Heterogeneous communication network physical layer security resource joint allocation method, system and device based on multi-hop attention hypergraph convolution and medium

Abstract

The invention belongs to the technical field of heterogeneous communication network security, and discloses a heterogeneous communication network physical layer security resource joint allocation method, system, equipment and medium based on multi-hop attention hypergraph convolution, and the method comprises the following steps: updating a node feature matrix to define a propagation operator of hypergraph convolution, and then according to the hyperedge type and the propagation operator, determining the hyperedge type according to the hyperedge type and the propagation operator; updating the multi-hop node feature matrix to carry out attention aggregation to obtain final node embedding, embedding the final node of any legal user node into an input multi-head task, distributing a base station access decision by multiple heads, and calculating the transmitting power and beam forming vector of a base station selected by the base station access decision to any legal user, and finally, based on a secrecy rate optimization problem, taking a negative value of the total secrecy rate as a loss function to guide physical layer security resource joint allocation. The system, the equipment and the medium are used for implementing the method. According to the invention, the physical layer security performance of the secret communication system is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of heterogeneous communication network security technology, specifically relating to a method, system, device, and medium for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution. Background Technology

[0002] In the field of modern wireless communication security, with the rapid advancements in quantum computing and high-performance distributed computing technologies, traditional high-level encryption algorithms based on computational complexity (such as RSA and ECC) are facing the potential threat of being "brute-forced" or "suboptimal." This reliance on pre-shared keys and complex mathematical problems has revealed serious lag and vulnerability in an era of overflowing computing power. Therefore, there is an urgent need to shift the security paradigm and utilize the inherent non-cloning nature and random physical characteristics of wireless channels (such as decorrelation of multipath fading, masking effects of thermal noise, and spatial degrees of freedom of multi-antenna systems) to build intrinsic security mechanisms.

[0003] When attempting to model the wireless topology secure communication optimization problem using existing graph convolutional networks (GCNs), ordinary graphs typically only focus on the characteristics of directly adjacent nodes. However, interference in the wireless environment is global. Traditional single-hop aggregation mechanisms struggle to capture the latent interference patterns of second-neighbors or more distant nodes. Simply increasing network depth can easily lead to over-smoothing, causing node characteristics to lose their distinctiveness, resulting in mediocre decisions on spectrum selection and power allocation, and an inability to adapt to dynamically changing wireless channels.

[0004] Patent application CN116056214A discloses a wireless communication resource allocation method based on a deep unfolded distributed graph neural network. However, because this patent application still uses a binary edge modeling method based on the channel gain matrix when constructing the graph topology, it can only depict the interference relationship between pairs of nodes. When representing the cooperation / competition relationship of multiple nodes such as base stations, multiple users, and malicious interference sources in heterogeneous cellular networks, it needs to approximate the relationship through a large number of binary edge stacks, resulting in high redundancy in topology representation, low feature aggregation efficiency, and the adjacency matrix is ​​sensitive to node displacement. In dynamic scenarios, the graph structure changes frequently, which destroys the smoothness of the loss function and hinders model convergence.

[0005] Perera T et al. published "Graph Neural Networks for Physical-Layer Security in Multi-User Flexible-Duplex Networks" (2024 IEEE International Conference on Communications (ICC), DOI: 10.1109 / ICC51166.2024.10622731). This study is the first to introduce graph neural networks into the physical layer security beamforming design of flexible duplex networks. It jointly optimizes resource allocation through unsupervised learning strategies to maximize the security rate and extends to scenarios where eavesdropper channel state information is missing. However, because the graphing in this paper still relies on the channel gain matrix (CSI) or Euclidean distance metric, it does not explicitly model the angle domain threat correlation under the condition of ambiguous passive eavesdropper location. In the directional beamforming scenario, it is difficult to effectively characterize the security constraint relationship of the beam angle dimension rather than the distance dimension, resulting in insufficient representation ability of ambiguous eavesdropping threats. The generalization performance of the model is limited by the accuracy of the graphing prior. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention aims to provide a method, system, device, and medium for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution. In a complex secure communication hypergraph, attention aggregation is performed through hypergraph convolution and multi-hop node feature matrices. Without relying on traditional computational complexity for security, it achieves accurate characterization and joint suppression of complex wireless interference relationships and potential leakage paths. Simultaneously, it optimizes secure communication for legitimate users through spectrum selection, power allocation, and beamforming, thereby significantly improving the security performance, adaptability, and practical deployment feasibility of heterogeneous communication networks.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A joint allocation method for physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution includes the following steps: Step 1: Based on the node feature matrix and node association matrix, perform message passing, update the node feature matrix, and then apply the associative law of matrix multiplication to define the propagation operator of hypergraph convolution. Step 2: Update the multi-hop node feature matrix according to the hyperedge type and the propagation operator of the hypergraph convolution; Step 3: Perform attention aggregation on the multi-hop node feature matrix to obtain the final node embedding; Step 4, add any legitimate user node The final node embeds the input spectrum allocation head, power allocation head, and beamforming head. The spectrum allocation head, power allocation head, and beamforming head jointly allocate base station access decisions. Base station access decision Selected base stations For any legitimate user Transmit power and beamforming vector; Step 5, based on base station access decision Base station access decision Selected base stations For any legitimate user The transmit power and beamforming vector are used to calculate the security rate; Step 6: Propose the security rate optimization problem, and use the negative value of the total security rate as the loss function to guide the joint allocation of physical layer security resources in step 4.

