A Fuzzy Trust Enhancement Method for Wireless Sensor Networks Based on Game Theory and Comparison Learning

Abstract

This invention discloses a fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning, relating to the field of wireless network security technology. The invention constructs a dataset by collecting node interaction data and calculating packet loss rate, latency ratio, and evidence credibility. A network-level trust matrix is ​​generated through interval type-II fuzzy algorithm inference, Karnik-Mendel type reduction, and defuzzification. A MoCo contrastive learning model is used to perform self-supervised representation learning of node trust relationships, obtaining node suspicion scores. A non-cooperative game utility function and a volunteer dilemma game model are constructed, and the suspicion scores, remaining energy, and communication costs are input to solve for a hybrid strategy Nash equilibrium, obtaining the cluster head competition probability. Based on this probability, cluster head election and secure clustering are completed, and an intra-cluster trust supervision mechanism is constructed to achieve fuzzy trust enhancement in the network. This invention improves network transmission reliability and security and has broad application value.

Description

Technical Field

[0001] This invention relates to the field of wireless network security technology, specifically to a fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning. Background Technology

[0002] Wireless sensor networks (WSNs) are widely used in environmental monitoring, industrial safety, and medical surveillance. Nodes complete tasks through autonomous sensing and collaborative transmission. However, they are typically deployed in open, harsh, or unattended environments, where node resources are limited, communication links are susceptible to interference, and they are vulnerable to internal attacks such as data tampering, identity forgery, and selective forwarding by malicious nodes, seriously threatening network reliability and data integrity. Existing security mechanisms rely on encryption and authentication, which are insufficient to defend against internal attacks and have high computational overhead, making them unsuitable for resource-constrained sensor nodes, such as:

[0003] Chinese patent (CN113891275A) discloses a trust model for underwater wireless sensor networks based on transfer learning. This scheme uses transfer learning to build a trust model, reducing the dependence on labeled data. However, it only relies on local interaction features for trust calculation, lacks global relational structure representation, and cannot reflect the overall trust status of nodes in the network. Anomaly detection relies on preset rules, has poor adaptability, and cannot dynamically identify new attack behaviors. The trust result and clustering decision are independent of each other, making it difficult to form a security-driven clustering mechanism.

[0004] Chinese patent (CN116546498A) discloses a trust assessment method for underwater wireless sensor networks based on variable membership functions. This scheme uses fuzzy logic for trust assessment and improves adaptability by adjusting the membership function. However, it only uses a type of fuzzy system, which has low tolerance for observation uncertainty and link fluctuations and is prone to misjudgment in complex environments. Anomaly identification relies on manually set thresholds, which is not intelligent enough. Furthermore, it does not effectively combine trust status with cluster structure optimization, so cluster security cannot be guaranteed.

[0005] Chinese patent (CN118042467A) discloses a wireless sensor network security protection method based on trust measurement. This scheme achieves node security discrimination through trust measurement and can filter some malicious nodes. However, the trust modeling dimension is single and does not fully integrate multi-dimensional behavioral characteristics, so the trust assessment is not comprehensive enough. It lacks structured learning and analysis of trust relationships and cannot identify complex behaviors such as coordinated attacks and disguised attacks. It does not establish a dynamic security clustering mechanism, so it is difficult to balance the overall robustness and energy efficiency of the network.

[0006] In summary, existing trust assessment and clustering methods for wireless sensor networks cannot simultaneously meet the requirements of high security, high energy efficiency, and high robustness. Therefore, there is an urgent need for a trust assessment technology to address the problems of weak uncertainty handling capabilities, insufficient trust representation, poor robustness of anomaly detection, and disconnect between security and clustering decisions in existing technologies. Summary of the Invention

[0007] To address the aforementioned technical problems, this application discloses a fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning, specifically including:

[0008] Obtain the packet loss rate (DPR), latency ratio (DLR), and evidence credibility (EVI) of node interactions to construct a node interaction dataset;

[0009] Based on the node interaction dataset, after trust inference is performed by the interval type II fuzzy algorithm IT2-FLS, a network-level trust matrix is ​​generated through Karnik-Mendel type reduction and defuzzification.

[0010] Based on the network-level trust matrix, a MoCo contrastive learning model is constructed to perform self-supervised representation learning of node trust relationships and output node suspiciousness scores.

[0011] A volunteer dilemma game model is constructed by using the utility function of non-cooperative game. The node suspicion score, residual energy and communication cost are input, and the mixed strategy Nash equilibrium is solved to obtain the cluster head competition probability.

[0012] Cluster head election and secure clustering are accomplished based on cluster head competition probability, thereby achieving fuzzy trust enhancement.

