A covert semantic communication method and system

Abstract

The application discloses a kind of covert semantic communication method and system, the method is configured with standard power antenna and low power antenna in sending end, standard power antenna sends open QPSK explicit signal, low power antenna sends the covert signal obtained by semantic encoder and covert signal generator according to whether there is covert service;Covert signal is executed block, and is filled up and is superimposed with explicit signal in the same time slot, and composite receiving signal is formed by wireless channel transmission;Average radius is calculated based on the average amplitude of receiving symbol I / Q component in receiving end, and composite receiving symbol is proportionally contracted to origin to recover covert component, and original data is reconstructed by semantic decoder again.Training stage introduces monitoring end discrimination network, and WGAN-GP framework is used to carry out confrontation matching to covert signal distribution and pseudo-noise distribution, and combined with reconstruction loss, transmitting end and receiving end are optimized jointly, so that covert signal is close to pseudo-noise in statistical distribution, while maintaining the reconstruction quality of receiving end and the availability of open communication link.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a covert semantic communication method and system. Background Technology

[0002] In scenarios such as the Internet of Things (IoT), unmanned systems, mobile terminals, emergency communications, and military communications, a large amount of sensitive information needs to be transmitted via wireless links. Traditional encryption technologies can protect the communication content, but the communication behavior itself may still be detected by external observers, thereby exposing the presence, location, or mission status of the communicating parties. Therefore, simply hiding the content without hiding the communication behavior is insufficient to meet the requirements of high-security scenarios.

[0003] The goal of covert communication is to transmit information under conditions where it is difficult for the monitoring end to determine whether communication has occurred. Existing covert communication schemes often embed the covert signal into background noise or legitimate traffic, but three types of problems still exist in real wireless environments: First, the limited power of the covert signal leads to insufficient transmission rate and semantic recovery quality; second, real wireless channels experience time-varying fading, and the covertness obtained under ideal assumptions such as AWGN is prone to failure; third, if the receiver relies on accurate channel state information for cancellation or demodulation, it will significantly increase system complexity and reduce robustness in slow fading environments such as Rayleigh and Rician.

[0004] Deep joint source-channel coding can directly map image and other data into continuous symbols suitable for channel transmission, exhibiting good semantic preservation capabilities even at low signal-to-noise ratios. Generative adversarial networks (GANs) can generate signals that statistically approximate the target distribution through distribution matching. Combining these two approaches and designing a transmission architecture where explicit public and covert signals coexist can potentially improve the anti-detection capability and reconstruction quality of covert communication while ensuring the availability of public communication links. Currently, there is a lack of a partially covert communication scheme for time-varying fading channels that can recover covert information without precise CSI and achieve statistical consistency in service presence / absence status through pseudo-noise placeholders. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a covert semantic communication method and system. The method configures a standard power antenna and a low-power antenna at the transmitting end. The standard power antenna transmits a publicly available QPSK explicit signal, while the low-power antenna transmits a covert signal obtained by a semantic encoder and a covert signal generator, or transmits pseudo-noise with the same power constraint as the covert signal, depending on the presence of covert services. The covert signal or pseudo-noise is segmented, padded, and superimposed on the explicit signal in the same time slot, then transmitted via a wireless channel to form a composite received signal. At the receiving end, without requiring precise CSI, the average radius is calculated based on the average amplitude of the I / Q components of the received symbol, and the composite received symbol is proportionally shrunk towards the origin to recover the covert components. The original data is then reconstructed by the semantic decoder. During the training phase, a monitoring-end discrimination network is introduced, employing the WGAN-GP framework to perform adversarial matching between the covert signal distribution and the pseudo-noise distribution. Combined with reconstruction loss, the transmitting and receiving ends are jointly optimized to make the covert signal statistically close to the pseudo-noise, while maintaining the reconstruction quality at the receiving end and the availability of the public communication link.

