-domain waveform diagram of a real QPSK signal in an embodiment of the present invention.

[0054] Figure 6 This is a constellation diagram of the generated signal in an embodiment of the present invention.

[0055] Figure 7 This is a constellation diagram of a real QPSK signal in an embodiment of the present invention. Detailed Implementation

[0056] To better illustrate the purpose, technical solution, and beneficial effects of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0057] Example 1:

[0058] To verify the effectiveness of the covert communication waveform generation method based on environmental signal feature learning proposed in this invention, this embodiment conducts simulation experiments in a typical wireless communication scenario.

[0059] In the experiment, a QPSK modulated signal was selected as the environmental background signal for learning, with a sampling rate of 102400Hz, a carrier frequency of 800Hz, and a symbol rate of 800 baud. During each transmission, the system transmits 16 bits of binary data. The network structure and related parameter settings are shown in Table 1.

[0060] Table 1 Network Parameter Settings

[0061] Module hyperparameters Value Encoder (3 sub-generators) Fully connected layers 1 Shared transposed convolutional layers 5 decoder Number of modulation layers 5 Number of convolutional layers 5 Phase Shuffle Layer 5 Discriminator / Classifier Shared convolutional layers 5 Fully connected layer 1 All networks Hidden layer activation function LeakyReLU Optimizer Adam

[0062] This embodiment discloses a covert communication waveform generation method based on environmental signal feature learning. The overall process includes two stages: waveform generation and modeling, and information recovery. The specific implementation steps are as follows:

[0063] Step 1: Construct a covert communication waveform generation architecture based on a multi-generator adversarial network and decoupled training. The encoder employs a multi-generator parallel structure, consisting of three sub-generators. This aims to capture multiple patterns in the distribution of real-world signals, enhancing the diversity of generated waveforms and preventing all generators from converging to the same pattern. This ensures that the generated camouflaged waveforms can cover different statistical characteristics of the real signal, reducing the risk of being identified by non-cooperative parties through pattern detection. The decoder is used to recover the original information from the received camouflaged waveforms. The discriminator distinguishes whether the input signal comes from a camouflaged waveform generated by the encoder or a real-world signal. The classifier identifies which sub-generator in the encoder produced the input camouflaged waveform. By maximizing the classifier's recognition loss, each sub-generator is forced to generate differentiated signals, further improving the diversity of generated waveforms.

[0064] Step 2: Design the network structure of the encoder, decoder, discriminator, and classifier; construct the network structure of the encoder, decoder, discriminator, and classifier according to the parameter settings shown in Table 1.

[0065] Step 2.1, Encoder Network Structure; The encoder network structure is as follows: Figure 2 As shown, the model consists of three parallel sub-generators. Each sub-generator contains an input mapping layer (a fully connected layer that maps a 16-dimensional low-dimensional noise vector to a 256×64=16384-dimensional high-dimensional feature space) and a transposed convolutional layer with shared parameters (kernel size 25, stride 4, upsampling the high-dimensional features to a target signal length of 1024 points). All sub-generators share the parameters of these transposed convolutional layers, so that the overall structure and time-frequency characteristics of the generated signal are determined by the shared layers, and the differences between different generators are only reflected in the initial mapping of the input features. Through this parameter-sharing design, the number of model parameters is reduced by about 66%, effectively reducing the risk of overfitting, while forcing all generators to learn their respective patterns in a unified feature space, significantly improving the stability and convergence speed of multi-generator collaborative training.

[0066] Step 2.2, Decoder Network Structure; The decoder network structure is as follows: Figure 3 As shown, the structure used to accurately reconstruct the original information from the received signal consists of multiple convolutional layers, an equal number of phase shuffling layers, an equal number of modulation layers, and fully connected layers.

[0067] The convolutional layer consists of 5 layers (kernel size 25, stride 4), used to extract multi-scale features from the signal step by step. Its parameter selection is symmetrical with the transposed convolutional layer in the encoder to ensure the matching of the receptive field, thereby enabling effective inverse mapping back to the original information space.

