Semantic communication method, transmitter, receiver, semantic communication system

Abstract

This disclosure provides a semantic communication method, a transmitter, a receiver, and a semantic communication system, relating to the field of semantic communication. The specific implementation scheme is as follows: Under the guidance of an intelligent agent, semantic extraction is performed on the image source to obtain semantic features; under the guidance of the intelligent agent, channel environment information is sensed, and based on the sensed channel environment information, the semantic features are jointly encoded using source and channel methods to generate channel symbols; under the guidance of the intelligent agent, security risks and system performance constraints are sensed, and based on the sensed security risks and system performance constraints, the channel symbols are encrypted to generate encrypted channel symbols and send them into the wireless channel.

Description

Technical Field

[0001] This disclosure relates to the field of semantic communication technology. Specifically, this disclosure relates to a semantic communication method applied to a transmitting end of a semantic communication system, a semantic communication method applied to a receiving end of a semantic communication system, a semantic communication system transmitting end, a semantic communication system receiving end, a semantic communication system, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Semantic communication is a new communication paradigm that focuses on the meaning of information rather than the precise transmission of bits. It aims to convey the "meaning" of information in an efficient and reliable manner, rather than pursuing precise bit-level copying as in traditional communication.

[0003] By extracting and transmitting core semantics and discarding redundant details, semantic communication reduces bandwidth requirements, improves reliability in noisy environments, and provides a new direction for future communication development. Summary of the Invention

[0004] This disclosure provides a semantic communication method applied to the sending end of a semantic communication system, a semantic communication method applied to the receiving end of a semantic communication system, a semantic communication system sending end, a semantic communication system receiving end, a semantic communication system, an electronic device, and a computer-readable storage medium.

[0005] According to a first aspect of this disclosure, a semantic communication method based on an intelligent agent is provided, applied to the sending end of a semantic communication system, the semantic communication system including a sending end, a receiving end, and a communication channel connecting the sending end and the receiving end, the method comprising: Under the guidance of intelligent agents, semantic features are extracted from image sources. Under the guidance of the intelligent agent, channel environment information is perceived, and based on the perceived channel environment information, the semantic features are jointly encoded by the source and channel to generate channel symbols; Under the guidance of the intelligent agent, security risks and system performance constraints are perceived. Based on the perceived security risks and system performance constraints, the channel symbols are encrypted, and the encrypted channel symbols are sent into the wireless channel.

[0006] According to a second aspect of this disclosure, a semantic communication method based on an intelligent agent is provided, applied to a receiving end of a semantic communication system, the semantic communication system including a sending end, a receiving end, and a communication channel connecting the sending end and the receiving end, the method comprising: Under the guidance of the intelligent agent, security risks and system performance constraints are perceived. Based on the perceived security risks and system performance constraints, the encrypted channel symbols received from the wireless channel are decrypted to recover the channel symbols. Under the guidance of the intelligent agent, the channel environment information is perceived, and based on the perceived channel environment information, the decrypted channel symbols are jointly decoded from the source and channel to obtain semantic features. Under the guidance of the intelligent agent, semantic features are used to restore the semantic image and reconstruct it.

[0007] According to a third aspect of this disclosure, a transmitter of a semantic communication system is provided, the semantic communication system including a transmitter, a receiver, and a communication channel connecting the transmitter and the receiver, the transmitter including: The semantic extraction module is used to extract semantic features from image sources under the guidance of an intelligent agent. The channel coding module is used to sense channel environment information under the guidance of the intelligent agent, and to perform source-channel joint coding on the semantic features based on the sensed channel environment information to generate channel symbols. The encrypted transmission module is used to perceive security risks and system performance constraints under the guidance of the intelligent agent, encrypt the channel symbols according to the perceived security risks and system performance constraints, generate encrypted channel symbols, and send them into the wireless channel.

[0008] According to a fourth aspect of this disclosure, a receiver for a semantic communication system is provided, the semantic communication system including a transmitter, a receiver, and a communication channel connecting the transmitter and the receiver, the receiver comprising: The decryption transmission module is used to perceive security risks and system performance constraints under the guidance of the intelligent agent, and decrypt the encrypted channel symbols received from the wireless channel to recover the channel symbols based on the perceived security risks and system performance constraints. The channel decoding module is used to sense channel environment information under the guidance of the intelligent agent, and to perform joint source-channel decoding on the decrypted channel symbols based on the sensed channel environment information to obtain semantic features. The semantic recovery module is used to perform semantic recovery on the semantic features and reconstruct the image under the guidance of the intelligent agent.

[0009] According to a fifth aspect of this disclosure, a semantic communication system is provided, comprising: the aforementioned transmitting end, the aforementioned receiving end, and a communication channel connecting the transmitting end and the receiving end.

[0010] According to a sixth aspect of this disclosure, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to at least one of the aforementioned processors; wherein, The memory stores instructions that can be executed by at least one processor, and the instructions are executed by at least one processor to enable at least one processor to perform the semantic communication method.

[0011] According to a seventh aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause a computer to perform the semantic communication method described above.

[0012] According to the eighth aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the semantic communication method described above.

[0013] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0014] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is a flowchart illustrating a semantic communication method applied to the sending end of a semantic communication system according to an embodiment of this disclosure; Figure 2 This is a flowchart illustrating some steps of another semantic communication method applied to the sending end of a semantic communication system, provided in an embodiment of this disclosure. Figure 3 This is a flowchart illustrating some steps of another semantic communication method applied to the sending end of a semantic communication system, provided in an embodiment of this disclosure. Figure 4 This is a flowchart illustrating some steps of another semantic communication method applied to the sending end of a semantic communication system, provided in an embodiment of this disclosure. Figure 5 This is a flowchart illustrating a semantic communication method applied to the receiving end of a semantic communication system according to an embodiment of this disclosure; Figure 6 This is a flowchart illustrating some steps of another semantic communication method applied to the receiving end of a semantic communication system, provided in an embodiment of this disclosure. Figure 7 This is a flowchart illustrating some steps of another semantic communication method applied to the receiving end of a semantic communication system, provided in an embodiment of this disclosure. Figure 8 This is a flowchart illustrating some steps of another semantic communication method applied to the receiving end of a semantic communication system, provided in an embodiment of this disclosure. Figure 9This is a schematic diagram of the structure of the sending end of a semantic communication system provided in an embodiment of this disclosure; Figure 10 This is a schematic diagram of the structure of the sending end of a semantic communication system provided in an embodiment of this disclosure; Figure 11 This is a block diagram of an electronic device used to implement the semantic communication method of the embodiments of this disclosure. Detailed Implementation

[0015] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0016] In some related technologies, the sending end extracts semantic information from the source signal through a semantic recognition model, performs querying and expansion processing with the help of public and private knowledge graph libraries, compresses the semantic information into semantic code, and sends it; the receiving end combines the shared public knowledge graph, the exclusive private knowledge graph, and context information to perform decompression and recovery operations on the received semantic code to achieve the reconstruction of complete semantics.

[0017] The aforementioned implementation focuses on idealized semantic information compression and representation, but fails to consider channel environment factors (such as noise, interference, and fading), making it difficult to guarantee transmission reliability under complex wireless channels. Furthermore, due to the lack of security mechanisms along the transmission path, the encoded semantic information is in plaintext, making it vulnerable to malicious eavesdropping or tampering, posing a risk to information security.

[0018] The semantic communication method applied to the sending end of a semantic communication system, the semantic communication method applied to the receiving end of a semantic communication system, the semantic communication system sending end, the semantic communication system receiving end, the semantic communication system, the electronic device, and the computer-readable storage medium provided in the embodiments of this disclosure are intended to solve at least one of the above-mentioned technical problems of the prior art.

[0019] The semantic communication method provided in this disclosure can be executed by electronic devices such as terminal devices or servers. Terminal devices can be in-vehicle devices, user equipment (UE), mobile devices, user terminals, terminals, cellular phones, cordless phones, personal digital assistants (PDAs), handheld devices, computing devices, in-vehicle devices, wearable devices, etc. The method can be implemented by a processor calling computer-readable program instructions stored in memory. Alternatively, the method can be executed by a server.

[0020] Figure 1 A flowchart illustrating a semantic communication method applied to the sending end of a semantic communication system, provided by an embodiment of this disclosure, is shown.

[0021] Semantic communication systems are a novel communication architecture that focuses on conveying the "meaning" (semantics) of information, unlike traditional communication systems that only focus on error-free bit transmission. Its core idea is to extract semantic content useful for the task from the source information at the sending end, and then reconstruct information with the same or similar meaning based on this semantic content at the receiving end, rather than simply reproducing the original data bit by bit. It includes a sending end, a receiving end, and a communication channel connecting the sending and receiving ends.

[0022] The sending end is the source-side device or functional module of the semantic communication system, responsible for understanding and extracting semantics from the original source information, and transforming the semantics into a form suitable for transmission.

[0023] The receiver is the destination device or functional module of the semantic communication system, which is responsible for recovering semantics from the received signal and reconstructing information useful to the user or downstream tasks based on the semantics.

[0024] A communication channel is a physical or logical transmission medium connecting a sender and a receiver, used to carry signals to be transmitted. In semantic communication, the channel transmits encoded semantic feature signals, rather than the original data bit stream.

[0025] like Figure 1 As shown in the embodiments of this disclosure, the semantic communication method applied to the sending end of a semantic communication system may include steps S110, S120, and S130.

[0026] S110. Under the guidance of the intelligent agent, semantic extraction is performed on the image source to obtain semantic features.

[0027] Among them, intelligent agent refers to artificial intelligence module deployed at the sending end and / or receiving end, which has the ability to perceive environmental information (including channel status, security risks, system performance constraints, etc.), make decisions and control the communication process.

[0028] In some possible implementations, the intelligent agent can be based on reinforcement learning, deep learning, or a rule engine, and internally maintains a channel model, a security policy library, and performance evaluation functions. In the embodiments of this disclosure, the intelligent agent at the sending end and the intelligent agent at the receiving end can be the same agent (sharing state), or they can be multiple agent instances that are synchronized with each other.

[0029] Semantic features are low-dimensional vectors extracted from the original image source that represent the core meaning of the image content. Compared with the original pixel data, semantic features significantly reduce the amount of data while retaining information in the image that is crucial for communication objectives (such as object recognition and scene understanding).

[0030] In some possible implementations, the intelligent agent first obtains the objective of the current task (e.g., image classification, object detection, or scene reconstruction). Depending on the task requirements, the intelligent agent selects or configures a deep neural network (e.g., a convolutional neural network or a visual Transformer) that encodes the raw input image (such as an RGB pixel matrix) into a low-dimensional semantic feature vector, i.e., semantic features.

[0031] In some specific implementations, during the semantic extraction process, the intelligent agent can determine the dimension of the features (e.g., compressing from 224×224×3 of the original image to a 256-dimensional vector) and the specific structure of the extraction network.

[0032] For example, suppose the image source is The intelligent agent invokes a pre-trained semantic extraction network. ,in These are the network parameters. The network outputs semantic features. ,in, The intelligent agent can dynamically adjust the value of d based on channel bandwidth or task accuracy requirements. For example, when the channel quality is good and bandwidth is sufficient, a larger d is chosen to retain more details; when the channel quality is poor, a smaller d is chosen to reduce transmission load.

[0033] By extracting semantic information, high-dimensional image data is compressed into low-dimensional semantic features, significantly reducing the amount of data required for subsequent encoding and transmission, thereby saving bandwidth resources. Simultaneously, because only core semantic information is retained, sensitivity to channel noise and interference is reduced, improving transmission robustness.

[0034] S120. Under the guidance of the intelligent agent, the channel environment information is perceived, and the semantic features are jointly encoded by the source and channel based on the perceived channel environment information to generate channel symbols.

[0035] The channel environment information keyi includes signal-to-noise ratio (SNR), multipath fading parameters, interference power, and channel time-varying characteristics. In some possible implementations, the intelligent agent can acquire this information in real time through pilot signals, historical feedback, or a channel estimation module.

[0036] Source-channel joint coding is a coding method that jointly optimizes semantic feature compression and channel protection (such as noise immunity and fading immunity). Unlike traditional separate source coding and channel coding, joint coding directly outputs a symbol sequence suitable for channel transmission and can dynamically adjust the compression rate and redundancy according to the channel state.

[0037] Channel symbols are complex or real number sequences generated through joint coding of the source and channel, ready to be transmitted through the physical channel. The statistical properties of the channel symbols (such as power and distribution) are matched with the channel conditions to resist distortion during transmission.

[0038] In some possible implementations, the intelligent agent perceives current channel environment information (such as signal-to-noise ratio and fading depth) in real time. Then, the intelligent agent adjusts the parameters or coding strategy of the source-channel joint encoder based on this information.

[0039] In some specific implementations, the joint encoder can be a deep neural network (which can share some layers with the semantic extraction network or be a separate network), whose input is the semantic feature z and whose output is the channel symbol sequence. (or The intelligent agent can control the encoder's output length m (i.e., bit rate) and redundancy. For example, when the channel quality is poor, the intelligent agent commands the encoder to increase the output length (i.e., increase redundancy) to introduce more error correction capabilities; when the channel quality is good, it shortens the output length to improve transmission efficiency.

[0040] For example, suppose the joint encoder network is ,in This is an adjustable parameter. The intelligent agent selects a parameter configuration based on the perceived channel environment information c (e.g., signal-to-noise ratio SNR = 10dB). Then, channel symbols are generated. For example, when the SNR is low, the smart agent settings... This increases the output symbol length m by 50% and employs a more robust modulation scheme (implicit in the network output). In implementation, the joint encoder can map z to points on the complex plane using a fully connected network or a convolutional network.

[0041] Joint source-channel coding can dynamically adjust the coding strategy according to channel conditions, achieving an optimal balance between source compression and channel protection. Compared with fixed coding methods, this method can maintain low semantic distortion even when the channel changes, significantly improving transmission reliability.