[0006] The technical solution adopted by this invention to solve its technical problem is as follows:

[0007] Step 1: Obtain the data to be secretly transmitted And through semantic encoding and covert signal generation at the sending end, a network is generated. Generate complex covert signals ;

[0008] Step 2: Configure a standard power antenna and a low-power antenna at the transmitting end. The standard power antenna is used to transmit explicit QPSK signals. The low-power antenna transmits multiple covert signals when covert services are present. Sending pseudo-noise when no covert service is present ;

[0009] Step 3: Perform block segmentation and padding on the concealed signal or pseudo-noise to obtain the number of symbols that can be transmitted in a single time slot. Matching block signal ;

[0010] Step 4: Block signal With explicit QPSK signal The composite received signal is obtained by superimposing the signals in the same or corresponding time slots and transmitting them through a wireless channel. ;

[0011] Step 5: The receiving end receives the composite signal. The average radius is calculated from the average amplitudes of the in-phase and quadrature components, and the covert block signal is reconstructed proportionally without prior acquisition or explicit estimation of precise Channel State Information (CSI). ; for hidden block signals Removing padding symbols and concatenating them yields the recovered hidden sequence. via semantic decoding network Obtain reconstructed data ;

[0012] Step 6: During the training phase, the discriminant network is monitored. Distinguish between the distribution of hidden signals and pseudo-noise distributions, and optimize based on adversarial loss and reconstruction loss. , and This makes the concealed signal statistically approximate pseudo-noise.

[0013] Preferably, the complex concealment signal is:

[0014]

[0015] in , Data Dimensions ,in , , These represent the number of channels, height, and width, respectively.

[0016] Preferably, the segmentation and padding includes dividing the complex masking signals. By each piece The complex number symbols are divided into multiple hidden blocks. When the last block is insufficient... When the symbol is 1, use pseudo-noise that follows a complex Gaussian distribution. Complete; when there is no covert service, the low-power antenna transmits by The resulting pure noise block;

[0017] The block signal is represented as:

[0018]

[0019] in, Represents complex concealed signals The first in There are several complex symbols, i = 1, 2, ..., k; k represents the total number of complex symbols contained in the complex covert signal; N represents the number of complex symbols that can be transmitted in a single time slot; when k is not divisible by N, the portion of the last covert block with fewer than N symbols uses pseudo-noise. Complete;

[0020] When no covert operations are present, the low-power antenna does not remain silent but instead transmits a pure noise block composed of pseudo-noise to simulate hardware noise or low-power background disturbances at the transmitting end. The explicit QPSK signal is represented as:

[0021]

[0022] in, This represents the i-th explicit QPSK complex modulation symbol transmitted within a time slot, where i = 1, 2, ..., N; each Select from the preset QPSK constellation symbol set;

[0023] The explicit QPSK signal and the covert block signal are superimposed in the corresponding time slot to form a composite transmission signal:

[0024]

[0025] In wireless channels, composite received signals are uniformly represented as:

[0026]

[0027] in, For channel coefficients, It is additive complex Gaussian noise.

[0028] Preferably, step 5 specifically comprises:

[0029] A radial recovery method based on the average radius is adopted;

[0030] First, the receiver calculates the average absolute value of the in-phase and quadrature components of the composite received symbol:

[0031]

[0032] in, This represents the total number of complex received symbols that participated in the statistics; Indicates the in-phase component of the received symbol. This represents the average of the absolute values ​​of the in-phase components of the received symbol. Indicates the orthogonal components of the received symbol. This represents the average of the absolute values ​​of the received symbol quadrature components;

[0033] Then the average radius of the receiving constellation points is calculated:

[0034]

[0035] For any received symbol, the instantaneous amplitude is denoted as:

[0036]

[0037] The received symbol is then radially shrunk proportionally towards the origin to obtain the recovered hidden block component:

[0038]

[0039] After removing padding symbols from the recovery blocks and concatenating them in sequence, the recovered hidden sequence is obtained. :

[0040]

[0041] in, For semantic decoding networks, To reconstruct the data.

[0042] Preferably, step 6 specifically comprises:

[0043] The monitoring end receives the same composite signal And through parameters as Discriminant network Output decision statistics:

[0044]

[0045] The judgment rule is: when If a hidden signal is detected, the monitoring terminal will determine that a normal signal is present; otherwise, it will be determined as a normal signal. For the threshold;

[0046] The WGAN-GP framework is used for training, and the concealment loss at the sending end is defined as:

[0047]

[0048] in, This indicates that the signal is received according to a concealed signal distribution. Composite received signal Perform mathematical expectation calculation; This represents the probability distribution of composite received signals when covert services are present.

