Semantic communication system design method oriented to task security

Abstract

The invention discloses a task security-oriented semantic communication system design method, which relates to the technical field of communication, and comprises the following implementation steps of: acquiring a data set, and dividing the data set into a training set and a test set; preprocessing the data; constructing a channel environment, and designing a related transmission architecture of the security semantic communication system, including a sending end, a receiving end and an eavesdropping end; pre-training is combined to construct a security semantic communication system; loading a pre-trained parameter model, and finely tuning the training model by using a generative adversarial network loss function and a security semantic feature loss function; inputting test set data into the network; outputting a result and verifying the effectiveness of the model. According to the method, the image reconstruction quality and the classification accuracy of the eavesdropping end are reduced by utilizing the generative adversarial network and the security semantic feature loss function, the performance security of a downstream AI task is ensured from the semantic perspective, and the application of semantic communication in an eavesdropping scene is expanded.

Description

Technical Field

[0001] The present invention relates to the field of communication technology, and in particular to a task safety-oriented semantic communication system design method. Background Art

[0002] As communication technology continues to evolve towards 6G, traditional communication technology based on Shannon's information theory is gradually approaching its performance boundaries and is increasingly unable to meet the stringent requirements of 6G in terms of capacity and energy efficiency. As a novel communication paradigm, semantic communication is considered to be the key to solving the capacity bottleneck of traditional communication systems, thus attracting widespread attention from researchers. Unlike traditional communication systems, semantic communication uses deep learning technology to encode input signals into representations containing semantic information, and then performs semantic interpretation at the receiving end.

[0003] At present, with the rapid development of artificial intelligence technology, semantic communication research based on deep learning has made significant progress in some aspects. However, in semantic communication, the inherent openness and easy accessibility of wireless channels make the transmitted semantic information easy to be intercepted by eavesdroppers. This not only endangers the privacy of the sender, but also brings significant security risks to legitimate users. Therefore, in order to ensure the security of semantic communication systems, it is necessary to implement strong security mechanisms to prevent the leakage of semantic information.

[0004] In order to ensure the security of semantic communication, researchers have proposed a variety of semantic security schemes. For example, a new pixel-level secure mean square error loss function is used to train semantic encoders and decoders, which can flexibly balance communication efficiency and privacy leakage to achieve secure semantic communication for image transmission. However, the above work only considers the semantic communication security of image reconstruction tasks, but does not consider the security of downstream AI tasks. In contrast, another method that uses simultaneous transmission and reflection reconfigurable intelligent surface (STAR-RIS) in semantic communication systems to achieve privacy protection and reduce the success rate of eavesdroppers' tasks is to convert signals into task-level interference specifically for eavesdroppers to reduce the success rate of downstream AI tasks of eavesdroppers, but this method is implemented from a physical layer perspective rather than a semantic perspective. It can be seen that none of the above methods can solve the problem of semantic communication security. Therefore, based on the above discussion, it is necessary to design a semantic communication system from a semantic perspective to ensure the performance security of downstream AI tasks. Summary of the invention

[0005] The present invention aims to solve one of the technical problems existing in the related art at least to a certain extent.

[0006] One object of the present invention is to provide a task-safety-oriented semantic communication system design method, to establish a task-safety-oriented semantic communication system based on a generative adversarial network, to integrate the generative adversarial network into the semantic communication system, and to use the semantic-level information required by downstream AI tasks to design a safety loss function to train the semantic communication system, thereby ensuring the performance safety of downstream AI tasks and achieving the goal of task-oriented safe semantic communication.

[0007] In order to achieve the above-mentioned object, the present invention provides a task-safety-oriented semantic communication system design method, comprising: S1. Obtain an image data set, divide the data set into a training set and a test set, and perform data preprocessing on the images in the training set and the test set respectively; S2. Build a channel environment and design a semantic communication model consisting of the transmission architecture of the sender, receiver, and eavesdropper. S3,semantic communication model constructed by joint pre-training using training set; S4, loading the pre-trained semantic communication model, and fine-tuning the parameters of the semantic communication model using the security semantic feature loss function; S5. Verify the effectiveness of the trained semantic communication model through the test set and obtain a semantic communication system oriented to task safety.

