Task-oriented security semantic communication method and system
By designing a pre-encoder and optimizing feature encoding in the edge inference model, the task-oriented secure semantic communication method solves the channel correlation and task efficiency problems of physical layer security technology in complex communication tasks, and achieves improved task performance and classification accuracy while ensuring security.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-08
- Publication Date
- 2026-03-10
AI Technical Summary
Existing physical layer security technologies suffer from strong channel correlation, limited anti-interference capabilities, and a lack of consideration for task-specific requirements when facing complex communication tasks, making it difficult to balance data security and task efficiency.
A task-oriented secure semantic communication approach is adopted. By establishing an edge inference model, designing a pre-encoder, introducing a maximum coding rate to reduce data representation, compressing data into a low-dimensional subspace, combining deep neural networks for feature extraction and encoding, optimizing the feature encoder to balance receiver classification performance and eavesdropper data recovery capability, constructing an objective function and introducing weight coefficients to ensure that the transmission power does not exceed the maximum limit.
While ensuring security, it improves task performance, enhances classification accuracy, reduces the ability of eavesdroppers to recover data, and strengthens the security and robustness of information transmission, especially maintaining high-precision task execution when channel quality fluctuates.
Smart Images
Figure CN121644185A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a task-oriented secure semantic communication method and system, belonging to the field of wireless communication technology. Background Technology
[0002] With the rapid development of information technology, especially in key areas such as the Internet of Things, smart healthcare, and autonomous driving, AI-driven edge devices are constantly emerging, prompting the continuous transmission and processing of personal data. Therefore, information security has become crucial, and preventing data leakage has become an urgent challenge. Physical layer security technology increases the difficulty for eavesdroppers to decode information by controlling channel interference, signal-to-noise ratio, and signal transmission methods, thereby enabling legitimate receivers to successfully decode information while preventing eavesdroppers from obtaining valuable data [Y Wang, S Guo, Y Deng, H Zhang, and Y Fang. Privacy-preserving task-oriented semantic communications against modelinversion attacks. IEEE Transactions on Wireless Communications, 2024, 23(8):10150-10165, ]. This method has significant advantages in preventing eavesdropping. However, it still has certain limitations when facing modern complex communication tasks, such as strong channel correlation, limited anti-interference capabilities, and a lack of consideration for task-specific requirements. In addition, traditional physical layer security technologies usually do not consider the specific needs of tasks, so how to optimize task efficiency while ensuring data security remains an unsolved problem.
[0003] In contrast, task-oriented requirements are becoming increasingly important in edge inference across various application scenarios, such as image recognition, speech recognition, and intelligent system control. To address the current demands for edge task execution and information security, Task-Oriented Semantic Communication (ToSC) offers a novel solution [W Yang, H Du, Z Q. Liew, WYB Lim, Z. Xiong, D Niyato, X Chi, X Shen, and C Miao, Semanticcommunications for future internet: Fundamentals, applications, and challenges. IEEE Communications Surveys and Tutorials, 2022, 25(1), 213–250]. ToSC systems employ a deep neural network-source-channel coding architecture to perform inference tasks. This paradigm enables ToSC systems to efficiently extract crucial information from high-dimensional task-related data, avoiding the transmission of redundant and irrelevant data. Specifically, ToSC systems can intelligently process information according to task requirements; therefore, even if information is eavesdropped on, the eavesdropper can only obtain limited semantic information and cannot fully recover the original data. Furthermore, task-oriented semantic communication can be combined with various physical layer security technologies to further enhance data protection. For example, differential privacy technology provides robust protection for sensitive information by adding noise to the joint source-channel coding output of deep neural networks. While excessive noise can effectively protect sensitive information, it inevitably affects task inference performance, i.e., reduces the usability of transmitted data. In task-oriented semantic communication systems, although existing security algorithms play a crucial role in data protection [A Xiong, T Wang, N Li, and S Jha. Towards effective differential privacy communication for users' data sharing decision and comprehension. IEEE Symposium on Security and Privacy, 2020, 392-410.], they still have certain limitations in balancing task requirements and security. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention proposes a task-oriented secure semantic communication method from the perspective of the communication physical layer. Specifically, this invention uses a pre-encoder designed based on maximum coding rate reduction to compress data representation in a low-dimensional subspace, thereby achieving compact task-specific semantic feature extraction. This improves task performance while ensuring security, better meeting the needs of task-oriented semantic communication systems.