[0042] S130. Under the guidance of the intelligent agent, the security risks and system performance constraints are perceived. Based on the perceived security risks and system performance constraints, the channel symbols are encrypted, and the encrypted channel symbols are generated and sent to the wireless channel.

[0043] Among them, security risk refers to the possibility that the wireless channel may be eavesdropped on, tampered with or interfered with. In some possible implementations, it can be assessed by intelligent agents by analyzing indicators such as abnormal bit error rate, fluctuations in received signal strength, and unauthorized access attempts.

[0044] System performance constraints may include maximum allowable latency, computing resources (such as computing power and memory) at the sending / receiving end, and power consumption limits. Intelligent agents need to balance security, reliability, and efficiency when making decisions.

[0045] Encrypting channel symbols refers to the process of converting channel symbols into ciphertext through a series of reversible transformations (such as noise addition, obfuscation, and permutation). The encryption in this embodiment is not traditional bit-level encryption, but rather an encryption operation performed directly on analog domain or continuous value symbols, and its strength can be controlled by a combination of operation steps and transformation types.

[0046] In some possible implementations, the intelligent agent perceives current security risks (e.g., the presence of eavesdroppers, whether the channel is being monitored) and system performance constraints (e.g., maximum allowable encryption latency, decryption computing power at the receiver). Based on this information, the intelligent agent selects an encryption strength (which may include a combination of transformation types and the number of operation steps). The encryption process employs a predefined transformation flow, where each step can be an operation such as Gaussian noise addition, feature obfuscation, or dynamic permutation. The encryption operation acts directly on continuous-valued channel symbols, rather than discrete bits. The encrypted symbol... It maintains the same dimension as x, but the statistical properties are perturbed, making it difficult to recover the original semantic information from the ciphertext.

[0047] For example, suppose the channel symbol sequence The predefined encryption process includes L optional transformation steps, each step corresponding to a transformation type. The intelligent agent is configured based on the security risk level *r* and latency constraints. Select transformation type combination and the number of operation steps K, where, Then execute in sequence: Add noise: ,in, The noise is Gaussian, and the noise variance is set by the intelligent agent based on security risks.

[0048] Confusion: This introduces linear combinations of adjacent symbols.

[0049] Permutation: Randomly arrange the sequence x. Reorder.

[0050] The ciphertext is obtained after K steps. Finally, The data is transmitted through a wireless channel. If the intelligent agent perceives a very high security risk and the latency allows, it can choose K=L and use all transformations; if the channel is relatively secure or latency is sensitive, it can choose a smaller K or even skip encryption (K=0).

[0051] By encrypting channel symbols with adjustable strength in the analog domain, the security of semantic communication is ensured while avoiding the high computational overhead and bit error propagation associated with traditional encryption. The intelligent agent can dynamically adjust the encryption strength based on real-time security risks and system constraints, achieving an adaptive balance between security and efficiency. Furthermore, due to the joint design of encryption and channel coding, the ciphertext still possesses resistance to channel noise.

[0052] In the semantic communication method applied to the transmitter of a semantic communication system provided in this disclosure embodiment, under the unified guidance of an intelligent agent, the three stages of semantic extraction, source-channel joint coding, and encryption are coordinated and adaptive. Semantic extraction significantly compresses the data volume, and joint coding further optimizes the compression-to-protection ratio. Joint coding dynamically adjusts redundancy according to channel conditions, effectively combating noise and fading. An encryption mechanism with adjustable strength prevents eavesdropping and is compatible with channel coding. The intelligent agent perceives environmental changes in real time, balancing transmission efficiency, reliability, and security without manual intervention.

[0053] The semantic communication method for the sending end of a semantic communication system provided in the embodiments of this disclosure will be described in detail below.

[0054] In some possible implementations, semantic extraction and source-channel joint coding use the same deep neural network, with an intelligent agent dynamically selecting the network's coding parameters based on channel environment information. By sharing the network, end-to-end joint optimization of feature extraction and channel mapping is achieved, reducing system complexity and improving overall performance.

[0055] Figure 2 The diagram illustrates a flowchart of an implementation method that achieves end-to-end joint optimization of feature extraction and channel mapping through a shared network, as shown below. Figure 2 As shown, steps S210, S220, and S230 may be included.

[0056] S210. Input the image source into a deep neural network, and the deep neural network outputs a low-dimensional semantic feature vector.

[0057] In this embodiment of the disclosure, the deep neural network is a multi-layer neural network that can be trained end-to-end, such as a convolutional neural network (CNN), a residual network (ResNet), or a Transformer.

[0058] In some possible implementations, the network performs two tasks simultaneously: mapping the original image into low-dimensional semantic features (semantic extraction), and further mapping the semantic features into a symbol sequence suitable for channel transmission (source-channel joint coding). Different layers of the network can be responsible for different functions: shallow layers extract local texture and edge information, middle layers form semantic abstractions, and deep layers (or specific branches) complete the modulation mapping to channel symbols.

[0059] In other words, semantic extraction and source-channel joint coding share the same set of network parameters, rather than being two separate networks. Specifically, the first few layers of the network output intermediate semantic features, which can be used as semantic output for monitoring or task analysis, or fed into subsequent layers to generate channel symbols. During training, the network undergoes end-to-end optimization using a joint loss function (e.g., a weighted sum of image reconstruction error and channel transmission bit error rate) to ensure that semantic features and channel symbols are mutually adapted. During inference, the original image passes through the network once, sequentially obtaining semantic features and channel symbols, without requiring the invocation of two separate models.

[0060] Low-dimensional semantic feature vectors are representation vectors output by intermediate layers of a network, and their dimension (i.e., the length of the feature vector) is much smaller than the number of pixels in the original image. The low-dimensionality characteristic comes from the bottleneck structure of the network (such as pooling layers, strided convolutions, or attention-based dimensionality reduction). Intelligent agents can also dynamically adjust the dimension of the bottleneck layer according to bandwidth constraints (e.g., by controlling the number of output neurons in the fully connected layer).

[0061] In some specific implementations, the intelligent agent first feeds the original image source into a deep neural network. The first few layers of the deep neural network (e.g., convolutional layers + activation layers + pooling layers) extract features from the image (i.e., the original image source) step by step, abstracting low-level features such as edges, corners, and textures from pixel-level RGB values, and then combining them into high-level semantic information such as object parts and scene layout. At a certain intermediate layer of the deep neural network (usually after a fully connected layer or a global pooling layer), a fixed-length vector, i.e., a low-dimensional semantic feature vector, is output. This feature vector should retain as much information as possible that is crucial to subsequent communication tasks (such as image reconstruction and object recognition), while discarding details that are irrelevant to the task (such as lighting changes and small noises).

[0062] In some possible implementations, intelligent agents can guide deep neural networks to select the dimension of semantic features: for example, when channel bandwidth is ample, a larger dimension is selected to retain more details; when bandwidth is limited, a smaller dimension is selected to ensure that basic semantics can be transmitted.

[0063] For example, suppose a deep neural network is ,in, These are the network parameters. An example of a deep neural network structure is shown below: Input layer: receives images .

[0064] Multiple convolutional blocks: Each convolutional block contains convolutional layers, batch normalization, ReLU activation, and pooling layers, progressively reducing spatial resolution and increasing the number of channels.

[0065] Global average pooling layer: converts feature maps into vectors.

[0066] Fully connected layer: outputs a semantic feature vector of dimension d. ,in, This represents the parameters of the front and middle layers of the network.

[0067] The intelligent agent can adjust the number of output neurons in the fully connected layer based on the current bandwidth requirements. For example, if the current channel bandwidth is 10 kHz, the intelligent agent chooses d=256; if the bandwidth is 5 kHz, it chooses d=128. This selection is achieved by the intelligent agent sending configuration instructions to the network, and the deep neural network dynamically prunes or expands the output dimension according to the instructions (using adjustable fully connected layers or masking techniques).

[0068] By extracting low-dimensional semantic features through a single network, the need to train and maintain a separate semantic extraction model is avoided, reducing computational and storage overhead. Since deep neural networks are jointly optimized end-to-end, semantic features naturally adapt to subsequent source-channel joint coding, resulting in better protection of semantically important components during channel mapping. The low-dimensionality effectively reduces transmission bandwidth requirements while improving robustness in low signal-to-noise ratio environments.

[0069] S220. Based on the channel environment information perceived by the intelligent agent, select the encoding parameters of the deep neural network.

[0070] S230. Input the semantic features into the deep neural network and map the semantic features into channel symbols according to the encoding parameters.

[0071] The encoding parameters are a set of adjustable parameters in a deep neural network, including network weights, biases, batch normalization parameters, etc. In this embodiment, the specific values ​​of the encoding parameters are dynamically selected by the intelligent agent based on perceived channel environment information (such as signal-to-noise ratio, multipath fading, and interference level).

[0072] In some possible implementations, the intelligent agent can maintain multiple pre-trained sets of network parameters (corresponding to different channel states) or generate parameters in real time through a hypernetwork. The encoding parameters determine the mapping function from semantic features to channel symbols, thus affecting the compression ratio and robustness of the transmission.

[0073] In some possible implementations, the intelligent agent perceives channel environment information in real time (including signal-to-noise ratio, multipath delay spread, Doppler shift, interference power, etc.). Based on this information, the intelligent agent selects an optimal set of encoding parameters from the candidate parameter set for subsequent parts of the deep neural network (i.e., the mapping part from semantic features to channel symbols).

[0074] This part of the network can be viewed as a modulation encoder, whose input is the semantic feature vector z output from step S210, and whose output is the channel symbol sequence x. The coding parameters determine the shape of the mapping function. For example, when the channel conditions are poor, parameters that can generate more redundant symbols are selected (i.e., increasing the length of x or introducing a more robust constellation diagram); when the channel conditions are good, parameters that are efficient in compression are selected (reducing the length of x and increasing the transmission rate). The semantic feature vector z is then fed into the latter half of the network (which can be several fully connected layers or recurrent layers), and mapped to complex channel symbols according to the selected coding parameters.

[0075] Assume the latter half of the deep neural network is ,in, These are the encoding parameters. The entire network F can be written as... The intelligent agent maintains an encoded parameter mapping table or parameter generator. Where c is the channel environment information vector (e.g., including signal-to-noise ratio (SNR), delay spread, etc.). Upon sensing c, the intelligent agent performs computation. . Specifically: If the SNR is high (e.g., >20dB), select the smart agent. This parameter makes the output symbol length of G... It employs high-order modulation implicit mapping (such as each symbol carrying 4 bits of semantic information).

[0076] If the SNR is moderate (e.g., 10-20 dB), choose Output length Medium protection.

[0077] If the SNR is low (e.g., <10 dB), select Output length To increase redundancy, repetitive coding or low-order modulation is used.

[0078] Then, the semantic feature z is input into G to calculate the channel symbol. Each element of x can be a complex number, represented by two real numbers, one real and one imaginary. In implementation, the output layer of G can be a fully connected layer outputting 2m real numbers, which are then reassembled into m complex symbols. The intelligent agent can also adjust the transmit power (as part of the coding parameters) according to the channel conditions to meet power constraints.

[0079] To ensure the end-to-end performance of the entire network F, the network needs to be trained in advance under various channel conditions, so that it can adapt to different... The network can reconstruct high-quality images. During training, the loss function includes the reconstruction error of semantic features and the image reconstruction error after the channel symbols pass through the simulated channel.

[0080] By dynamically selecting coding parameters based on the channel environment through an intelligent agent, the same deep neural network can adaptively adjust the compression ratio and redundancy without switching models. Compared to traditional fixed coding methods, the embodiments of this disclosure significantly improve the transmission quality of the system under different channel conditions.

[0081] For example, when the signal-to-noise ratio (SNR) is as low as 5dB, image recognizability can still be guaranteed by increasing redundancy; when the SNR is as high as 30dB, high-speed transmission can be achieved by reducing redundancy. At the same time, since semantic features and channel symbols share network parameters, information mismatch between feature extraction and channel coding is avoided, further improving end-to-end semantic fidelity.

[0082] In summary, the embodiments of this disclosure achieve semantic feature extraction and source-channel coding through a single deep neural network. The shared network reduces the number of parameters and computational load, making it suitable for edge device deployment. Joint training of semantic extraction and source-channel coding avoids the performance loss inherent in traditional separate designs, offering significant advantages, especially under high compression ratios and poor channel conditions. Semantic features and channel symbols can be obtained with a single forward propagation, eliminating the need for serially calling multiple models, making it suitable for real-time communication scenarios. The intelligent agent dynamically adjusts coding parameters based on the channel environment, ensuring stable transmission quality under various channel conditions.

[0083] Figure 3 This illustrates a flowchart of another implementation method for joint source-channel coding of semantic features based on perceived channel environment information, as shown below. Figure 3 As shown, steps S310 and S320 may be included.

[0084] S310. Based on the channel environment information perceived by the intelligent agent, adjust the initial coding strategy to obtain the adjusted coding strategy.

[0085] The channel environment information is the same as defined above, including signal-to-noise ratio (SNR), multipath fading depth, Doppler shift, interference signal power, and channel time-varying characteristics. The intelligent agent acquires this information in real time through channel estimation, pilot measurement, and historical feedback.

[0086] The initial coding strategy is the coding strategy currently being used by the transmitter before changes occur in the channel environment.

[0087] Coding strategy is a broad concept that encompasses all the rules and parameters required to convert semantic features into channel symbols. Specifically, it can include: the quantization method of semantic features (such as scalar quantization and vector quantization); the source compression ratio (i.e., the compression ratio from semantic features to channel symbols); the channel coding scheme (such as code rate, code pattern, and interleaving method); the modulation method (such as constellation mapping of QPSK and 16QAM); and the power allocation strategy (the transmit power of each symbol).

[0088] The initial encoding strategy can be the default strategy or a strategy retained from the last adjustment. The smart agent internally maintains the state of the current encoding strategy.