[0013] Preferably, the node interaction dataset specifically comprises: after the wireless sensor network is deployed, setting a fixed observation period for each node. For all its neighboring nodes Real-time interactive data collection is performed, and the collected content includes: nodes. To the node Total number of data packets sent, nodes Successfully received node Number of data packets sent, nodes To the node Actual transmission delay, node With nodes The number of effective interactions within the observation period T; the node is calculated based on packet transmission and reception data. To the node Forwarding packet loss rate (DPR); based on nodes To the node The latency ratio (DLR) is calculated based on the baseline latency and the actual transmission latency; the evidence credibility index (EVI) is calculated based on the number of effective interactions and a preset minimum effective interaction threshold; after normalizing the three indices, the data for each node is then processed. Normalized As a set of three-dimensional feature vectors, the feature vectors of all node pairs are integrated to form a complete node interaction dataset.

[0014] Preferably, the network-level trust matrix is ​​specifically: for the input The three feature parameters are defined into interval type II fuzzy sets by the interval type II fuzzy algorithm IT2-FLS. Each feature parameter is divided into three fuzzy semantic levels: low, medium, and high. Each fuzzy semantic level corresponds to a set of interval type II membership functions, forming upper and lower membership intervals.

[0015] An interval-type fuzzy rule base is constructed based on the fuzzy semantic level combination of three input feature parameters. Each rule corresponds to an interval-type fuzzy trust output set. Interval-type fuzzy inference is performed based on the constructed fuzzy rule base. The upper and lower membership information of the three input feature parameters is fused to calculate the global interval-type fuzzy output of the trust degree.

[0016] Based on the type-II fuzzy output set, Karnik-Mendel type reduction is performed to transform the interval type-II fuzzy set into a type-I fuzzy set. Then, defuzzification is completed using the centroid method to obtain the nodes. For nodes scalar trust value ; Traverse all node pairs in the network to generate a network-level trust matrix.

[0017] Preferably, the Karnik-Mendel type reduction process specifically involves: based on the type II fuzzy output set, iteratively solving for the switching point to gradually eliminate the uncertainty redundancy of the interval type II fuzzy set until a preset convergence condition is met, transforming the interval type II fuzzy trust output set into a type I fuzzy set, and obtaining the reduced type I fuzzy membership function. The iterative formula is as follows:

[0018]

[0019] in, For the first The switching point at +1 iteration. Let T be the number of discretized sampling points for the trust level. For the first The trust level output value for each sampling point. , The lower and upper membership functions of the interval type II fuzzy domain with trust degree T are respectively the values ​​at the th... The values ​​of each sampling point The iteration terminates when the value is less than the preset convergence threshold. After convergence, a reduced type I fuzzy membership function is obtained.

[0020] Preferably, the defuzzification using the centroid method specifically involves: applying the centroid method to the type-1 fuzzy set after KM type reduction, transforming the fuzzy trust semantics into scalar trust values ​​in the [0,1] interval. ;

[0021] Traverse all node pairs in the wireless sensor network ,in For the node that initiates the interaction, For nodes that receive interactions, each node is paired with... Corresponding scalar trust value As matrix elements, construct a network-level trust matrix.

[0022] Preferably, the MoCo contrastive learning model specifically comprises a query encoder, a key encoder, and a queue memory.

[0023] Input a network-level trust matrix, and the framework will process each node in the matrix. The corresponding row vector serves as the initial trust relationship feature vector for that node. The initial trust relationship feature vector of each node is subjected to multi-view augmentation processing to generate two different view features for the same node, which are used as inputs to the query encoder and key encoder respectively, and the output is the query feature. Bond features ;

[0024] Key features Store in queue memory The queue memory maintains a fixed capacity and calculates query characteristics. With all key features in the queue memory Similarity, based on key features of the same node As positive samples, the key features of other nodes in the queue As negative samples, the InfoNCE loss function is used to optimize the parameters of the query encoder and key encoder. The optimization is iterative to maximize the similarity of the features encoded by different views of the same node and minimize the similarity of the features encoded by different nodes.

[0025] Input the feature vector of the initial trust relationship between nodes into the trained query encoder, and output the node. Trust relationship representation vector Calculate the self-consistency outliers and memory outliers in the representation vector of each node, and fuse the two indices to obtain the node's self-consistency outliers. Suspicion rating.

[0026] Preferably, the volunteer dilemma game model specifically involves: inputting the normalized node suspicion score, remaining energy, and communication cost to construct a non-cooperative game utility function for each node; based on the constructed utility function, with the goal of having a cluster head node and being a low-suspicion, high-energy node, establishing a payoff equilibrium equation for node strategy selection; and obtaining the mixed-strategy Nash equilibrium solution for each node by solving the equation, which is the cluster head competition probability of the node.