[0005] The technical solution of this invention is as follows: A task-oriented secure semantic communication method includes: (1) Establish a task-oriented edge reasoning model; (2) Based on the edge reasoning model and physical layer security theory established in step (1), design a precoder to balance the classification performance of the expected receiver and the data recovery capability of the eavesdropper, introduce the loss of the communication system, and construct the objective function; (3) Based on the objective function obtained by the transformation in step (2), an appropriate weight coefficient is introduced. The class function is considered by taking into account the separability of inter-class features and the correlation of intra-class features. This ensures that the target receiver achieves the best classification accuracy and prevents the eavesdropper from recovering the transmitted data. In addition, the system's transmission power is constrained to ensure that it does not exceed the predefined maximum limit, thereby ensuring the effective allocation of communication resources. When facing the classification task, the received signal is classified by the classifier and the classification accuracy is calculated.
[0006] (4) Based on the established task-oriented edge reasoning model, semantic communication security is achieved, including: inputting the original data to be transmitted; the transmitter extracts and encodes the data; the encoded features are normalized by power and then transmitted; the receiver performs task reasoning based on the received features, so that the target receiver achieves the best classification accuracy, while preventing eavesdroppers from recovering the transmitted data, thus achieving data security protection.
[0007] According to a preferred embodiment of the present invention, the task-oriented edge inference model includes a transmitter and a receiver; at the resource-constrained transmitter end, the feature encoder extracts deep features from the original image according to the preset task objective, generating a low-dimensional feature vector; transmitted via a wireless link, at the receiver end, a classifier decodes and understands the received low-dimensional feature vector, and directly outputs the final task result.
[0008] According to a preferred embodiment of the present invention, the specific implementation process of step (2) includes: For the input image, i.e., the raw data to be transmitted, firstly, a feature encoder is used to extract feature vectors. In single-user scenarios, linear analog modulation is used to transmit features, and the feature vector is... Reshape into a complex vector , To represent complex numbers, the specific transformation method is as shown in equation (I): (I); In formula (I), The imaginary unit, D satisfy D= 2 F ,Right now D For the corresponding feature vector The dimension of F; F is a complex vector. Dimensions , This represents a slicing operation, where 1:F indicates taking a vector. The first to the Fth elements form a new real number vector; F+1:D means taking the (F+1)th to the Dth elements to form another new real number vector. For edge devices, set For a linear precoder, where, N t This indicates the number of transmitting antennas; then the transmitted signal vector... Generated from equation (II): (II); definition N t The number of receiving antennas, Given the channel matrix between terminals, the received signal vector... Represented as Equation (III): (III); in, It is noise generated during the transmission process of the communication system; For the transmit signal vector of the linear precoder V Assuming the maximum average transmit power constraint is satisfied, the specific expression is (IV): (IV); In equation (IV), for The covariance matrix, P Indicates the maximum average power; represents the square of the 2-norm of s, represents the average total power of the transmitted signal s, tr() represents the trace of a matrix, tr is an operator in linear algebra used to calculate the sum of all elements on the main diagonal of a square matrix; Assume that a communication system simultaneously has both a legitimate receiver and an eavesdropper, and that the legitimate receiver and the eavesdropper experience different channel conditions; assume... The communication channel matrix represents the legitimate receiver. This represents the communication channel matrix of the eavesdropper; Construct loss functions separately for legitimate receivers and eavesdroppers; the loss function for legitimate receivers is as follows: For example, equation (V): (V); The eavesdropper's loss function (VI): (VI); In equations (V) and (VI), and They represent the eigenvectors respectively. Overall covariance matrix and category The covariance matrix within; ;in ,and p j Represents the situation in a given channel matrix or At that time, the prior probability of the signal category, This represents the average noise power.