[0049] The reconstruction loss at the receiver is defined as:

[0050]

[0051] in, This represents the set of input samples in the current training batch. Indicates the current training batch The average expectation is calculated from the input samples. This represents the original input data. This represents the reconstructed data output by the semantic decoding network at the receiving end. Indicates the dimension of the input data;

[0052] The total loss at the sending end is:

[0053]

[0054] in, and All are weighting coefficients;

[0055] The training loss at the receiver is:

[0056]

[0057] The WGAN-GP loss at the monitoring end is:

[0058]

[0059] in, Indicates the distribution of pseudo-noise reception. This represents the sample distribution obtained by interpolating between the true distribution and the generated distribution. The weights are used for gradient penalty.

[0060] The monitoring end approximates a 1-Lipschitz discriminator to estimate and The sender minimizes the Wasserstein distance between them. make Towards near.

[0061] Preferably, the training phase employs an alternating optimization strategy, in which the monitoring-end discrimination network is updated at least once in each training cycle, and then the sending-end and receiving-end networks are updated, so as to form an adversarial balance between stealth and reconstruction quality.

[0062] Preferably, the Jensen-Shannon divergence, Wasserstein distance, Willie detection accuracy, and publicly available QPSK link bit error rate are used. and reconstruction quality At least one of the indicators is used to evaluate or select parameters for the covert semantic communication method.

[0063] Preferably, the pseudo noise With covert signals Having the same average power constraint ensures that low-power antennas exhibit a consistent power pattern whether covert services are present or not.

[0064] Preferably, the wireless channel includes an AWGN channel, a Rayleigh slow fading channel, a Rician slow fading channel, or a combination thereof.

[0065] A covert semantic communication system includes: a sender, a receiver, and a monitoring unit;

[0066] The transmitting end includes a standard power antenna, a low power antenna, a semantic coding and covert signal generation module, a pseudo noise generation module, a block filling module, and a visible-covert superposition transmission module;

[0067] The receiving end includes a composite signal receiving module, a hidden component recovery module that does not require precise CSI, and a semantic decoding module;

[0068] The monitoring terminal includes a discrimination network module and a threshold decision module.

[0069] The beneficial effects of this invention are as follows:

[0070] 1) A dual-antenna structure that transmits public QPSK signals through a standard power antenna and covert signals or pseudo-noise through a low-power antenna hides covert communication within a legitimate communication context, reducing the probability of detection by the monitoring end.

[0071] 2) Low-power antennas still transmit pseudo noise when there is no covert service. The power of the occupant signal is consistent with that of the covert signal, and the distribution is matched through adversarial training, which can avoid significant statistical differences between the "service-available / service-free" states.

[0072] 3) The receiver adopts a radial recovery method based on I / Q average amplitude and average radius to recover the concealed component without the need for precise instantaneous CSI, which is suitable for AWGN, Rayleigh and Rician channels.

[0073] 4) Combining DeepJSCC-style semantic coding with WGAN-GP-style adversarial distribution matching improves concealment while maintaining the quality of semantic data reconstruction such as images.

[0074] 5) By using parameters such as the transmitter's "SNR", the channel's "SNR", the reconstruction loss weight, and the concealment loss weight, engineering adjustments can be made between the public link "BER", the concealment detection accuracy, and the reconstruction "PSNR". Attached Figure Description

[0075] Figure 1 This is a block diagram of the overall structure of the covert semantic communication system of the present invention.

[0076] Figure 2 This is a flowchart of the covert semantic communication method of the present invention.

[0077] Figure 3 This is a schematic diagram of the adversarial training structure of Alice, Bob, and Willie in an embodiment of the present invention.

[0078] Figure 4 The following are schematic diagrams of the recovery of the hidden component based on the average radius of the receiver in an embodiment of the present invention: (a) schematic diagram of the receiver composite constellation, and (b) schematic diagram of radial contraction recovery.