[0027] Preferably, the non-cooperative game utility function is specifically defined as follows: the local cluster head selection of node i is considered as a two-strategy game, where each participating node can choose to become the cluster head, denoted as... Or it may not become a cluster head, denoted as ;

[0028] When node selection At that time, the cluster structure will inevitably be successfully formed, and its effect is: When a node is selected And the rest At least one node is selected. Earn income If the rest Select all nodes Then suffer losses ;

[0029] Each node has a probability ∈[0,1] selection With probability choose Then for the choice The nodes, the rest Select all nodes The probability is Therefore, choose The expected utility is:

[0030]

[0031] in, For nodes The total number of nodes in the local competitive domain. The probability of a node electing a cluster head. When no node is the cluster head in this round, the node... The losses suffered The public benefits resulting from the successful formation of the cluster structure, The expected utility of a node not running for cluster head;

[0032] choose The expected utility is independent of other nodes and is always 0. , For nodes The additional energy cost of serving as the cluster head.

[0033] Preferably, the benefit equilibrium equation is: nodes select competing cluster heads. At times, higher energy consumption and communication costs are required, so choosing a non-competitive cluster head is preferable. At this time, it does not need to bear this cost, but it needs to rely on other nodes to become cluster heads to ensure the normal operation of the network. The condition for the hybrid strategy Nash equilibrium is: nodes Choose strategy With strategy Their utility is equal, that is Substituting the utility function to solve for the cluster head competition probability, we derive the equilibrium solution formula:

[0034] Let the equilibrium probability be Substituting into

[0035]

[0036]

[0037] Thus, the nodes are obtained. The equilibrium probability of choosing to become the cluster head is:

[0038]

[0039] Substituting the energy and security mappings, we obtain the closed form:

[0040]

[0041] in, To balance the probability, The additional energy cost of the cluster head relative to its members corresponds to the cost in the game theory. , To achieve cluster head-free loss by combining the number of local nodes with the security posture, For nodes The security gain / risk index of the area;

[0042] When the cluster head has relatively higher energy consumption, that is As the size increases, the cost of the cluster head rises. The tendency is to decrease; when the security situation worsens, that is... Increase or local scale No cluster head loss when enlarged. Rise, thereby driving Increasing the size makes the network more prone to generating cluster heads, thus ensuring cluster structure formation and improving robustness.

[0043] Preferably, the cluster head election and secure clustering specifically involve: comparing the cluster head competition probability of each node with a preset threshold to filter candidate cluster head nodes; for nodes in the candidate cluster head set, combining their cluster head competition probability with the trust value in the network-level trust matrix, and using a weighted sorting method to complete the cluster head election; using a clustering strategy combining distance priority and trust priority to assign all ordinary member nodes to the corresponding cluster heads to complete secure clustering; after clustering is completed, based on the network-level trust matrix, constructing an intra-cluster trust submatrix to strengthen trust supervision between intra-cluster nodes.

[0044] Compared with the prior art, the technical solution of this application has the following technical effects:

[0045] This invention generates a network-level trust matrix by combining a type II fuzzy algorithm with KM type reduction and defuzzification. This effectively addresses uncertainties such as wireless link noise and observation fuzziness, significantly reduces the false positive rate, and improves the robustness of trust inference.

[0046] This invention achieves self-supervised representation learning of node trust relationships based on MoCo contrastive learning. It can automatically discover abnormal node patterns and output suspiciousness scores without manual annotation. It has a stronger ability to identify covert attacks and intermittent attacks, and improves the intelligence level and generalization ability of anomaly detection.

[0047] This invention constructs a volunteer dilemma game model and solves the Nash equilibrium to obtain the cluster head competition probability, achieving a dynamic balance between security factors and energy and communication costs. Compared with traditional random or simply energy-priority cluster head election, it can effectively suppress highly suspicious nodes from being elected as cluster heads and improve cluster security.

[0048] This invention completes secure clustering based on cluster head contention probability and establishes an intra-cluster trust supervision mechanism, realizing closed-loop collaboration of trust assessment, anomaly detection, cluster head election and clustering maintenance. While improving network security, it optimizes energy consumption distribution, extends network lifespan, and balances security, energy efficiency and transmission reliability.

[0049] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0050] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0052] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0053] Figure 1 The flowchart shows the overall process of a fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning.

[0054] Figure 2 This is a schematic diagram of a fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning.

