Semantic anti-interference communication method

Abstract

The invention discloses a semantic anti-interference communication method, which designs a simulation algorithm applied to a wireless image transmission scene, improves the reliability and generalization ability of semantic communication in a low signal-to-noise ratio and strong interference scene by constructing a cooperative anti-interference mechanism of a signal domain and a semantic domain, and avoids extra channel resource consumption. The method is suitable for communication scenes of communication terminals such as unmanned aerial vehicles in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and specifically relates to a semantic anti-interference communication method. Background Technology

[0002] With the rise of the sixth-generation mobile communication system (6G), the continuous emergence of smart terminals has driven profound changes in communication needs and business scenarios. Various highly dynamic and highly interfered diverse scenarios (such as autonomous driving of vehicles and drone inspection of complex terrain) have posed new challenges to communication systems. They not only need to achieve accurate data transmission, but also need to ensure the core goal of "meaningful" information delivery. In particular, how to effectively transmit semantic information in a complex and interfering environment has become crucial.

[0003] In recent years, unmanned aerial vehicles (UAVs), as typical air-space-ground integrated intelligent terminals, have been widely used in critical missions such as disaster emergency response, urban inspection, logistics transportation, environmental monitoring, and military reconnaissance. In such mission scenarios, UAVs typically face low-altitude, dynamic, and highly interference-prone communication environments with complex and variable channel conditions, which can easily lead to distortion or even interruption of semantic information such as images and videos during transmission, seriously affecting the reliability and security of mission execution.

[0004] While traditional communication methods possess bit-level error correction capabilities, they struggle to guarantee the accurate transmission of critical semantic content. The rise of semantic communication technology offers a new solution, using deep networks to semantically compress and restore multimedia data such as images and speech. By training the model, it enhances the expressive efficiency and anti-interference performance of communication. However, existing semantic communication systems still suffer from insufficient robustness in the complex real-world environments of drones, especially at low signal-to-noise ratios, where the semantic features extracted by deep models are easily corrupted by interference, leading to inaccurate reconstruction of key information.

[0005] Therefore, improving the adaptability of semantic communication systems in highly interfering application scenarios such as drones has become an important research direction. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a semantic anti-interference communication method and designs a simulation algorithm for wireless image transmission scenarios. By constructing a collaborative anti-interference mechanism between the signal domain and the semantic domain, the reliability and generalization ability of semantic communication are improved in low signal-to-noise ratio and strong interference scenarios, while avoiding additional channel resource consumption. This method is suitable for communication scenarios of communication terminals such as UAVs in complex environments.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows:

[0008] Step 1: Construct a basic model of a cross-domain anti-interference communication system based on Wyner-Ziv theory;

[0009] A communication structure including a channel encoder, a channel decoder, and a semantic autoencoder is established. By cascading the channel decoder and the semantic autoencoder, a Turbo structure is constructed in the signal domain and the semantic domain, enabling the two to form an information processing interaction loop. Among them, the channel decoder realizes error correction of digital signals, the semantic autoencoder performs noise reduction processing on deep semantics, and through multiple rounds of iterative exchange of cross-domain information, the collaborative elimination of complex interference is achieved.

[0010] Step 2: Design a semantic autoencoder based on the Swing Transformer;

[0011] Step 2-1: Select the network structure skeleton;

[0012] The Swing Transformer is used as the basic backbone network of the semantic autoencoder.

[0013] Step 2-2: Design and Coding Phase;

[0014] At the encoding end, the input image data is divided into non-overlapping small patches and transformed into a token sequence through an embedding layer. This sequence is then input into the multi-stage Swing Transformer module to extract semantic feature vectors, compress them, and retain key semantic information.

[0015] Steps 2-3: Design the decoding stage:

[0016] The decoding end structure is symmetrical to the encoding end, and the image is gradually recovered through upsampling and the Swing Transformer module;

[0017] Step 3: Construct an iterative anti-interference mechanism for cross-domain information exchange;

[0018] The cascaded semantic autoencoder and the channel decoder form a Turbo loop, which gradually eliminates interference through multiple rounds of cross-domain information exchange. In each iteration, the channel decoder first processes the received signal, uses the error correction capability provided by the traditional channel code to complete the initial denoising, and then inputs the intermediate results into the semantic autoencoder for semantic-level denoising. The output generates side information and feeds it back to the channel decoder to assist the next round of decoding.