[0089] The adjusted coding strategy is a new strategy obtained by the intelligent agent after modifying one or more parameters of the initial coding strategy based on the perceived channel environment information.

[0090] The goal of the adjustment is to make the new coding strategy more suitable for the current channel conditions. For example, when the channel deteriorates, channel coding redundancy is increased or a lower-order modulation is used; when the channel improves, redundancy is reduced to increase the transmission rate. The adjusted coding strategy will be used for subsequent mapping of semantic features to channel symbols.

[0091] In some possible implementations, coding strategy adjustment is a decision-making process executed by an intelligent agent. Its input can be the currently perceived channel environment information and the initial coding strategy, and its output is the adjusted coding strategy. The adjustment can employ a rule engine (e.g., a pre-defined mapping table between signal-to-noise ratio thresholds and coding strategies), a reinforcement learning model, or an optimization algorithm. The adjustment can be incremental (e.g., fine-tuning the code rate based on the initial strategy) or replacement-based (switching to a completely different coding scheme).

[0092] In some specific implementations, the intelligent agent continuously monitors channel environment information. When a change in the channel environment is detected (e.g., the signal-to-noise ratio drops beyond a preset threshold, or the fading depth increases significantly), the intelligent agent initiates a coding strategy adjustment process.

[0093] First, the intelligent agent acquires the currently used initial coding strategy (e.g., by reading the strategy identifier or parameter set from internal memory). Then, based on the perceived channel environment information, the intelligent agent calculates a new coding strategy suitable for the current channel according to predefined adjustment rules or decision models. The adjustment can involve one or more dimensions of the coding strategy, such as modifying the code rate of the channel coding, changing the modulation order, adjusting the quantization precision of semantic features, or reallocating the transmit power of each semantic component. The adjusted coding strategy is stored as the new current strategy and used in subsequent mapping steps.

[0094] The following provides a specific implementation example using signal-to-noise ratio (SNR) as the primary channel environment information.

[0095] In some specific implementations, the intelligent agent internally maintains an encoding strategy table, as shown in the table below: Assume the initial coding strategy is Strategy B (medium channel quality). The intelligent agent acquires the SNR estimate in real time. When the SNR drops to 8 dB, the intelligent agent detects that the threshold condition of SNR < 10 dB has been triggered, and therefore decides to adjust the coding strategy to Strategy C. The adjustment process includes: The compression ratio of semantic features was changed from 1:2 to 1:1 (i.e. no compression, each semantic dimension corresponds to one channel symbol). Reduce the channel coding rate from 1 / 2 to 1 / 3 (to increase redundancy); Change the modulation scheme from QPSK to BPSK (to improve noise immunity).

[0096] The adjusted encoding strategy (strategy C) is stored in the strategy register of the smart agent and sent to the encoder for execution.

[0097] If the channel environment gradually improves and the SNR rises back to 15 dB, the intelligent agent can adjust back to strategy B, or even further back to strategy A. The intelligent agent can also employ more refined continuous adjustments, such as continuously changing the code rate according to a linear function of the SNR (rather than discrete switching). For example, setting the code rate... The intelligent agent calculates and updates the bitrate parameters in real time based on this formula.

[0098] By dynamically adjusting the coding strategy through intelligent agents, the system can quickly adapt to changes in the channel environment, avoiding performance losses caused by using outdated or inappropriate coding strategies. Compared to fixed coding strategies, this method improves transmission efficiency when channel quality is good and enhances error correction capabilities when channel quality is poor, thus maintaining high semantic transmission reliability in changing wireless environments. Furthermore, the adjustment process requires no manual intervention, achieving automated operation and maintenance.

[0099] S320. Based on the adjusted coding strategy, the semantic features are mapped to channel symbols.

[0100] In some possible implementations, after obtaining the adjusted coding strategy, the encoder converts the semantic feature vector into a channel symbol sequence according to the rules of that strategy. The mapping process can be a deterministic algorithm (such as linear transformation, quantization plus channel coding) or a nonlinear neural network mapping. That is, the encoder receives semantic features as input, performs the corresponding transformation according to the parameters specified in the adjusted coding strategy (such as compression ratio, code rate, modulation constellation, and power allocation coefficients), and outputs a symbol sequence suitable for transmission over the physical channel. These symbols can be real numbers, complex numbers, or integer sequences, depending on the modulation scheme.

[0101] Continuing with the example above, assume the adjusted encoding strategy is Strategy C (low SNR scenario). The encoder performs the following operations: Semantic feature preparation: Let the semantic feature vector be... Strategy C specifies a compression ratio of 1:1, therefore no additional dimensionality reduction is performed on z. If the original z dimension d is large, all dimensions can be retained.

[0102] Channel coding: A convolutional code with a code rate of 1 / 3 is used. Each element in the semantic feature vector z is treated as an information symbol (which may require quantization or soft information processing beforehand). In practice, each component of z is usually converted into a binary representation before channel coding, resulting in a 3d-dimensional encoded bitstream.

[0103] Modulation: BPSK modulation is used, mapping each coded bit to a real number symbol of +1 or -1. Therefore, the final channel symbol sequence x has a length of 3d, with each symbol being ±1.

[0104] Power adjustment: Strategy C may also specify a transmit power coefficient, such as multiplying the symbol amplitude by a coefficient. To meet the transmit power constraint.

[0105] To ensure the real-time performance of the mapping process, the intelligent agent can pass the adjusted encoding strategy as a parameter to the encoder, which then dynamically loads the corresponding algorithm or configuration. If the encoder is based on deep learning, the adjusted encoding strategy can be reflected in the switching of neural network weights (i.e., Figure 2 The corresponding implementation method can also be reflected in the adjustment of parameters in the post-processing of the network output (such as scaling, quantization step size, etc.).

[0106] Mapping based on the adjusted coding strategy ensures that the generation of channel symbols matches the current channel environment, thereby improving the semantic recovery quality after decoding at the receiver.

[0107] This disclosure provides a flexible and efficient source-channel joint coding method by dynamically adjusting the coding strategy through an intelligent agent. Figure 2Compared to other deep neural network parameter selection schemes, the embodiments disclosed herein focus more on policy-level adjustments and are not limited to the underlying implementation, thus making them applicable to a wider range of system architectures. It can achieve "channel-aware coding" by sensing the channel environment in real time and adjusting the coding strategy. When channel conditions change drastically (such as deep fading or sudden interference), it can quickly switch to a more robust coding strategy to maintain communication continuity. It can be implemented based on traditional coding theory or combined with machine learning methods, facilitating engineering deployment and upgrades.

[0108] Figure 4 The diagram illustrates a flowchart of one implementation method for encrypting channel symbols based on perceived security risks and system performance constraints, such as... Figure 4 As shown, steps S410 and S420 may be included.

[0109] S410. Based on the security risks and system performance constraints perceived by the intelligent agent, select the transformation type combination and the number of operation steps, wherein the transformation type combination is an ordered sequence used to specify the transformation type used in each step of the predefined encryption process.

[0110] Specifically, security risk refers to the possibility that a wireless channel may be eavesdropped on, tampered with, or interfered with. Smart agents can assess security risks using various indicators, such as: the presence of unauthorized receiving devices in the channel (through ranging or authentication), abnormally high bit error rates (potentially indicating interference or eavesdropping), geographical location information (e.g., vehicles entering sensitive areas), and historical attack records. Security risks can be quantified as a level (e.g., low, medium, high) or a continuous value (0 to 1). Smart agents determine encryption strength based on the security risk level: the higher the risk, the more likely they are to choose more complex combinations of transformations with more steps.

[0111] System performance constraints refer to the limitations of a communication system in terms of latency, computing resources, and power consumption. Specifically, these may include: Maximum allowable latency: Encryption operations increase latency. For real-time communication (such as autonomous driving and telemedicine), the latency must be less than the threshold.

[0112] Receiver computing resources: Decryption requires the inverse operation. If the receiver's computing power is limited (such as IoT sensors), the encryption strength cannot be too high, otherwise the receiver will not be able to decrypt in real time.

[0113] Power consumption budget: Encryption and decryption consume electrical energy. For battery-powered devices, a trade-off between security and battery life is necessary.

[0114] When making encryption decisions, intelligent agents must meet these constraints to avoid system unavailability due to excessive encryption.

[0115] The transformation type combination is an ordered sequence in which each element represents the transformation type to be used in the corresponding step of a predefined encryption process. A predefined encryption process can contain multiple steps, and each step can independently select its transformation type.

[0116] The combination of transformation types determines the overall behavior of the encryption process. For example, the combination [Gaussian noise addition, feature obfuscation, dynamic permutation] means that step 1 performs Gaussian noise addition, step 2 performs feature obfuscation, and step 3 performs dynamic permutation. Different combinations will produce different encryption effects, and the length of the combination is the actual number of operation steps performed. The smart agent can choose combinations of any length (from 1 step to the maximum number of steps) as needed, and can also choose to reuse a certain transformation type (such as [noise addition, noise addition, obfuscation]).

[0117] The number of operation steps refers to the number of steps in the predefined encryption process executed during the actual encryption process. The number of operation steps K is a positive integer and does not exceed the maximum number of steps L in the predefined encryption process.

[0118] The number of operation steps determines the computational overhead and encryption strength: more steps generally result in higher ciphertext security, but also greater latency and computational resource consumption. Intelligent agents select an appropriate number of operation steps based on system performance constraints (such as maximum allowable latency) and the level of security risk. For example, a larger K can be chosen when the security risk is high and latency is ample; a smaller K is chosen when the system is latency-sensitive or the receiver's computing power is limited.

[0119] A predefined encryption process is a set of predefined, sequentially executed encryption transformation steps, each step corresponding to a specific transformation operation. A predefined encryption process can be viewed as an encryption "pipeline," where each position (step) can carry a different transformation type. The process itself has a fixed framework, but the specific transformation type executed at each step is specified by the combination of transformation types. For example, if the maximum number of steps in a predefined encryption process is L=5, then the process includes steps 1 to 5. If the smart agent selects a combination length of 3, and the combination is [A, B, C], then step 1 will be executed using A, step 2 using B, step 3 using C, and steps 4 and 5 will be skipped. The design of predefined encryption processes makes encryption operations modular and configurable.

[0120] In some possible implementations, the intelligent agent first obtains the current security risk level and system performance constraints (such as maximum latency and receiver computing power). Then, based on a predefined strategy or a trained decision model, the intelligent agent selects a combination of transformation types (i.e., an ordered sequence specifying the transformation type at each step of the encryption process) and a number of operation steps, K. The selection principle is to maximize security while satisfying system performance constraints. For example, when the security risk is high and system resources are sufficient, a longer combination (such as multiple combinations including noise addition, obfuscation, and permutation) and a larger K are selected; when the security risk is low or resources are limited, a simplified combination (such as noise only) and a smaller K are selected. The combination of transformation types can be selected from a pre-defined candidate combination library or dynamically generated by the intelligent agent.

[0121] In some specific implementations, the smart agent internally maintains an encryption policy decision table or decision model. Below is an example of rule-based decision-making.

[0122] First, define the security risk level: Low risk: No suspicious devices were detected in the channel, the bit error rate was normal, and the system was in a safe environment.

[0123] Medium risk: There may be unauthorized access attempts or abnormal fluctuations in the bit error rate.

[0124] High risk: Detection of a clear eavesdropping device (e.g., a non-cooperative receiver is found through signal angle of arrival measurement), or the system requests the transmission of highly confidential information.

[0125] Define system performance constraints: Latency sensitive: End-to-end latency requirement <10ms, total encryption / decryption latency must be <2ms.

[0126] Normal: Latency requirement <100ms, encryption latency can be moderate.

[0127] Limited computing power: The receiver is a low-power IoT device that only supports simple calculations.

[0128] The intelligent agent selects the combination of transformation types and the number of steps based on the combination of risks and constraints. For example: The intelligent agent assesses the current state in real time and looks up the corresponding combination of K and transformation type in a table. For finer control, continuous functions can be used; for example, setting a security risk value. The maximum number of steps allowed by the delay constraint Then choose Then, the transformation types for the first K steps are selected randomly or in a predetermined order (e.g., fixed order: noise addition, obfuscation, permutation, noise addition, obfuscation, etc.). The generation of transformation type combinations can also be achieved using a reinforcement learning model, which can be dynamically optimized at runtime.

[0129] After selection, the smart agent sends the combination of K and transformation type to the encryption module and notifies the receiver (via a shared smart agent or out-of-band signaling) of the encryption parameters used so that the receiver can decrypt. In a preferred embodiment of this disclosure, the sender and receiver share the same smart agent, so the receiver can infer the encryption parameters based on perceived security risks and performance constraints without additional communication.

[0130] By dynamically selecting encryption strength and strategies based on security risks and system performance constraints through intelligent agents, security and efficiency can be automatically balanced in different scenarios. Unnecessary computational overhead is avoided in secure environments or when resources are limited; sufficient protection is provided when risks are high or confidentiality requirements are stringent. This adaptive mechanism enhances both the level of intelligence and practicality.

[0131] In some possible implementations, the transformation type includes one or more of Gaussian noise addition, feature obfuscation, and dynamic permutation.

[0132] Gaussian noise addition refers to superimposing independent and identically distributed Gaussian noise onto the channel symbol. The mean of the noise is typically 0, and the variance is determined by the intelligent agent based on the security risk level. The noise addition operation introduces random differences between the ciphertext and the original symbol, increasing the difficulty for an eavesdropper to recover the original symbol. Because the noise is random and unpredictable, even if the eavesdropper knows about the noise addition operation, they cannot completely remove the noise. The receiver, knowing the statistical characteristics of the noise or the specific noise sequence (if a synchronous pseudo-random number generator is used), can perform denoising operations (such as Wiener filtering or subtraction of known noise). In implementation, Gaussian noise addition can be expressed as... , Configured by the smart agent.