[0079] Figure 5 This is a schematic diagram illustrating the matching of concealed signal distribution and pseudo-noise distribution, and the Willie decision method in an embodiment of the present invention. Detailed Implementation

[0080] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0081] This invention aims to address the problems of low data rate, fragile recovery, reliance on precise CSI, and significant differences in signal statistical characteristics between the presence and absence of covert services in existing covert communication in actual wireless channels. It proposes a covert semantic communication method and system that enables low-power covert branches to maintain similar power and statistical distribution in both the state of transmitting covert data and transmitting pseudo-noise, and enables the receiver to recover covert semantic information without the need for precise instantaneous CSI.

[0082] The transmitting end, Alice, is configured with a standard power antenna and a low-power antenna: the standard power antenna continuously transmits the publicly available QPSK explicit signal; the low-power antenna transmits a covert signal generated by a neural network when covert services are present, and transmits pseudo-noise with the same or approximately the same average power when covert services are not present. Therefore, even if the external monitoring end, Willie, can receive the composite signal, it is difficult to determine the existence of covert services solely based on changes in the power or distribution of the low-power branch.

[0083] The transmitting end inputs the data to be transmitted into the semantic encoder and the covert signal generator to obtain a complex covert vector; then it performs block segmentation and padding according to the number of time slot symbols to align the covert components with the public QPSK signal in time structure. After the composite signal is transmitted through wireless channels such as AWGN, Rayleigh, or Rician, Bob at the receiving end calculates from the composite received signal... The average amplitude and average radius of the components are calculated, and each received symbol is proportionally shrunk towards the origin to approximately remove the influence of the QPSK explicit components and slow fading amplitude, thereby obtaining the recovered hidden block signal.

[0084] Example:

[0085] (I) System Composition (see) Figure 1 );

[0086] The system in this embodiment includes a transmitter Alice 100, a receiver Bob 200, and a monitoring unit Willie 300. The transmitter Alice 100 includes a semantic coding and covert signal generation module 102, a pseudo-noise generation module 106, a segmentation and padding module 104, a standard power antenna 108, a low-power antenna 105, and a visible / covert overlay module 109. The receiver Bob 200 includes a composite signal receiving module, a covert component recovery module 201, and a semantic decoding module 202. The monitoring unit Willie 300 includes a discrimination network 301 and a threshold decision module 302.

[0087] The standard power antenna 108 is used to transmit the QPSK explicit signal 107 for public services, while the low-power antenna 105 is used to transmit the covert signal 103 or spurious noise. Since the spurious noise and the covert signal are subject to the same average power constraint and are further distributed-matched through adversarial training, the Willie 300 has difficulty distinguishing the statistical differences of the low-power branch in the two states.

[0088] (ii) Semantic coding and covert signal generation at the transmitting end;

[0089] Given the data to be transmitted Taking image data as an example, ,in , , These represent the number of channels, height, and width, respectively. Alice uses the following parameters: Semantic encoding and hidden signal generation network Will Mapped to Complex hidden vector:

[0090]

[0091] The original data dimensions are:

[0092]

[0093] Bandwidth compression ratio is defined as:

[0094]

[0095] Among them, the smaller This indicates a higher degree of communication resource compression. The covert signal generation network may include convolutional layers, PReLU activation layers, Tanh amplitude constraint layers, and complex sign rearrangement layers. The Tanh amplitude constraint layer and output coefficients... Together, they control the amplitude of the concealed signal to make it consistent with the amplitude range of the pseudo-noise.

[0096] (iii) Segmentation, padding, and overlay of explicit and implicit data;

[0097] Assume a single time slot can transmit A complex number symbol. Alice will... According to each Each symbol is divided into a group; if the last group is insufficient... If there are 1 symbol, then it is composed of pseudo-noise. Complete, among which It follows a complex Gaussian noise distribution. The block signal can be represented as:

[0098]

[0099] When covert operations are not present, the low-power antenna does not remain silent but instead transmits pure noise blocks composed of pseudo-noise to simulate hardware noise or low-power background disturbances at the transmitter. An explicit QPSK signal can be represented as:

[0100]

[0101] The explicit QPSK signal and the covert block signal are superimposed in the corresponding time slot to form a composite transmission signal:

[0102]

[0103] In wireless channels such as AWGN, Rayleigh, or Rician, the composite received signal is uniformly represented as:

[0104]

[0105] in, For channel coefficients, It is additive complex Gaussian noise. For slow fading channels, It can remain approximately unchanged within a time slot or several symbol periods.

