Underwater communication method based on neural network model

Abstract

The invention provides an underwater communication method based on a neural network model, the neural network model at least comprises a watermark embedding layer module and a watermark extraction layer module, and the method comprises the following steps: obtaining a first carrier signal, and embedding watermark information into the first carrier signal through the watermark embedding layer module to obtain a second carrier signal; and transmitting the second carrier signal through a transmitting end in an underwater environment, receiving the second carrier signal through a receiving end in the underwater environment, and extracting watermark information from the received second carrier signal through the watermark extraction layer module.

Description

Technical Field

[0001] The present invention relates to the technical field of underwater communication, and particularly to an underwater communication method based on a neural network model. Background Art

[0002] Covert underwater acoustic communication, as an effective solution to the inherent security problems of underwater acoustic communication, has become a research hotspot in recent years. At present, although neural network-based technologies have achieved remarkable results in air acoustic fields such as speech and natural environment audio, in the underwater acoustic field, especially in the application of covert underwater acoustic communication, signal generation and embedding methods based on deep neural networks are still in the research stage, and there are significant differences in physical characteristics between underwater acoustic signals and air acoustic signals. Therefore, how to use neural network technology to embed signals according to the characteristics of underwater acoustic signals and improve the concealment and robustness of underwater acoustic communication has become an urgent problem to be solved.

[0003] Therefore, an underwater communication method based on a neural network model is needed. Summary of the Invention

[0004] The object of the present invention is to provide an underwater communication method based on a neural network model, which embeds watermark information into ship radiated noise to address the security risks in existing covert underwater communication technologies.

[0005] To achieve the above object, in a first aspect, the present invention provides an underwater communication method based on a neural network model. The neural network model at least includes a watermark embedding layer module and a watermark extraction layer module. The method includes:

[0006] Obtain a first carrier signal, and embed watermark information into the first carrier signal through the watermark embedding layer module to obtain a second carrier signal;

[0007] Transmit the second carrier signal through a transmitter in an underwater environment, receive the second carrier signal at a receiver in the underwater environment, and extract watermark information from the received second carrier signal through the watermark extraction layer module.

[0008] Preferably, the first carrier signal is ship radiated noise.

[0009] Specifically, the watermark embedding layer module includes a watermark encoder and a carrier encoder;

[0010] Embedding watermark information into the first carrier signal through the watermark embedding layer module to obtain a second carrier signal includes:

[0011] The first amplitude spectrum and the first phase information corresponding to the first carrier signal are obtained through short-time Fourier transform. Based on a plurality of preset convolutional kernels, the first amplitude spectrum is amplified, and the first amplitude spectrum feature corresponding to the amplified first amplitude spectrum is obtained through the carrier encoder. The first watermark feature corresponding to the watermark information is obtained through the watermark encoder and the first amplitude spectrum feature;

[0012] A fusion feature is obtained according to the first watermark feature and the first amplitude spectrum, and a second carrier signal with a watermark is constructed according to the fusion feature and the first phase information.

[0013] Specifically, the neural network model further includes an embedding network layer;

[0014] Obtaining a fusion feature according to the first watermark feature and the first amplitude spectrum, and constructing a second carrier signal with a watermark according to the fusion feature and the first phase information, includes:

[0015] The first watermark feature is repeatedly copied according to the number of time frames of the first amplitude spectrum to obtain a second watermark feature of the same size as the first amplitude spectrum feature; the second watermark feature and the amplified first amplitude spectrum are concatenated with the first amplitude spectrum feature to obtain a concatenated feature, and the concatenated feature is input into the embedding network layer to obtain a fusion feature. According to the fusion feature and the first phase information, an inverse short-time Fourier transform is used to construct a second carrier signal with a watermark.

[0016] Specifically, the watermark extraction module includes a watermark extractor and a watermark decoder;

[0017] Extracting watermark information from the received second carrier signal through the watermark extraction layer module includes:

[0018] The second amplitude spectrum and the second phase information corresponding to the second carrier signal are obtained through short-time Fourier transform. The third watermark feature corresponding to the second amplitude spectrum is obtained through the watermark extractor, and the watermark information corresponding to the third watermark feature is obtained through the watermark decoder.