[0133] Feature obfuscation refers to the linear or nonlinear mixing of elements in a channel symbol sequence with other elements in the sequence to disrupt the independence and characteristic structure between symbols. Obfuscation can be achieved by applying a perturbation to each symbol that depends on other symbols, for example: Alternatively, matrix multiplication can be performed on the entire sequence. Where M is an invertible confusion matrix. The confusion operation makes it difficult for an eavesdropper to infer the statistical pattern of the original symbols from the ciphertext, while the receiver can de-confuse it (multiply by ). Recover the original sequence. The degree of confusion is determined by parameters. The condition number of matrix M determines the level of obfuscation, allowing the intelligent agent to select the obfuscation strength based on security requirements.

[0134] Dynamic permutation refers to rearranging the positions of elements in a channel symbol sequence according to a random order. A permutation can be represented as... ,in, It is a bijective function that returns an index set to itself. Dynamic permutation implies a permutation. The key can vary over time or session and is generated by an intelligent agent according to a security policy (e.g., using a pseudo-random number generator with a session key). The permutation operation does not change the value of the symbols but shuffles their order, making it impossible for an eavesdropper to correlate the ciphertext location with the original semantic structure. The receiver then uses the inverse permutation (i.e., applying...) Restore the original order.

[0135] S420. Using the channel symbol as the initial state, according to the predefined encryption process, and based on the combination of transformation types, execute the encryption transformation steps corresponding to the operation steps in sequence, wherein each encryption transformation step adopts the transformation type at the corresponding position in the combination of transformation types.

[0136] In some possible implementations, the channel symbol sequence is received. This serves as the initial state. Then, following the predefined encryption process, steps 1, 2, ..., K are executed sequentially. During each step, based on the transformation type at the corresponding position in the transformation type combination, the corresponding transformation operation is applied to the input data of that step (i.e., the output of the previous step). The output of each step serves as the input for the next step. After executing K steps, the encrypted channel symbol sequence is finally obtained. The entire process can be represented as: in, This refers to the transformation operation corresponding to the k-th step (such as Gaussian noise addition, feature obfuscation, or dynamic permutation). These transformation operations are all reversible so that the receiver can decrypt and recover the original channel symbols.

[0137] The following is a specific implementation example. Assume that the maximum number of steps in the predefined encryption process is L=3, the smart agent selection is K=3, the transformation type combination is [Gaussian noise, feature confusion, dynamic permutation], and the length of the input channel symbol sequence x is m.

[0138] Step 1: Gaussian noise addition The intelligent agent selects the noise variance based on the security risk level. For example, choose when the risk is medium. =0.1, selected when there is high risk. =0.5.

[0139] Generate independent zero-mean Gaussian noise sequences .

[0140] calculate .

[0141] If needed, the seed of the noise generator can be bound to the shared key, so that the receiving end can synchronously generate the same noise sequence, thereby achieving noise reduction.

[0142] Step 2: Feature Confusion Using a linear confusion method: Define a confusion matrix. This matrix is ​​invertible and has a controllable condition number. The intelligent agent can choose a simple tridiagonal matrix, for example: in, To obfuscate the strength parameter, the intelligent agent sets it based on security risks (e.g., high risk). Medium risk ).

[0143] calculate This operation mixes each symbol with its neighbors, making it difficult for an eavesdropper to distinguish them.

[0144] The receiving end uses an inverse matrix for decryption. It can be restored immediately.

[0145] Step 3: Dynamic replacement The intelligent agent generates a random permutation It operates on sequence indexes. For example, when m=4, it generates... This means that the first element of the new sequence is taken from the second element of the original sequence, and so on.

[0146] calculate ,in, (or (Depending on the agreement).

[0147] The substitution operation can be implemented by rearranging an index with minimal computation. The receiving end restores the order by inverting the substitution.

[0148] After the above three steps, the encrypted channel symbol is obtained. It is then sent to the wireless channel for transmission. If the number of operation steps K=2, only the first two steps are executed, and the third step is skipped.

[0149] In practical systems, to further enhance security, key-controlled parameters (such as keys for noise variance, confusion matrices, and permutation sequences) can be used at each step. These keys can be generated by the smart agent based on a shared secret and do not need to be transmitted over the channel. If the sender and receiver share the same smart agent, the keys can be automatically synchronized between the two ends.

[0150] By sequentially performing multiple reversible transformations, the channel symbols are subjected to multi-level scrambling in the analog domain. Compared to traditional bit-level encryption, this offers the following advantages: Low complexity: Operations apply directly to consecutive value symbols, avoiding bit string encoding / decoding and padding, resulting in low computational overhead.

[0151] Noise resistance compatibility: The encrypted ciphertext can still be transmitted through the wireless channel, and the decryption process has a certain tolerance to channel noise (for example, the noise addition operation can be fused with the channel noise and processed jointly by the receiver).

[0152] Adjustable strength: By selecting different combinations of transformation types and operation steps, a smooth transition from lightweight obfuscation to high-strength encryption can be achieved.

[0153] Transparency: The encryption operation is transparent to the upper-layer semantic encoding, meaning that the joint encoder does not need to be aware of the encryption details, which facilitates system integration.

[0154] The encryption method provided in the above-disclosed embodiments, under the unified control of the intelligent agent, achieves the following: Dynamically selects encryption strength and strategy based on real-time perceived security risks and system resources, avoiding over-encryption or under-encryption. Furthermore, the encryption operation is designed for channel symbols and seamlessly integrated with source-channel joint coding, without affecting the error correction performance of the coding. The transformation operations (noise addition, obfuscation, and permutation) are simple and can all be completed in linear time, making them suitable for real-time communication. The combination of multiple transformations makes it difficult for attackers to recover the original semantic features from the ciphertext. Even if some transformation parameters are known, the unknown number of steps and combinations make cracking difficult. By selecting the number of operation steps, only lightweight encryption (such as permutation) can be enabled on computing-limited devices, ensuring basic security while avoiding power consumption spikes.

[0155] In some possible implementations, the sender and receiver share the same smart agent. The smart agent senses channel environment information and determines the encryption strategy, which is used for both encryption at the sender and decryption at the receiver.

[0156] Sharing the same smart agent means that the sender and receiver do not each have independently running smart agent instances, but rather share the same smart agent entity or its state. This smart agent can be deployed in any of the following locations: Cloud server: Both the sending and receiving ends communicate with this cloud-based intelligent agent to obtain decision-making instructions.

[0157] Network edge nodes (such as base stations and edge servers): Intelligent agents are centrally run by edge nodes, serving multiple terminals simultaneously.

[0158] One of the sending or receiving ends: The intelligent agent is deployed on one end, but its state (including perception information, decision model parameters, internal memory, etc.) is synchronized to the other end in a secure manner, so that the two ends logically have the "same" agent.

[0159] In physical implementation, "sharing" can manifest as both ends maintaining completely identical copies of the smart agent and keeping their states consistent through periodic synchronization; or it can manifest as a centralized smart agent simultaneously issuing instructions to both ends. In the embodiments of this disclosure, the core feature of sharing the same smart agent is that the decision information (such as channel environment perception results, security risk assessment, and encryption strategy selection) used by the sending end and the receiving end during encryption and decryption is the same, and there is no need to exchange this information through the channel through which data is being transmitted.

[0160] The encryption strategy is the encryption scheme determined by the intelligent agent for this communication. It specifically includes the combination of transformation types and the number of operation steps, as well as possible additional parameters (such as the noise variance of Gaussian noise, the confusion matrix of feature confusion, and the random permutation seed for dynamic permutations). The encryption strategy determines how the sender converts channel symbols into ciphertext and how the receiver restores ciphertext to channel symbols. Because the intelligent agent guides the operations of both ends simultaneously, the receiver does not need to wait for the sender to inform it of the encryption strategy through the channel; instead, it can independently (or through interaction with the intelligent agent) learn the same strategy.

[0161] The intelligent agent acquires channel environment information through various methods. For example, it can obtain uplink channel measurement reports from the transmitter, downlink channel measurement reports from the receiver, or perform estimations using channel reciprocity. Since the transmitter and receiver share the same intelligent agent, the agent can fuse information from both ends to form a more comprehensive view of the channel state, thereby enabling better encryption strategy decisions.

[0162] The encryption strategy determined by the smart agent is symmetric: the sending end uses the strategy to perform encryption, and the receiving end uses the same strategy (or its reverse process) to perform decryption. Because the smart agent is shared, the two ends can automatically align encryption parameters without exchanging keys or strategy identifiers over the communication link. This not only saves bandwidth but also avoids the risk of parameters being eavesdropped on or tampered with during transmission.

[0163] In some possible implementations, before or during communication, the intelligent agent continuously senses channel environment information (such as signal-to-noise ratio, interference level, multipath fading, etc.) and security risks (such as the presence of eavesdroppers). Based on this information, the intelligent agent determines an encryption strategy according to predefined decision logic (such as rule tables, reinforcement learning models, etc.). The specific content of this strategy might include, for example, selecting a combination of transform types such as [Gaussian noise addition, feature obfuscation, dynamic permutation], with K=3 operation steps, and specific parameters for each transform. Then, the intelligent agent simultaneously sends this encryption strategy to both the sender and receiver. The sender uses this strategy to encrypt channel symbols; the receiver uses the corresponding inverse strategy to decrypt the received ciphertext. Since both ends obtain the strategy from the same intelligent agent, there is no need to transmit any encryption parameters in the communication channel, thus avoiding additional overhead and security vulnerabilities.

[0164] By using a shared intelligent agent, strong encryption is achieved while avoiding the overhead of key distribution and parameter synchronization in traditional encryption. The intelligent agent can sense channel changes in real time and adjust the encryption strategy accordingly, while both ends automatically synchronize without additional signaling. Even if an attacker completely controls the data channel, they cannot obtain the encryption strategy and can only blindly crack the ciphertext, significantly improving security. When there are multiple senders and receivers in the network, a single global intelligent agent can be shared to achieve unified encryption strategy management and simplify system operation and maintenance.

[0165] Figure 5 The diagram illustrates a flowchart of one implementation of a semantic communication method applied to a receiving end of a semantic communication system according to an embodiment of this disclosure. Figure 5 As shown, steps S510, S520, and S530 may be included.

[0166] S510, under the guidance of the intelligent agent, perceives security risks and system performance constraints, and decrypts the encrypted channel symbols received from the wireless channel to recover the channel symbols based on the perceived security risks and system performance constraints.

[0167] The intelligent agent is an artificial intelligence module shared or synchronized with the sending end. In this embodiment, the intelligent agent of the receiving end can be the same agent as the intelligent agent of the sending end (shared state), or it can be multiple agent instances synchronized with each other.

[0168] Intelligent agents possess the ability to perceive environmental information (including channel status, security risks, system performance constraints, etc.), make decisions, and control the communication process. The intelligent agent at the receiving end can infer the encryption strategy (such as transformation type combinations and operation steps) adopted by the sending end based on its perceived security risks and system performance constraints, thus eliminating the need to receive encryption parameters through the communication channel.

[0169] Security risk refers to the possibility of the wireless channel being eavesdropped on, tampered with, or interfered with. The receiver can assess security risks by analyzing indicators such as abnormal received signal quality, unauthorized access attempts, and sudden increases in the bit error rate. Because the receiver and transmitter share the same intelligent agent, the security risks perceived by the receiver are highly consistent with those perceived by the transmitter, thus enabling the determination of the same encryption strategy.

[0170] System performance constraints include available computing resources at the receiver (such as CPU / GPU computing power and memory size), power consumption limits, and maximum allowable decryption latency. The intelligent agent needs to select an appropriate decryption strength based on these constraints during decryption to avoid system congestion or timeouts caused by overly complex decryption operations.

[0171] Decryption is the inverse process of encryption, referring to the operation of recovering the original channel symbols from the received ciphertext (encrypted channel symbols) through a series of reversible transformations. Decryption operations include denoising, deobfuscation, and inverse permutation, and their specific types and number of operation steps match the encryption process at the sending end. Decryption can be performed directly in the analog domain (continuous-value symbols) without conversion to a bitstream.

[0172] In some possible implementations, the receiver's intelligent agent first perceives the current security risk level and system performance constraints (e.g., the receiver's current CPU load and the maximum allowed decryption latency). Based on this information, the intelligent agent determines the encryption strategy adopted by the sender, including the transformation type combination and the number of operation steps K. Since the sender and receiver share the same intelligent agent (or maintain state synchronization), the receiver can independently make the same encryption strategy decision as the sender without needing to receive encryption parameters via a wireless channel. Then, the intelligent agent guides the decryption module to perform denoising, deobfuscation, and inverse permutation operations sequentially, following the reverse process of encryption, to decrypt the received ciphertext. Restore to channel symbol x.

[0173] For example, suppose the receiver receives an encrypted channel symbol sequence via a wireless channel. The intelligent agent perceives a medium level of security risk, a system latency constraint of less than 20ms, and sufficient computing power at the receiving end. The intelligent agent determines the following based on the same decision rules (e.g., table lookup or model inference) as the sending end: number of operation steps K=3, transformation type combination [Gaussian noise addition, feature obfuscation, dynamic permutation], and specific parameters for each transformation. Note that these parameters are completely identical to those used by the sending end during encryption.

[0174] The decryption process is performed in reverse order of encryption (i.e., reversing the encryption process from the last step): inverse permutation, deobfuscation, and noise reduction.

[0175] If the number of operation steps K=2, only the inverse permutation and descrambling are performed (assuming that noise addition and obfuscation were performed during encryption), skipping the noise. If K=0, no decryption operation is required, and the received symbol is directly considered to be the channel symbol.

[0176] The receiving end intelligent agent can also dynamically adjust the decryption strength based on the current security risks: if it senses an increase in risk, it can proactively request the sending end to increase the number of encryption steps and adjust the decryption process accordingly.