[0009] According to a preferred embodiment of the present invention, the specific implementation process of step (3) includes: Introducing a maximum power constraint further controls the resource consumption of the communication system; The final optimization objective function For equation (VII): (VII); In equation (VII), by introducing appropriate weighting coefficients... In two objectives and Achieving balance between them; P represents maximum power; A method based on the alternating block gradient projection algorithm is used to calculate... include and about The gradient is given by equation (VIII): (VIII); The expression obtained is equation (IX): (IX); In equation (IX), as well as , The first part of the objective function The gradient is used to maximize the signal quality received by legitimate receivers; This represents the second part of the objective function. The gradient is used to minimize the quality of the signal received by the eavesdropper; gradient and The calculation is performed by differentiating the corresponding terms with respect to V, and the computation is simplified by utilizing the properties of trace operations and the block structure of the problem. The calculated gradient is then used to iteratively update V, ensuring a balance between maximizing the signal quality of the legitimate receiver and minimizing the signal quality of the eavesdropper. When the transmit power of the edge device reaches... When, function To obtain the maximum value, the gradient The projection is given by equation (X): (X); In formula (X), ; Therefore, update for As shown in equation (XI): (XI); In formula (XI), Then, by selecting the appropriate value And set Obtain the optimized and ; Repeat the above process until convergence, approximating the optimal linear precoder V.
[0010] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the task-oriented secure semantic communication method.
[0011] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the task-oriented secure semantic communication method.
[0012] A task-oriented secure semantic communication system, comprising: The edge inference model building module is configured to: build a task-oriented edge inference model; The objective function construction module is configured to: design a precoder to balance the expected classification performance of the receiver and the data recovery capability of the eavesdropper based on the established edge inference model and physical layer security theory, introduce the loss of the communication system, and construct the objective function; The weighting coefficient introduction module is configured to: introduce appropriate weighting coefficients based on the transformed objective function, considering the separability of inter-class features and the correlation of intra-class features to achieve optimal classification accuracy for the target receiver while preventing eavesdroppers from recovering transmitted data; furthermore, constraints ensure that the system's transmission power does not exceed a predefined maximum limit, thereby ensuring the effective allocation of communication resources; and for classification tasks, classify the received signals using a classifier and calculate the classification accuracy.
[0013] The semantic communication security implementation module is configured to implement semantic communication security based on the established task-oriented edge reasoning model, including: inputting the raw data to be transmitted; the transmitter extracting and encoding features from the data; transmitting the encoded features after power normalization; and the receiver performing task reasoning based on the received features, so that the target receiver can achieve the best classification accuracy, while preventing eavesdroppers from recovering the transmitted data, thereby achieving data security protection.
[0014] The beneficial effects of this invention are as follows: To address the high risk of eavesdropping during information transmission, the tendency for enhanced security to compromise classification accuracy, and the potential weakening of security protection due to channel quality fluctuations, this invention proposes a task-oriented secure semantic communication method and system. This method improves classification accuracy by optimizing the feature encoder. The semantic communication system employs this method to ensure the classifier maintains high accuracy on the edge server while effectively reducing the ability of eavesdroppers to intercept signals, thereby enhancing the security of information transmission. Through exploring the balance between task performance and security capabilities, the system investigates the adjustment level of physical layer security, demonstrating its robustness. It outperforms traditional benchmark methods that maximize security capacity, further validating its effectiveness. This invention exhibits good reliability under various classifier and channel limitations. Attached Figure Description
[0015] Figure 1 This is an architecture diagram of the task-oriented edge reasoning model of this invention; Figure 2 This is a schematic diagram showing the relationship between the classification accuracy and weight coefficients of a reasonable receiver and an eavesdropper in the secure semantic communication method proposed in this invention at signal-to-noise ratios of -10dB and 0dB. Figure 3 This is a schematic diagram showing the relationship between the classification accuracy and signal-to-noise ratio of the secure semantic communication method proposed in this invention and the traditional maximum security capacity (MSC) on the classifier MAP. Figure 4 This is a schematic diagram showing the relationship between the classification accuracy and signal-to-noise ratio of the secure semantic communication method proposed in this invention and the traditional maximum security capacity (MSC) on the classifier NN; Figure 5This is a schematic diagram showing the relationship between the classification accuracy and signal-to-noise ratio of the secure semantic communication method proposed in this invention and the traditional maximum security capacity (MSC) on the classifier NM. Detailed Implementation
[0016] The present invention will be further defined below with reference to the accompanying drawings and embodiments, but is not limited thereto.