[0019] Preferably, step 1 specifically includes:

[0020] Step 1-1: When the receiving end possesses non-causal side information Y and is given a distortion requirement D, the minimum transmission rate R of the Wyner-Ziv theoretical system satisfies the corresponding rate-distortion function relationship:

[0021]

[0022] In the formula, U represents the compressed information of the source sequence X, and p(u|x) represents the encoder compression process. Let represent the reconstruction function of the decoder, d(·) represent the distortion metric function, I(·) represent the Shannon mutual information; E[·] represent the mathematical expectation of the distortion between the source sequence and the reconstructed sequence, u represent the value of the compressed information, x represent the value of the source sequence, and y represent the known noncausal side information observations at the receiver. This represents the reconstructed sequence obtained at the receiving end;

[0023] Steps 1-2: Source sequence X n First, compressed information U is generated by the encoder and transmitted at a rate R. At the decoding end, when reconstructing the source sequence, the non-causal auxiliary information sequence Y is... n Provide effective supplementary information.

[0024] Preferably, step 2 specifically comprises:

[0025] Step 2-1: Construct a novel semantic autoencoder structure using the Swing Transformer block as the core backbone network;

[0026] In each encoding / decoding stage i, the feature sequence undergoes k pairs of feature extraction processes using Swin Transformer Blocks; the kth... i +1 block input feature sequence From the kth i-1 Output of Block The result obtained through the W-MSA module is represented as follows:

[0027]

[0028] Where LN(·) represents layer normalization operation; MLP(·) represents multilayer perceptron module;

[0029] Subsequently intermediate variables After further processing by the SW-MSA module and other layers, the k-th... i The final output features of the Block The expression is as follows:

[0030]

[0031] Step 2-2: Design the coding stage for wireless image transmission tasks;

[0032] For an RGB image input to a semantic autoencoder First, it was divided into several Non-overlapping image patches, each image patch is treated as a token, and these tokens are arranged in order from top left to bottom right to obtain a token sequence. After linear embedding, N1 SwinTransformer blocks are applied to these l1 tokens; these l1 tokens are then fed into multiple stages of processing, each stage undergoing a halving of the height and width and an increase in dimension, where each stage consists of a downsampling layer and a subsequent set of N blocks. i The Swing Transformer block is composed of;

[0033] Steps 2-3: Design a decoding stage symmetrical to the encoder, reconstructing the image through upsampling layers and Swing Transformer blocks;

[0034] Steps 2-4: Train the semantic autoencoder using the dataset.

[0035] Preferably, step 3 specifically comprises:

[0036] Step 3-1: Transform the semantic autoencoder Output connection back to channel decoder This constructs a Turbo loop; channel decoder Handling prior log-likelihood ratio (LLR) a Output posterior log-likelihood ratio LLR p , represented as:

[0037]

[0038] Step 3-2: The semantic autoencoder bases the input LLR... p After semantic denoising, then LLR a The output is given to the channel decoder as prior information to optimize the next round of decoding;

[0039]

[0040] The gain can be expressed by the following formula:

[0041]

[0042] In the formula, This refers to information recovered through cross-domain information exchange using a semantic autoencoder. This indicates information that is recovered independently by the channel decoder only; and All are calculated from the mutual information between the transmitted information and the recovered information.

[0043] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to enable the electronic device to perform the above-described UAV semantic anti-jamming communication method.

[0044] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described semantic anti-jamming communication method for unmanned aerial vehicles.

[0045] A chip includes a processor for calling and running a computer program from a memory, causing a device equipped with the chip to perform the aforementioned UAV semantic anti-jamming communication method.

[0046] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the above-described UAV semantic anti-jamming communication method.

[0047] The beneficial effects of this invention are as follows:

[0048] This invention provides a semantic anti-interference communication method based on cross-domain information exchange. By combining the core ideas of deep learning and the Wyner-Ziv theorem, a novel decoding end structure is designed. Syntax communication anti-interference technology and semantic communication denoising method are integrated to construct an iterative decoding method that alternately removes noise in the semantic domain and the signal domain. Without adding extra channel resources, the robustness of the communication system and the signal recovery quality are improved. Attached Figure Description

[0049] Figure 1 This is a system model structure diagram of the Wyner-Ziv theory.