[0055] Figure 3 This is an architecture diagram of the MoCo contrastive learning model in this application;

[0056] Figure 4 This is a membership function graph for each input in the embodiments of this application;

[0057] Figure 5 This is a network lifecycle data diagram of each method in the embodiments of this application;

[0058] Figure 6 This is a comparison chart of the total data packets of each method in the embodiments of this application under a malicious node;

[0059] Figure 7 This is a graph showing packet loss data for each method in the embodiments of this application at different proportions of malicious nodes;

[0060] Figure 8 This is a cluster head packet loss data diagram for each method in the embodiments of this application at different proportions of malicious nodes;

[0061] Figure 9 This is a graph showing the malicious node delay forwarding data for each method in the embodiments of this application at different proportions of malicious nodes;

[0062] Figure 10 This is a graph showing the real-time data transmission and energy efficiency ratio data of each method in the embodiments of this application. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0064] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0065] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0066] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0067] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0068] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0069] Example 1 mainly describes a fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning, such as... Figure 1 As shown, it specifically includes:

[0070] Obtain the packet loss rate (DPR), latency ratio (DLR), and evidence credibility (EVI) of node interactions to construct a node interaction dataset;

[0071] Based on the node interaction dataset, after trust inference is performed by the interval type II fuzzy algorithm IT2-FLS, a network-level trust matrix is ​​generated through Karnik-Mendel type reduction and defuzzification.

[0072] Based on the network-level trust matrix, a MoCo contrastive learning model is constructed to perform self-supervised representation learning of node trust relationships and output node suspiciousness scores.

[0073] A volunteer dilemma game model is constructed by using the utility function of non-cooperative game. The node suspicion score, residual energy and communication cost are input, and the mixed strategy Nash equilibrium is solved to obtain the cluster head competition probability.

[0074] Cluster head election and secure clustering are accomplished based on cluster head competition probability, thereby achieving fuzzy trust enhancement.

[0075] Furthermore, the node interaction dataset specifically involves setting a fixed observation period for each node after the wireless sensor network is deployed. For all its neighboring nodes Real-time interactive data collection is performed, and the collected content includes: nodes. To the node Total number of data packets sent, nodes Successfully received node Number of data packets sent, nodes To the node Actual transmission delay, node With nodes The number of effective interactions within the observation period T; the node is calculated based on packet transmission and reception data. To the node Forwarding packet loss rate (DPR); based on nodes To the node The latency ratio (DLR) is calculated using the baseline latency (preset by the channel bandwidth, transmission distance, and node hardware parameters during network deployment) and the actual transmission latency. Based on the number of effective interactions and a preset minimum effective interaction threshold, the evidence credibility index (EVI) is calculated. After normalizing the three indices, the data for each node is... Normalized As a set of three-dimensional feature vectors, the feature vectors of all node pairs are integrated to form a complete node interaction dataset.

[0076] Furthermore, the network-level trust matrix specifically refers to: the trust matrix for the input... Three feature parameters are defined into interval type II fuzzy sets using the interval type II fuzzy algorithm IT2-FLS. Each feature parameter is divided into three fuzzy semantic levels: low, medium, and high. Each fuzzy semantic level corresponds to a set of interval type II membership functions, forming upper and lower membership intervals. An interval type II fuzzy rule base (3×3×3 structure) is constructed based on the combination of fuzzy semantic levels of the three input feature parameters. Each rule corresponds to an interval type II fuzzy trust output set. Interval type II fuzzy inference is performed based on the constructed fuzzy rule base, fusing the upper and lower membership information of the three input feature parameters to calculate the global interval type II fuzzy output of the trust level. The formula is as follows:

[0077]

[0078]

[0079] in, , Let T be the interval type II fuzzy upper and lower membership functions for the trust level T. They are respectively Input value, , Let T be the upper and lower membership functions of the trust level T in the fuzzy rule base. The output value of the trust level T is where To find the minimum value of the maximum;

[0080] Based on the type-II fuzzy output set, Karnik-Mendel type reduction is performed to transform the interval type-II fuzzy set into a type-I fuzzy set. Then, defuzzification is completed using the centroid method to obtain the nodes. For nodes scalar trust value ; Traverse all node pairs in the network to generate a network-level trust matrix.

[0081] Furthermore, the Karnik-Mendel type reduction process is as follows: Based on the type II fuzzy output set, the uncertainty redundancy of the interval type II fuzzy set is gradually eliminated by iteratively solving the switching point until the preset convergence condition is met. The interval type II fuzzy trust output set is then transformed into a type I fuzzy set, resulting in the reduced type I fuzzy membership function. The iterative formula is as follows:

[0082]

[0083] in, For the first The switching point at +1 iteration. Let T be the number of discretized sampling points for the trust level. For the first The trust level output value for each sampling point. , The lower and upper membership functions of the interval type II fuzzy domain with trust degree T are respectively the values ​​at the th... The values ​​of each sampling point The iteration terminates and converges, yielding a reduced form of fuzzy membership function. .

[0084] Furthermore, defuzzification is achieved using the centroid method. Specifically, for the type I fuzzy set after KM type reduction, the centroid method is used for defuzzification, transforming the fuzzy trust semantics into scalar trust values ​​in the [0,1] interval. The formula is:

[0085]

[0086] in, The output value of the trust level T; scalar trust value The size directly represents the node For nodes The level of trust, The higher the value, the higher the level of trust.