[0019] Specifically, obtaining the third watermark feature corresponding to the second amplitude spectrum through the watermark extractor and obtaining the watermark information corresponding to the third watermark feature through the watermark decoder includes:

[0020] The watermark feature of each time frame of the second amplitude spectrum is extracted through the watermark extractor, the watermark features are averaged along the time axis to obtain a third watermark feature, and the third watermark feature is subjected to feature mapping through the watermark decoder and the decoded watermark information is output.

[0021] Preferably, the neural network model further includes a distortion layer module, which is used in the training stage of the neural network. The distortion layer module receives the second carrier signal obtained by embedding the watermark information into the first carrier signal through the watermark embedding layer module, applies an analog interference to the second carrier signal in the underwater transmission environment, and inputs the second carrier signal after the interference is applied into the watermark extraction layer module to extract the watermark information.

[0022] Specifically, the neural network model is trained by means of adversarial training.

[0023] In a second aspect, the present invention provides an electronic device, including: a processor, a memory, and computer program instructions stored on the memory and running on the processor. When the processor executes the computer program instructions, it is used to implement the method described in the first aspect.

[0024] In a third aspect, the present invention provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the method described in the first aspect. Description of the Drawings

[0025] Figure 1 It is a flowchart of an underwater communication method based on a neural network model provided by an embodiment of the present invention;

[0026] Figure 2 It is a schematic diagram of the neural network model structure provided by an embodiment of the present invention;

[0027] Figure 3 It is a flowchart of the watermark embedding layer provided by an embodiment of the present invention;

[0028] Figure 4 It is a flowchart of the watermark extraction layer provided by an embodiment of the present invention;

[0029] Figure 5 It is a diagram of simulation environment parameter settings provided by an embodiment of the present invention;

[0030] Figure 6 It is a curve graph of the change of BER with SNR at different transmission rates provided by an embodiment of the present invention;

[0031] Figure 7 It is a signal-to-noise ratio graph at different transmission rates provided by an embodiment of the present invention. Detailed Embodiments

[0032] Next, through the drawings and embodiments, the technical solutions of the present invention will be further described in detail.

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be described below with reference to the accompanying drawings. It should be noted that like reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0034] In the description of the embodiments of the present invention, words such as "exemplary", "for example", or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design solution described as "exemplary", "for example", or "for instance" in the embodiments of the present invention should not be construed as being more preferred or having more advantages than other embodiments or design solutions. Rather, the use of words such as "exemplary", "for example", or "for instance" is intended to present relevant concepts in a specific manner.

[0035] Existing covert underwater communication technologies can generally be divided into two categories: one is a low-detection-rate system, such as a spread-spectrum communication system, which aims to prevent communication signals from being detected, thereby ensuring the security of the communication platform; the other is a low-recognition-rate system, which aims to prevent communication signals from being cracked and ensure information security even if the signals are detected. Low-detection-rate systems usually use artificial signals that do not exist in the ocean environment, so they may be detected by passive sonar, especially when the receiver is close to the transmitter. The low-recognition-rate system, on the other hand, uses inherent acoustic signals in the ocean environment, such as the calls of marine mammals and ship radiated noise, to conceal communication signals. However, existing low-recognition-rate systems usually use manual modulation or parameter embedding based on signal models to transmit information, resulting in a large difference between the communication signal and the carrier wave, thereby reducing the concealment effect.

[0036] The present invention proposes an underwater communication method based on a neural network model, which embeds watermark information into the carrier signal to address the security risks in existing covert underwater communication technologies. This method performs end-to-end watermark embedding and extraction through a neural network model, and improves the robustness of the communication system in a complex ocean environment through a multiple distortion training strategy. At the same time, a feature fusion method is adopted in the watermark embedding layer, enabling the watermark information to effectively resist signal loss and distortion interference while maintaining a low bit error rate. Compared with traditional manual modulation or signal model parameter embedding methods, the present invention can more effectively embed information into ship radiated noise, improving the security and concealment of underwater communication.

[0037] Specifically, according to an underwater communication method based on a neural network model provided by an embodiment of the present invention, the neural network model at least includes a watermark embedding layer module and a watermark extraction layer module. Figure 1 The flowchart of an underwater communication method based on a neural network model provided by an embodiment of the present invention is asFigure 1 As shown, the method at least includes the following steps: Step S101: Obtain a first carrier signal, and embed watermark information into the first carrier signal through the watermark embedding module to obtain a second carrier signal; Step S102: Transmit the second carrier signal through a transmitting end in an underwater environment, receive the second carrier signal at a receiving end in the underwater environment, and extract the watermark information from the received second carrier signal through the watermark extraction module.