[0068] The phase shuffling layer consists of 5 layers, located after the convolutional layers, and employs a hierarchical and differentiated deployment strategy. Specifically: a phase shuffling layer with a shift factor of n=2 is deployed after the first three convolutional layers, a phase shuffling layer with a shift factor of n=1 is deployed after the fourth convolutional layer, and no phase shuffling layer is deployed after the fifth convolutional layer. The working method of each phase shuffling layer is as follows: first, a random integer shift amount Δ is generated, with a value ranging from [-n, n] (where n is a preset shift factor). Then, the entire feature map is cyclically shifted to the left or right by |Δ| time steps. Data shifted out of bounds is discarded, and missing positions are filled using reflection filling.

[0069] The modulation layer comprises five linear transform layers, which combine the output of the phase shuffle layer with the waveform from the encoder output. Specifically, after each phase shuffle layer, a modulation signal is generated by passing the waveform from the encoder output through an independent linear layer and added element-wise to the output of the phase shuffle layer. This allows the decoder to dynamically adjust the feature extraction process according to different input waveforms, greatly enhancing its information reconstruction capability and adaptability to complex environments.

[0070] The fully connected layer is located at the end of the network. Its function is to map the feature vector after the aforementioned convolution, phase shuffling and modulation processing into a reconstruction sequence of the same length as the encoder input (in this embodiment, a 16-dimensional binary vector), thus completing the final reconstruction of the information.

[0071] Step 2.3 Discriminator and Classifier Network Structure

[0072] The discriminator and classifier share a feature extraction module consisting of five convolutional layers. Both need to extract high-level semantic features from the original signal. Sharing this module reduces the total number of parameters, prevents overfitting, and improves the learning efficiency and generalization ability of both sub-tasks. The discriminator is followed by a fully connected layer after this shared module, outputting a scalar truth score. The classifier is connected to a separate fully connected classification layer, outputting a 3-dimensional probability vector representing the probability distribution of the input signal belonging to each sub-generator.

[0073] Step 3: Define the objective function for each network;

[0074] Step 3.1, Discriminator Network Optimization Objective: Based on Wasserstein distance and introducing a gradient penalty term, as shown in Equation (1). By maximizing the score of the real signal and minimizing the score of the generated signal, the discriminator can effectively distinguish between real and fake signals; the gradient penalty term forces the discriminator to satisfy the 1-Lipschitz continuity condition, ensuring the accuracy of Wasserstein distance calculation.

[0075] Step 3.2, Classifier network optimization objective: As shown in formula (2), minimize the negative log-likelihood loss for predicting the source of the generated signal, and train the classifier to accurately distinguish which sub-generator the generated signal comes from. The classifier is only trained on the generated signal and does not involve real environment signals.

[0076] Step 3.3, Encoder network optimization objective; as shown in formula (3), it consists of two parts: adversarial loss (maximizing the discriminator's score on the generated signal) and classifier confusion loss (maximizing the classifier's recognition loss, even if the classifier has difficulty recognizing the source). The two losses work together to enable the encoder to generate high-fidelity waveforms while maintaining diversity and avoiding pattern collapse.

[0077] Step 3.4, Decoder network optimization objective; as shown in formula (4), based on reconstruction loss and cycle consistency loss The system is structured with a balance factor a=1. Cyclic consistency loss strengthens the bidirectional mapping relationship through an encoding-decoding-re-encoding closed-loop structure, improving the reliability of information recovery.

[0078] Step 4: Decoupled two-stage joint training: This embodiment adopts a decoupled training strategy, dividing the training process into a signal camouflage capability training stage and an information reconstruction capability training stage.

[0079] Phase 1: Signal camouflage capability optimization; in this phase, only the encoder, discriminator, and classifier are trained. Through adversarial training, the encoder learns to map a standard Gaussian distributed random sampled input into a camouflaged waveform that highly resembles the real environment signal. For example... Figure 4 and Figure 5 As shown, the generated signal and the real QPSK signal have a high degree of similarity in the time domain waveform, making them difficult to distinguish by waveform morphology. Furthermore, as... Figure 6 and Figure 7 As shown, the constellation distribution of the generated signal is basically consistent with that of the real QPSK signal, with a clear phase structure and a reasonable discrete distribution. This indicates that the generated signal also highly matches the real signal in terms of modulation characteristics, thus verifying the effectiveness of the training in this stage in terms of signal camouflage.