[0087] Traverse all node pairs in the wireless sensor network ( , ),in For the node that initiates the interaction, For the nodes that receive the interaction, each node pair ( , ) corresponding scalar trust value As matrix elements, construct a network-level trust matrix.

[0088] Furthermore, the network-level trust matrix is ​​N×N dimensional, where N is the total number of network nodes, and each node has a trust level of 1; if a node Not a node The neighboring nodes have no interactive trust relationship, corresponding to =0, forming a complete network-level trust matrix.

[0089] Furthermore, a preferred approach in the reasoning process is to employ the Mamdani reasoning method, combined with the credibility of the evidence. Dynamic softening inference output: when When the evidence is sufficient, retain the original output of the rule-based reasoning; when... When the inference output is less than 0.8 (insufficient evidence), the inference output is adjusted to the neutral trust range to suppress misjudgments of trust caused by small samples or noise.

[0090] Furthermore, the MoCo contrastive learning model specifically includes a query encoder, a key encoder, and a queue memory.

[0091] Input a network-level trust matrix, and the framework will process each node in the matrix. The corresponding row vector serves as the initial trust relationship feature vector for that node. The initial trust relationship feature vector of each node is subjected to multi-view augmentation processing to generate two different view features for the same node, which are used as inputs to the query encoder and key encoder respectively, and the output is the query feature. Bond features ;

[0092] Key features Store in queue memory The queue memory maintains a fixed capacity and calculates query characteristics. With all key features in the queue memory Similarity, based on key features of the same node As positive samples, the key features of other nodes in the queue As negative samples, the InfoNCE loss function is used to optimize the parameters of the query encoder and key encoder, as shown in the formula:

[0093]

[0094] in, For nodes The contrastive learning loss value, This is a temperature parameter, with a value ranging from 0.05 to 0.1, used to adjust the discrimination of similarity. The set of all key features in the queue memory. for Key features in;

[0095] Iterative optimization aims to maximize the similarity of features encoded from different views of the same node, and minimize the similarity of features encoded from different nodes.

[0096] Input the feature vector of the initial trust relationship between nodes into the trained query encoder, and output the node. Trust relationship representation vector Calculate the self-consistency outliers and memory outliers in the representation vector of each node, and fuse the two indices to obtain the node's self-consistency outliers. The suspiciousness score is calculated using the following formula:

[0097]

[0098]

[0099]

[0100] Among them, is the suspiciousness score of node , is the weight coefficient for balancing the influence of the two indicators, with a value range of 0.4 - 0.6, is the self - consistency anomaly term of node , is the memory bank outlier term of node , is the view enhancement times of node , is the query feature encoding after the th view enhancement of node is the trust relationship characterization vector of node , ,) is for calculating the cosine similarity.

[0101] Furthermore, the enhancement method of the multi - view enhancement process adopts the combination of random masking and feature perturbation: randomly mask some trust values in the feature vector (the masking ratio is 10% - 20%), and at the same time add a small random perturbation to the un - masked trust values (the perturbation range is [-0.05, 0.05]), ensuring that the two view features originate from the same node but have slight differences, meeting the requirements of positive samples for contrastive learning.

[0102] Furthermore, the MoCo contrastive learning model adopts a three - level hierarchical structure of dual encoders + queue memory bank. The input data of its input layer includes the pre - processed initial trust relationship feature vector of the node , as well as the two view features generated after multi - view enhancement; among them, the query feature is passed into the query encoder of the encoding layer, and the key feature is passed into the key encoder of the encoding layer;

[0103] The encoding layer is located between the input layer and the memory bank layer, and is the core calculation layer of the framework. It includes a query encoder and a key encoder. The structures of the two are exactly the same and the parameters are updated synchronously. Both adopt a lightweight fully - connected network architecture, and the hierarchy is input interface → hidden layer → output interface. The network structure of the encoding layer is a 2 - layer fully - connected network. The number of neurons in the hidden layer has a value range of 64 - 128, and the output layer (feature encoding dimension d) has a value range of 32 - 64, and satisfies d < N (N is the total number of network nodes); the dimension of the input interface is the same as that of the view feature; the hidden layer uses the ReLU activation function, and the output has no activation function;

[0104] The memory bank layer adopts a first - in - first - out FIFO queue structure, and the fixed capacity has a value range of 2 - 5 times the number of network nodes N.

[0105] Furthermore, the volunteer dilemma game model is as follows: after normalizing the node suspicion score, remaining energy, and communication cost, input them to construct the non-cooperative game utility function for each node; based on the constructed utility function, with the existence of cluster head nodes and the goal of them being low-suspicion, high-energy nodes, establish the payoff equilibrium equation for node strategy selection; by solving the equation, the mixed strategy Nash equilibrium solution for each node is obtained, and this equilibrium solution is the cluster head competition probability of the node.