[0008] A further preferred technical solution of the present invention is that in step S1, data preprocessing is performed on the images in the training set, including: Fill the images in the training set with zero values ​​for a certain number of pixels, and then randomly crop out fixed-size areas; The cropped image is horizontally flipped with a preset probability, and the rest of the image remains the same; Randomly block a fixed-size area on the image; Normalize the images in the training set.

[0009] Preferably, in step S1, data preprocessing is performed on the images in the test set, including: The images in the test set are normalized using the same mean and standard deviation as the images in the training set.

[0010] Preferably, the semantic communication model composed of the transmission architecture of the sending end, the receiving end and the eavesdropping end in step S2 is constructed as follows: After the image at the sender is semantically encoded and channel encoded, the feature map is output. The feature map is perturbed by the generative network. After the perturbation is added, the dimension of the feature map remains unchanged, which is expressed as: ; in, is the semantic encoder, is the input image, are the neural network parameters of the semantic encoder: is the channel encoder, are the neural network parameters of the channel encoder; To generate the network, Neural network parameters for generating the network; The input signal is then transmitted through the constructed AWGN channel, and the signals received by the receiving end and the eavesdropping end are expressed as: ; ; in, and Respectively represent the signals received by the receiving end and the eavesdropping end, is the input signal of the channel, and denote the independent and identically distributed Gaussian noises at the receiving end and the eavesdropping end respectively; The signals received by the receiving end and the eavesdropping end are reconstructed into images through the channel decoder and the semantic decoder. The image reconstruction process is expressed as: ; ; in, and Reconstruct the image for the receiving end and the eavesdropping end respectively; and They are channel decoder and semantic decoder respectively. and are the neural network parameters of the channel decoders at the receiving end and the eavesdropping end, respectively. and are the neural network parameters of the semantic decoder at the receiving end and the eavesdropping end, respectively.

[0011] Preferably, step S3 uses the semantic communication model constructed by the training set combined with pre-training, specifically: using a pixel-level mean square error loss function to pre-train the semantic communication model composed of the transmission architecture of the transmitter, the receiver and the eavesdropper, and the pixel-level mean square error loss function is expressed as: ; Among them, MSE represents mean square error calculation.

[0012] Preferably, the step S4 loads the pre-trained semantic communication model and uses the security semantic feature loss function to fine-tune the parameters of the semantic communication model, specifically including: The semantic encoder and channel encoder at the sender, as well as the semantic decoder and channel decoder at the receiver and eavesdropper are fine-tuned using a secure semantic feature loss function; When the classification task is used as the downstream AI task, the original image, the receiving end, and the eavesdropping end reconstructed image are respectively input into the feature extraction network. The feature extraction network has been pre-trained for the classification task. The security semantic feature loss function is expressed as: ; in, For the The original image feature map, and Respectively The receiving end reconstructs the image and The eavesdropping end reconstructs the image feature map, is the task weight that controls the balance, is the amount of data in each batch.

[0013] Preferably, the receiving end of the semantic communication model has a discriminant network corresponding to the generative network of the sending end. In step S2, after the signals received at the receiving end and the eavesdropping end are reconstructed through the channel decoder and the semantic decoder, the reconstructed image is input into the discriminant network of the receiving end, and adversarial training is performed with the generative network.

[0014] Preferably, when fine-tuning the parameters of the semantic communication model in step S4, the method further includes: The generative network at the sender and the discriminative network at the receiver are jointly trained using the generative adversarial network loss function. The optimization goal of the generative network is to maximize the output of the discriminative network at the receiving end to reconstruct the image, thereby deceiving the discriminative network; the optimization goal of the discriminative network is to make the output of the original image as large as possible and the output of the reconstructed image at the receiving end as small as possible to distinguish it from the original image; the loss function of the generative adversarial network is expressed as: ; ; in, is the loss function of the generated network, is the loss function of the discriminant network, represents the discriminant network, is the neural network parameter of the discriminant network, To take the mean value, is the hyperparameter of the regularization term, To limit the disturbance amplitude Norm.

[0015] Preferably, step S5 verifies the validity of the trained semantic communication model through a test set, including: From a semantic perspective, FPSNR is used as a performance evaluation indicator, expressed as: ; in, Represents the maximum activation value of the original image feature map; It represents the mean square error based on the feature level and is expressed as , is the feature map extracted from the original image after passing through the feature extraction network. It is the feature map extracted after the reconstructed image passes through the feature extraction network.