[0050] Figure 2 It is a pair of consecutive Swing Transformer block structures.

[0051] Figure 3 It is a semantic autoencoder network structure based on the Swing Transformer.

[0052] Figure 4 It is a semantic anti-interference system model that includes a Turbo loop structure.

[0053] Figure 5 It is a semantic anti-interference communication system architecture oriented towards image transmission.

[0054] Figure 6 It is a semantic autoencoder network structure based on CNN.

[0055] Figure 7 The comparison shows the BER, ED, and PSNR indices of different systems under different SNR channel conditions. (a): Comparison of BER indices under different SNR channel conditions; (b): Comparison of ED indices under different SNR channel conditions; (c): Comparison of PSNR indices under different SNR channel conditions.

[0056] Figure 8 The comparison of BER iterative gain for different systems under different channel SNR conditions is as follows: (a) SNR = -5dB, (b) SNR = -2dB, (c) SNR = 0dB, (d) SNR = 5dB.

[0057] Figure 9 The comparison of ED iterative gain for different systems under different channel SNR conditions is as follows: (a) SNR = -5dB, (b) SNR = -2dB, (c) SNR = 0dB, (d) SNR = 5dB.

[0058] Figure 10 The comparison of PSNR iterative gain of different systems under different channel SNR conditions is as follows: (a) SNR = -5dB, (b) SNR = -2dB, (c) SNR = 0dB, (d) SNR = 5dB.

[0059] Figure 11 These are the reconstruction comparison results of the original image by different systems under extreme channel conditions: (a): original image, (b): reconstruction comparison under SNR=-5dB condition, (c): reconstruction comparison under SNR=5dB condition. Detailed Implementation

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] This invention provides a semantic anti-interference communication method and designs a simulation algorithm for wireless image transmission scenarios. By constructing a collaborative anti-interference mechanism between the signal domain and the semantic domain, the reliability and generalization ability of semantic communication are improved in low signal-to-noise ratio and strong interference scenarios, while avoiding additional channel resource consumption. It is suitable for communication scenarios of communication terminals such as UAVs in complex environments.

[0062] Step 1: Construct a basic model for a cross-domain anti-interference communication system based on the Wyner-Ziv theory. This theory states that, given non-causal side information at the receiver, the transmission rate at the encoder can be reduced while satisfying certain distortion constraints. This invention adopts an auxiliary side information collaborative reconstruction approach, establishing a general communication structure including a channel encoder, a channel decoder, and a semantic autoencoder. In the system model, by cascading the channel decoder and the semantic autoencoder, a Turbo structure is constructed between the signal and semantic domains, forming an information processing interaction loop. The channel decoder primarily corrects digital signal errors, while the semantic autoencoder focuses on denoising deep semantics. Through multiple rounds of iterative exchange of cross-domain information, collaborative elimination of complex interference is achieved.

[0063] Step 2: A semantic autoencoder based on Swin Transformer was designed, which has stronger global modeling and semantic extraction performance compared with traditional CNN networks.

[0064] Step 2-1: Select the network structure skeleton:

[0065] The Swing Transformer was chosen as the backbone network for the semantic autoencoder. This structure employs a sliding window mechanism and a multi-scale hierarchical modeling strategy, which can effectively capture cross-regional and long-distance semantic correlations in multimodal data such as images. It overcomes the limitations of traditional CNNs in modeling global semantics and provides reliable assurance for signal recovery in complex interference environments.

[0066] Step 2-2: Design and Coding Phase

[0067] At the encoding end, the input image data is divided into non-overlapping patches and transformed into a "token" sequence through an embedding layer. This sequence is then input into a multi-stage Swing Transformer module to extract semantic feature vectors, compress and retain key semantic information, ensuring a deep expression of the structural and semantic elements in the signal sequence. This allows for signal reconstruction with sufficient semantic prior information in subsequent interference environments.

[0068] Steps 2-3: Design the decoding stage:

[0069] The decoding end has a symmetrical structure to the encoding end, and the image is gradually recovered through upsampling and the Swing Transformer module. The decoding process can reconstruct high-quality image semantics in interference environments, maintaining the consistency of image contours, structure, and key content, and providing stable auxiliary information (side information) for subsequent channel iterative decoding.