[0038] First, in step S101, a first carrier signal can be obtained, and watermark information is embedded into the first carrier signal through the watermark embedding module to obtain a second carrier signal. In different embodiments, the first carrier signal can be different audio carrier signals. In a specific embodiment, the first carrier signal can be ship radiated noise.

[0039] Figure 2 Schematic diagram of the neural network model structure provided by the embodiment of the present invention. As Figure 2 shown, the neural network model is divided into two parts: a watermark embedding layer and a watermark extraction layer. In the training stage, the neural network model can also include a distortion layer. Among them, the watermark embedding layer is used to implement the watermark embedding operation, the watermark extraction layer is used to implement the watermark extraction operation, and the distortion layer simulates the distortion caused by the signal during the transmission process in the underwater environment. As Figure 2 shown, the training of the entire watermark embedding and extraction model is completed end-to-end: In the embedding stage, the audio carrier signal and the watermark information respectively pass through the carrier encoder and the watermark encoder to obtain the carrier feature (amplitude spectrum feature) and the watermark feature, and are simultaneously input into the embedding network layer for the fusion of the watermark feature and the carrier feature. The output fusion feature is reconstructed back to the audio to obtain the watermarked audio carrier signal, and then input into the distortion layer to simulate the interference of the signal during the transmission process in the underwater environment. Finally, in the watermark extraction layer, the signal passing through the distortion layer is extracted and decoded to obtain the watermark information to complete the communication process.

[0040] In one embodiment, the watermark embedding layer module may include a watermark encoder and a carrier encoder; embedding watermark information into the first carrier signal through the watermark embedding layer module, comprising: obtaining a first amplitude spectrum and first phase information corresponding to the first carrier signal through short-time Fourier transform, amplifying the first amplitude spectrum based on a plurality of preset convolutional kernels, and obtaining a first amplitude spectrum feature corresponding to the amplified first amplitude spectrum through the carrier encoder, and obtaining a first watermark feature corresponding to the watermark information through the watermark encoder and the first amplitude spectrum feature; specifically, the neural network model further includes an embedding network layer; repeating and copying the first watermark feature according to the number of time frames of the first amplitude spectrum to obtain a second watermark feature of the same size as the first amplitude spectrum feature; splicing the second watermark feature, the amplified first amplitude spectrum and the first amplitude spectrum feature to obtain a splicing feature, inputting the splicing feature into the embedding network layer to obtain a fusion feature, and constructing a watermark-bearing second carrier signal through inverse short-time Fourier transform according to the fusion feature and the first phase information.

[0041] Figure 3 The flowchart of the watermark embedding layer provided by the embodiment of the present invention is as Figure 3 shown. In the watermark embedding stage, the model takes a watermark vector and a carrier signal as inputs. First, the audio carrier signal is transformed into the frequency domain through short-time Fourier transform (STFT) to obtain an amplitude spectrum and phase information, where the amplitude spectrum serves as the carrier for watermark embedding, and the phase information is only used for subsequent audio reconstruction operations. The carrier encoder is responsible for extracting the features of the amplitude spectrum, and the watermark encoder encodes the watermark information to be embedded into a feature vector. Since the signal received at the receiving end is distorted in the time domain in a complex marine environment, in order to enhance the robustness of the watermark to time-domain operations, the encoded watermark features are repeatedly copied (Repeat) according to the number of time frames of the amplitude spectrum to obtain features of the same size as the carrier amplitude spectrum features, and then are spliced with the original amplitude spectrum in the channel dimension with the carrier features, so that the watermark information is added to each time frame, ensuring that accurate watermark information can still be decoded in the case of some signal segments being lost. The spliced features are input into the embedding network layer to fuse the watermark features and the amplitude spectrum features to generate a watermark-bearing amplitude spectrum, and the spliced features are input into the embedding network layer to obtain a fusion feature. Finally, in combination with the original phase information, an audio carrier signal with a watermark is reconstructed through inverse short-time Fourier transform (ISTFT).