[0016] Another aspect of the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, wherein the computer instructions enable a computer to execute the above-mentioned task-safety-oriented semantic communication system design method.

[0017] Yet another aspect of the present invention provides an electronic device, comprising: a processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other via the communication bus, and the processor calls the logic instructions in the memory to execute the above-mentioned task-safety-oriented semantic communication system design method.

[0018] Yet another aspect of the present invention provides a computer program product, comprising a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer executes the above-mentioned task-safety-oriented semantic communication system design method.

[0019] Beneficial effects: The present invention discloses a method for designing a semantic communication system for task security, targeting the information security issues in the image transmission process. The method constructs a secure semantic communication architecture including a transmitter, a receiver and an eavesdropper, and uses semantic coding, channel coding and generative adversarial network technology to perform feature extraction and perturbation processing on the image. During the transmission process, the present invention uses an AWGN channel to simulate the actual communication environment, and performs signal processing and image reconstruction at the receiver and the eavesdropper, respectively. By designing a secure semantic feature loss function and a generative adversarial network loss function, the present invention realizes joint optimization training of the system, effectively improves the image reconstruction quality of the legitimate receiver, and reduces the reconstruction effect of the eavesdropper. Finally, the present invention uses the feature-level peak signal-to-noise ratio indicator to verify the effectiveness of the model, providing a new solution for secure image transmission, and achieving information security while ensuring communication quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1A flowchart of a method for designing a task-safety-oriented semantic communication system according to the present invention; Figure 2 It is a framework diagram of the semantic communication model in the present invention; Figure 3 This is a model framework diagram of the sending end in Embodiment 1 of the present invention; Figure 4 is a model framework diagram of the receiving end in embodiment 1 of the present invention; Figure 5 It is a comparison diagram of FPSNR values ​​of the method of the present invention and other methods under different signal-to-noise ratio conditions of AWGN channel; Figure 6 This is a comparison chart of classification accuracy between the method of the present invention and other methods under different signal-to-noise ratio conditions of the AWGN channel. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings in the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments, and they should not be understood as limitations on the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0022] Combine the following Figure 1-Figure 6 The present invention describes a method for designing a semantic communication system oriented to task safety.

[0023] Embodiment 1: This embodiment provides a method for designing a semantic communication system for task safety. The overall process of the method is as follows: Figure 1 As shown, including: S1. Obtain an image data set, divide the data set into a training set and a test set, and perform data preprocessing on the images in the training set and the test set respectively; S2. Build a channel environment and design a semantic communication model consisting of the transmission architecture of the sender, receiver, and eavesdropper. S3,semantic communication model constructed by joint pre-training using training set; S4, loading the pre-trained semantic communication model, and designing a comprehensive loss function to fine-tune the parameters of the semantic communication model; S5. Verify the effectiveness of the trained semantic communication model through the test set and obtain a semantic communication system oriented to task safety.

[0024] Each step is described in detail below.

[0025] In step S1, a data set is first obtained, which can be directly obtained from the Internet or constructed by oneself as needed. Then the data set is divided into a training set and a test set.

[0026] Preprocess the data set, including preprocessing of the training set and preprocessing of the test set.

[0027] For the images in the training set, we first fill the four pixels with zero values ​​around the image, and then randomly crop out a 32×32 area. Then, we flip the image horizontally with a probability of 50%, keep the remaining 50% as it is, and randomly block a 16×16 square area on the image. Finally, we normalize the image.

[0028] The images in the test set are only normalized to the same mean and standard deviation as those in the training set.

[0029] Step S2 is to build the channel environment and design the transmission architecture related to the semantic communication model, such as Figure 2 As shown, it includes a sending end, a receiving end and an eavesdropping end.

[0030] In this embodiment, the sending end is as follows Figure 3 As shown in the figure, it consists of a semantic encoder, a channel encoder, and a generator network. The semantic encoder is used to extract the feature map of the input training set image, and the channel encoder reduces the dimension of the feature map output by the semantic encoder for subsequent transmission in the physical channel. Then the output feature map passes through a generator network composed of four 3×3 convolutional layers, and the generator network adds perturbations to it. After adding perturbations, the dimension of the feature map remains unchanged. The whole process is expressed as: ; in, is the semantic encoder, is the input image, are the neural network parameters of the semantic encoder: is the channel encoder, are the neural network parameters of the channel encoder; To generate the network, are the neural network parameters for generating the network.