[0106] Furthermore, the utility function for the non-cooperative game is as follows: The local cluster head selection of node i is considered a two-strategy game, where each participating node can choose to become the cluster head, denoted as... Or it may not become a cluster head, denoted as ;

[0107] When node selection At that time, the cluster structure will inevitably be successfully formed, and its effect is: When a node is selected And the rest At least one node is selected. Earn income If the rest Select all nodes Then suffer losses ;

[0108] Each node has a probability ∈[0,1] selection With probability choose Then for the choice The nodes, the rest Select all nodes The probability is Therefore, choose The expected utility is:

[0109]

[0110] in, For nodes The total number of nodes in the local competitive domain. The probability of a node electing a cluster head. When no node is the cluster head in this round, the node... The losses suffered The public benefits resulting from the successful formation of the cluster structure, The expected utility of a node not running for cluster head;

[0111] choose The expected utility is independent of other nodes and is always 0. , For nodes The additional energy cost of serving as the cluster head.

[0112] Furthermore, the payoff equilibrium equation is as follows: Nodes compete for cluster heads. At times, higher energy consumption and communication costs are required, so choosing a non-competitive cluster head is preferable. At this time, it does not need to bear this cost, but it needs to rely on other nodes to become cluster heads to ensure the normal operation of the network. The condition for the hybrid strategy Nash equilibrium is: nodes Choose strategy With strategy Their utility is equal, that is Substituting the utility function to solve for the cluster head competition probability, we derive the equilibrium solution formula:

[0113] Let the equilibrium probability be Substituting into

[0114]

[0115]

[0116] Thus, the nodes are obtained. The equilibrium probability of choosing to become the cluster head is:

[0117]

[0118] Substituting the energy and security mappings, we obtain the closed form:

[0119]

[0120] in, To balance the probability, The additional energy cost of the cluster head relative to its members corresponds to the cost in the game theory. , To achieve cluster head-free loss by combining the number of local nodes with the security posture, For nodes The security gain / risk index of the area;

[0121] When the cluster head has relatively higher energy consumption, that is As the size increases, the cost of the cluster head rises. The tendency is to decrease; when the security situation worsens, that is... Increase or local scale No cluster head loss when enlarged. Rise, thereby driving Increasing the size makes the network more prone to generating cluster heads, thus ensuring cluster structure formation and improving robustness.

[0122] Furthermore, the cluster head election and secure clustering are as follows: the cluster head competition probability of each node is compared with a preset threshold to screen candidate cluster head nodes; for nodes in the candidate cluster head set, a weighted sorting method is used to complete the cluster head election by combining their cluster head competition probability with the trust value in the network-level trust matrix; a clustering strategy combining distance priority and trust priority is adopted to assign all ordinary member nodes to the corresponding cluster heads to complete secure clustering; after clustering is completed, an intra-cluster trust submatrix is ​​constructed based on the network-level trust matrix to strengthen trust supervision between intra-cluster nodes.

[0123] Furthermore, the cluster head election specifically involves: The primary weight is used (0.6~0.8), and the secondary weight is the average of the node's own trust value (0.2~0.4). The cluster head comprehensive score for each candidate node is calculated using the following formula:

[0124]

[0125] in, Cluster head candidate node The overall score, The weighting coefficient for the score. For nodes The probability of cluster head contention. For nodes Average trust value;

[0126] Sort the nodes by comprehensive score from highest to lowest, and select the top K nodes as the official cluster heads (K is the preset number of cluster heads, which is determined according to the network size, usually 10% to 20% of the total number of nodes).

[0127] Furthermore, secure clustering specifically involves: calculating the communication distance between each ordinary node and all formal cluster heads; and combining the cluster head nodes... Trust value compute nodes With cluster head The cluster fit is calculated using the following formula:

[0128]

[0129] in, For ordinary nodes With cluster head Cluster fit For the fit weighting coefficient, For nodes With cluster head Normalized value of communication distance For nodes For cluster head Trust value;

[0130] Ordinary nodes are assigned to the cluster corresponding to the cluster head with the highest suitability, ensuring that the trust between each member node and the cluster head is high and the communication distance is short, thereby improving the reliability and security of intra-cluster transmission.

[0131] Furthermore, trust supervision specifically involves: updating the trust value of each member node within the cluster with the cluster head and other members in real time; if a member node's suspiciousness score is... Exceeding the preset threshold ( If a node is identified as suspicious, it will be isolated or expelled by the cluster head to prevent malicious nodes from affecting cluster security. At the same time, the cluster head competition probability is recalculated periodically (observation period T), the cluster head is re-elected, and the cluster structure is adjusted to achieve dynamic maintenance of the clusters and continuously ensure the fuzzy trust enhancement effect.