[0008] Step 1 specifically includes the following steps: definition For any node The node feature matrix; The feature extraction process is decomposed into node-to-hyperedge aggregation and hyperedge-to-node update; The node-to-hyperedge aggregation: each hyperedge As a local observer, aggregate each hyperedge Features of connected nodes are used to generate a hyperedge representation matrix. The polymerization process is represented as: in, This represents the node adjacency matrix. Represents the node degree matrix, Represents the hypermarginality matrix. This indicates that node-level pre-normalization is performed. This aggregation process maps node domain features to the hyperedge space, reflecting the characteristics of each hyperedge. The status of collective interference within the coverage area; The hyperedge-to-node update: A node updates its hyperedge by sensing the hyperedge it belongs to. The spatial disturbance field represents the update of the node feature matrix. The updated node feature matrix is ​​represented as The update process is represented as follows: in, diagonal superside The weight matrix is ​​pre-normalized at the node level. This update process ensures that nodes can effectively access multiple overlapping hyperedges. Extract information from it while maintaining numerical stability; Represent the hyperedge matrix Substitute the updated node feature matrix Perform normalized symmetric propagation operator: By applying the associative law of matrix multiplication to extract all intermediate transformation matrices, a propagation operator for hypergraph convolution is defined. : Step 2 specifically includes the following steps: For each type of superedge Define independent propagation operators for the first... K-hop and the types of super-edges are Propagation operator of hypergraph convolution Represented as: in, This indicates that node-level prenormalization is being performed. Represents the hypermarginality matrix. It is learnable and targeted at the first The types of jump and super-edge are The hyperedge weight matrix, Given the set of hyperedge types of the input hypergraph, For the superedge type The correlation matrix; No. Jump update of node feature matrix Represented as: Initial node feature matrix , For any node The node feature matrix, The maximum number of aggregate hops, Indicates the first Jumping to the side of the super league The trainable weight matrix, This represents the activation function.

[0009] Step 3 specifically includes the following steps: Constructing vector dimension mapping This represents a parameterized single-hidden-layer multilayer perceptron (MLP), which reduces the original feature dimension. input vector Mapped to feature dimension The output vector Θ contains all the weight matrices and bias vectors. Represents the set of trainable parameters. Indicates the input layer bias term. Indicates the output layer bias term; Indicates shape as OK The weight matrix of the input layer of the column. Indicates shape as OK The weight matrix of the output layer of the column, Indicates the input layer bias term. Indicates the output layer bias term; A query-based multi-hop attention mechanism is employed, which focuses on the number of hops from the initial input to the maximum aggregation number. The nodes are represented by weighted fusion, and any node query vector From any node Initial node features Generate to provide foundational knowledge: in, This indicates that it has a dedicated function for querying any node. Initial node features A single-hidden-layer multilayer perceptron (MLP) with independent parameters; Then, calculate any node In the Jump node features Correlation score between To identify the feature scale that contributes most to the regression objective: in, It is a scaling factor, expressed as the square root of the hidden layer dimension. It is a predefined temperature hyperparameter that applies the Softmax function to the set of correlation scores for all hops. To obtain any node Then Attention Convergence Coefficient of Jump : Finally, any node Multi-hop information is integrated and refined through a fusion layer to capture high-order feature interactions while dynamically balancing local interference and global topological context, ultimately resulting in node embedding. The calculation is as follows: in, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically designed for multi-hop aggregation.

[0010] In step 4, the spectrum allocation head manages any legitimate user. Base station access decision , Indicates the total number of base stations: probability distribution vector pass generate: This refers to a single-hidden-layer multilayer perceptron (MLP) that has independent parameters specifically for base station selection. Represents any final legitimate user Node embedding; The power allocation head determines the base station access decision for the power allocation task. Selected base stations For the first legitimate user Transmission power : in, Base station Maximum transmission power, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically for power distribution. Indicates the final first legal user Node embedding, Indicates the final second legal user Node embedding, This represents the total number of legitimate users. Represents an indicator function, when The indicator function takes the value 1 when the condition is met, and 0 otherwise. The power divider ensures that the desired power is met in a differentiable manner. This effectively maps the unconstrained network output onto the power simplex; The beamforming head generation is based on the base station access decision. Selected base stations For any legitimate user Beamforming vector : Satisfying the unit norm constraint ; in, , Indicates flattened output: in, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically used for beamforming vector generation. ], Any legitimate user Beamforming vector The real part, Any legitimate user Beamforming vector The imaginary part, , Indicates length is real vectors, Represents any base station The number of antennas.