[0031] The input signal is then transmitted through the constructed AWGN channel, and the signals received by the receiving end and the eavesdropping end are expressed as: ; ; in, and Respectively represent the signals received by the receiving end and the eavesdropping end, is the input signal of the channel, and represent the independent and identically distributed Gaussian noise at the receiving end and the eavesdropping end respectively.

[0032] The receiving end of this embodiment is as follows Figure 4 As shown in the figure, it is composed of a channel decoder, a semantic decoder and a discriminant network. The channel decoder increases the dimension of the received signal and then inputs it into the semantic decoder for image reconstruction. The reconstructed image also needs to be input into the discriminant network to prepare for adversarial training of the subsequent joint generation network. The architecture of the eavesdropping end is basically the same as that of the receiving end, but only has a channel decoder and a semantic decoder. The image reconstruction process can be expressed as: ; ; in, and Reconstruct the image for the receiving end and the eavesdropping end respectively; and They are channel decoder and semantic decoder respectively. and are the neural network parameters of the channel decoders at the receiving end and the eavesdropping end, respectively. and are the neural network parameters of the semantic decoder at the receiving end and the eavesdropping end, respectively.

[0033] Step S3 is to construct a semantic communication model using the training set joint pre-training.

[0034] Based on the semantic communication model constructed in step S2, the semantic communication model framework including the transmitter, receiver and eavesdropper is pre-trained using the pixel-level mean square error (MSE) loss function to ensure the image reconstruction quality of the receiver and the eavesdropping ability of the eavesdropper. The pixel-level mean square error loss function is expressed as: ; Among them, MSE represents mean square error calculation.

[0035] Step S4 is to load the pre-trained model and fine-tune the training model using the generative adversarial network loss function and the security semantic feature loss function.

[0036] Based on the obtained pre-trained model, training and fine-tuning are performed on different architecture parts using specialized loss functions.

[0037] For the semantic encoder and channel encoder at the sending end, and the semantic decoder and channel decoder at the receiving end and the eavesdropping end, a security semantic feature loss function composed of feature maps is used for fine-tuning. In this embodiment, the classification task is taken as an example of a downstream AI task, and the original image, the reconstructed image at the receiving end and the eavesdropping end are respectively input into the feature extraction network, which is pre-trained for the classification task in advance. The security semantic feature loss function is expressed as: ; in, For the The original image feature map, and Respectively The receiving end reconstructs the image and The eavesdropping end reconstructs the image feature map, is the task weight that controls the balance, is the amount of data in each batch.

[0038] The generative network at the sender and the discriminative network at the receiver are jointly trained using the generative adversarial network loss function. The principle is that for the perturbed feature map generated by the generative network at the sender, the goal is to jointly perform adversarial training with the discriminative network at the receiver to ensure that the generated perturbation has minimal impact on the image reconstruction task at the receiver, while maintaining the image reconstruction quality at the receiver as much as possible. However, since the eavesdropper lacks a discriminative network, the perturbation will reduce the image reconstruction quality at the eavesdropping end. The optimization goal of the generative network is to maximize the output of the discriminative network at the receiver to reconstruct the image, thereby deceiving the discriminative network; the optimization goal of the discriminative network is to make the output of the original image as large as possible and the output of the reconstructed image at the receiver as small as possible to distinguish it from the original image. Therefore, the generative adversarial network loss function is expressed as: ; ; in, is the loss function of the generated network, is the loss function of the discriminant network, represents the discriminant network, is the neural network parameter of the discriminant network; To take the mean value, is the hyperparameter of the regularization term, To limit the disturbance amplitude Norm.

[0039] The ultimate training goal is to ensure that the perturbations introduced by the generator network reduce the image reconstruction quality of the eavesdropping end while maintaining the image reconstruction quality of the receiving end as much as possible. Finally, the overall comprehensive loss function of the model is expressed as: .

[0040] Step S5, input the test set data into the network, output the results and verify the effectiveness of the model.

[0041] For the evaluation of reconstructed image quality, due to the use of a feature-level secure semantic feature loss function training model framework, the traditional pixel-level PSNR (peak signal-to-noise ratio) may not be suitable for evaluating the image reconstruction quality under this model. In view of this, a new performance indicator FPSNR is proposed from a semantic perspective, expressed as: ; in, Represents the maximum activation value of the original image feature map; It represents the mean square error based on the feature level and is expressed as , is the feature map extracted from the original image after passing through the feature extraction network. It is the feature map extracted after the reconstructed image passes through the feature extraction network. For the evaluation of classification tasks, the classification accuracy index is used, which is defined as the ratio of the number of correctly predicted samples to the total number of samples.