[0070] Step 3: Construct an iterative anti-interference mechanism for cross-domain information exchange. A cascaded semantic autoencoder and channel decoder form a Turbo loop, gradually eliminating interference through multiple rounds of cross-domain information exchange. In each iteration, the channel decoder first processes the received signal, utilizing the error correction capability provided by the traditional channel code to complete preliminary denoising. The intermediate results are then input into the semantic autoencoder for semantic-level denoising, and the output generates side information which is fed back to the channel decoder to assist in the next round of decoding.

[0071] Step 4: A simulation system suitable for UAV image transmission was designed to verify the effectiveness and potential of the proposed method. The system considers different signal-to-noise ratio conditions under AWGN channels, uses the CIFAR-10 image dataset for training and testing of the semantic autoencoder, introduces LDPC channel coding and reconstruction procedures, and evaluates the system's image restoration quality, bit error rate, and other metrics through multiple rounds of Turbo decoding iterations.

[0072] Example:

[0073] This invention proposes a semantic anti-interference communication method based on cross-domain information exchange, and designs a simulation algorithm based on this method for wireless image transmission scenarios, achieving better anti-interference performance than traditional systems and improving signal reception quality.

[0074] Step S1: Perform relational modeling based on the Wyner-Ziv theorem, the theoretical foundation of the proposed method.

[0075] Step S101: When the receiving end possesses non-causal side information Y and is given a distortion requirement D, the minimum transmission rate R of the Wyner-Ziv theoretical system should satisfy the corresponding rate-distortion function relationship:

[0076]

[0077] In the formula, U represents the compressed information of the source sequence X, and p(u|x) represents the encoder compression process. Let d(·) represent the reconstruction function of the decoder, d(·) represent the distortion metric function, and I(·) represent ordinary Shannon mutual information.

[0078] Step S102: As Figure 1 As shown, the source sequence X n First, the codeword is compressed into a codeword U by the encoder and transmitted under a rate constraint R. At the decoding end, when reconstructing the source sequence, the non-causal auxiliary information sequence Y... n It can provide effective supplementary information.

[0079] Step S2: A novel semantic autoencoder is constructed using the Swin Transformer, which has stronger semantic global modeling and feature extraction capabilities than CNN, as the core backbone network, serving as an important component of the semantic anti-interference system.

[0080] Step S201: Use as follows Figure 2 The Swin Transformer block shown serves as the core backbone network for constructing a novel semantic autoencoder structure. As a sequence-to-sequence network structure, the Swin Transformer block is built by replacing the multi-head self-attention module in the standard Transformer block with window-based module (W-MSA) and shifted window module (SW-MSA). The window partitioning translates adjacent self-attention layers, enabling connections between these layers and significantly enhancing the model's feature modeling capabilities.

[0081] In each encoding / decoding stage i, the feature sequence undergoes k pairs of feature extraction processes using Swin Transformer Blocks; the kth... i +1 block input feature sequence From the kth i-1 Output of Block The result obtained through the W-MSA module is represented as follows:

[0082]

[0083] Where LN(·) represents layer normalization operation; MLP(·) represents multilayer perceptron module;

[0084] Subsequently intermediate variables After further processing by the SW-MSA module and other layers, the k-th... i The final output features of the Block The expression is as follows:

[0085]

[0086] Step S202: Design the coding stage for wireless image transmission tasks.

[0087] For an RGB image input to a semantic autoencoder First, they will be divided into... Non-overlapping image patches, each of which can be considered a "token", are arranged in order from top left to bottom right to form a token sequence. After linear embedding, N1 Swing Transformer blocks are applied to these l1 "tokens". These N1 Swing Transformer blocks and the image patch embedding layer are generally referred to as the first stage. In actual implementation, the image patch partitioning and linear embedding operations are implemented together using a single convolutional layer. Next, these tokens are fed into multiple stages for processing, undergoing a halving of height and width and an increase in dimension at each stage. Each stage consists of a downsampling layer and a subsequent set of N... i It consists of a SwingTransformer block.

[0088] Step S203: Design a decoding stage symmetrical to the encoder, reconstructing the image through upsampling layers and Swing Transformer blocks.