[0011] Step 5 specifically includes the following steps: Base stations in heterogeneous communication networks Based on the channel environment, select the first legitimate user. Establish downlink channel, As the first legitimate user Base station access decision, The first legitimate user Received signal The model is as follows: in, Representing base stations For the first legitimate user The transmission power, Indicates base station To the first legitimate user distance, Indicates base station To the first legitimate user The channel vector, This is the path loss index. This indicates the conjugate transpose. For base stations For the first legitimate user Beamforming vector of the downlink channel, First legitimate user Received signal, Indicates the second legitimate user. This represents the total number of legitimate users. Represents an indicator function, when The indicator function takes a value of 1 when the condition is met, and 0 otherwise. Representing base stations For the second legitimate user The transmission power, For base stations For the second legitimate user The beamforming vector of the downlink channel satisfies , Second legitimate user Received signal, malicious interference The transmission power of the omnidirectional jamming signal. Indicates malicious interference To the first legitimate user distance, Indicates malicious interference To the first legitimate user The channel vector, Malicious interference When sending interference signals, the signals satisfy... , Represents the mathematical expectation. Additive white Gaussian noise (AWGN); Based on the first legitimate user Received signal Modeling signal-to-interference-plus-noise ratio (SINR), which is expressed as: in, The power of Gaussian white noise. As the first legitimate user The total disturbance term is represented as: eavesdropper For the first legitimate user Received signal The signal received during eavesdropping The model is as follows: in, Indicates base station To the eavesdropper distance, Representing base stations To the eavesdropper The channel vector, Indicates malicious interference The transmission power, Indicates malicious interference To the eavesdropper distance, Representing malicious interference To the eavesdropper The channel vector; Based on eavesdroppers For the first legitimate user Received signal The signal received during eavesdropping Modeling signal-to-interference-plus-noise ratio (SINR), which is expressed as: in, Representing the eavesdropper The first legitimate user of eavesdropping Total disturbance term received at that time: First legitimate user The secure rate is defined as the difference between the legitimate rate and the eavesdropping rate: in, ; According to the first legitimate user The method for calculating the confidentiality rate involves iterating over any legitimate user. The rate of confidentiality, This represents the total number of legitimate users; any legitimate user... The security rate is expressed as .

[0012] The specific steps of step 6 are as follows: Jointly optimize base stations For any legitimate user Beamforming vector of downlink channel Any legitimate user Base station access decision and base stations For any legitimate user Transmission power A confidentiality rate optimization problem is proposed, and the optimization problem is modeled as follows: in, Indicates the total number of base stations. Base station Maximum transmit power; By minimizing the task loss function To optimize the overall security rate : The variance of the confidentiality rate is calculated as follows: in, This represents the total number of legitimate users. The average security rate is calculated as follows: A joint allocation system for physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution includes: The message passing module performs message passing based on the node feature matrix and the node association matrix, updates the node feature matrix, and then applies the associative law of matrix multiplication to define the propagation operator of hypergraph convolution. The feature matrix update module updates the feature matrix of multi-hop nodes based on the type of hyperedge and the propagation operator of hypergraph convolution. The multi-hop attention aggregation module performs attention aggregation on the feature matrix of multi-hop nodes to obtain the final node embedding; The decision header module will include any legitimate user node. The final node embeds the input spectrum allocation head, power allocation head, and beamforming head. The spectrum allocation head, power allocation head, and beamforming head jointly allocate base station access decisions. Base station access decision Selected base stations For any legitimate user Transmit power and beamforming vector; The confidentiality rate calculation module calculates the data based on the base station access decision. Base station access decision Selected base stations For any legitimate user The transmit power and beamforming vector are used to calculate the security rate; The problem optimization module proposes a security rate optimization problem and uses the negative value of the total security rate as a loss function to guide the joint allocation of physical layer security resources.

[0013] A device for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution includes: Memory: Used to store computer programs that implement a joint allocation method for physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution; Processor: A method for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution when executing the computer program.

[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a method for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses a hypergraph structure for message passing, which avoids the structural redundancy of the traditional binary edge model, enhances the ability of this invention to model various relationships at the semantic level, reduces the noise caused by graph structure redundancy, and when network conditions change, the hypergraph structure requires only less topological structure change than the graph structure to accurately characterize the network environment again.

[0016] 2. This invention employs a spectrum allocation head, a power allocation head, and a beamforming head to optimize the allocation of physical layer security resources. It achieves joint optimization of beamforming, power allocation, and spectrum selection, thereby improving the physical layer security performance of heterogeneous communication networks. At the same time, it has the ability to make resource allocation decisions with fast reasoning, significantly reducing computational complexity and shortening decision latency.

[0017] 3. This invention employs a multi-hop semantic recursive aggregation mechanism, which effectively captures implicit interference and potential information leakage paths caused by second-neighbors and more distant nodes without blindly increasing network depth. This avoids the over-smoothing problem that easily occurs in traditional deep graph convolutional networks. By introducing a multi-hop semantic aggregation mechanism, this invention can effectively capture implicit interference caused by distant nodes. Compared with the method of only performing single-hop information aggregation, it has a stronger global interference perception capability, thereby maintaining the distinguishability of node features and decision sensitivity.

[0018] In summary, this invention, with hypergraph convolutional networks as its core, effectively improves the physical layer security performance of secure communication systems by jointly optimizing beamforming, power allocation, and spectrum selection. It is compatible with multiple types of hypergraph topologies, supports multi-hop semantic aggregation, and has fast inference capabilities. It outperforms traditional methods in terms of scalability, interference perception accuracy, and computational efficiency, and has strong practical value and engineering deployment potential, filling a technological gap in this field. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method described in this invention.

[0020] Figure 2 The method described in this invention is used to calculate different maximum aggregation hop counts. Hyperparameter sensitivity analysis plot.

[0021] Figure 3 This is a training convergence curve of the method described in this invention.

[0022] Figure 4 This is a comparison chart of the loss convergence performance of the method described in this invention and existing solutions.

[0023] Figure 5The method described in this invention differs from existing solutions in different eavesdropping scenarios. Density and legitimate users Robustness comparison chart under various conditions.

[0024] Figure 6 The method described in this invention differs from existing solutions in different legitimate users. Numbers and malicious interference Comparison chart of scalability under different numbers. Detailed Implementation

[0025] The present invention will now be described in detail with reference to the accompanying drawings.