[0177] By using a shared intelligent agent, the receiving end can determine the decryption strategy without receiving encryption parameters, avoiding the overhead of parameter transmission and the risk of eavesdropping. The decryption operations (inverse permutation, inverse matrix multiplication, and subtraction) all have linear complexity, are computationally fast, and meet real-time communication requirements. The denoising operation during decryption can be jointly processed with channel-induced distortion without amplifying noise. Based on the receiver's computing power constraints, the intelligent agent can choose to skip some decryption steps (for example, if the receiver's computing power is weak, it can require the sender to use only permutation encryption, and the receiver only needs to perform inverse permutation), ensuring system availability.

[0178] S520. Under the guidance of the intelligent agent, the channel environment information is perceived. Based on the perceived channel environment information, the decrypted channel symbols are jointly decoded by the source and channel to obtain semantic features.

[0179] The channel environment information includes signal-to-noise ratio (SNR), multipath fading, Doppler shift, and interference power. The receiver can acquire this information in real time through pilot signals and a channel estimation module. Due to channel reciprocity, the downlink channel environment information perceived by the receiver is highly consistent with the uplink channel information perceived by the transmitter.

[0180] Source-channel joint decoding is the reverse process of source-channel joint coding. It receives the channel symbol x (which may contain channel noise and distortion) after transmission and decryption over the wireless channel, and outputs an estimate of the original semantic features z. It is typically implemented using deep neural networks, and utilizes semantic structure information and channel error mode information to intelligently resist and compensate for channel impairments.

[0181] The semantic features are defined identically to those at the sending end; they are low-dimensional vectors representing the core meaning of the image. The goal of joint decoding at the receiving end is to recover a vector that is as consistent as possible with the semantic features extracted at the sending end, providing a foundation for subsequent semantic recovery.

[0182] In some possible implementations, the intelligent agent at the receiver senses channel environment information (such as signal-to-noise ratio and fading depth) in real time. Based on this information, the intelligent agent selects or configures the parameters or decoding strategy of the joint decoder. The joint decoder can be a deep neural network (which can share some layers with the semantic recovery network or be a separate network), whose input is the decrypted channel symbol sequence x, and whose output is semantic features. The intelligent agent can dynamically adjust the decoding method based on channel quality: for example, when the channel quality is poor, a more robust decoding algorithm is used (such as increasing the number of iterations or using soft decision); when the channel quality is good, fast decoding is used to reduce latency.

[0183] In some possible implementations, the receiver uses the same deep neural network (or symmetric network) as the transmitter for joint decoding. The intelligent agent determines the encoding parameters used by the transmitter based on the perceived channel environment information (the receiver can independently infer these parameters due to the shared intelligent agent), and then configures the corresponding parameters of the decoding network to match the decoding process with the encoding process.

[0184] For example, suppose the channel symbol sequence obtained after decryption is The receiver's intelligent agent obtains the current signal-to-noise ratio (SNR) of 12dB by measuring the pilot signal and detects that the channel is experiencing frequency-flat fading. The intelligent agent determines, based on a predefined strategy (consistent with the transmitter), that the transmitter uses joint coding based on a deep neural network, with the coding parameters being... (Corresponding to a medium channel quality configuration). The receiver's intelligent agent therefore selects the decoding parameters. (and Matching can mean having the same network weights or symmetrical weights.

[0185] Joint decoder network Receive x as input, output semantic feature estimate The network can be an inverse network that shares weights with the joint encoder at the transmitting end (e.g., if the encoder is a linear transform, the decoder can be a pseudo-inverse; if the encoder is a nonlinear neural network, the decoder can be the decoding part of a symmetric autoencoder). In practical implementations, an end-to-end trained autoencoder structure can be used: the encoder is at the transmitting end, the decoder is at the receiving end, and the two are jointly trained to minimize the reconstruction error of semantic features.

[0186] If the transmitter uses a non-neural network coding strategy, the receiver will use a matching decoding strategy accordingly. For example, if the transmitter selects strategy C based on SNR (low SNR scenario: compression ratio 1:1, code rate 1 / 3, BPSK modulation), the receiver's intelligent agent will select the corresponding decoding strategy C: first perform BPSK soft demodulation, then perform channel decoding at a code rate of 1 / 3, and finally reassemble the decoded bitstream into a semantic feature vector.

[0187] The decoding parameters are dynamically adjusted based on channel environment information, enabling high semantic feature recovery accuracy under various channel conditions. Regardless of whether it is based on deep neural networks or traditional coding strategies, the receiver can achieve the corresponding decoding under the guidance of the intelligent agent.

[0188] S530: Under the guidance of the intelligent agent, semantic features are restored to reconstruct the image.

[0189] Semantic restoration refers to the process of mapping low-dimensional semantic feature vectors back to high-dimensional image data. It is the inverse operation of semantic extraction. Unlike traditional decompression, semantic restoration does not pursue pixel-level precise reconstruction, but rather aims to preserve semantic information (such as object category, location, and scene structure) to generate visually plausible and semantically consistent images.

[0190] The reconstructed image is the final output image from the receiving end. It can be an image that is not exactly the same as the original image in terms of pixels but is semantically similar. For example, in autonomous driving scenarios, the reconstructed image should be able to clearly display key objects such as vehicles, pedestrians, and traffic signs, and should not affect decision-making even if some texture details are lost.

[0191] In some possible implementations, the intelligent agent at the receiving end will jointly decode the semantic features. The input is a semantic recovery network. This network is typically a deep neural network (e.g., a deconvolutional network, a generator of a generative adversarial network, or a diffusion model), and its output is a reconstructed image. The intelligent agent can select different semantic recovery networks or adjust recovery parameters based on current task requirements (such as target detection accuracy requirements and display resolution). In some possible implementations, the semantic recovery network shares the same deep neural network (i.e., the decoding part of the autoencoder) with the semantic extraction network at the sending end, thereby achieving end-to-end joint optimization.

[0192] Assume the semantic feature estimation of the joint decoding output is The receiving end intelligent agent invokes the pre-trained semantic recovery network. ,in, These are the network parameters. The network outputs the reconstructed image. The structure of the semantic recovery network is symmetrical to that of the semantic extraction network at the sending end: for example, if the extraction network uses convolutional and pooling layers, the recovery network uses transposed convolutional and upsampling layers; if the extraction network uses a Transformer encoder, the recovery network can use a Transformer decoder.

[0193] In a specific example, the semantic recovery network is a deep deconvolutional neural network: Input: semantic feature vector First, the image is mapped to a feature map using a fully connected layer. Then, it sequentially passes through multiple deconvolution blocks (transposed convolution + batch normalization + ReLU) to progressively upsample the spatial resolution. Finally, the reconstructed image is output. .

[0194] The intelligent agent can dynamically adjust the resolution of the output image based on the performance of the receiving display device or the application requirements. For example, if the receiving end is a mobile phone screen, a 224×224 image is reconstructed; if it is a high-definition television, it can request higher-dimensional semantic features from the sending end and then restore it to a 1920×1080 image. The intelligent agent can also decide whether to enable post-processing enhancements (such as super-resolution, deblocking filtering, etc.) based on channel conditions.

[0195] Semantic recovery networks are specifically trained on semantic information, prioritizing the reconstruction of task-critical content while ignoring less important details, thus maintaining semantic intelligibility even with extremely low bandwidth. Intelligent agents can dynamically adjust output resolution as needed, adapting to different display devices and application scenarios.

[0196] In the semantic communication method applied to the receiver of a semantic communication system provided in this disclosure embodiment, the receiver's intelligent agent automatically matches the encryption strength of the sender based on perceived security risks and system constraints, without requiring additional signaling. It dynamically adjusts the decoding strategy using channel environment information, effectively resisting channel noise and fading, and ensuring the accuracy of semantic feature recovery.

[0197] The semantic communication method for the receiver of a semantic communication system provided in this embodiment will be described in detail below.

[0198] In some possible implementations, the intelligent agent at the receiving end and the intelligent agent at the sending end are the same intelligent agent. The intelligent agent at the receiving end determines the combination of transformation types and the number of operation steps adopted by the sending end based on the security risks and system performance constraints it perceives.

[0199] Here, "same intelligent agent" refers to a logical connection where the sending and receiving ends share a unified artificial intelligence decision-making entity. In physical implementation, this can be represented as: A centralized intelligent agent is deployed on a cloud or edge server, simultaneously issuing decision instructions to both the sender and receiver. The sending and receiving ends each deploy a smart agent instance. The two instances are initially in the same state and maintain parameter consistency through low-frequency synchronization mechanisms (such as sharing a random seed and periodically exchanging state digests). The intelligent agent is fully deployed on one end (such as the sending end), but the decision rules and necessary states are pre-configured in a secure manner on the other end.

[0200] Regardless of the implementation, the core feature is that the perceived information (security risks, system performance constraints) and decision-making logic used by both ends are consistent. Therefore, both ends can independently arrive at the same encryption strategy decision without exchanging strategy parameters through the communication channel before each encryption.

[0201] The security risks and system performance constraints perceived by the receiver are independently measured or assessed by the receiver's intelligent agent. Security risks include: abnormal signals detected by the receiver (such as unauthorized transmission sources, sudden increases in interference power), abnormal bit error rate, and geographical location risks. System performance constraints include: the receiver's currently available computing resources (CPU / GPU utilization, memory availability), power budget, and maximum allowable decryption latency.

[0202] Since the sender and receiver operate within the same communication environment, the security risks and system performance constraints perceived by both are highly correlated (for example, if an eavesdropper is present in the channel, both ends can detect the anomaly; if the system requires low latency, both ends must adhere to the same latency budget). Therefore, the results perceived independently by the receiver are sufficient to infer the encryption strategy employed by the sender.

[0203] The receiving intelligent agent, based on perceived information, runs the exact same decision-making algorithm (such as a rule table, decision tree, or neural network model) as the sending end to calculate the encryption strategy. This strategy includes a combination of transformation types (an ordered sequence specifying the transformation type at each step) and the number of operation steps, K. Since the decision-making algorithm and input information are identical, the result calculated by the receiving end will inevitably be consistent with the strategy chosen by the sending end. Therefore, the receiving end can obtain the encryption parameters without receiving any additional signaling and perform the decryption operation accordingly.

[0204] In some possible implementations, the sending and receiving ends have established a shared intelligent agent mechanism before or during communication (e.g., both ends obtain the same intelligent agent model and initial state from the same authentication center). When the receiving end receives the encrypted channel symbols, it does not directly rely on the parameters transmitted by the sending end through the data channel, but instead activates its local intelligent agent. The intelligent agent first performs environmental awareness: measuring the security risks in the current channel (e.g., detecting eavesdroppers through spectrum sensing, analyzing bit error rate fluctuations) and system performance constraints (e.g., reading the current CPU load, calculating the available decryption time budget). Then, the intelligent agent inputs this awareness information into a predefined encryption strategy decision function. This decision function is completely consistent with the decision function used by the sending end (e.g., the same threshold table, the same neural network weights). The output of the decision function is the combination of transformation types and the number of operation steps. The receiving end then uses these parameters to perform the inverse decryption transformation.

[0205] By sharing a smart agent, encryption policy parameters do not need to be transmitted over the wireless channel, completely saving this bandwidth resource. This is particularly effective in scenarios involving large volumes of short data packets (such as the Internet of Things). Because the encryption parameters are not transmitted over the channel, attackers cannot obtain any information about the encryption policy through eavesdropping; they can only attempt brute-force attacks on the ciphertext, significantly improving security. The receiving end can begin decryption immediately without waiting for the sending end to send parameters, reducing the impact of round-trip time (RTT) and benefiting real-time services. The receiving end can dynamically adjust the decryption strength based on its perceived security risks and computing power constraints. Even if the sending end cannot update the policy in time for some reason, the receiving end can adapt autonomously. Even if the data channel is completely interrupted, as long as the control channel (or the synchronization link of the shared smart agent) is normal, the receiving end can still correctly decrypt the received data, enhancing the system's resilience.

[0206] Figure 6 A flowchart illustrating one implementation of decrypting encrypted channel symbols received from a wireless channel is shown, such as... Figure 6 As shown, steps S610 and S620 may be included.

[0207] S610. Obtain the transformation type combination and operation steps used by the sending end.

[0208] The transformation type combination is the same as that used during encryption at the sending end. It is an ordered sequence used to specify the inverse operation type corresponding to each step in the predefined decryption process. For example, if the transformation type combination used during encryption at the sending end is [Gaussian noise, feature obfuscation, dynamic permutation], then the corresponding inverse operation combination for decryption at the receiving end is [denoising, deobfuscation, inverse permutation]. The length of the transformation type combination is the number of operation steps K. The receiving end obtains this combination through a shared intelligent agent, for example.

[0209] The number of operation steps is K, which is the number of encryption transformation steps actually performed by the sending end. The receiving end needs to perform the same number of inverse operation steps during decryption. The number of operation steps can be obtained directly from the length of the transformation type combination, or it can be obtained separately. During decryption, the receiving end performs the inverse operation of step K, the inverse operation of step K-1, ..., the inverse operation of step 1 in the reverse order of encryption (i.e., complete reverse execution).

[0210] In some possible implementations, the receiving end must know the transformation type combination and operation steps used by the sending end during encryption before starting decryption. As mentioned earlier, this can be achieved in several ways: the receiving end's intelligent agent and the sending end's intelligent agent are the same agent (shared state), and the receiving end independently determines the agent based on its perceived security risks and system performance constraints; or it can be generated through a pre-agreed key derivation function combined with channel measurement results; or a set of default strategies can be pre-configured during system initialization. The embodiments of this disclosure preferably adopt the shared intelligent agent method, that is, the receiving end's intelligent agent is determined independently without needing to receive parameters through the communication channel.

[0211] In some specific implementations, the receiving end's intelligent agent perceives the current security risk level and system performance constraints, runs the same decision model as the sending end, and obtains the transformation type combination [Gaussian noise, feature obfuscation, dynamic permutation] and the number of operation steps K=3. This information is stored in the register of the decryption module for use in subsequent steps.