[0017] Example 1 A task-oriented secure semantic communication method includes: (1) Establish a task-oriented edge reasoning model; (2) Based on the edge reasoning model and physical layer security theory established in step (1), design a precoder to balance the expected classification performance of the receiver and the data recovery capability of the eavesdropper, introduce the loss of the communication system, and construct the objective function (Equation (VII)). (3) Based on the objective function obtained by the transformation in step (2), an appropriate weight coefficient is introduced. The class function is considered by taking into account the separability of inter-class features and the correlation of intra-class features. This ensures that the target receiver achieves the best classification accuracy and prevents the eavesdropper from recovering the transmitted data. In addition, the system's transmission power is constrained to ensure that it does not exceed the predefined maximum limit, thereby ensuring the effective allocation of communication resources. When facing the classification task, the received signal is classified by the classifier and the classification accuracy is calculated.
[0018] (4) Based on the established task-oriented edge reasoning model, semantic communication security is achieved, including: inputting the original data to be transmitted; the transmitter extracts and encodes the data; the encoded features are normalized by power and then transmitted; the receiver performs task reasoning based on the received features, so that the target receiver achieves the best classification accuracy, while preventing eavesdroppers from recovering the transmitted data, thus achieving data security protection.
[0019] Example 2 The difference between the task-oriented secure semantic communication method described in Embodiment 1 and the method described in Embodiment 1 is as follows: like Figure 1 As shown, the task-oriented edge inference model includes a transmitter and a receiver. The task-oriented edge inference model adopts an end-to-end collaborative design architecture, the core of which lies in establishing an intelligent system that communicates for the task. At the resource-constrained transmitter end, the feature encoder (feature extractor) performs deep feature extraction on the original image according to the preset task objective (such as classification), generating a highly compressed, task-related low-dimensional feature vector. This feature is then transmitted via a wireless link after channel coding. At the receiver end, a classifier (such as a neural network) decodes and understands the received low-dimensional feature vector, directly outputting the final task result (such as a classification label).
[0020] The specific implementation process of step (2) includes: For the input image, i.e., the raw data to be transmitted, firstly, a feature encoder is used to extract feature vectors. The feature encoder consists of a encoder with learnable parameters. The neural network implementation; the feature encoder is a designed neural network whose core mission is to map high-dimensional, redundant raw input data into a low-dimensional, dense semantic feature space (i.e., feature vector z). This feature vector contains all the key information needed to complete downstream tasks (such as classification), while discarding irrelevant details. Specifically, the transformation involves taking an existing image and performing a series of matrix multiplications and nonlinear transformations to obtain the final feature vector z representing the image features. In single-user scenarios, to adapt to wireless transmission, linear analog modulation (LINEAM) is used to transmit features, and the feature vector... Reshape into a complex vector , To represent complex numbers, the specific transformation method is as shown in equation (I):
[0021] (I); In formula (I), The imaginary unit, D satisfy D= 2 F ,Right now D For the corresponding feature vector The dimension of F; F is a complex vector. Dimensions , This represents a slicing operation, where 1:F indicates taking a vector. The first to the Fth elements form a new real number vector; F+1:D means taking the (F+1)th to the Dth elements to form another new real number vector. For edge devices, set The linear precoder is a crucial module situated between channel coding and antenna transmission. Its core function is to perform a linear transformation (i.e., multiply by a matrix) on the signal vector to be transmitted, thereby optimizing the spatial orientation of the transmitted signal to better adapt to channel conditions, ultimately improving transmission efficiency and reliability. N t This indicates the number of transmitting antennas; then the transmitted signal vector... Generated from equation (II):
[0022] (II); definition N tThe number of receiving antennas was set to 4 in the experiment. Given the channel matrix between terminals, the received signal vector... Represented as Equation (III): (III); in, It is noise generated during the transmission process of the communication system; For the transmit signal vector of the linear precoder V Assuming the maximum average transmit power constraint is satisfied, the specific expression is (IV): (IV); In equation (IV), for The covariance matrix, P This represents the maximum average power; introducing this transmission power constraint helps control device power consumption, prevent interference caused by excessive power usage, and ensure reliable signal transmission. represents the square of the 2-norm of s. The square of the 2-norm of a vector represents the total power carried by that vector. E is an abbreviation for Expectation, representing the average total power of the transmitted signal s. tr() represents the trace of a matrix. tr is an abbreviation for Trace, which is an operator in linear algebra used to calculate the sum of all elements on the main diagonal of a square matrix (a matrix with the same number of rows and columns).