[0026] like Figure 1 As shown, a method for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution includes the following steps: Step 1: Based on the node feature matrix and node association matrix, perform message passing, update the node feature matrix, and then apply the associative law of matrix multiplication to define the propagation operator of hypergraph convolution. Step 2: Update the multi-hop node feature matrix according to the hyperedge type and the propagation operator of the hypergraph convolution; Step 3: Perform attention aggregation on the multi-hop node feature matrix to obtain the final node embedding; Step 4, add any legitimate user node The final node embeds the input spectrum allocation head, power allocation head, and beamforming head. The spectrum allocation head, power allocation head, and beamforming head jointly allocate base station access decisions. Base station access decision Selected base stations For any legitimate user Transmit power and beamforming vector; Step 5, based on base station access decision Base station access decision Selected base stations For any legitimate user The transmit power and beamforming vector are used to calculate the security rate; Step 6: Propose the security rate optimization problem, and use the negative value of the total security rate as the loss function to guide the joint allocation of physical layer security resources in step 4.

[0027] Step 1 specifically includes the following steps: definition For any node The node feature matrix; In order to capture base stations Any legitimate user Malicious interference and eavesdroppers The complex interactions between them decompose the feature extraction process into node-to-hyperedge aggregation and hyperedge-to-node update; The node-to-hyperedge aggregation: each hyperedge As a local observer, aggregate each hyperedge Features of connected nodes are used to generate a hyperedge representation matrix. The polymerization process is represented as: in, This represents the node adjacency matrix. Represents the node degree matrix, Represents the hypermarginality matrix. This indicates that node-level pre-normalization is performed. This aggregation process maps node domain features to the hyperedge space, reflecting the characteristics of each hyperedge. The status of collective interference within the coverage area; The hyperedge-to-node update: A node updates its hyperedge by sensing the hyperedge it belongs to. The spatial disturbance field represents the update of the node feature matrix. The updated node feature matrix is ​​represented as The update process is represented as follows: in, diagonal superside The weight matrix is ​​pre-normalized at the node level. This update process ensures that nodes can effectively access multiple overlapping hyperedges. Extract information from it while maintaining numerical stability; Represent the hyperedge matrix Substitute the updated node feature matrix Perform normalized symmetric propagation operator: By applying the associative law of matrix multiplication to extract all intermediate transformation matrices, a propagation operator for hypergraph convolution is defined. : Symmetric normalized structure This ensures that the feature norm remains stable even as the network topology becomes more complex. By restricting the eigenvalues ​​of the operator to a stable range, smooth gradient flow is facilitated.

[0028] Step 2 specifically includes the following steps: To capture heterogeneous physical dependencies at different spatial scales, for each hyperedge type Define independent propagation operators for the first... K-hop and the types of super-edges are Propagation operator of hypergraph convolution Represented as: in, This indicates that node-level prenormalization is being performed. Represents the hypermarginality matrix. It is learnable and targeted at the first The types of jump and super-edge are The hyperedge weight matrix, Given the set of hyperedge types of the input hypergraph, For the superedge type The correlation matrix, defined by convolution kernels under different hyperedges and different hop counts, enables the method described in this invention to extract features with clear physical semantic distinctions, thus avoiding confusion; No. Jump update of node feature matrix Represented as: Initial node feature matrix , For any node The node feature matrix, The maximum aggregation hop count represents the maximum receptive field size. Indicates the first Jumping to the side of the super league The trainable weight matrix, The activation function is represented by the Leaky ReLU linear rectifier unit in this embodiment. In other embodiments, other activation functions may include ReLU, ELU, SELU, PReLU, RReLU, GELU, Swish, etc.

[0029] Step 3 specifically includes the following steps: Constructing vector dimension mapping This represents a parameterized single-hidden-layer multilayer perceptron (MLP), which reduces the original feature dimension. input vector Mapped to feature dimension The output vector Θ contains all the weight matrices and bias vectors. Represents the set of trainable parameters. Indicates the input layer bias term. Indicates the output layer bias term; Indicates shape as OK The weight matrix of the input layer of the column. Indicates shape as OK The weight matrix of the output layer of the column, Indicates the input layer bias term. Indicates the output layer bias term; To adaptively integrate cross-layer structural information, this invention employs a query-based multi-hop attention mechanism, which focuses on the number of hops from the initial input to the maximum aggregation number. The nodes are represented by weighted fusion, and any node query vector From any node Initial node features Generate to provide foundational knowledge: in, This indicates that it has a dedicated function for querying any node. Initial node features A single-hidden-layer multilayer perceptron (MLP) with independent parameters.

[0030] Then, calculate any node In the Jump node features Correlation score between To identify the feature scale that contributes most to the regression objective: in, It is a scaling factor, expressed as the square root of the hidden layer dimension. It is a predefined temperature hyperparameter. To formalize the weight allocation, the Softmax function is applied to the set of correlation scores for all hops. To obtain any node Then Attention Convergence Coefficient of Jump : Finally, any node Multi-hop information is integrated and refined through a fusion layer to capture high-order feature interactions while dynamically balancing local interference and global topological context, ultimately resulting in node embedding. The calculation is as follows: in This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically designed for multi-hop aggregation.

[0031] In step 4, the spectrum allocation head manages any legitimate user. Base station access decision , Indicates the total number of base stations: probability distribution vector pass generate: This refers to a single-hidden-layer multilayer perceptron (MLP) that has independent parameters specifically for base station selection. Represents any final legitimate user Node embedding; To achieve end-to-end optimization, this invention distinguishes between the training and inference processes. During training, this invention uses discrete interference indicators... Relaxed to a differentiable inner product Then, during inference, it is restored to any legitimate user of the discrete downlink channel. Base station access decisions.