[0132] This embodiment details a fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning. It constructs a dataset by collecting node interaction data and calculating packet loss rate, latency ratio, and evidence credibility. A network-level trust matrix is ​​generated through interval type-II fuzzy algorithm inference, Karnik-Mendel type reduction, and defuzzification. A MoCo contrastive learning model is used to perform self-supervised representation learning of node trust relationships, yielding node suspicion scores. A non-cooperative game utility function and a volunteer dilemma game model are constructed, and the suspicion scores, remaining energy, and communication costs are input to solve for a hybrid strategy Nash equilibrium, obtaining the cluster head competition probability. Based on this probability, cluster head election and secure clustering are performed, and an intra-cluster trust supervision mechanism is constructed to achieve fuzzy trust enhancement in the network.

[0133] Example 2, based on Example 1, describes in detail the experiment of testing this method in an open monitoring area of ​​100m × 100m, as follows:

[0134] One hundred wireless sensor nodes were deployed randomly within the monitoring area. The initial energy of each node was 2J, the communication radius was 15m, and the observation period was set to 10s. The number of packets sent was set to 1000 per cycle, the number of successfully received packets ranged from 500 to 990, the actual transmission delay was 0.1 to 0.8s, the reference delay was 0.3s, the minimum effective interaction threshold was 20 times, and the number of effective interactions ranged from 15 to 30 times.

[0135] The calculated packet loss rate (DPR) ranged from 0.01 to 0.5, the latency ratio (DLR) ranged from 0.33 to 2.67, and the evidence credibility (EVI) ranged from 0.75 to 1.5. After normalization, all values ​​fell within the [0,1] interval, forming a node interaction dataset. The packet loss rate (DPR), latency ratio (DLR), and evidence credibility were denoted as... For input variables Fuzzification is performed by defining three fuzzy sets: low, medium, and high. The membership degrees of these sets are then calculated using a piecewise linear membership function, resulting in the following: Figure 4 The membership function graphs for each input are shown below, where Figure 4 (a) is the membership function of DPR, and the corresponding formula is:

[0136]

[0137]

[0138]

[0139] Figure 4 (b) is the membership function of 2)DLR, and the formula is:

[0140]

[0141]

[0142]

[0143] Figure 4 (c) is the membership function of EVI, and the formula is:

[0144]

[0145]

[0146]

[0147] Construct a 3×3×3 fuzzy rule base, and perform Karnik-Mendel type reduction (iterative convergence threshold). The process involves 15-22 iterations, centroid defuzzification, obtaining scalar trust values, and constructing a 100×100 dimensional network-level trust matrix.

[0148] The node suspicion score was obtained using the MoCo contrastive learning framework. The trust matrix row vectors were used as initial features, with a multi-view enhancement mask ratio of 15% and a perturbation range of [-0.05, 0.05]. The framework queue capacity was 300, the temperature parameter was 0.07, and the training was iterated for 300 rounds until the loss function converged to below 0.08. The output node suspicion score was then used. Normal node ≤0.3, malicious node ≥0.7; will Input the remaining energy (0.1~0.9 after normalization) and communication cost (0.2~0.8 after normalization), and set the utility function weights. , , , Solving for the hybrid strategy Nash equilibrium yields the cluster head competition probability pᵢ, ranging from 0 to 0.8, for low-suspicion, high-energy nodes. The cluster head election threshold is preset to 0.5. Candidate nodes are screened and selected as the 10 official cluster heads through weighted sorting. The clustering fitness weight β = 0.5. After clustering, an intra-cluster trust submatrix is ​​constructed, and a suspicious node judgment threshold of 0.7 is set. The nodes are isolated.

[0149] In the network, nodes are randomly deployed and the aforementioned clustering communication mechanism is adopted. The cluster head is elected through competition and is responsible for data aggregation within the cluster and forwarding to the aggregation node. Node energy consumption follows a common wireless power consumption model, and data interaction is executed iteratively in rounds. A certain proportion of malicious nodes are set. The attack scenario can implement internal attacks such as selective packet loss and delayed forwarding during the forwarding process to simulate malicious collaboration and masquerading behavior in an open environment. Comparative methods GSCTO, DST_WOA, and ACTAR were selected, and repeated experiments were conducted under the same network size, communication load, and attack configuration to obtain the following results: Figures 5-10 Experimental data graph:

[0150] like Figure 5 The network lifecycle data graphs for each method shown demonstrate that, from a network lifecycle perspective, this method significantly delays the network's entry into the concentrated death phase, resulting in a smoother overall curve. In this set of experiments, the final node death round was approximately 3.4 × 10³ rounds, while the comparative methods generally ended earlier (e.g., GSCTO around 3.0 × 10³ rounds). This verifies that, when malicious behavior and link uncertainty are superimposed, this method can more effectively avoid invalid communication and energy waste caused by untrusted cluster heads and forwarding nodes, thereby extending the stable operating range of the network.