[0023] The above is a classic transmitter-channel-receiver model. The edge device is the signal processing and control center of the transmitter, the transmitting antenna is responsible for signal conversion and transmission, the wireless channel is the intermediate propagation medium, a physical environment, the receiving antenna receives the transmitted signals, and the terminal receiver is the signal processing and execution unit.
[0024] Assume that a communication system simultaneously has both a legitimate receiver and an eavesdropper, and that the legitimate receiver and the eavesdropper experience different channel conditions; assume... The communication channel matrix represents the legitimate receiver. This represents the communication channel matrix of the eavesdropper; the challenge here is how to ensure the security of communication while maintaining reliable data transmission to the legitimate recipient in the presence of an eavesdropper.
[0025] To address this issue, loss functions are constructed separately for legitimate receivers and eavesdroppers. The loss function for legitimate receivers is... For example, equation (V): (V); The eavesdropper's loss function (VI): (VI); In equations (V) and (VI), and They represent the eigenvectors respectively. Overall covariance matrix and category The covariance matrix within; ;in ,and p j Represents the situation in a given channel matrix or At that time, the prior probability of the signal category, This represents the average noise power.
[0026] The specific implementation process of step (3) includes: Specifically, this invention combines the channel matrices of the receiver and the eavesdropping end, as well as the covariance matrices of the signal and interference, to comprehensively consider the channel conditions of both the receiver and the eavesdropper. Furthermore, a maximum power constraint is introduced to further control the resource consumption of the communication system.
[0027] The final optimization objective function For equation (VII): (VII); In equation (VII), by introducing appropriate weighting coefficients... In two objectives and Achieving balance between them; P represents maximum power; A novel approach based on the Alternative Block Gradient Projection (ABGP) algorithm is employed. This algorithm iteratively updates each variable using gradient information, thereby progressively optimizing each variable and improving overall optimization efficiency. ABGP is particularly suitable for optimization problems involving multidimensional parameters, and its block-based processing effectively reduces computational complexity. include and about The gradient is given by equation (VIII):
[0028] (VIII); The expression obtained is equation (IX): (IX); In equation (IX), as well as , The first part of the objective function The gradient is used to maximize the signal quality received by legitimate receivers; This represents the second part of the objective function. The gradient is used to minimize the quality of the signal received by the eavesdropper; gradient and The calculation is performed by differentiating the corresponding terms with respect to V, and the computation is simplified by utilizing the properties of trace operations and the block structure of the problem. The calculated gradient is then used to iteratively update V, ensuring a balance between maximizing the signal quality of the legitimate receiver and minimizing the signal quality of the eavesdropper. When the transmit power of the edge device reaches... When, function To achieve the maximum value, the gradient is... The projection is given by equation (X): (X); In formula (X), ; Therefore, update for As shown in equation (XI): (XI); In formula (XI), Then, by selecting the appropriate value And set Obtain the optimized and ; Repeat the above process until convergence, approximating the optimal linear precoder V.
[0029] Figure 2 The classification performance of the receiver and the eavesdropper varies with the confidentiality weight parameter. β The relationship between ∈[0,1] reflects the relative importance of safety and task utility in precoder optimization. Larger values... β This value implies a greater focus on reducing information leaks to eavesdroppers, thereby enhancing physical layer security. From Figure 2 It can be observed that, with β As the density increases, the receiver's classification accuracy typically decreases due to enhanced suppression of semantic feature leakage; conversely, the eavesdropper's accuracy also decreases, demonstrating improved security. This trade-off reflects the flexibility of the method in adapting to different security-performance requirements. Furthermore, better channel conditions (e.g., 0 dB) consistently yield higher classification accuracy than poorer conditions (e.g., -10 dB), highlighting the impact of physical layer signal quality on task inference.
[0030] exist Figure 3 ,Figure 4 and Figure 5 In the middle, fixed β =0.5, and classification accuracy was evaluated over a wide signal-to-noise ratio range of -20 dB to 10 dB using three different classifiers (MAP, NN, and NM). The inventive method consistently outperformed the MSC baseline scheme in all configurations. Figure 3 The results show that when the signal-to-noise ratio reaches 10 dB, the MCP achieves a classification accuracy of over 90% at the receiver, demonstrating its ability to maintain high task fidelity under good channel conditions. Especially with the MAP classifier, the MCP provides more stable and accurate semantic extraction, resulting in superior task performance. Figure 4 and Figure 5 The results show that when using NN and NM classifiers, the proposed method maintains its performance advantage over MSC throughout the entire signal-to-noise ratio range, further verifying the effectiveness of the proposed precoding strategy in maintaining semantic integrity to support downstream inference tasks. Figures 3-5 The classification performance of the eavesdropper was simultaneously demonstrated, showing that at all signal-to-noise ratio levels, the invented method significantly reduces the classification accuracy of the eavesdropper compared to the MSC scheme. This result clearly confirms that the invented method can effectively distort the semantic structure of transmission features in non-target eavesdropping channels, thereby significantly reducing the risk of successful reverse inference.