[0106] (iv) The receiver does not require accurate CSI covert component recovery;

[0107] In the conventional processing method, Bob can first demodulate the QPSK explicit signal, and then from... The covert component is extracted by canceling the open signal. However, in Rayleigh or Rician fading channels, if Bob cannot obtain accurate instantaneous CSI, QPSK demodulation and cancellation errors will significantly affect the recovery of the covert signal. Therefore, this invention employs a radial recovery method based on the average radius.

[0108] First, Bob calculates the average absolute value of the in-phase and quadrature components of the composite received symbol:

[0109]

[0110] in, The number of received symbols is given. Then, the average radius of the receiving constellation points is calculated:

[0111]

[0112] For any received symbol, the instantaneous amplitude is denoted as:

[0113]

[0114] The received symbol is then radially shrunk proportionally towards the origin to obtain the recovered hidden block component:

[0115]

[0116] like Figure 4 As shown, this method utilizes the average radius characteristics formed by a QPSK constellation under slow fading conditions to approximately strip away the principal components of the explicit signal, thereby obtaining low-power concealed components. After removing padding symbols from the recovered blocks and concatenating them sequentially, the recovered concealed sequence is obtained. .

[0117]

[0118] in, For semantic decoding networks, To reconstruct the data.

[0119] (v) Willie discrimination and threshold decision;

[0120] Willie receives the same composite signal And through parameters as Discriminant network Output decision statistics:

[0121] ;

[0122] The judgment rule is: when When the signal is active, Willie determines that a hidden signal exists; otherwise, it is considered a normal signal. Threshold This can be obtained by taking the mean, quantiles, or optimal threshold of the validation set from statistics of a large number of spurious noise samples and hidden samples. This is because Alice uses adversarial training to distribute the received hidden signal. Approximating the pseudo-noise receiver distribution Therefore, Willie's detection accuracy approaches that of random guessing.

[0123] (vi) Adversarial training objectives;

[0124] Willie receives the same composite signal And through parameters as Discriminant network Output decision statistics:

[0125] ;

[0126] To improve concealment and training stability, this invention employs the WGAN-GP framework for training. Alice's concealment loss is defined as:

[0127]

[0128] Bob's reconstruction loss is defined as:

[0129]

[0130] Alice's total loss is:

[0131]

[0132] Bob's training loss is:

[0133]

[0134] Willie's WGAN-GP loss is:

[0135]

[0136] in, This indicates the distribution of covert signal reception. Indicates the distribution of pseudo-noise reception. This represents the sample distribution obtained by interpolating between the true distribution and the generated distribution. Gradient penalty weights. Willie approximates the 1-Lipschitz discriminator to estimate... and Alice minimizes the Wasserstein distance between them. make Towards near.

[0137] (vii) Performance evaluation and parameter setting;

[0138] In one alternative implementation, the sending end Defined as the average power of the public QPSK signal Average power of concealed signal The ratio:

[0139]

[0140] Receiver Channel Defined as:

[0141]

[0142] Covert signal at the receiver It can be represented as:

[0143]

[0144] exist and When all are much greater than 1, its dB is approximately:

[0145]

[0146] Image reconstruction quality can be measured by peak signal-to-noise ratio. measure:

[0147]

[0148] Stealth can be measured using metrics such as Willie detection accuracy, Jensen-Shannon divergence, and Wasserstein distance. In one embodiment, the JS divergence between the stealth signal and the spurious noise can be as low as 0.003267, indicating a high degree of similarity in their statistical distributions; the Willie detection accuracy is close to 0.5, suggesting that the monitoring end is unlikely to outperform random guessing. The bit error rate of the publicly disclosed QPSK link is also discussed. It can be used to assess the interference of covert signals on public services; in engineering, it can be adjusted... Output coefficients The loss weights represent a trade-off between concealment and the quality of the public link.