[0042] The effect of the task-safety-oriented semantic communication system constructed by the present invention is further illustrated below in conjunction with simulation experiments.

[0043] 1. Simulation conditions and parameter settings: The simulation experiment of the present invention is carried out based on the simulation platform of Python3.9, Pytorch1.13.0+cu117. The dataset used is the CIFAR-10 dataset, which contains 60,000 32×32 color images, divided into 10 categories, 6,000 images in each category, which are divided into 50,000 training sets (including 5 sub-batches) and 10,000 test sets. The pre-trained ResNet-18 is used as the feature extraction network for classification tasks and the classifier for downstream AI tasks. In the simulation of this embodiment, Figure 3 and Figure 4As shown in the figure, the semantic encoder consists of a 5×5 convolutional layer and three residual blocks, the semantic decoder consists of a 5×5 convolutional layer, three residual blocks and a 3×3 convolutional layer, and the discriminant network consists of 2 3×3 convolutional layers and a fully connected layer. The channel encoder and channel decoder are a 3×3 convolutional neural network with 64 neurons in each layer. Both use the ReLU activation function, and the semantic decoder and channel decoder architectures of the receiving end and the eavesdropping end are consistent. For the channel environment, the channel environment conditions of the receiving end and the eavesdropping end are set to be consistent in this embodiment.

[0044] In the pre-training stage, the system is pre-trained using the pixel-level MSE loss function, using the Adamw optimizer with a learning rate of , after the training loss value converges, load the obtained parameters and then fine-tune the network. For the security semantic feature loss function, Set to 0.5, the learning rate ; For the generative adversarial network, the learning rate of the generative network is , the learning rate of the discriminant network is , the data size of each batch is 256.

[0045] 2. Simulation content: In the simulation of the present invention, the comparison benchmark has a receiving end and an eavesdropping end. Figure 5 The relationship between the FPSNR value and signal-to-noise ratio of the receiving end and the eavesdropping end of different benchmarks in the AWGN channel environment is demonstrated. Figure 5 The horizontal axis represents different signal-to-noise ratios (dB), ranging from -6 to 9dB, with a step size of 3dB, and the vertical axis represents the FPSNR value. The broken line marked with a rectangle represents the FPSNR value curve of the receiving end using the method of the present invention, the broken line marked with a circle represents the FPSNR value curve of the eavesdropping end using the method of the present invention, and the broken lines marked with a triangle and an inverted triangle represent the FPSNR value curves of the receiving end and the eavesdropping end based on the pixel-level MSE security loss function, respectively. The pixel-based benchmark system architecture is consistent with the system architecture of this embodiment, except that a pixel-based security loss function is used, and the channel environment and parameter settings are consistent with the method proposed in the present invention. It can be seen from Figure 5It can be seen that as the signal-to-noise ratio increases, the FPSNR value of the eavesdropping end of the method of the present invention is always lower than the FPSNR value of the eavesdropping end based on the pixel level. This is because the pixel-based benchmark does not take into account the semantic-level information required for the classification task. Further, at each signal-to-noise ratio, the FPSNR value of the receiving end of the method of the present invention is always about 4dB higher than the FPSNR value of the eavesdropping end of the method of the present invention. It can also be seen that compared with the pixel-based benchmark, the FPSNR value of the receiving end of the method of the present invention is always higher than the FPSNR value of the receiving end based on the pixel level. The above results show that the method of the present invention effectively reduces the semantic information of the reconstructed image of the eavesdropping end, reduces its image reconstruction quality, and also ensures the image reconstruction quality of the receiving end. It also shows that the pixel-based security loss function benchmark is not as good as the method of the present invention.