[0032] The power allocation head determines the base station access decision for the power allocation task. Selected base stations For the first legitimate user Transmission power : in, Base station Maximum transmission power, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically for power distribution. Indicates the final first legal user Node embedding, Indicates the final second legal user Node embedding, This represents the total number of legitimate users. Represents an indicator function, when The indicator function takes the value 1 when the condition is met, and 0 otherwise. The power divider ensures that the desired power is met in a differentiable manner. This effectively maps the unconstrained network output onto the power simplex; The beamforming head generation is based on the base station access decision. Selected base stations For any legitimate user Beamforming vector : Satisfying the unit norm constraint ; in, , Indicates flattened output: in, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically used for beamforming vector generation. ], Any legitimate user Beamforming vector The real part, Any legitimate user Beamforming vector The imaginary part, , Indicates length is real vectors, Represents any base station The number of antennas.

[0033] Step 5 specifically includes the following steps: Base stations in heterogeneous communication networks Based on the channel environment, select the first legitimate user. Establish downlink channel, As the first legitimate user Base station access decision, The first legitimate user Received signal The model is as follows: in, Representing base stations For the first legitimate user The transmission power, Indicates base station To the first legitimate user distance, Indicates base station To the first legitimate user The channel vector, This is the path loss index. This indicates the conjugate transpose. For base stations For the first legitimate user Beamforming vector of the downlink channel, First legitimate user Received signal, Indicates the second legitimate user. This represents the total number of legitimate users. Represents an indicator function, when The indicator function takes a value of 1 when the condition is met, and 0 otherwise. Representing base stations For the second legitimate user The transmission power, For base stations For the second legitimate user The beamforming vector of the downlink channel satisfies , Second legitimate user Received signal, malicious interference The transmission power of the omnidirectional jamming signal. Indicates malicious interference To the first legitimate user distance, Indicates malicious interference To the first legitimate user The channel vector, Malicious interference When sending interference signals, the signals satisfy... , Represents the mathematical expectation. Additive white Gaussian noise (AWGN); Based on the first legitimate user Received signal Modeling signal-to-interference-plus-noise ratio (SINR), which is expressed as: in, The power of Gaussian white noise. As the first legitimate user The total disturbance term is represented as: eavesdropper For the first legitimate user Received signal The signal received during eavesdropping The model is as follows: in, Indicates base station To the eavesdropper distance, Representing base stations To the eavesdropper The channel vector, Indicates malicious interference The transmission power, Indicates malicious interference To the eavesdropper distance, Representing malicious interference To the eavesdropper The channel vector; Based on eavesdroppers For the first legitimate user Received signal The signal received during eavesdropping Modeling signal-to-interference-plus-noise ratio (SINR), which is expressed as: in, Representing the eavesdropper The first legitimate user of eavesdropping Total disturbance term received at that time: First legitimate user The secure rate is defined as the difference between the legitimate rate and the eavesdropping rate: in, ; According to the first legitimate user The method for calculating the confidentiality rate involves iterating over any legitimate user. The rate of confidentiality, This represents the total number of legitimate users; any legitimate user... The security rate is expressed as .

[0034] The specific steps of step 6 are as follows: Jointly optimize base stations For any legitimate user Beamforming vector of downlink channel Any legitimate user Base station access decision and base stations For any legitimate user Transmission power A confidentiality rate optimization problem is proposed, and the optimization problem is modeled as follows: in, Indicates the total number of base stations. Base station Maximum transmit power; The optimization problem is a mixed-integer nonlinear programming (MINLP) task. Its nonconvexity stems not only from the objective function in log-difference form, but also from the eavesdropper's optimal eavesdropping strategy and the high-dimensional variable coupling resulting from the total interference. Furthermore, due to… The discrete constraints cause the combinatorial search space to vary with the number of legitimate users. The number grows exponentially. Traditional iterative methods (such as successive convex approximation or branch and bound methods) often face problems of high computational complexity and sensitivity to initialization in dynamic environments.

[0035] To address this, this invention proposes a joint allocation method for physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution (DyHGNN) to capture high-order interference patterns and eavesdropping risks in complex networks and achieve efficient real-time resource allocation. The proposed dynamic hypergraph neural network is not based on a specific hypergraph modeling scheme and can perform hypergraph convolution on any network, thus filling the gap in the lack of hypergraph neural networks for secure communication optimization in complex communication networks.

[0036] By minimizing the task loss function To optimize the overall security rate : The variance of the confidentiality rate is calculated as follows: in, This represents the total number of legitimate users. The average security rate is calculated as follows: While maximizing the overall security rate, a variance penalty term is introduced to ensure the fairness of security performance among multiple users and to avoid the network falling into a local optimum where the security performance of some users is sacrificed in exchange for an overall increase.

[0037] A joint allocation system for physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution includes: The message passing module performs message passing based on the node feature matrix and the node association matrix, updates the node feature matrix, and applies the associative law of matrix multiplication to define the propagation operator of hypergraph convolution, which is used to implement step 1 of the method described in this invention. The feature matrix update module updates the multi-hop node feature matrix according to the hyperedge type and the propagation operator of the hypergraph convolution to implement step 2 of the method described in this invention; The multi-hop attention aggregation module performs attention aggregation on the multi-hop node feature matrix to obtain the final node embedding, which is used to implement step 3 of the method described in this invention. The decision header module will include any legitimate user node. The final node embeds the input spectrum allocation head, power allocation head, and beamforming head. The spectrum allocation head, power allocation head, and beamforming head jointly allocate base station access decisions. Base station access decision Selected base stations For any legitimate user The transmit power and beamforming vector are used to implement step 4 of the method described in this invention; The confidentiality rate calculation module calculates the data based on the base station access decision. Base station access decision Selected base stations For any legitimate user The transmit power and beamforming vector are used to calculate the security rate, which is used to implement step 5 of the method described in this invention; The problem optimization module proposes a security rate optimization problem and uses the negative value of the total security rate as a loss function to guide the joint allocation of physical layer security resources, thereby implementing step 6 of the method described in this invention.