[0031] Example 3 A computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a task-oriented secure semantic communication method as described in Embodiment 1 or 2.
[0032] Example 4 A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a task-oriented secure semantic communication method as described in Embodiment 1 or 2.
[0033] Example 5 A task-oriented secure semantic communication system, comprising: The edge inference model building module is configured to: build a task-oriented edge inference model; The objective function construction module is configured to: design a precoder to balance the expected classification performance of the receiver and the data recovery capability of the eavesdropper based on the established edge inference model and physical layer security theory, introduce the loss of the communication system, and construct the objective function (Equation (VII)). The weighting coefficient introduction module is configured to: introduce appropriate weighting coefficients based on the transformed objective function, considering the separability of inter-class features and the correlation of intra-class features to achieve optimal classification accuracy for the target receiver while preventing eavesdroppers from recovering transmitted data; furthermore, constraints ensure that the system's transmission power does not exceed a predefined maximum limit, thereby ensuring the effective allocation of communication resources; and for classification tasks, classify the received signals using a classifier and calculate the classification accuracy.
[0034] The semantic communication security implementation module is configured to implement semantic communication security based on the established task-oriented edge reasoning model, including: inputting the raw data to be transmitted; the transmitter extracting and encoding features from the data; transmitting the encoded features after power normalization; and the receiver performing task reasoning based on the received features, so that the target receiver can achieve the best classification accuracy, while preventing eavesdroppers from recovering the transmitted data, thereby achieving data security protection.
Claims
1. A task-oriented secure semantic communication method, characterized in that, Comprise: (1) establish a task-oriented edge inference model; (2) according to the edge inference model established in step (1) and the physical layer security theory, design a precoder to balance the classification performance of the intended receiver and the data recovery ability of the eavesdropper, introduce the loss of the communication system, and construct the objective function; (3) based on the objective function obtained by transforming step (2), introduce appropriate weight coefficients, consider the separability of inter-class features and the correlation of intra-class features; ensure that the target receiver achieves the best classification accuracy, while preventing the eavesdropper from recovering the transmitted data; in addition, the constraint ensures that the transmission power of the system does not exceed the predefined maximum limit, thereby ensuring the effective allocation of communication resources; in the face of classification tasks, the received signals are classified by the classifier and the classification accuracy is calculated; (4) based on the established task-oriented edge inference model, semantic communication security is realized, including: inputting the original data to be transmitted; the transmitter performs feature extraction and encoding on the data; the encoded features are normalized in power and transmitted; the receiver performs task inference based on the received features, so that the target receiver achieves the best classification accuracy, while preventing the eavesdropper from recovering the transmitted data, realizing the security protection of the data.
2. The task-oriented secure semantic communication method of claim 1, wherein, The task-oriented edge inference model includes a transmitter and a receiver; at the resource-limited transmitter end, the feature encoder extracts deep features from the original image according to the preset task target to generate a low-dimensional feature vector; through wireless link transmission, at the receiver end, the received low-dimensional feature vector is decoded and understood by the classifier, and the final task result is directly output.