[0080] Step 3.2: Information Reconstruction Capability Training Phase: After completing the signal camouflage training, the encoder parameters are frozen, and only the decoder is trained. The specific process is as follows: The binary sequence is input into the encoder to obtain the camouflaged waveform, and then this waveform is input into the decoder for information recovery, and the reconstruction loss is calculated. Simultaneously, the decoding result is input back into the encoder to generate the reconstructed signal, and the cycle consistency loss is calculated to further constrain the mapping relationship between the encoder and decoder. Through the above training, the decoder can accurately recover the original information from complex camouflaged waveforms, thereby ensuring the reliability of communication.

[0081] Step 5: Covert Communication Implementation: After model training is completed, covert communication is implemented during the inference phase. The specific process is as follows: The transmitting end inputs the 16-bit binary sequence to be transmitted into the trained encoder to generate a camouflaged waveform and transmit it through a wireless channel. The generated signal is statistically highly consistent with the real environment signal and does not possess obvious identifiable features. After receiving the signal, the receiving end inputs it into the trained decoder, which can accurately recover the original binary data, thus achieving covert communication. Furthermore, this embodiment also verifies the signal analysis capabilities of non-cooperative parties. When using the traditional QPSK demodulation method to process intercepted signals, only surface-level bit information can be obtained, and the true hidden information cannot be recovered. Only the decoder trained by this invention can correctly parse the hidden data in the signal. Therefore, this invention not only achieves highly realistic camouflage at the signal level but also achieves deep concealment at the information level. Even if the signal is intercepted, it is difficult to detect and crack, thereby significantly improving the concealment and security of the communication system.

[0082] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating covert communication waveforms based on environmental signal feature learning, characterized in that: Includes the following steps: Step 1: Construct a covert communication waveform generation network based on multi-generator adversarial network and decoupled training. The network includes an encoder, a decoder, a discriminator and a classifier. Step 2: Design the network structure of the encoder, decoder, discriminator, and classifier; Step 3: Define the objective function for each network; Step 4: Perform two-stage joint training of the covert communication waveform generation network using the objective function from Step 3, decoupled from the target function. Step 5: After completing the training of the covert communication waveform generation network, the inference phase is initiated to achieve highly covert and reliable communication. At the transmitting end, the secret information to be transmitted is input into the pre-trained encoder. The encoder generates a disguised waveform that is highly consistent with the statistical characteristics of the environmental background signal as the transmission signal and transmits it to the wireless channel. At the legitimate receiving end, the received signal is input into the pre-trained decoder. The decoder reconstructs the original transmission information, completing the covert and reliable communication. The highly consistent statistical characteristics refer to a consistency level exceeding a preset threshold.

2. The method as described in claim 1, characterized in that: The encoder network structure consists of multiple parallel sub-generators. Each sub-generator contains an input mapping layer and multiple transposed convolutional layers with shared parameters. The input mapping layer is a fully connected layer. The multiple transposed convolutional layers are used to progressively upsample high-dimensional features to the target signal length, and all sub-generators share the parameters of these transposed convolutional layers.

3. The method as described in claim 1, characterized in that: The decoder network structure consists of multiple convolutional layers, an equal number of phase shuffling layers, an equal number of modulation layers, and fully connected layers. The phase shuffling layers are located after the convolutional layers. Phase shuffling layers with a shift factor of n1 are deployed after the first three convolutional layers, phase shuffling layers with a shift factor of n2 are deployed after the fourth convolutional layer, and no phase shuffling layers are deployed after the fifth convolutional layer; n1 > n2; the selection of convolutional layer parameters is symmetrical to that of the transposed convolutional layer in the encoder. The modulation layer contains multiple linear transformation layers, which combine the output of the phase shuffle layer with the waveform from the encoder output. That is, after each phase shuffle layer, the waveform from the encoder output is used to generate a modulation signal through an independent linear layer, and then added element by element to the output after the phase shuffle layer. Fully connected layers are used to map the feature vectors of the input itself into a reconstructed sequence of the same length as the encoder input; The working method of the phase shuffle layer is as follows: Each phase shuffling layer first generates a random integer shift amount Δ, with a value range of [-n, n], where n is a preset shift factor. Then, the entire feature map is cyclically shifted to the left or right by |Δ| time steps. Data that shifts out of the boundary is discarded, and missing positions are filled using reflection filling. All training samples in the same batch are shifted by the same amount Δ before entering the phase shuffling module and passing through multiple phase shuffling layers in sequence to achieve phase shift.