[0038] A device for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution includes: Memory: Used to store computer programs that implement a joint allocation method for physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution; Processor: Used to implement a joint allocation method for physical layer security resources of heterogeneous communication networks based on multi-hop attention hypergraph convolution when executing the computer program.

[0039] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a method for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution.

[0040] Experimental verification The simulation was conducted in a large-scale heterogeneous communication network with a range of 2km × 2km. Base stations The deployment follows a fixed triangular topology, with the remaining nodes (eavesdroppers) Malicious interference and any legitimate user For each sample location, the coordinates of 30% of the nodes are randomly changed based on the previous sample to simulate limited disturbances generated by the real communication environment. Channel conditions are modeled according to the 3GPP TR 38.901 specification.

[0041] The default parameter settings are shown in the table below: parameter Value Network topology parameters Number of base stations 3 Antenna configuration (base station / legitimate user) 64 / 1 Number of legitimate users / malicious interferers 30 / 2 Number of eavesdroppers 3 Channel Model carrier frequency 3.5 GHz Path loss index 3.0–4.0 stochastic fluctuations Nakagami parameter m 2 Shadow fading standard deviation σn 8.0 dB Base station transmit power 46 dBm Interferer transmission power 30 dBm Safety parameters DyHGNN parameters Number of layers / Hidden dimensions / Number of attention heads 2 / 64 / 2 Number of jumps K 2 Total number of parameters ~156,553 GCN parameters Number of layers / Hidden dimensions 2 / 64 Total number of parameters ~150,000 DQN parameters Hidden Dimensions / Experience Replay Pool 64 / 10000 Epsilon (Initial / Minimum / Decrease) 1.0 / 0.01 / 0.995 Discount factor γ 0.99 Update frequency (target network / Q network) 100 / every 4 steps Total number of parameters ~150,000 Training parameters Initial learning rate 0.001 Learning rate decay method Exponential decay (γ=0.95) Attenuation frequency Each epoch Batch size / Number of training epochs optimizer selection of activation function 8 / 50 AdamLeakyReLU Dataset configuration Total number of samples 10,000 Dataset partitioning (training / validation / testing) 50% / 25% / 25% Sample size (training / validation / testing) 5,000 / 2,500 / 2,500 To ensure fair comparison, the baseline model is defined as follows: all models have similar parameter sizes and share the same embedding layer and task-specific output header: Based on ordinary Figure 2 Generic Edge Modeling (GCN): A baseline model based on a standard graph. In comparative experiments, the hypergraph topology will be transformed into a standard graph through clique expansion. This baseline is used to evaluate the necessity of high-order relation modeling in wireless networks.

[0042] Topology-Independent Neural Network Modeling (DQN): A benchmark for traditional reinforcement learning methods where node features are flattened and processed by a Q-network based on a multilayer perceptron (MLP). This benchmark is used to evaluate the performance improvements brought by spatial topology awareness in dynamic environments.

[0043] Figure 2 The invention was compared with different maximum aggregation hop counts. Under the default training parameter settings, the confidentiality rate comparison shows that the proposed method (DyHGNN) achieves optimal performance with 2 hops, demonstrating the effectiveness of multi-hop attention aggregation. This is because the current network depth dictates that 2-order relation aggregation can describe the complete network association. However, performance degrades as the number of hops is blindly increased due to oversmoothing, indicating that the number of hops should not be blindly increased. In subsequent experiments, the number of hops was set to 2.

[0044] Figure 3 The training convergence curve of the proposed method (DyHGNN) is shown. Under default training parameter configurations and a user count of 40, the proposed method was trained for 50 epochs. The results show that the proposed method can quickly converge to a relatively optimal security rate in the early stages of training (the first 5 epochs), and after the 20th epoch, the training loss and validation loss tend to be highly consistent and remain stable, demonstrating good convergence and generalization ability.

[0045] Figure 4 The convergence performance of the baseline methods (GCN and DQN) and the present invention was compared, and the experimental conditions were the same. Figure 2 In terms of training stability, this invention leverages the inherent stability of the hypergraph topology, significantly enhancing the robustness of the training process; while GCN is more sensitive to dynamic changes in the graph topology, easily causing oscillations in the loss curve during training, leading to unstable convergence. Regarding convergence quality, DQN's convergence accuracy is limited because it cannot effectively capture the correlation features between nodes; GCN and this invention, however, can fully utilize the graph structure's topological information to more accurately guide the gradient descent direction, ultimately converging to a lower loss value. Among these, this invention performs best due to the high-order modeling capabilities of the hypergraph.

[0046] Figure 5 and Figure 6 The performance of the baseline methods (GCN and DQN) and the present invention was compared under different heterogeneous communication network parameter configurations.

[0047] Figure 5 By cross-scanning eavesdropper density (1.0–5.0) and the number of legitimate users (10, 30, 50), the physical layer security capabilities of various methods in a 30% node perturbation scenario were systematically evaluated. The results show that even under 30% node perturbation, this invention can still effectively learn physical layer security optimization patterns in high eavesdropper density environments, fully validating the significant advantages of hypergraph convolution over traditional graph neural network methods under topological perturbation conditions.