[0042] For example, when encoding an audio carrier signal, its amplitude spectrum is generally encoded to obtain amplitude spectrum features (carrier features). The size of each dimension of the amplitude spectrum is, for example, (1, 1025, 87), where the first dimension 1 is the batch size, the second dimension 1025 is the frequency dimension, and the third dimension 87 is the time dimension (representing the number of time frames). First, the convolution channel dimension (abbreviated as the channel dimension) is amplified to obtain a size of each dimension of (1, 1, 1025, 87). At this time, the second dimension becomes the channel dimension. The carrier encoder uses 64 groups (each 3×3) of convolution kernels to operate on the channel dimension, so that the output of the channel dimension is mapped to a 64-dimensional channel dimension, obtaining amplitude spectrum features with a size of each dimension of (1, 64, 1025, 87). Not only does the frequency dimension change, but the channel dimension also becomes 64. For the watermark information, it is encoded into a dimension size that can be concatenated with the amplitude spectrum features. For example, for 30-bit watermark information, the length of each dimension is, for example, (1, 30). First, it is mapped to 1025 bits through a linear layer to obtain watermark information with a dimension of (1, 1025). To improve the robustness of the neural network model, the watermark information can be aligned with the amplitude spectrum features in each frame (doing so is to ensure that even if the received audio at the receiving end has losses in the time dimension, the watermark information in a small section of the audio carrier signal with a watermark can be decoded because each frame has complete watermark information and there is no worry about the audio being truncated). Therefore, such watermark features are repeatedly copied 87 times in the time dimension to obtain a dimension of (1, 1025, 87), and then the channel dimension is expanded and set to 1 to obtain new watermark features with a dimension of (1, 1, 1025, 87). In this way, the watermark features after watermark encoding and the amplitude spectrum features after carrier signal encoding can be concatenated along the channel dimension. To enable the model to focus on the original amplitude spectrum information, the amplitude spectrum with a size of each dimension of (1, 1, 1025, 87) can also be concatenated to obtain concatenated features with a dimension of (1, 66, 1025, 87). Passing the concatenated features through the embedding network layer obtains fused features with a dimension of (1, 1, 1025, 87), which contain watermark information. Inverse transformation of it obtains the audio carrier signal with a watermark.

[0043] Thereafter, in step S102, the second carrier signal can be transmitted by the transmitting end in an underwater environment, and the second carrier signal can be received by the receiving end in the underwater environment. The watermark extraction module extracts watermark information from the received second carrier signal.

[0044] In one embodiment, the watermark extraction module may include a watermark extractor and a watermark decoder; the watermark information is extracted from the received second carrier signal through the watermark extraction layer module. Specifically, the second amplitude spectrum and the second phase information corresponding to the second carrier signal can be obtained through short-time Fourier transform. The watermark features of each time frame of the second amplitude spectrum are extracted by the watermark extractor, the third watermark features are obtained by averaging the watermark features along the time axis, and the third watermark features are subjected to feature mapping by the watermark decoder to output the decoded watermark information.

[0045] Figure 4 The flowchart of the watermark extraction layer provided by the embodiment of the present invention is shown in Figure 4 As shown, the watermarked audio carrier signal is converted into an amplitude spectrum and a phase spectrum through short-time Fourier transform (STFT), and the signal amplitude spectrum is input into the watermark extractor. Since the signal is distorted by some complex environments during the input process, in order to avoid damage to some segments of the signal and affect the overall watermark information extraction, the opposite operation of the watermark embedding layer is adopted. The watermark extractor extracts the watermark features of each time frame of the amplitude spectrum and averages these watermark features along the time axis, which improves the robustness of the watermark extraction operation. Finally, the obtained watermark features are input into the watermark decoder for feature mapping to output the decoded watermark information. The entire watermark extraction process is optimized by calculating the mean square error loss between the decoded watermark and the original watermark to ensure the accurate extraction of the watermark.

[0046] In one embodiment, the neural network model further includes a distortion layer module. The distortion layer module is used in the training stage of the neural network. It receives the second carrier signal obtained by embedding the watermark information into the first carrier signal through the watermark embedding layer module, applies simulated interference to the second carrier signal in the underwater transmission environment, and inputs the second carrier signal after the interference is applied into the watermark extraction layer module to extract the watermark information.