[0212] By acquiring parameters through self-synchronization, signaling overhead is avoided, security is enhanced, and latency is reduced.

[0213] S620. Following a predefined decryption process that is the reverse of the encryption process, the decryption transformation steps corresponding to the number of operation steps are executed sequentially according to the combination of transformation types, wherein each decryption transformation step adopts the inverse operation of the transformation type at the corresponding position in the combination of transformation types.

[0214] The predefined decryption process is an inverse transformation process corresponding to the predefined encryption process at the sending end. The predefined decryption process defines the maximum number of steps L and the types of inverse operations that can be performed at each step. During actual execution, the receiving end executes the corresponding inverse operations sequentially in the first K steps of the process based on the acquired combination of transformation types. The decryption process is the reverse mapping of the encryption process, ensuring that the decrypted data can be restored to the channel symbols before encryption.

[0215] In this context, the inverse operation refers to a mathematical or logical operation that is the opposite of the encryption transformation operation. In this embodiment, the inverse operation includes one or more of denoising, deobfuscation, and inverse permutation.

[0216] Denoising is the inverse operation of Gaussian noise addition. If Gaussian noise n is added during encryption, then the noise sequence needs to be subtracted during decryption (i.e., ... Denoising requires the receiver to generate the same noise sequence as the transmitter (e.g., by sharing a random seed). If the noise sequence cannot be perfectly synchronized, statistical denoising methods (such as Wiener filtering) can be used, but synchronous noise subtraction is preferred to ensure lossless recovery.

[0217] Deobfuscation is the inverse operation of feature obfuscation. If linear obfuscation was performed during encryption... Therefore, decryption requires multiplying by the inverse of the confusion matrix. ,Right now The confusion matrix M is usually invertible, and the inverse matrix can be pre-computed or solved using a fast algorithm.

[0218] Inverse permutation is the inverse operation of dynamic permutation. If a permutation is applied to the sequence during encryption... ,Right now Decryption requires the application of inverse permutation. ,Right now Inverse permutations can be achieved by rearranging the indices, with extremely low computational complexity.

[0219] The encrypted channel symbol sequence received by the receiver from the wireless channel. It may contain noise and distortion introduced by channel transmission, as well as transformations added by encryption operations. The decryption process aims to eliminate the effects of encryption transformations and recover the original joint encoder output symbol x.

[0220] In some possible implementations, the receiving end decryption module reverses the encryption process step by step, starting from the last step. Specifically, if the encryption process sequentially executes step 1, step 2, ..., step K, then the decryption process sequentially executes the inverse operation of step K, step K-1, ..., step 1. The type of the inverse operation at each step is determined by the transformation type at the corresponding position (mapped from back to front) in the transformation type combination. For example, if the transformation type combination is [T1, T2, T3] (K=3), then the decryption order is: first execute the inverse operation of T3, then the inverse operation of T2, and finally the inverse operation of T1. This reverse execution ensures the complete recovery of the data.

[0221] In some possible implementations, if the transformation type combination is [Gaussian noise addition], then the decryption inverse operation is [denoising]. If the combination is [feature obfuscation], then the inverse operation is [deobfuscation]. If the combination is [dynamic permutation], then the inverse operation is [inverse permutation]. If the combination is [Gaussian noise addition, feature obfuscation], then the inverse operation order is [deobfuscation, denoising] (note the reverse order). If the combination is [feature obfuscation, dynamic permutation, Gaussian noise addition], then the inverse operation order is [denoising, inverse permutation, deobfuscation]. In the specific code implementation, an inverse operation function table can be defined, calling the corresponding functions sequentially according to the combination identifier.

[0222] Specifically, suppose the received ciphertext sequence is The obtained transformation type combination is [Gaussian noise, feature obfuscation, dynamic permutation], K=3. The predefined maximum number of steps in the decryption process is L=3.

[0223] Step 1 (the inverse of step 3 in encryption): Inverse permutation Obtaining the permutation used during encryption (Generated by the smart agent based on the shared seed, consistent with the sender).

[0224] For ciphertext Applying inverse permutation , to obtain the sequence ,satisfy .

[0225] For example, if m=4, Then the inverse permutation (That is, the original 3rd element is now in the 1st position). The result after calculation is... .

[0226] Step 2 (the inverse of step 2 of encryption): Deobfuscation Obtain the confusion matrix M used for encryption (generated by the smart agent based on security risks, and consistent with the sender). M is an invertible matrix, such as a tridiagonal matrix. .

[0227] calculate You can directly use the pre-computed inverse matrix, or solve the system of linear equations (because M is sparse, the chasing method can be used to solve it quickly).

[0228] For example, when At that time, it can be calculated in advance. The expression or numerical matrix is ​​then used to perform matrix multiplication.

[0229] Step 3 (the inverse of step 1 in encryption): Denoising Obtain the Gaussian noise sequence n added during encryption (generated by the smart agent using the same seed, consistent with the sender). Noise variance. Known.

[0230] Calculate the original channel symbols .

[0231] Since the noise is pseudo-randomly generated, the receiver can generate the exact same noise sequence using the same pseudo-random number generator (PRNG) and the same seed, thus achieving precise subtraction.

[0232] After the above three steps, the channel symbol x before encryption is successfully recovered. If K=2, only the inverse permutation and deobfuscation are performed (assuming that encryption only performed obfuscation and permutation), skipping the denoising step. If K=1, only the corresponding single inverse operation is performed (e.g., only the inverse permutation).

[0233] To handle additional noise (unencrypted noise) introduced during channel transmission, a lightweight filtering operation (such as a minimum mean square error filter) can be added after the denoising step, but this is not a necessary feature of the present invention.

[0234] Since the decryption process is the exact inverse operation of the encryption process, under ideal channel conditions (no transmission distortion), the original channel symbols can be completely recovered, achieving lossless decryption. The inverse operation corresponds one-to-one with the encryption operation, facilitating hardware implementation and pipelined processing. Because the decryption process is strictly reversed, even minor errors in some steps (such as noise generator offsets) will not lead to catastrophic errors and can be corrected through subsequent channel decoding.

[0235] Through the above implementation, as long as the encryption parameters are known, decryption can accurately recover the original channel symbols without affecting the subsequent source channel decoding performance. The computational complexity of the inverse operation is the same as that of the encryption operation, both being linear, and it does not become a system bottleneck. It supports inverse operations with arbitrary combinations and arbitrary numbers of steps, adapting to dynamically adjusted encryption strength at the sending end. The decryption process does not amplify channel noise (the denoising step only subtracts known noise) and can be jointly optimized with channel decoding. Even if an attacker knows the decryption algorithm, they still cannot decrypt the ciphertext because they do not know the transformation type combination and the number of operation steps (not transmitted through the channel).

[0236] Figure 7 A flowchart illustrating one implementation of joint source-channel decoding of decrypted channel symbols is shown, such as... Figure 7 As shown, steps S710 and S720 may be included.

[0237] S710 The intelligent agent at the receiving end obtains the encoding parameters of the deep neural network used by the transmitting end, and inputs the decrypted channel symbols into the deep neural network. Based on the decoding parameters corresponding to the encoding parameters, the channel symbols are decoded into semantic features. The encoding parameters are determined by the intelligent agent at the receiving end based on the perceived channel environment information.

[0238] Channel environment information refers to the channel state perceived by the receiver, including signal-to-noise ratio, multipath fading, and interference level. Due to channel reciprocity (especially in TDD systems), the downlink channel information perceived by the receiver is highly consistent with the uplink channel information perceived by the transmitter. The receiver's intelligent agent uses this information to determine the coding parameters used by the transmitter, thereby ensuring that the decoding parameters match the coding parameters.

[0239] The deep neural network uses the same or symmetrical network structure as the deep neural network used at the transmitting end. In this embodiment, the network performs three roles simultaneously: at the transmitting end, it is used for semantic extraction and joint source-channel coding; at the receiving end, it is used for joint source-channel decoding and semantic recovery. The network can be designed as an autoencoder structure, where the encoder part is at the transmitting end and the decoder part is at the receiving end. The decoder network at the receiving end can share some parameters with the encoder network at the transmitting end (e.g., if the channel is symmetrical, the same weights can be used; or a separate but structurally symmetrical decoder network can be used). Preferably, the transmitting and receiving ends use two copies of the same neural network model, maintaining the same network architecture and weight parameters (obtained through joint training).

[0240] Encoding parameters are the network parameters (such as weight matrices, bias vectors, batch normalization parameters, etc.) used by the transmitting deep neural network to map semantic features to channel symbols. The encoding parameters determine the mapping function from input semantic features to output channel symbols. The transmitting intelligent agent selects a set of encoding parameters based on perceived channel environment information (such as signal-to-noise ratio, fading) (e.g., selecting one from multiple pre-trained parameter sets, or generating them in real-time through a hypernetwork). Since the receiving end shares the same intelligent agent with the transmitting end, it can independently determine the encoding parameters used by the transmitting end based on the same perceived channel environment information. The receiving end does not need to receive these parameters; instead, it calculates or finds them itself.

[0241] Decoding parameters are network parameters corresponding to encoding parameters, used to map received channel symbols back to semantic features. Ideally, if the channel is ideal and the network is invertible, the decoding parameters can be the inverse of the encoding parameters (e.g., for a linear layer, the decoding parameters can be a pseudo-inverse of the encoding parameters). In practical deep neural networks, decoding parameters are typically a different set of parameters obtained through end-to-end training, but with a symmetrical relationship between them. For example, in an autoencoder, the weights of the decoder can be the transpose of the encoder weights (weight binding). In embodiments of this disclosure, the receiving-end intelligent agent can derive the corresponding decoding parameters based on the determined encoding parameters (e.g., selected from a pre-trained set of paired parameters, or generated using the same rules).

[0242] The obtained semantic features are defined identically to those at the sending end, and are low-dimensional vectors representing the core meaning of the image. The goal of joint decoding at the receiving end is to recover a vector that is as consistent as possible with the semantic features extracted by the sending end. The semantic features are then fed into a semantic recovery network to reconstruct the image.

[0243] In some possible implementations, the receiver's intelligent agent first perceives the current channel environment information (e.g., obtaining a signal-to-noise ratio estimate by measuring pilot signals, or obtaining fading depth by analyzing received signals). Since it shares the same intelligent agent with the transmitter (or maintains state synchronization), the receiver can run the same coding parameter selection algorithm as the transmitter, thereby determining the deep neural network coding parameters actually used by the transmitter during coding.

[0244] For example, the sending-end intelligent agent selects the encoding parameter set based on SNR=12dB. The receiving end's intelligent agent also sensed an SNR of approximately 12dB, and therefore determined the encoding parameters to be... Next, the receiving end retrieves the decoding parameters corresponding to the encoding parameters from local storage. If the network uses weight binding (e.g., the decoder weights are transposes of the encoder weights), the decoding parameters can be directly calculated from the encoding parameters; otherwise, a mapping table between the encoding and decoding parameters needs to be pre-stored. Then, the receiver inputs the decrypted channel symbol sequence xx into the deep neural network, which uses the decoding parameters... Perform forward propagation and output semantic feature estimates. Because the decoding parameters match the encoding parameters at the sending end, and the network is jointly trained end-to-end, this decoding process can effectively resist channel noise and distortion, recovering high-quality semantic features.

[0245] Specifically, assuming the receiver and transmitter use the same deep neural network model, this model has an encoder part (used by the transmitter) and a decoder part (used by the receiver). The entire network structure is an autoencoder: encoder Input image I, output semantic features z and channel symbols x (in fact, semantic features can be intermediate representations, and channel symbols are the final output; for simplification, the encoder can be considered to directly output channel symbols).

[0246] decoder Input channel symbol x, output reconstructed image In the joint decoding step, we only need to obtain the intermediate semantic features. This feature can be the output of a certain layer of the decoder.

[0247] Specifically, the workflow of a deep neural network G at the receiver is as follows: the input channel symbol x is processed through several layers (joint decoding part) to obtain semantic features. The reconstructed image is then obtained after several layers (semantic restoration). The entire G is jointly trained with the transmitting end's coding network F, enabling G to recover the image from noisy channel symbols.

[0248] The receiver, through a shared intelligent agent and awareness of the channel environment, can obtain the encoding parameters without receiving additional signaling, thus enabling the use of correct decoding parameters. Joint training of the decoding and encoding parameters allows the decoder to adapt to channel distortion, resulting in semantic feature recovery quality superior to traditional separate decoding. Using the same deep neural network for decoding avoids switching between multiple models, saving storage and computational resources. Furthermore, because the decoding parameters are selected based on the channel environment, the system maintains stable semantic feature recovery accuracy under different channel conditions.

[0249] S720. Input semantic features into a deep neural network, and output a reconstructed image from the deep neural network.

[0250] In some possible implementations, in obtaining semantic features Subsequently, the receiving end continues to use the later part (semantic recovery part) of the same deep neural network G to map semantic features to the reconstructed image. This part of the network typically consists of deconvolutional layers, upsampling layers, or Transformer decoding layers. Its input is low-dimensional semantic features, and its output is high-dimensional image data.

[0251] Since the joint decoding part belongs to the same network, the entire forward propagation at the receiver can be completed in one go: the channel symbols are input into the network, semantic features are obtained after several layers, and the image is directly output after several more layers. Alternatively, the same network can be called in two steps (first obtaining intermediate features, then continuing forward propagation). The advantage of using the same network is that joint decoding and semantic recovery are jointly optimized during training, which avoids error accumulation caused by separate training and reduces the complexity of model deployment.

[0252] In some possible implementations, in step S710, we have already obtained As the output of the network's intermediate layer. Now, Continue inputting into subsequent layers of the network (semantic recovery part).

[0253] The input to the entire network G is the channel symbol x, and the output is... During training, we use the original image I as supervision to minimize reconstruction losses (such as MSE and perceptual loss). Therefore, the network automatically learns to directly recover images from noisy channel symbols, with its intermediate layers... Nature becomes a representation of semantic features.