[0046] Figure 6 The figure shows the relationship between the classification accuracy and signal-to-noise ratio of the method of the present invention and the pixel-level benchmark method in the AWGN channel environment. Figure 6 The vertical axis represents the classification accuracy, and the other settings are the same as Figure 5 Stay consistent. Figure 6 It can be seen that when the signal-to-noise ratio is 9dB, the classification accuracy of the receiving end of the method of the present invention reaches about 80%, while the eavesdropping end of the method of the present invention performs poorly, with a classification accuracy of less than 20%. And the classification accuracy of the eavesdropping end based on the pixel-level benchmark always remains at a relatively high level at each signal-to-noise ratio. This shows that the pixel-level benchmark cannot guarantee the performance security of the classification task, and as the signal-to-noise ratio decreases, the classification accuracy of the receiving end of the method of the present invention still maintains a considerable accuracy performance, while the classification accuracy of the eavesdropping end of the method of the present invention always remains at a low level, less than 20%. This proves that the method of the present invention effectively reduces the classification accuracy of the eavesdropping end and realizes the security of the semantic communication system in the classification task.

[0047] Based on the above simulation results and analysis, the task-oriented safe semantic communication system design method based on generative adversarial networks proposed in the present invention can effectively ensure the performance safety of downstream AI tasks, which enables the present invention to be better applied in actual communication scenarios.

[0048] Embodiment 2: This embodiment provides a non-transitory computer-readable storage medium on which computer instructions are stored. The computer instructions enable a computer to execute a method for designing a semantic communication system for task safety. The method comprises the following steps: S1. Obtain an image data set, divide the data set into a training set and a test set, and perform data preprocessing on the images in the training set and the test set respectively; S2. Build a channel environment and design a semantic communication model consisting of the transmission architecture of the sender, receiver, and eavesdropper. S3,semantic communication model constructed by joint pre-training using training set; S4, loading the pre-trained semantic communication model, and designing a comprehensive loss function to fine-tune the parameters of the semantic communication model; S5. Verify the effectiveness of the trained semantic communication model through the test set and obtain a semantic communication system oriented to task safety.

[0049] Embodiment 3: This embodiment provides an electronic device, which may include: a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor may call logic instructions in the memory to execute a task-safety-oriented semantic communication system design method, which includes the following steps: S1. Obtain an image data set, divide the data set into a training set and a test set, and perform data preprocessing on the images in the training set and the test set respectively; S2. Build a channel environment and design a semantic communication model consisting of the transmission architecture of the sender, receiver, and eavesdropper. S3,semantic communication model constructed by joint pre-training using training set; S4, loading the pre-trained semantic communication model, and designing a comprehensive loss function to fine-tune the parameters of the semantic communication model; S5. Verify the effectiveness of the trained semantic communication model through the test set and obtain a semantic communication system oriented to task safety.

[0050] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., which can store program code.

[0051] Embodiment 4: This embodiment provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can perform a task-safety-oriented semantic communication system design method, which includes the following steps: S1. Obtain an image data set, divide the data set into a training set and a test set, and perform data preprocessing on the images in the training set and the test set respectively; S2. Build a channel environment and design a semantic communication model consisting of the transmission architecture of the sender, receiver, and eavesdropper. S3,semantic communication model constructed by joint pre-training using training set; S4, loading the pre-trained semantic communication model, and designing a comprehensive loss function to fine-tune the parameters of the semantic communication model; S5. Verify the effectiveness of the trained semantic communication model through the test set and obtain a semantic communication system oriented to task safety.

[0052] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0053] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for designing a semantic communication system for task safety, characterized in that: include: S1. Obtain an image data set, divide the data set into a training set and a test set, and perform data preprocessing on the images in the training set and the test set respectively; S2. Build a channel environment and design a semantic communication model consisting of the transmission architecture of the sender, receiver, and eavesdropper. S3,semantic communication model constructed by joint pre-training using training set; S4, loading the pre-trained semantic communication model, and fine-tuning the parameters of the semantic communication model using the security semantic feature loss function; S5. Verify the effectiveness of the trained semantic communication model through the test set and obtain a semantic communication system oriented to task safety.

2. The method for designing a task-safety-oriented semantic communication system according to claim 1, characterized in that: In step S1, data preprocessing is performed on the images in the training set, including: Fill the images in the training set with zero values ​​for a certain number of pixels, and then randomly crop out fixed-size areas; The cropped image is horizontally flipped with a preset probability, and the rest of the image remains the same; Randomly block a fixed-size area on the image; Normalize the images in the training set.

3. The method for designing a task-safety-oriented semantic communication system according to claim 2, characterized in that: In step S1, data preprocessing is performed on the images in the test set, including: The images in the test set are normalized using the same mean and standard deviation as the images in the training set.