[0047] For example, as shown in Figure 2 As shown, the distortion layer is used to simulate the interference suffered by the signal in the real transmission environment. The distortion layer is added to the entire training process to train an end-to-end model that can resist the interference of complex underwater environments. First, multiple simulated underwater acoustic channel impulse responses are generated using the Bellhop model according to the settings of different environmental parameters (such as communication distance, transducer and receiver depth, etc.). Secondly, in order to simulate the ocean noise in the real transmission environment, white noise with a certain signal-to-noise ratio is added. During the training stage, whenever a sample passes through the distortion layer, the underwater acoustic channel is randomly sampled and white noise is added to simulate different transmission environments to ensure that the model learns the characteristics of the underwater acoustic channel.

[0048] As shown in Figure 3As shown, in one embodiment, to achieve the concealment of watermark embedding, during the training phase, an encoding loss function and a discriminator network are also added to the watermark embedding layer. The encoding loss is optimized using the mean squared error loss function. The discriminator network is used to determine whether the audio carrier signal embedded with the watermark can effectively "fool" the discriminator so that it cannot distinguish the watermarked audio from the original audio, thereby improving the concealment of the watermark. Through adversarial training, the discriminator and the watermark embedding network are jointly optimized, prompting the embedding process to generate a more natural and imperceptible watermark, which helps reduce audio distortion. Therefore, in a specific embodiment, the neural network model can also be trained through adversarial training. For example, the Encoder can be used as the generator (that is, encoding the watermark into the audio carrier signal), and the Discriminator can be used as the discriminator, and the two play against each other to achieve high-quality watermark embedding. Among them, the goal of the encoder is to embed the watermark information into the audio while ensuring that the watermarked audio is as consistent with the original audio as possible to deceive the discriminator. The goal of the discriminator is to distinguish whether an audio has a watermark. Finally, under the combined action of the two, the encoder can generate a sufficiently concealed watermarked audio that is difficult for the discriminator to distinguish. Therefore, the loss function of the training process is mainly composed of the audio reconstruction loss and the watermark decoding loss, which respectively measure the fidelity of the audio quality and the accuracy of watermark extraction. The final generation loss is mainly the weighted sum of these two parts. The discriminator loss also needs to calculate the gradient backpropagation for iterative update, mainly to enable the discriminator to better distinguish whether an audio has a watermark. The generation and discrimination together make the generated audio of better quality and higher decoding accuracy.

[0049] In one embodiment, the method based on the neural network model for underwater communication provided in the embodiments of the present invention is verified on the DeepShip dataset. This dataset specifically includes 265 different ships, divided into four types: oil tankers, tugboats, passenger ships, and cargo ships. All the data in the DeepShip dataset is divided into a training set, a validation set, and a test set at a ratio of 90:5:5. During the training phase, 500 different underwater acoustic channel impulse responses are generated using the Bellhop model. To simulate the interference of ocean noise, white noise with a signal-to-noise ratio (SNR) of -5 dB is introduced. Regarding the analysis and evaluation of indicators, since in the actual application scenario of covert underwater acoustic communication, more attention is paid to the information transmission rate, the accuracy of information decoding at the receiving end, and the concealment of signal transmission. Therefore, in the experimental phase of the present invention, the bit error ratio (BER) under different signal-to-noise ratio conditions is used to measure the decoding performance of the system. Figure 5 The figure shows the parameter settings of the simulation environment provided in the embodiments of the present invention, as Figure 5As shown, in order to evaluate the performance of the system, environmental parameters, including receiver parameters, etc., can also be set during the test phase of the model. Figure 6 The graph shows the variation of BER with SNR at different transmission rates provided by the embodiment of the present invention, as Figure 6 shown. Under the condition of -5dB, the present invention achieves a communication rate of 50 bits per second (bps), and the bit error rate is 10 -3 , especially at a communication rate of 30bps, a bit error rate of 10 -4 is achieved. The results show that although the bit error rate decreases with the increase of the transmission rate, the present invention shows excellent performance under the condition of a transmission rate within 50bps. Figure 7 The graph shows the signal-to-noise ratio at different transmission rates provided by the embodiment of the present invention, as Figure 7 shown, the signal-to-noise ratio (SNR) between the carrier wave and the watermark signal at different communication rates.

[0050] According to an embodiment of another aspect, a computer-readable medium is also provided, including a computer program stored thereon, and the computer executes the above method when running.