3. The task-oriented secure semantic communication method of claim 1, wherein, The specific implementation process of step (2) includes: For the input image, i.e., the original data to be transmitted, first, a feature vector is extracted using a feature encoder ; in a single-user scenario, linear analog modulation is used to transmit the features, and the feature vector is reshaped into a complex vector , denotes a complex number, and the specific transformation is as shown in formula (I): (I)； In formula (I), is an imaginary unit, D satisfies D= 2 F , i.e. D is a corresponding feature vector with dimension D; F is a complex number vector with dimension F, , represents a slice operation, 1:F represents taking the first F elements of the orientation vector to form a new real number vector; F+1:D represents taking the F+1th to Dth elements of it to form another new real number vector; For edge devices, set is a linear precoder, where N t denotes the number of transmit antennas; the transmit signal vector is generated by equation (II): (I); and (II). Definitions N t is the number of receive antennas, is the channel matrix between terminals, then the received signal vector is represented as equation (III): (III); wherein is the noise generated in the transmission process of the communication system; For the transmit signal vector of the linear precoder V , assuming that the maximum average transmit power constraint is satisfied, is specifically represented by expression (IV): (IV); In formula (IV), is the covariance matrix, P denotes the maximum average power; denotes the square of the 2-norm of s, denotes the average total power of the transmitted signal s, and tr() denotes the trace of a matrix, tr is an operator in linear algebra, which is used to calculate the sum of all elements on the main diagonal of a square matrix.
4. The task-oriented secure semantic communication method of claim 2, wherein, The communication system is assumed to exist simultaneously with a legitimate receiver and an eavesdropper, and the legitimate receiver and the eavesdropper experience different channel conditions, respectively. a communication channel matrix representing the legitimate receiver, a communication channel matrix representing the eavesdropper; The loss function is constructed for the legitimate receiver and the eavesdropper respectively, and the loss function of the legitimate receiver is is of formula (V): (V)； The loss function of the eavesdropper (VI): (WE); (V) and (VI) respectively represent the overall covariance matrix of the feature vectors and respectively represent the overall covariance matrix of the feature vectors and the covariance matrix within the class ; ; where , and p j represents the prior probability of the signal class given the channel matrix or , denotes the average noise power.
5. The task-oriented secure semantic communication method of claim 1, wherein, The specific implementation process of step (3) includes: Introducing the maximum power constraint to further control the resource consumption of the communication system; Final optimization objective function for formula (VII): (VII); In formula (VII), a balance is achieved between two objectives by introducing appropriate weighting factors and P represents the maximum power.
6. The task-oriented secure semantic communication method of claim 5, wherein, A method based on the alternating block gradient projection algorithm is used to calculate... include and about The gradient is given by equation (VIII): (VIII); The expression is formula (IX): (IX); In formula (IX), and , denotes the gradient of the first part of the objective function for maximizing the signal quality received by the legitimate receiver; denotes the gradient of the second part of the objective function for minimizing the signal quality received by the eavesdropper; gradient and The gradient is computed by taking the derivative of the corresponding term with respect to V and simplifying the computation using properties of the trace operation and the block structure of the problem; the computed gradient is then used to iteratively update V, ensuring a balance between maximizing the signal quality of the legitimate receiver and minimizing the signal quality of the eavesdropper; When the transmission power of the edge device reaches the function takes the maximum value, and the gradient is projected as formula (X): (X)； in formula (X), ; Accordingly, the update is as shown in formula (XI): (XI); in formula (XI), then, by choosing appropriate values and setting obtaining an optimized and ; Repeat the above process until convergence, and approach the optimal linear precoder V. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor executes the computer program to realize the steps of the task-oriented secure semantic communication method of any one of claims 1-6.
8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the task-oriented secure semantic communication method of any one of claims 1-6.
9. A task-oriented secure semantic communication system, characterized by, Comprise: The edge inference model establishment module is configured to establish a task-oriented edge inference model; The objective function construction module is configured to design a precoder to balance the classification performance of the intended receiver and the data recovery ability of the eavesdropper according to the established edge inference model and the physical layer security theory, introduce the loss of the communication system, and construct the objective function; The weight coefficient introduction module is configured to introduce appropriate weight coefficients based on the transformed objective function, consider the separability of inter-class features and the correlation of intra-class features; ensure that the target receiver achieves the best classification accuracy, while preventing the eavesdropper from recovering the transmitted data; in addition, the constraint ensures that the transmission power of the system does not exceed the predefined maximum limit, thereby ensuring the effective allocation of communication resources; in the face of classification tasks, the received signals are classified by the classifier and the classification accuracy is calculated; The semantic communication security implementation module is configured to implement semantic communication security based on the established task-oriented edge inference model, including: inputting original data to be transmitted; a transmitter performing feature extraction and encoding on the data; performing power normalization on the encoded features and then transmitting; a receiver performing task inference based on the received features, so that the target receiver achieves the best classification accuracy, while preventing an eavesdropper from recovering the transmitted data, thereby achieving security protection of the data.