4. The method as described in claim 1, characterized in that: The discriminator and classifier share a feature extraction module consisting of multiple convolutional layers. The discriminator is connected to a fully connected layer after the shared module, and outputs a scalar value to determine the authenticity of the input signal. The classifier is connected to an independent classification fully connected layer after the shared module, and its output dimension is equal to the number of sub-generators. The output result is the probability distribution of the input signal belonging to each sub-generator.

5. The method as described in claim 4, characterized in that: Define the objective function for each network, and assume that the real-world signal follows a distribution. , recorded as The encoder input is a random noise vector. It follows a standard Gaussian distribution. ,remember encoder Depend on Child generator Composition, each sub-generator Input Mapped to generated signal Discriminator Output a scalar representing the authenticity score of the input signal; classifier Output a A probability vector of dimension represents the logarithmic probability distribution of the input signal belonging to each sub-generator.

6. The method as described in claim 5, characterized in that: The optimization objective of the discriminator network is: in, The first term is the negative value of the true signal loss, which is minimized. Equivalent to maximizing The first term improves the discriminator's score for the real signal; the second term is a positive value of the generated signal loss, which is minimized. This is equivalent to reducing the discriminator's score on the generated signal; the first two terms together constitute the original form of the Wasserstein distance; the third term is the gradient penalty term, which forces the discriminator to satisfy the 1-Lipschitz continuity condition by constraining the gradient norm of the discriminator at the interpolation point to be close to 1.

7. The method as described in claim 6, characterized in that: The optimization objective of the classifier network is: in, This indicates that the classifier determines the generated signal. Belongs to the The probability of each sub-generator class.

8. The method as described in claim 7, characterized in that: The optimization objective of the encoder network is: The first term is the adversarial loss; the second term is the classifier confusion loss.

9. The method as described in claim 8, characterized in that: The optimization objective of the decoder network is: in, To balance the hyperparameters of reconstruction loss and cycle consistency loss, reconstruction loss Defined as the mean squared error between the original message and the message reconstructed by the decoder: In the formula, The bitstream input to the encoder is used to train the decoder. For the encoder to send messages Mapped to a masquerading waveform, Decoder The encoder input is used by the decoder to reconstruct the waveform from the masquerading. Cycle consistency loss Defined as the mean square error between the original generated signal and the signal re-encoded from the decoder output: The loss will cause the decoder to reconstruct the message. Enter the encoder again Generate a cyclic signal And calculate its relationship with the original generated signal. The error.

10. The method as described in claim 9, characterized in that: The implementation method for step four is as follows: Phase 1: Optimizing signal camouflage capabilities; sampling latent vectors from a standard Gaussian distribution. As a simulated representation of secret information, it is input to the encoder. The sub-generator generates candidate camouflage waveforms. The generated camouflaged waveforms are input into the discriminator and classifier respectively; based on the discriminator loss... Update the discriminator parameters based on the classifier loss. Update classifier parameters; Then, with the discriminator and classifier parameters fixed, the joint loss of the encoder is calculated. Update the encoder parameters and train the encoder to generate high-quality spoofed waveforms that can both fool the discriminator and confuse the classifier; repeat the above process until the encoder model converges. Phase Two: Optimization of Information Reconstruction Capabilities; A batch of binary bit streams m is generated as the secret message to be transmitted, and input into an encoder with frozen parameters to convert it into a masquerade waveform. Input the masquerading waveform into the decoder and output the reconstructed message. Calculate reconstruction loss The message reconstructed by the decoder is input again into the same frozen encoder to generate a fake signal. Calculate the cycle consistency loss between the original generated signal and the forged signal. According to the decoder joint loss Update the decoder parameters until the decoder model converges.

Images