[0149] (viii) Deployment and implementation methods;

[0150] This invention can be implemented using a software-defined radio platform, a general-purpose radio frequency transceiver platform, an edge intelligent communication device, or an embedded communication terminal. The transmitting end deploys a semantic coding and covert signal generation network, a spurious noise generator, and block-filling logic; the receiving end deploys covert component recovery logic and a semantic decoding network; during the training phase, a Willie discriminant network can be deployed for offline or online adversarial training, and during the deployment phase, the Willie discriminant network can also be used as a covert monitoring and parameter calibration module.

[0151] When the training data is images, the input data can be normalized to [0,1] and then encoded into complex symbols by a convolutional network. The receiver decoder can use deconvolutional layers, PReLU activation layers, and sigmoid output layers to recover the image. For text, sensor data, or task semantic data, the corresponding semantic encoder and decoder can be replaced, as long as they can satisfy the function of mapping the data to complex hidden symbols and being recovered by the receiver.

[0152] This invention can be applied to scenarios such as low-observable communication in the Internet of Things, covert data transmission from unmanned platforms, emergency communication, secure communication for mobile terminals, and military and public safety communication. This solution does not require hiding public communication links; instead, it overlays covert data onto the public communication background and utilizes pseudo-noise occupancy and generative adversarial distribution matching to reduce the detection probability at the monitoring end. It has strong engineering feasibility and scalability.

Claims

1. A method for covert semantic communication, characterized in that, Includes the following steps: Step 1: Obtain the data to be secretly transmitted And through semantic encoding and covert signal generation at the sending end, a network is generated. Generate complex covert signals ; Step 2: Configure a standard power antenna and a low-power antenna at the transmitting end. The standard power antenna is used to transmit explicit QPSK signals. The low-power antenna transmits multiple covert signals when covert services are present. Sending pseudo-noise when no covert service is present ; Step 3: Perform block segmentation and padding on the concealed signal or pseudo-noise to obtain the number of symbols that can be transmitted in a single time slot. Matching block signal ; Step 4: Block signal With explicit QPSK signal The composite received signal is obtained by superimposing the signals in the same or corresponding time slots and transmitting them through a wireless channel. ; Step 5: The receiving end receives the composite signal. The average radius is calculated from the average amplitudes of the in-phase and quadrature components, and the covert block signal is recovered proportionally without prior acquisition or explicit estimation of precise channel state information (CSI). ; for hidden block signals Removing padding symbols and concatenating them yields the recovered hidden sequence. via semantic decoding network Obtain reconstructed data ; Step 6: During the training phase, the discriminant network is monitored. Distinguish between the distribution of hidden signals and pseudo-noise distributions, and optimize based on adversarial loss and reconstruction loss. , and This makes the concealed signal statistically approximate pseudo-noise.

2. The covert semantic communication method according to claim 1, characterized in that, The complex concealment signal is: ； in , Data Dimensions ,in , , These represent the number of channels, height, and width, respectively.

3. The covert semantic communication method according to claim 2, characterized in that, The segmentation and padding include the segmentation of complex concealed signals. By each piece The complex number symbols are divided into multiple hidden blocks. When the last block is insufficient... When the symbol is 1, use pseudo-noise that follows a complex Gaussian distribution. Complete; when there is no covert service, the low-power antenna transmits by The resulting pure noise block; The block signal is represented as: ； in, Represents complex concealed signals The first in There are several complex symbols, i = 1, 2, ..., k; k represents the total number of complex symbols contained in the complex covert signal; N represents the number of complex symbols that can be transmitted in a single time slot; when k is not divisible by N, the portion of the last covert block with fewer than N symbols uses pseudo-noise. Complete; When no covert operations are present, the low-power antenna does not remain silent but instead transmits a pure noise block composed of pseudo-noise to simulate hardware noise or low-power background disturbances at the transmitting end. The explicit QPSK signal is represented as: ； in, This represents the i-th explicit QPSK complex modulation symbol transmitted within a time slot, where i = 1, 2, ..., N; each Select from the preset QPSK constellation symbol set; The explicit QPSK signal and the covert block signal are superimposed in the corresponding time slot to form a composite transmission signal: ； In wireless channels, composite received signals are uniformly represented as: ； in, For channel coefficients, It is additive complex Gaussian noise.