[0254] In actual operation, the receiving end only needs to perform one network forward propagation: If you want to obtain semantic features separately, you can set the output interface in the middle layer of the network.

[0255] In a practical implementation, the first half of the network can be defined to output semantic features, and the second half can receive these features and output an image. When called, the first half is executed first, and then the second half is executed. However, the two share network parameters and structure definitions and are still regarded as the same network.

[0256] Through the above implementation, joint decoding and semantic recovery are jointly optimized in the same network, eliminating error propagation from intermediate representations (semantic features) and improving reconstruction quality. A single forward propagation is sufficient to obtain the reconstructed image, eliminating the need to serially call two independent models. Compared to using two independent networks, the shared network reduces the number of parameters and memory usage, making it beneficial for deployment on resource-constrained devices. Since the network aims to reconstruct the image as its final goal, semantic features adaptively encode information important for reconstruction, prioritizing the recovery of this information during channel decoding. This ensures that the reconstructed image maintains semantic intelligibility even at low signal-to-noise ratios.

[0257] Figure 8 A flowchart illustrating one implementation of joint source-channel decoding of decrypted channel symbols is shown, such as... Figure 8 As shown, steps S810 and S820 may be included.

[0258] In the S810, the transmitting and receiving ends share the same intelligent agent. The intelligent agent at the receiving end adjusts the initial decoding strategy corresponding to the initial coding strategy based on the channel environment information it perceives, thus obtaining the adjusted decoding strategy.

[0259] The channel environment information is the same as that defined at the transmitting end, including signal-to-noise ratio (SNR), multipath fading parameters, and interference power. The receiving end acquires this information in real time through channel estimation, pilot measurement, and other methods. In semantic communication systems, uplink and downlink channels are reciprocal; the downlink channel information perceived by the receiving end can reflect the uplink channel state, thereby inferring the environment perceived by the transmitting end.

[0260] The initial decoding strategy refers to the decoding strategy currently being used by the receiver before changes occur in the channel environment. Decoding strategy is a broad concept, encompassing all the rules and parameters required to convert received channel symbols into semantic features. Specifically, it may include: Demodulation methods (such as soft demodulation or hard decision strategies for BPSK, QPSK, and 16QAM); Channel decoding schemes (such as Viterbi decoding of convolutional codes, iterative decoding parameters, code rate, and interleaving methods for LDPC codes); Source decompression methods (such as inverse quantization, inverse transform, and semantic feature reconstruction algorithms); Equilibrium and disturbance suppression strategies (such as zero-forcing equilibrium and minimum mean square error equilibrium).

[0261] The initial decoding strategy can be the system default strategy or a strategy retained from the last adjustment. The receiving end's intelligent proxy internally maintains the state of the current decoding strategy.

[0262] The adjusted decoding strategy is a new strategy obtained by the receiver's intelligent agent modifying one or more parameters of the initial decoding strategy based on the perceived channel environment information. The goal of the adjustment is to make the new decoding strategy match the transmitter's adjusted coding strategy, thereby correctly recovering semantic features. For example, when the transmitter increases channel coding redundancy or switches to lower-order modulation due to channel degradation, the receiver should adjust its decoding strategy accordingly, adopting a corresponding lower-order demodulation and a decoding algorithm with higher error correction capabilities. Since the transmitter and receiver share the same intelligent agent, the channel environment information perceived by the receiver is highly consistent with that of the transmitter, thus enabling the receiver to make independent adjustment decisions synchronized with the transmitter.

[0263] Decoding strategy adjustment refers to a decision-making process executed by the receiver's intelligent agent. The inputs are the currently perceived channel environment information and the initial decoding strategy, and the output is the adjusted decoding strategy. Adjustment can employ a rule engine (e.g., a pre-defined mapping table between a signal-to-noise ratio threshold and the decoding strategy), a reinforcement learning model, or the same optimization algorithm as the transmitter. Because they share the same intelligent agent, the rules for adjusting the receiver's decoding strategy are completely consistent with the rules for adjusting the transmitter's encoding strategy, ensuring strategy synchronization between the two ends.

[0264] In some possible implementations, the receiver and transmitter share the same intelligent agent (logically identical, but physically it can be a centralized or distributed synchronization instance). The receiver's intelligent agent senses channel environment information in real time (such as signal-to-noise ratio, fading depth, and interference level). Based on this information, the intelligent agent decides whether the current decoding strategy needs to be adjusted. The basis for adjustment is exactly the same as that for the transmitter's coding strategy adjustment (e.g., the same SNR threshold, the same strategy mapping table). The receiver's intelligent agent obtains the currently used initial decoding strategy (which corresponds to the initial coding strategy currently used by the transmitter) and then calculates the new decoding strategy parameters based on the channel environment information. Since the transmitter also adjusts its coding strategy based on the same sensed information, the adjusted decoding strategy naturally matches the transmitter's adjusted coding strategy. The receiver does not need to receive any signaling regarding coding strategy adjustments through the communication channel; synchronization is achieved entirely through the shared intelligent agent and independent sensing.

[0265] The following example of decoding strategy adjustment uses signal-to-noise ratio (SNR) as the primary channel environment information. Assume that the transmitter and receiver have pre-agreed a mapping rule between encoding and decoding strategies, as shown in the table below: Assume the initial decoding strategy is strategy B' (corresponding to medium channel quality). The receiver's intelligent agent measures the SNR in real time. When the SNR drops to 8dB, the intelligent agent detects that the threshold condition of SNR < 10dB has been triggered, and therefore decides to adjust the decoding strategy to strategy C'. The adjustment process includes: Demodulation method: Switch from QPSK software demodulation to BPSK software demodulation; Channel decoding: Switch from Viterbi decoding with a code rate of 1 / 2 to Viterbi decoding with a code rate of 1 / 3; Source decompression: Switch from a decompression ratio of 2:1 (that is, recovering 2 channel symbols into 1 semantic feature component) to a decompression ratio of 1:1 (that is, each channel symbol directly corresponds to a semantic feature component).

[0266] The adjusted decoding strategy (strategy C') is stored in the smart agent's strategy register and sent to the decoder for execution.

[0267] If the channel environment gradually improves and the SNR rises back to 15 dB, the receiver's intelligent agent can adjust back to strategy B', or even further to strategy A'. The adjustment can be discrete (table lookup switching) or continuous (e.g., linearly adjusting the decoder's iteration count or soft-decision quantization precision based on the SNR). Since the transmitter and receiver use the same threshold and strategy tables, the transmitter will also synchronously adjust its encoding strategy to strategy C', thus ensuring matching between encoding and decoding.

[0268] For asymmetric channels (such as FDD systems), uplink and downlink channel conditions may differ. In this case, the downlink channel information perceived by the receiver cannot be directly used to infer the uplink coding strategy. However, since a common intelligent agent is shared, this agent can fuse uplink and downlink measurement reports (e.g., the transmitter reports uplink channel quality, and the intelligent agent synchronizes this information to the receiver), enabling the receiver to understand the channel environment perceived by the transmitter. In this case, "based on its own perceived channel environment information" can be extended to based on comprehensive information provided by the intelligent agent (including the receiver's own measurements and the transmitter's measurements obtained through the control channel).

[0269] S820. According to the adjusted decoding strategy, the channel symbols are decoded into semantic features.

[0270] In some possible implementations, after obtaining the adjusted decoding strategy, the receiver's decoder converts the channel symbol sequence into a semantic feature vector according to the rules of the strategy. The decoding process is the inverse operation of the transmitter's encoding process, specifically including: demodulation (mapping channel symbols to soft bits or hard decision bits), channel decoding (correcting transmission errors), and source decompression (restoring the decoded bitstream to a semantic feature vector). The decoder receives the decrypted channel symbol x as input and, based on the parameters specified in the adjusted decoding strategy (such as demodulation method, decoding rate, and decompression ratio), performs the corresponding inverse transformation, outputting a semantic feature estimate. .

[0271] Continuing with the example above, assume the adjusted decoding strategy is strategy C' (low SNR scenario). The decoder performs the following operations: Demodulation: BPSK soft demodulation is used. For each received channel symbol... (A real number, affected by noise), calculate its log-likelihood ratio (LLR) or directly output the soft bit value. For example, for BPSK, ,in, This is for noise variance estimation.

[0272] Channel decoding: A Viterbi decoder with a code rate of 1 / 3 is used. The input is a soft bit sequence, and the output is a decoded bit stream. Since the semantic feature components are converted into a bit stream of 3 times the length during encoding (through convolutional coding), the decoded bit stream recovers the same length as the original semantic feature bit representation.

[0273] Source decompression: The decoded bitstream is converted into a floating-point vector according to a predetermined quantization scheme (such as uniform quantization) to obtain semantic feature estimates. Since the compression ratio is 1:1, each channel symbol (after demodulation and decoding) corresponds to one semantic feature component. The dimension dd is equal to the total number of bits in the decoded bitstream divided by the number of quantization bits for each component.

[0274] In terms of implementation, the decoder can be a configurable software module that calls different subroutines (such as different demodulators and decoders) based on the decoding strategy identifier. The intelligent agent passes the adjusted decoding strategy to the decoder as parameters, and the decoder dynamically loads the corresponding algorithm configuration.

[0275] The receiver and transmitter share the same intelligent agent, automatically matching encoding and decoding strategies based on the same channel awareness and adjustment rules, without the need for signaling interaction. The decoding strategy is dynamically adjusted according to the channel environment, prioritizing the correct recovery of semantic features in poor channel conditions and maximizing decoding efficiency in good channel conditions. It does not rely on specific deep neural networks and can be implemented using software or hardware modules based on existing communication protocol stacks, making it easy to deploy and upgrade. Decoding strategy adjustment mainly involves parameter configuration switching and classic algorithm calls, resulting in low computational overhead and suitability for real-time communication and low-power devices.

[0276] Based on and Figure 1 The method shown follows the same principle. Figure 9 This diagram illustrates the structure of a transmitter in a semantic communication system according to an embodiment of this disclosure. As described above, the semantic communication system includes a transmitter, a receiver, and a communication channel connecting the transmitter and the receiver. Figure 9 As shown, the transmitter 90 of the semantic communication system may include: The semantic extraction module 910 is used to extract semantic features from image sources under the guidance of an intelligent agent. The channel coding module 920 is used to sense channel environment information under the guidance of the intelligent agent, and to perform source-channel joint coding on semantic features based on the sensed channel environment information to generate channel symbols. The encrypted transmission module 930 is used to perceive security risks and system performance constraints under the guidance of the intelligent agent, encrypt the channel symbols according to the perceived security risks and system performance constraints, generate encrypted channel symbols, and send them into the wireless channel.

[0277] In the transmitting end of the semantic communication system provided in this disclosure, under the unified guidance of the intelligent agent, the three stages of semantic extraction, source-channel joint coding, and encryption are coordinated and adaptive. Semantic extraction significantly compresses the data volume, and joint coding further optimizes the compression-to-protection ratio. Joint coding dynamically adjusts redundancy according to channel conditions, effectively combating noise and fading. The encryption mechanism with adjustable strength prevents eavesdropping and is compatible with channel coding. The intelligent agent perceives environmental changes in real time, balancing transmission efficiency, reliability, and security without manual intervention.

[0278] In some possible implementations, the encrypted transmission module is specifically used to: select a combination of transformation types and a number of operation steps based on the security risks and system performance constraints perceived by the intelligent agent, wherein the combination of transformation types is an ordered sequence used to specify the transformation type used in each step of the predefined encryption process; take the channel symbol as the initial state, and execute the encryption transformation steps corresponding to the number of operation steps in sequence according to the predefined encryption process and the combination of transformation types, wherein each encryption transformation step adopts the transformation type at the corresponding position in the combination of transformation types.

[0279] In some possible implementations, the transformation type includes one or more of Gaussian noise addition, feature obfuscation, and dynamic permutation.

[0280] In some possible implementations, the channel coding module is specifically used to: select coding parameters for the deep neural network based on the channel environment information perceived by the intelligent agent; input semantic features into the deep neural network; and map the semantic features into channel symbols according to the coding parameters.

[0281] In some possible implementations, the semantic extraction module is specifically used to: input image information sources into a deep neural network, and output a low-dimensional semantic feature vector from the deep neural network.

[0282] In some possible implementations, the channel coding module is specifically used to: adjust the initial coding strategy based on the channel environment information perceived by the intelligent agent to obtain the adjusted coding strategy; and map semantic features to channel symbols based on the adjusted coding strategy.

[0283] In some possible implementations, the sender and receiver share the same smart agent. The smart agent senses channel environment information and determines the encryption strategy, which is used for both encryption at the sender and decryption at the receiver.

[0284] It is understood that the above-mentioned modules of the sending end of the semantic communication system in the embodiments of this disclosure have the ability to implement... Figure 1 The embodiments shown illustrate the functionality of corresponding steps in the semantic communication method applied to the sending end of the semantic communication system. This functionality can be implemented in hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above-described functions. These modules can be software and / or hardware, and each module can be implemented individually or integrated from multiple modules. For a detailed description of the functions of each module at the sending end of the semantic communication system, please refer to [link to relevant documentation]. Figure 1 The corresponding description of the semantic communication method applied to the sending end of the semantic communication system in the embodiments shown will not be repeated here.

[0285] Based on and Figure 5 The method shown follows the same principle. Figure 10 This diagram illustrates the structure of a receiver in a semantic communication system according to an embodiment of the present disclosure. As described above, the semantic communication system includes a transmitter, a receiver, and a communication channel connecting the transmitter and the receiver. Figure 10 As shown, the receiver 1000 of the semantic communication system may include: The decryption transmission module 1010 is used to perceive security risks and system performance constraints under the guidance of the intelligent agent, and to decrypt the encrypted channel symbols received from the wireless channel and recover the channel symbols based on the perceived security risks and system performance constraints. The channel decoding module 1020 is used to sense channel environment information under the guidance of the intelligent agent, and to perform source-channel joint decoding on the decrypted channel symbols based on the sensed channel environment information to obtain semantic features. The semantic recovery module 1030 is used to perform semantic recovery on the semantic features and reconstruct the image under the guidance of the intelligent agent.