[0048] Figure 6 The scalability of each method under a 30% node perturbation scenario was systematically evaluated using two dimensions: the number of legitimate users (10-50) and the number of malicious interferers (2, 6, 10). The results show that, under 30% node perturbation, this invention maintains a high level of physical layer security as the number of users increases, while DQN and GCN exhibit a more significant decrease in average user confidentiality rate as the number of users increases. Therefore, the experiments fully verify that hypergraph convolution has better scalability for practical deployment under topology perturbation conditions.

[0049] Finally, it should be noted that in the specification or claims of this application, relational terms such as first, second, etc., are used only to distinguish two similar objects, and do not necessarily require or imply any such actual relationship or order between these objects.

Claims

1. A method for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution, characterized in that, Includes the following steps: Step 1: Based on the node feature matrix and node association matrix, perform message passing, update the node feature matrix, and then apply the associative law of matrix multiplication to define the propagation operator of hypergraph convolution. Step 2: Update the multi-hop node feature matrix according to the type of hyperedge and the propagation operator of hypergraph convolution; Step 3: Perform attention aggregation on the multi-hop node feature matrix to obtain the final node embedding; Step 4, add any legitimate user node The final node embeds the input spectrum allocation head, power allocation head, and beamforming head. The spectrum allocation head, power allocation head, and beamforming head jointly allocate base station access decisions. Base station access decision Selected base stations For any legitimate user Transmit power and beamforming vector; Step 5, based on base station access decision Base station access decision Selected base stations For any legitimate user The transmit power and beamforming vector are used to calculate the security rate; Step 6: Propose the security rate optimization problem, and use the negative value of the total security rate as the loss function to guide the joint allocation of physical layer security resources in Step 4.

2. The method according to claim 1, characterized in that, Step 1 specifically includes the following steps: definition For any node The node feature matrix; The feature extraction process is decomposed into node-to-hyperedge aggregation and hyperedge-to-node update; The node-to-hyperedge aggregation: each hyperedge As a local observer, aggregate each hyperedge Features of connected nodes are used to generate a hyperedge representation matrix. The polymerization process is represented as: in, This represents the node adjacency matrix. Represents the node degree matrix, Represents the hypermarginality matrix. This indicates that node-level pre-normalization is performed. This aggregation process maps node domain features to the hyperedge space, reflecting the characteristics of each hyperedge. The status of collective interference within the coverage area; The hyperedge-to-node update: A node updates its hyperedge by sensing the hyperedge it belongs to. The spatial disturbance field represents the update of the node feature matrix. The updated node feature matrix is ​​represented as The update process is represented as follows: in, diagonal overside The weight matrix is ​​pre-normalized at the node level. This update process ensures that nodes can effectively access multiple overlapping hyperedges. Extract information from it while maintaining numerical stability; Represent the hyperedge matrix Substitute the updated node feature matrix Perform normalized symmetric propagation operator: By applying the associative law of matrix multiplication to extract all intermediate transformation matrices, a propagation operator for hypergraph convolution is defined. : 。 3. The method according to claim 1, characterized in that, Step 2 specifically includes the following steps: For each type of superedge Define independent propagation operators for the first... K-hop and the types of super-edges are Propagation operator of hypergraph convolution Represented as: in, This indicates that node-level prenormalization is being performed. Represents the hypermarginality matrix. It is learnable and targeted at the first The types of jump and super-edge are The hyperedge weight matrix, Given the set of hyperedge types of the input hypergraph, For the superedge type The correlation matrix; No. Jump update of node feature matrix Represented as: Initial node feature matrix , For any node The node feature matrix, The maximum number of aggregate hops, Indicates the first Jumping to the side of the Chinese Super League The trainable weight matrix, This represents the activation function.

4. The method according to claim 1, characterized in that, Step 3 specifically includes the following steps: Constructing vector dimension mapping This represents a parameterized single-hidden-layer multilayer perceptron (MLP), which reduces the original feature dimension. input vector Mapped to feature dimension The output vector Θ contains all the weight matrices and bias vectors. Represents the set of trainable parameters. Indicates the input layer bias term. Indicates the output layer bias term; Indicates shape as OK The weight matrix of the input layer of the column. Indicates shape as OK The weight matrix of the output layer of the column, Indicates the input layer bias term. Indicates the output layer bias term; A query-based multi-hop attention mechanism is employed, which focuses on the number of hops from the initial input to the maximum aggregation number. The nodes are represented by weighted fusion, and any node query vector From any node Initial node features Generate to provide foundational knowledge: in, This indicates that it has a dedicated function for querying any node. Initial node features A single-hidden-layer multilayer perceptron (MLP) with independent parameters; Then, calculate any node In the Jump node features Correlation score between To identify the feature scale that contributes most to the regression objective: in, It is a scaling factor, expressed as the square root of the hidden layer dimension. It is a predefined temperature hyperparameter that applies the Softmax function to the set of correlation scores for all hops. To obtain any node Then Attention Convergence Coefficient of Jump : Finally, any node Multi-hop information is integrated and refined through a fusion layer to capture high-order feature interactions while dynamically balancing local interference and global topological context, ultimately resulting in node embedding. The calculation is as follows: in, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically designed for multi-hop aggregation.