4. The covert semantic communication method according to claim 3, characterized in that, Step 5 specifically involves: A radial recovery method based on the average radius is adopted; First, the receiver calculates the average absolute value of the in-phase and quadrature components of the composite received symbol: ； in, This represents the total number of complex received symbols that participated in the statistics; Indicates the in-phase component of the received symbol. This represents the average of the absolute values ​​of the in-phase components of the received symbol. Indicates the orthogonal components of the received symbol. This represents the average of the absolute values ​​of the received symbol quadrature components; Then the average radius of the receiving constellation points is calculated: ； For any received symbol, the instantaneous amplitude is denoted as: ； The received symbol is then radially shrunk proportionally towards the origin to obtain the recovered hidden block component: ； After removing padding symbols from the recovery blocks and concatenating them in sequence, the recovered hidden sequence is obtained. : ； in, For semantic decoding networks, To reconstruct the data.

5. The covert semantic communication method according to claim 4, characterized in that, Step 6 specifically involves: The monitoring end receives the same composite signal And through parameters as Discriminant network Output decision statistics: ； The judgment rule is: when At that time, the monitoring terminal determined that a hidden signal existed; Otherwise, it is considered a normal signal; For the threshold; The WGAN-GP framework is used for training, and the concealment loss at the sending end is defined as: ； in, This indicates that the signal is received according to a concealed signal distribution. Composite received signal Perform mathematical expectation calculation; This represents the probability distribution of composite received signals when covert services are present. The reconstruction loss at the receiver is defined as: ； in, This represents the set of input samples in the current training batch. Indicates the current training batch The average expectation is calculated from the input samples. This represents the original input data. This represents the reconstructed data output by the semantic decoding network at the receiving end. Indicates the dimension of the input data; The total loss at the sending end is: ； in, and All are weighting coefficients; The training loss at the receiver is: ； The WGAN-GP loss at the monitoring end is: ； in, Indicates the distribution of pseudo-noise reception. This represents the sample distribution obtained by interpolating between the true distribution and the generated distribution. The weights are used for gradient penalty. The monitoring end approximates a 1-Lipschitz discriminator to estimate and The sender minimizes the Wasserstein distance between them. make Towards near.

6. The covert semantic communication method according to claim 5, characterized in that, The training phase employs an alternating optimization strategy, where the monitoring-end discrimination network is updated at least once in each training cycle, followed by updates to the sending-end and receiving-end networks, in order to achieve an adversarial balance between stealth and reconstruction quality.

7. The covert semantic communication method according to claim 6, characterized in that, The accuracy of Jensen-Shannon divergence, Wasserstein distance, Willie detection, and publicly available QPSK link bit error rate were used to measure these parameters. and reconstruction quality At least one of the indicators is used to evaluate or select parameters for the covert semantic communication method.

8. The covert semantic communication method according to claim 7, characterized in that, The pseudo noise With covert signals Having the same average power constraint ensures that low-power antennas exhibit a consistent power pattern whether covert services are present or not.

9. A covert semantic communication method according to claim 8, characterized in that, The wireless channels include AWGN channels, Rayleigh slow fading channels, Rician slow fading channels, or combinations thereof.

10. A covert semantic communication system employing the covert semantic communication method as described in claim 1, characterized in that, include: Sending end, receiving end, and monitoring end; The transmitting end includes a standard power antenna, a low power antenna, a semantic coding and covert signal generation module, a pseudo noise generation module, a block filling module, and a visible-covert superposition transmission module; The receiving end includes a composite signal receiving module, a hidden component recovery module that does not require precise CSI, and a semantic decoding module; The monitoring terminal includes a discrimination network module and a threshold decision module.

Images