[0286] In the receiver of the semantic communication system provided in this embodiment, the receiver's intelligent agent automatically matches the encryption strength of the sender based on perceived security risks and system constraints, without requiring additional signaling. It dynamically adjusts the decoding strategy using channel environment information, effectively resisting channel noise and fading, and ensuring the accuracy of semantic feature recovery.

[0287] In some possible implementations, the decryption transmission module is used to: obtain the transformation type combination and operation step number used by the sender; and, according to a predefined decryption process that is the reverse of the encryption process, sequentially execute the decryption transformation steps corresponding to the operation step number based on the transformation type combination, wherein each decryption transformation step adopts the inverse operation of the transformation type at the corresponding position in the transformation type combination.

[0288] In some possible implementations, the inverse operation includes one or more of denoising, deobfuscation, and inverse permutation.

[0289] In some possible implementations, the decryption transmission module is used to: have the same smart agent at the receiving end and the smart agent at the sending end, and determine the combination of transformation types and the number of operation steps adopted by the sending end based on the security risks and system performance constraints it perceives.

[0290] In some possible implementations, the channel decoding module is used for: the intelligent agent at the receiving end to obtain the encoding parameters of the deep neural network used by the sending end, and input the decrypted channel symbols into the deep neural network, and decode the channel symbols into semantic features according to the decoding parameters corresponding to the encoding parameters, wherein the encoding parameters are determined by the intelligent agent at the receiving end based on the perceived channel environment information.

[0291] In some possible implementations, the semantic recovery module is used to: input semantic features into a deep neural network, and output a reconstructed image from the deep neural network.

[0292] In some possible implementations, the channel decoding module is used for: the transmitter and receiver sharing the same intelligent agent; the receiver's intelligent agent adjusting the initial decoding strategy corresponding to the initial coding strategy based on the channel environment information it perceives, to obtain the adjusted decoding strategy; and decoding the channel symbols into semantic features according to the adjusted decoding strategy.

[0293] It is understood that the above-mentioned modules of the receiving end of the semantic communication system in the embodiments of this disclosure have the ability to implement... Figure 5 The embodiments shown illustrate the functionality of corresponding steps in the semantic communication method applied to the receiving end of the semantic communication system. This functionality can be implemented in hardware or by hardware executing corresponding software. The hardware or software includes one or more modules corresponding to the above-described functions. These modules can be software and / or hardware, and each module can be implemented individually or integrated from multiple modules. For a detailed description of the functions of each module at the receiving end of the semantic communication system, please refer to [link to relevant documentation]. Figure 5 The corresponding description of the semantic communication method applied to the receiving end of the semantic communication system in the embodiments shown in the illustration will not be repeated here.

[0294] In the technical solution disclosed herein, the collection, storage, use, processing, transmission, provision, disclosure, and application of user personal information comply with the provisions of relevant laws and regulations, necessary confidentiality measures have been taken, and there is no violation of public order and good morals.

[0295] In the technical solution disclosed herein, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0296] According to embodiments of this disclosure, this disclosure also provides a semantic communication system, an electronic device, a readable storage medium, and a computer program product.

[0297] The semantic communication system includes: a sending end of the semantic communication system, a receiving end of the semantic communication system, and a communication channel connecting the sending end and the receiving end.

[0298] Compared to existing technologies, this semantic communication system, under the unified guidance of an intelligent agent, achieves coordinated and adaptive processing across three stages: semantic extraction, joint source-channel coding, and encryption. Semantic extraction significantly compresses data volume, while joint coding further optimizes the compression-to-protection ratio. Joint coding dynamically adjusts redundancy based on channel conditions, effectively combating noise and fading. An encryption mechanism with adjustable strength prevents eavesdropping and is compatible with channel coding. The intelligent agent perceives environmental changes in real time, balancing transmission efficiency, reliability, and security without requiring manual intervention.

[0299] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0300] The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the semantic communication method provided in the embodiments of this disclosure.

[0301] Compared to existing technologies, this electronic device, under the unified guidance of an intelligent agent, achieves coordinated and adaptive processing across three stages: semantic extraction, joint source-channel coding, and encryption. Semantic extraction significantly compresses data volume, while joint coding further optimizes the compression-to-protection ratio. Joint coding dynamically adjusts redundancy based on channel conditions, effectively combating noise and fading. An encryption mechanism with adjustable strength prevents eavesdropping and is compatible with channel coding. The intelligent agent perceives environmental changes in real time, balancing transmission efficiency, reliability, and security without requiring manual intervention.

[0302] The readable storage medium is a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the semantic communication method provided in the embodiments of this disclosure.

[0303] Compared to existing technologies, this readable storage medium, under the unified guidance of an intelligent agent, achieves coordinated and adaptive processing across three stages: semantic extraction, joint source-channel coding, and encryption. Semantic extraction significantly compresses data volume, while joint coding further optimizes the compression-to-protection ratio. Joint coding dynamically adjusts redundancy based on channel conditions, effectively combating noise and fading. An encryption mechanism with adjustable strength prevents eavesdropping and is compatible with channel coding. The intelligent agent perceives environmental changes in real time, balancing transmission efficiency, reliability, and security without human intervention.

[0304] The computer program product includes a computer program that, when executed by a processor, implements the semantic communication method as provided in the embodiments of this disclosure.

[0305] Compared to existing technologies, this computer program product, under the unified guidance of an intelligent agent, achieves coordinated and adaptive processing across three stages: semantic extraction, joint source-channel coding, and encryption. Semantic extraction significantly compresses data volume, while joint coding further optimizes the compression-to-protection ratio. Joint coding dynamically adjusts redundancy based on channel conditions, effectively combating noise and fading. An encryption mechanism with adjustable strength prevents eavesdropping and is compatible with channel coding. The intelligent agent perceives environmental changes in real time, balancing transmission efficiency, reliability, and security without human intervention.

[0306] Figure 11 A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0307] like Figure 11 As shown, device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1102 or a computer program loaded into random access memory (RAM) 1103 from storage unit 1108. The RAM 1103 may also store various programs and data required for the operation of device 1100. The computing unit 1101, ROM 1102, and RAM 1103 are interconnected via bus 1104. Input / output (I / O) interface 1105 is also connected to bus 1104.

[0308] Multiple components in device 1100 are connected to I / O interface 1105, including: input unit 1106, such as keyboard, mouse, etc.; output unit 1107, such as various types of monitors, speakers, etc.; storage unit 1108, such as disk, optical disk, etc.; and communication unit 1109, such as network card, modem, wireless transceiver, etc. Communication unit 1109 allows device 1100 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0309] The computing unit 1101 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1101 performs the various methods and processes described above, such as semantic communication methods. For example, in some embodiments, the semantic communication method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 1108. In some embodiments, part or all of the computer program may be loaded and / or installed on device 1100 via ROM 1102 and / or communication unit 1109. When the computer program is loaded into RAM 1103 and executed by the computing unit 1101, one or more steps of the semantic communication method described above may be performed. Alternatively, in other embodiments, the computing unit 1101 may be configured to perform semantic communication methods by any other suitable means (e.g., by means of firmware).

[0310] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0311] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0312] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0313] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0314] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0315] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0316] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0317] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A semantic communication method based on intelligent agents, applied to the sending end of a semantic communication system, the semantic communication system including a sending end, a receiving end, and a communication channel connecting the sending end and the receiving end, the method comprising: Under the guidance of intelligent agents, semantic features are extracted from image sources. Under the guidance of the intelligent agent, channel environment information is perceived, and based on the perceived channel environment information, the semantic features are jointly encoded by the source and channel to generate channel symbols; Under the guidance of the intelligent agent, security risks and system performance constraints are perceived. Based on the perceived security risks and system performance constraints, the channel symbols are encrypted, and the encrypted channel symbols are sent into the wireless channel.

2. The method according to claim 1, wherein, The encryption of the channel symbols based on perceived security risks and system performance constraints includes: Based on the security risks and system performance constraints perceived by the intelligent agent, a combination of transformation types and the number of operation steps are selected, wherein the combination of transformation types is an ordered sequence used to specify the transformation type used in each step of a predefined encryption process; Using the channel symbol as the initial state, and following the predefined encryption process, the encryption transformation steps corresponding to the number of operation steps are executed sequentially according to the transformation type combination, wherein each encryption transformation step adopts the transformation type at the corresponding position in the transformation type combination.

3. The method according to claim 2, wherein, The transformation type includes one or more of Gaussian noise addition, feature obfuscation, and dynamic permutation.

4. The method according to claim 1, wherein, The step of performing source-channel joint coding on the semantic features based on the perceived channel environment information includes: Based on the channel environment information perceived by the intelligent agent, the encoding parameters of the deep neural network are selected; The semantic features are input into the deep neural network, and the semantic features are mapped to channel symbols according to the encoding parameters.

5. The method according to claim 4, wherein, The semantic extraction of the image source to obtain semantic features includes: The image source is input into the deep neural network, which outputs a low-dimensional semantic feature vector.

6. The method according to claim 1, wherein, The step of performing source-channel joint coding on the semantic features based on the perceived channel environment information includes: Based on the channel environment information perceived by the intelligent agent, the initial coding strategy is adjusted to obtain the adjusted coding strategy; According to the adjusted coding strategy, the semantic features are mapped to channel symbols.

7. The method according to claim 1, wherein the sending end and the receiving end share the same intelligent agent, the intelligent agent senses channel environment information and determines an encryption strategy, and the encryption strategy is used for both encryption at the sending end and decryption at the receiving end.

8. A semantic communication method based on intelligent agents, applied to the receiving end of a semantic communication system, the semantic communication system including a sending end, a receiving end, and a communication channel connecting the sending end and the receiving end, the method comprising: Under the guidance of the intelligent agent, security risks and system performance constraints are perceived. Based on the perceived security risks and system performance constraints, the encrypted channel symbols received from the wireless channel are decrypted to recover the channel symbols. Under the guidance of the intelligent agent, the channel environment information is perceived, and based on the perceived channel environment information, the decrypted channel symbols are jointly decoded from the source and channel to obtain semantic features. Under the guidance of the intelligent agent, semantic features are used to restore the semantic image and reconstruct it.

9. The method according to claim 8, wherein, The process of decrypting the encrypted channel symbols received from the wireless channel includes: Obtain the combination of transformation types and the number of operation steps used by the sending end; According to a predefined decryption process that is the reverse of the encryption process, the decryption transformation steps corresponding to the number of operation steps are executed sequentially according to the combination of transformation types, wherein each decryption transformation step adopts the inverse operation of the transformation type at the corresponding position in the combination of transformation types.

10. The method according to claim 9, wherein, The inverse operation includes one or more of the following: denoising, deobfuscation, and inverse permutation.

11. The method according to claim 9, wherein, The acquisition of the transformation type combination and operation steps used by the sending end includes: The intelligent agent at the receiving end and the intelligent agent at the sending end are the same intelligent agent. The intelligent agent at the receiving end determines the combination of transformation types and the number of operation steps adopted by the sending end based on the security risks and system performance constraints it perceives.

12. The method according to claim 8, wherein, The process of performing joint source-channel decoding on the decrypted channel symbols includes: The intelligent agent at the receiving end obtains the encoding parameters of the deep neural network used by the sending end, and inputs the decrypted channel symbol into the deep neural network. Based on the decoding parameters corresponding to the encoding parameters, the channel symbol is decoded into semantic features, wherein the encoding parameters are determined by the intelligent agent at the receiving end based on the perceived channel environment information.

13. The method according to claim 12, wherein, The semantic recovery of the semantic features includes: The semantic features are input into the deep neural network, and the deep neural network outputs a reconstructed image.

14. The method according to claim 8, wherein, The process of performing joint source-channel decoding on the decrypted channel symbols includes: The transmitting end and the receiving end share the same intelligent agent. The intelligent agent of the receiving end adjusts the initial decoding strategy corresponding to the initial encoding strategy based on the channel environment information it perceives, and obtains the adjusted decoding strategy. According to the adjusted decoding strategy, the channel symbols are decoded into semantic features.

15. A transmitting end of a semantic communication system, the semantic communication system comprising a transmitting end, a receiving end, and a communication channel connecting the transmitting end and the receiving end, the transmitting end comprising: The semantic extraction module is used to extract semantic features from image sources under the guidance of an intelligent agent. The channel coding module is used to sense channel environment information under the guidance of the intelligent agent, and to perform source-channel joint coding on the semantic features based on the sensed channel environment information to generate channel symbols. The encrypted transmission module is used to perceive security risks and system performance constraints under the guidance of the intelligent agent, encrypt the channel symbols according to the perceived security risks and system performance constraints, generate encrypted channel symbols, and send them into the wireless channel.

16. A receiving end of a semantic communication system, the semantic communication system comprising a sending end, a receiving end, and a communication channel connecting the sending end and the receiving end, the receiving end comprising: The decryption transmission module is used to perceive security risks and system performance constraints under the guidance of the intelligent agent, and decrypt the encrypted channel symbols received from the wireless channel to recover the channel symbols based on the perceived security risks and system performance constraints. The channel decoding module is used to sense channel environment information under the guidance of the intelligent agent, and to perform joint source-channel decoding on the decrypted channel symbols based on the sensed channel environment information to obtain semantic features. The semantic recovery module is used to perform semantic recovery on the semantic features and reconstruct the image under the guidance of the intelligent agent.

17. A semantic communication system, comprising: The transmitting end as described in claim 15; The receiving end as described in claim 16; And a communication channel connecting the sending end and the receiving end.

18. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-7 or 8-14.

19. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-7 or 8-14.

20. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-7 or 8-14.

Images