5. The method according to claim 1, characterized in that, In step 4, the spectrum allocation head manages any legitimate user. Base station access decision , Indicates the total number of base stations: probability distribution vector pass generate: This refers to a single-hidden-layer multilayer perceptron (MLP) that has independent parameters specifically for base station selection. Represents any final legitimate user Node embedding; The power allocation head determines the base station access decision for the power allocation task. Selected base stations For the first legitimate user Transmission power : in, Base station Maximum transmit power, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically for power distribution. Indicates the final first legal user Node embedding, Indicates the final second legal user Node embedding, This represents the total number of legitimate users. Represents an indicator function, when The indicator function takes the value 1 when the condition is met, and 0 otherwise. The power divider ensures that the desired power is met in a differentiable manner. This effectively maps the unconstrained network output onto the power simplex; The beamforming head generation is based on the base station access decision. Selected base stations For any legitimate user Beamforming vector : Satisfying the unit norm constraint ; in, , Indicates flattened output: in, This refers to a single-hidden-layer multilayer perceptron (MLP) with independent parameters specifically used for beamforming vector generation. ], Any legitimate user Beamforming vector The real part, Any legitimate user Beamforming vector The imaginary part, , Indicates length is real vectors, Represents any base station The number of antennas.

6. The method according to claim 1, characterized in that, Step 5 specifically includes the following steps: Base stations in heterogeneous communication networks Based on the channel environment, select the first legitimate user. Establish downlink channel, As the first legitimate user Base station access decision, The first legitimate user Received signal The model is as follows: in, Representing base stations For the first legitimate user The transmission power, Indicates base station To the first legitimate user distance, Indicates base station To the first legitimate user The channel vector, This is the path loss index. This indicates the conjugate transpose. For base stations For the first legitimate user Beamforming vector of the downlink channel, First legitimate user Received signal, Indicates the second legitimate user. This represents the total number of legitimate users. Represents an indicator function, when The indicator function takes a value of 1 when the condition is met, and 0 otherwise. Representing base stations For the second legitimate user The transmission power, For base stations For the second legitimate user The beamforming vector of the downlink channel satisfies , Second legitimate user Received signal, malicious interference The transmission power of the omnidirectional jamming signal. Indicates malicious interference To the first legitimate user distance, Indicates malicious interference To the first legitimate user The channel vector, Malicious interference When sending interference signals, the signals satisfy... , Represents the mathematical expectation. Additive white Gaussian noise (AWGN); Based on the first legitimate user Received signal Modeling signal-to-interference-plus-noise ratio (SINR), which is expressed as: in, The power of Gaussian white noise. As the first legitimate user The total disturbance term is represented as: eavesdropper For the first legitimate user Received signal The signal received during eavesdropping The model is as follows: in, Indicates base station To the eavesdropper distance, Representing base stations To the eavesdropper The channel vector, Indicates malicious interference The transmission power, Indicates malicious interference To the eavesdropper distance, Representing malicious interference To the eavesdropper The channel vector; Based on eavesdroppers For the first legitimate user Received signal The signal received during eavesdropping Modeling signal-to-interference-plus-noise ratio (SINR), which is expressed as: in, Representing the eavesdropper The first legitimate user of eavesdropping Total disturbance term received at that time: First legitimate user The secure rate is defined as the difference between the legitimate rate and the eavesdropping rate: in, ; According to the first legitimate user The method for calculating the security rate involves iterating over any legitimate user. The rate of confidentiality, Represents the total number of legitimate users; any legitimate user The security rate is expressed as .

7. The method according to claim 1, characterized in that, The specific steps of step 6 are as follows: Jointly optimize base stations For any legitimate user Beamforming vector of downlink channel Any legitimate user Base station access decision and base stations For any legitimate user Transmission power A confidentiality rate optimization problem is proposed, and the optimization problem is modeled as follows: in, Indicates the total number of base stations. Base station Maximum transmit power; By minimizing the task loss function To optimize the overall security rate : The variance of the confidentiality rate is calculated as follows: in, This represents the total number of legitimate users. The average security rate is calculated as follows: 。 8. A joint allocation system for physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution, characterized in that, The method described in any one of claims 1 to 7 includes: The message passing module performs message passing based on the node feature matrix and the node association matrix, updates the node feature matrix, and then applies the associative law of matrix multiplication to define the propagation operator of hypergraph convolution. The feature matrix update module updates the feature matrix of multi-hop nodes based on the type of hyperedge and the propagation operator of hypergraph convolution. The multi-hop attention aggregation module performs attention aggregation on the feature matrix of multi-hop nodes to obtain the final node embedding; The decision header module will include any legitimate user node. The final node embeds the input spectrum allocation head, power allocation head, and beamforming head. The spectrum allocation head, power allocation head, and beamforming head jointly allocate base station access decisions. Base station access decision Selected base stations For any legitimate user Transmit power and beamforming vector; The confidentiality rate calculation module calculates the data based on the base station access decision. Base station access decision Selected base stations For any legitimate user The transmit power and beamforming vector are used to calculate the security rate; The problem optimization module proposes a security rate optimization problem and uses the negative value of the total security rate as a loss function to guide the joint allocation of physical layer security resources.

9. A device for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution, characterized in that, include: Memory: for storing a computer program that implements the method for joint allocation of physical layer security resources of heterogeneous communication networks based on multi-hop attention hypergraph convolution as described in any one of claims 1 to 7; Processor: configured to implement the method for joint allocation of physical layer security resources of heterogeneous communication networks based on multi-hop attention hypergraph convolution as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method for joint allocation of physical layer security resources in heterogeneous communication networks based on multi-hop attention hypergraph convolution as described in any one of claims 1 to 7.

Images