[0089] Step S204: Train the semantic autoencoder using a suitable dataset;

[0090] The dataset used is the CIFAR-10 dataset, which contains 50,000 training images and 10,000 test images. During the data preprocessing stage, the dimensions of each image were adjusted to 3×96×96, where 3 represents the three color channels of the image (RGB), and 96×96 represents that the height and width of the image are both 96 pixels. Normalization and standardization were also performed to make the model more stable in training and convergence.

[0091] like Figure 3 As shown, the encoding / decoding stage of the semantic autoencoder is set to 2, the window size of the Swin Transformer block is set to 2, and the corresponding parameter configuration is [N1,N2]=[2,4],[C1,C2]=[128,256], where C1 and C2 represent the dimensions of the output features after the first and second stages, respectively. The optimizer chosen for training is Adam, and the learning rate is set to 1×10⁻⁶. -4 The batch size of the dataset was set to 64, and the number of training rounds was 100. The model was implemented using the PyTorch framework, and the training process was carried out under channel conditions where the signal-to-noise ratio (SNR) followed a uniform distribution from 0dB to 5dB. Computational acceleration was achieved on an NVIDIA GeForce RTX 2080Ti.

[0092] Step S3: Based on the Wyner-Ziv theory, the channel decoder and semantic autoencoder are cascaded to form a Turbo loop, and a cross-domain alternating interference cancellation mechanism between the signal domain and the semantic domain is constructed.

[0093] Step S301: As Figure 4 As shown, the semantic autoencoder Output connection back to channel decoder This constructs a Turbo loop. Channel decoder Handling prior log-likelihood ratio (LLR) a Output posterior log-likelihood ratio LLR p , represented as:

[0094]

[0095] The semantic autoencoder is based on the input LLR p After semantic denoising, then LLR a The output is given to the channel decoder as prior information to optimize the next round of decoding.

[0096]

[0097] Step S302: During channel decoding, interference introduced during transmission can be initially eliminated in the signal domain using the error correction capability of digital signal channel decoding. In subsequent processes, the semantic autoencoder utilizes its ability to model image semantic features and, through semantic correlation, further eliminates errors that channel decoding cannot correct at the semantic level. This semantic anti-interference benefit, achieved through information exchange across the physical signal domain and semantic domain, allows the reconstructed signal quality and decoding gain to gradually improve in each iteration until convergence is achieved or application requirements are met.

[0098] The performance improvement achieved through this decoding-end optimization is achieved without introducing additional channel bandwidth resources, and the gain can be expressed by the following formula:

[0099]

[0100] In the formula, This refers to information recovered through cross-domain information exchange using a semantic autoencoder. This indicates information that is recovered independently by the channel decoder. and Both can be calculated from the mutual information between the transmitted information and the recovered information.

[0101] Step S4: Based on the method of the present invention, design and implement a simulation system for UAV wireless image transmission scenarios.

[0102] Step S401: As Figure 5 As shown, a simulation system for wireless image transmission of UAVs is implemented. LDPC code is selected as the channel code, and the decoding rule is based on a soft decoding algorithm with log-likelihood ratio (LLR).

[0103] The raw RGB image is first processed by a general encoder before transmission. It is quantized into a binary sequence before being input to the channel decoder to accommodate LDPC code processing. The channel-decoded sequence is dequantized back into pixel data and input to a semantic autoencoder, whose subsequent output is converted back into a binary stream. Finally, this is fed back to the channel decoder to assist in the next iterative decoding process, until the maximum number of iterations is reached.

[0104] In the process, as an important component of realizing the essence of the Wyner-Ziv theorem, the trained semantic autoencoder learns semantic prior knowledge and sends the output containing side information back to the channel encoder to help it make more accurate bit decisions in the next iteration of decoding, thereby gradually correcting transmission errors.

[0105] simulation:

[0106] The simulated channel incorporates additive white Gaussian noise (AWGN) at different signal-to-noise ratios (SNR). The LDPC channel code parameters are set as follows: codeword length n = 900, and the number of check equations involved in each bit is d. v =2, and the number of bits contained in each check equation is d. c =3, and the maximum number of iterations is set to 7.