[0051] According to an embodiment of another aspect, a computing device is also provided, including a memory and a processor. An executable code is stored in the memory, and when the processor executes the executable code, the above method is implemented.

[0052] The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve the desired results. Additionally, the processes depicted in the figures do not necessarily require the particular order or sequential order shown to achieve the desired result. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0053] Those skilled in the art should further realize that the units and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. To clearly illustrate the interchangeability of hardware and software, the components and steps of the examples have been generally described according to their functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Skilled artisans can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of the present invention.

[0054] The steps of the methods or algorithms described in connection with the embodiments disclosed herein may be implemented by hardware, software modules executed by a processor, or a combination of both. The software modules may be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0055] The specific embodiments described above further elaborate on the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the protection scope of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An underwater communication method based on a neural network model, the neural network model at least including a watermark embedding layer module and a watermark extraction layer module, the method comprising: Obtain a first carrier signal, and embed watermark information into the first carrier signal through the watermark embedding layer module to obtain a second carrier signal; Transmit the second carrier signal through a transmitter in an underwater environment, receive the second carrier signal at a receiver in the underwater environment, and extract the watermark information from the received second carrier signal through the watermark extraction layer module.

2. The method according to claim 1, wherein The first carrier signal is ship radiated noise.

3. The method according to claim 1, wherein The watermark embedding layer module includes a watermark encoder and a carrier encoder; Embedding the watermark information into the first carrier signal through the watermark embedding layer module to obtain a second carrier signal, comprising: Obtain a first amplitude spectrum and first phase information corresponding to the first carrier signal through short-time Fourier transform, amplify the first amplitude spectrum based on a plurality of preset convolutional kernels, obtain a first amplitude spectrum feature corresponding to the amplified first amplitude spectrum through the carrier encoder, and obtain a first watermark feature corresponding to the watermark information through the watermark encoder and the first amplitude spectrum feature; Obtain a fusion feature according to the first watermark feature and the first amplitude spectrum, and construct a watermarked second carrier signal according to the fusion feature and the first phase information.

4. The method according to claim 3, wherein The neural network model further includes an embedding network layer; Obtain a fusion feature according to the first watermark feature and the first amplitude spectrum, and construct a watermarked second carrier signal according to the fusion feature and the first phase information, comprising: Repeat and copy the first watermark feature according to the number of time frames of the first amplitude spectrum to obtain a second watermark feature of the same size as the first amplitude spectrum feature; Splice the second watermark feature, the amplified first amplitude spectrum and the first amplitude spectrum feature to obtain a spliced feature, input the spliced feature into the embedding network layer to obtain a fusion feature, and construct a watermarked second carrier signal through inverse short-time Fourier transform according to the fusion feature and the first phase information.

5. The method according to claim 1, wherein, The watermark extraction module includes a watermark extractor and a watermark decoder; Extracting the watermark information from the received second carrier signal through the watermark extraction layer module, comprising: Obtain a second amplitude spectrum and second phase information corresponding to the second carrier signal through short-time Fourier transform, obtain a third watermark feature corresponding to the second amplitude spectrum through the watermark extractor, and obtain the watermark information corresponding to the third watermark feature through the watermark decoder.

6. The method according to claim 5, wherein obtaining the third watermark feature corresponding to the second amplitude spectrum through the watermark extractor and obtaining the watermark information corresponding to the third watermark feature through the watermark decoder, comprising: Extract the watermark feature of each time frame of the second amplitude spectrum through the watermark extractor, average the watermark features along the time axis to obtain a third watermark feature, and perform feature mapping on the third watermark feature through the watermark decoder and output the decoded watermark information.

7. The method according to claim 1, wherein The neural network model further includes a distortion layer module, which is configured to, during the training phase of the neural network, receive a second carrier signal obtained by embedding watermark information into a first carrier signal through a watermark embedding layer module, apply simulated interference in an underwater transmission environment to the second carrier signal, and input the second carrier signal with the interference applied thereto into the watermark extraction layer module to extract the watermark information.

8. The method according to claim 1, wherein the neural network model is trained by means of adversarial training.

9. An electronic device, comprising: A processor, a memory, and computer program instructions stored on the memory and running on the processor, wherein when the processor executes the computer program instructions, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium, wherein, Computer-executable instructions are stored in the computer-readable storage medium, and when the computer-executable instructions are executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images