4. The method for designing a task-safety-oriented semantic communication system according to claim 1, characterized in that: In step S2, the semantic communication model consisting of the transmission architecture of the sender, the receiver and the eavesdropper is constructed as follows: After the image at the sender is semantically encoded and channel encoded, the feature map is output. The feature map is perturbed by the generative network. After the perturbation is added, the dimension of the feature map remains unchanged, which is expressed as: ； in, is the semantic encoder, is the input image, are the neural network parameters of the semantic encoder: is the channel encoder, are the neural network parameters of the channel encoder; To generate the network, Neural network parameters for generating the network; The input signal is then transmitted through the constructed AWGN channel, and the signals received by the receiving end and the eavesdropping end are expressed as: ； ； in, and Respectively represent the signals received by the receiving end and the eavesdropping end, is the input signal of the channel, and denote the independent and identically distributed Gaussian noises at the receiving end and the eavesdropping end respectively; The signals received by the receiving end and the eavesdropping end are reconstructed into images through the channel decoder and the semantic decoder. The image reconstruction process is expressed as: ； ； in, and Reconstruct the image for the receiving end and the eavesdropping end respectively; and They are channel decoder and semantic decoder respectively. and are the neural network parameters of the channel decoders at the receiving end and the eavesdropping end, respectively. and are the neural network parameters of the semantic decoder at the receiving end and the eavesdropping end, respectively.

5. The method for designing a task-safety-oriented semantic communication system according to claim 4, characterized in that: Step S3 uses the training set to jointly pre-train the constructed semantic communication model, specifically: use the pixel-level mean square error loss function to pre-train the semantic communication model composed of the transmission architecture of the transmitter, receiver and eavesdropper. The pixel-level mean square error loss function is expressed as: ； Among them, MSE represents mean square error calculation.

6. The method for designing a task-safety-oriented semantic communication system according to claim 4, characterized in that: Step S4 loads the pre-trained semantic communication model and uses the security semantic feature loss function to fine-tune the parameters of the semantic communication model, specifically including: The semantic encoder and channel encoder at the sender, as well as the semantic decoder and channel decoder at the receiver and eavesdropper are fine-tuned using a secure semantic feature loss function; When the classification task is used as the downstream AI task, the original image, the receiving end, and the eavesdropping end reconstructed image are respectively input into the feature extraction network. The feature extraction network has been pre-trained for the classification task. The security semantic feature loss function is expressed as: ； in, For the The original image feature map, and Respectively The receiving end reconstructs the image and The eavesdropping end reconstructs the image feature map, is the task weight that controls the balance, is the amount of data in each batch.

7. The method for designing a task-safety-oriented semantic communication system according to claim 4, characterized in that: The receiving end of the semantic communication model has a discriminant network corresponding to the generative network of the sending end. In step S2, after the signals received at the receiving end and the eavesdropping end are reconstructed through the channel decoder and the semantic decoder, the reconstructed image is input into the discriminant network of the receiving end, and adversarial training is performed with the generative network.

8. The method for designing a task-safety-oriented semantic communication system according to claim 7, characterized in that: When fine-tuning the parameters of the semantic communication model in step S4, it also includes: The generative network at the sender and the discriminative network at the receiver are jointly trained using the generative adversarial network loss function. The loss function of the generative adversarial network is expressed as: ； ； in, is the loss function of the generated network, is the loss function of the discriminant network, represents the discriminant network, is the neural network parameter of the discriminant network, To take the mean value, is the hyperparameter of the regularization term, To limit the disturbance amplitude Norm.

9. The method for designing a task-safety-oriented semantic communication system according to claim 1, characterized in that: Step S5 verifies the effectiveness of the trained semantic communication model through a test set, including: From a semantic perspective, FPSNR is used as a performance evaluation indicator, expressed as: ； in, Represents the maximum activation value of the original image feature map; It represents the mean square error based on the feature level and is expressed as , is the feature map extracted from the original image after passing through the feature extraction network. It is the feature map extracted after the reconstructed image passes through the feature extraction network.

10. An electronic device, characterized in that: include: A processor, a communication interface, a memory and a communication bus, wherein the processor, the communication interface and the memory communicate with each other through the communication bus, and the processor calls the logic instructions in the memory to execute the task-safety-oriented semantic communication system design method described in any one of claims 1-9.

Images