[0107] Two systems were set up as control groups: one without SemantIC (without SemantIC) and the other using the method but employing a traditional CNN autoencoder scheme (with SemantIC). Figure 6 As shown, this is the semantic autoencoder structure of the CNN autoencoder scheme, consisting of {n F ,(w K ,h K ),s} specifies the parameters of the convolutional and deconvolutional layers, where n F Indicates the number of filters, (w K ,h K The kernel size is represented by , and the step size is s. Through calculation, an image with an original size of 3×96×96 becomes a feature map of size 16×23×23 after passing through the semantic encoder. The system described in this invention, which utilizes the Swin Transformer to construct a semantic autoencoder and implement semantic anti-interference, serves as the experimental group (With Swin-SemantIC). Bit error rate (BER), image pixel difference (ED), and peak signal-to-noise ratio (PSNR) are used as evaluation metrics. The same interference conditions are applied to the three systems mentioned above, and their performance is compared and evaluated.

[0108] First, the anti-interference performance of the three systems was compared. For example... Figure 7(a)–7(c) illustrate the relationship between the performance metrics and SNR of the three systems when transmitting images over an AWGN channel. It is clear that the Swin-SemantIC and SemantIC systems outperform the system without semantic anti-interference methods in all metrics. Swin-SemantIC further improves performance because, thanks to the stronger semantic modeling capabilities of the Swin Transformer, its autoencoder can more effectively eliminate interference in cross-domain Turbo loops, providing richer and more accurate side information for channel decoding, thus enhancing overall decoding performance.

[0109] Figures 8 to 10 This paper compares the iterative decoding gains of three systems under different signal-to-noise ratios (SNR). The results show that introducing a semantic anti-interference mechanism significantly improves the system's anti-interference capability in low SNR environments, effectively mitigating the impact of channel noise on information transmission quality. Furthermore, by improving the semantic autoencoder structure (e.g., using a SwingTransformer), the model's ability to model and express semantic features is further enhanced, enabling more effective elimination of semantic-level interference. This improves the overall performance of the interference cancellation stage in the Turbo loop, especially at extremely low SNR (e.g., -5dB), where the PSNR gain is almost twice that of the unimproved model. This result not only further verifies the effectiveness of the semantic anti-interference mechanism but also demonstrates its application potential under complex channel conditions. However, it is important to note that as the model's capabilities increase, system complexity and computational latency also increase accordingly, particularly in inference time and resource overhead in each iteration. Therefore, in practical deployments, the relationship between performance improvement and system overhead should be weighed according to the specific needs of the communication scenario.

[0110] Figure 11 The reconstruction results of sample images by three systems under two relatively extreme channel conditions (SNR = -5dB and SNR = 5dB) are presented. The results subjectively verify the significant role of the semantic anti-interference method in improving the quality of the received signal. In low SNR channel environments, the SemantIC system recovers image contours better than traditional methods, indicating its enhanced anti-interference performance under adverse channel conditions. Furthermore, the Swin-SemantIC system with optimized autoencoder further improves the visual quality of the reconstructed images. Figure 11 (b) It can be clearly observed that Swin-SemantIC not only significantly restores the vehicle body contour, but also greatly eliminates the thermal pixel noise within the yellow circle in the SemantIC system, verifying its stronger interference cancellation capability.

[0111] Under medium channel quality (SNR = 5dB), such as Figure 11As shown in (c), in the image reconstructed by the system without SemantIC, the hotspot pixels marked with yellow arrows are significantly corrected by the SemantIC method and no longer appear abrupt. Furthermore, Swin-SemantIC not only achieves this effect but also successfully eliminates the pixel noise marked with blue arrows, completing an optimization operation that the SemantIC system cannot perform. These phenomena fully demonstrate that semantic anti-interference methods based on cross-domain information exchange have significant advantages in improving the quality of received signals. By optimizing the semantic autoencoder structure, the potential of this method can be further explored, providing a promising solution for signal transmission in complex communication environments.

Claims

1. A semantic anti-interference communication method, characterized in that, Includes the following steps: Step 1: Construct a basic model of a cross-domain anti-interference communication system based on Wyner-Ziv theory; A communication structure including a channel encoder, a channel decoder, and a semantic autoencoder is established. By cascading the channel decoder and the semantic autoencoder, a Turbo structure of the signal domain and the semantic domain is constructed, so that the two form an information processing interaction loop. Among them, the channel decoder realizes error correction of digital signal bit, the semantic autoencoder performs noise reduction processing on deep semantics, and achieves collaborative elimination of complex interference through multi-round iterative exchange of cross-domain information. Step 2: Design a semantic autoencoder based on the Swing Transformer; Step 2-1: Select the network structure skeleton; The Swing Transformer is used as the basic backbone network of the semantic autoencoder. Step 2-2: Design and Coding Phase; At the encoding end, the input image data is divided into non-overlapping small patches and transformed into a token sequence through an embedding layer. This sequence is then input into the multi-stage Swing Transformer module to extract semantic feature vectors, compress them, and retain key semantic information. Steps 2-3: Design the decoding stage: The decoding end structure is symmetrical to the encoding end, and the image is gradually recovered through upsampling and the Swing Transformer module; Step 3: Construct an iterative anti-interference mechanism for cross-domain information exchange; The cascaded semantic autoencoder and the channel decoder form a Turbo loop, which gradually eliminates interference through multiple rounds of cross-domain information exchange. In each iteration, the channel decoder first processes the received signal, uses the error correction capability provided by the traditional channel code to complete the initial denoising, and then inputs the intermediate results into the semantic autoencoder for semantic-level denoising. The output generates side information and feeds it back to the channel decoder to assist the next round of decoding.

2. The semantic anti-interference communication method according to claim 1, characterized in that, Step 1 specifically includes: Step 1-1: When the receiving end possesses non-causal side information Y and is given a distortion requirement D, the minimum transmission rate R of the Wyner-Ziv theoretical system satisfies the corresponding rate-distortion function relationship: In the formula, U represents the compressed information of the source sequence X, and p(u|x) represents the encoder compression process. Let represent the reconstruction function of the decoder, d(·) represent the distortion metric function, I(·) represent the Shannon mutual information; E[·] represent the mathematical expectation of the distortion between the source sequence and the reconstructed sequence, u represent the value of the compressed information, x represent the value of the source sequence, and y represent the known noncausal side information observations at the receiver. This represents the reconstructed sequence obtained at the receiving end; Steps 1-2: Source sequence X n First, the encoder processes the compressed information U, which is then transmitted at a rate R. At the decoding end, when reconstructing the source sequence, the non-causal auxiliary information sequence Y is... n Provide effective supplementary information.

3. The semantic anti-interference communication method according to claim 2, characterized in that, Step 2 specifically involves: Step 2-1: Construct a novel semantic autoencoder structure using the Swing Transformer block as the core backbone network; In each encoding / decoding stage i, the feature sequence undergoes k pairs of feature extraction processes using Swin Transformer Blocks; the kth... i +1 block input feature sequence From the kth i-1 Output of Block The result obtained through the W-MSA module is represented as follows: Where LN(·) represents the layer normalization operation; MLP(·) represents the multilayer perceptron module; Subsequently intermediate variables After further processing by the SW-MSA module and other layers, the k-th... i The final output features of the Block The expression is as follows: Step 2-2: Design the coding stage for wireless image transmission tasks; For an RGB image input to a semantic autoencoder First, it was divided into several Non-overlapping image patches, each image patch is treated as a token, and these tokens are arranged in order from top left to bottom right to obtain a token sequence. After linear embedding, N1 SwinTransformer blocks are applied to these l1 tokens; these l1 tokens are then fed into multiple stages of processing, each stage undergoing a halving of the height and width and an increase in dimension, where each stage consists of a downsampling layer and a subsequent set of N blocks. i The Swing Transformer block is composed of; Steps 2-3: Design a decoding stage symmetrical to the encoder, reconstructing the image through upsampling layers and Swing Transformer blocks; Steps 2-4: Train the semantic autoencoder using the dataset.

4. The semantic anti-interference communication method according to claim 3, characterized in that, Step 3 specifically involves: Step 3-1: Transform the semantic autoencoder Output connection back to channel decoder This constructs a Turbo loop; channel decoder Handling prior log-likelihood ratio (LLR) a Output posterior log-likelihood ratio LLR p , is represented as: Step 3-2: The semantic autoencoder bases the input LLR... p After semantic denoising, then LLR a The output is given to the channel decoder as prior information to optimize the next round of decoding; The gain can be expressed by the following formula: In the formula, This refers to information recovered through cross-domain information exchange using a semantic autoencoder. This indicates information that is recovered independently by the channel decoder only; and All are calculated from the mutual information between the transmitted information and the recovered information.

5. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4.

7. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 4.

8. A computer program product, characterized in that, The computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the method as described in any one of claims 1 to 4.

Images