[0151] like Figure 6 The chart showing the comparison of total data packets for each method under malicious nodes demonstrates that as the proportion of malicious nodes increases from 10% to 50%, the total number of forwarded packets for each method decreases, but this method consistently maintains a higher service carrying capacity. For example, when the proportion of malicious nodes is 10%, the total number of data packets for this method reaches 264,130, significantly higher than ACTAR's 175,646. This verifies that this method reduces the probability of malicious nodes entering critical forwarding paths by anomaly representation and security constraints on trust relationships, thereby enabling the network to maintain a more stable data aggregation capability even when attacks intensify.

[0152] like Figure 7The packet loss data graphs shown for each method at different proportions of malicious nodes demonstrate that the increase in packet loss with the increase of the proportion of malicious nodes is an objective law. However, the packet loss growth of this method is slower and the overall level is lower. When the proportion of malicious nodes is 50%, the number of packets lost by this method is 2188, while ACTAR reaches 5233. This verifies that this method can more effectively avoid high-risk nodes in the cluster formation and forwarding decision-making stages, reduce the reliability loss caused by selective packet loss, and thus improve the robustness of data arrival.

[0153] like Figure 8 The data graphs of cluster head packet loss for each method at different proportions of malicious nodes are shown. This method has a stronger ability to suppress malicious packet loss. When the proportion of malicious nodes is 50%, the number of malicious packet loss is 74, while the ACTAR is 702. This verifies that this method can effectively limit malicious nodes from becoming cluster heads or key forwarding nodes, thereby compressing their attack influence radius from a mechanism perspective and making it difficult for attacks to spread on a large scale.

[0154] like Figure 9 The diagram shows the malicious node delay forwarding data for each method under different malicious node proportions. This method can still significantly reduce malicious delay events even with a majority of malicious nodes. For example, when the malicious node proportion is 10%, the number of malicious delays for this method is only 6, while ACTAR is 126. This phenomenon shows that the anomaly identification and security constraints of this method can effectively weaken the interference caused by delay masquerading, laying the foundation for subsequent improvement of timely transmission rate.

[0155] like Figure 10 (a) shows the timely data transmission of each method at different proportions of malicious nodes. As the proportion of malicious nodes increases, the timely transmission rate of each method decreases, but the decline of this method is the slowest and always the best. When the proportion of malicious nodes is 50%, this method still maintains 97.81%, while ACTAR drops to 94.90%. This shows that this method not only reduces packet loss, but also reduces the waiting and retransmission chain induced by malicious delay, making the end-to-end forwarding process more stable, so as to maintain a high quality of service even under high attack intensity.

[0156] like Figure 10 (b) shows the node energy efficiency ratio data of each method under different malicious node ratios. This method has higher energy efficiency under all malicious node ratios, and its advantage is more prominent under high malicious node ratios. For example, when the malicious node ratio is 50%, the energy efficiency of this method is 97.83%, while that of ACTAR is 94.89%. This shows that by suppressing untrusted nodes from playing a key role and reducing ineffective communication overhead, this method can achieve better energy utilization while ensuring reliability and timeliness, thereby extending the network's sustainable operation capability.

[0157] This embodiment details the experiment testing the proposed method in a 100m×100m open monitoring area. The experiment simultaneously compared it with existing methods GSCTO, DST_WOA, and ACTAR. The experimental results verify that the proposed method absorbs noise using interval membership, enabling selective packet loss and delayed forwarding to be more stably mapped to low trust. Contrastive learning learns the consistency of normal relationships from the global trust matrix, giving nodes deviating from the pattern higher suspicion levels and maintaining distinguishability even when the proportion of malicious activity increases. The game-theoretic mechanism embeds suspicion levels into cluster head competition, dynamically suppressing high-risk nodes from becoming cluster heads and reducing the attack amplification effect.

[0158] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning, characterized in that, include: Obtain the packet loss rate (DPR), latency ratio (DLR), and evidence credibility (EVI) of node interactions to construct a node interaction dataset; Based on the node interaction dataset, after trust inference is performed by the interval type II fuzzy algorithm IT2-FLS, a network-level trust matrix is ​​generated through Karnik-Mendel type reduction and defuzzification. Based on the network-level trust matrix, a MoCo contrastive learning model is constructed to perform self-supervised representation learning of node trust relationships and output node suspiciousness scores. A volunteer dilemma game model is constructed by using the utility function of non-cooperative game. The node suspicion score, residual energy and communication cost are input, and the mixed strategy Nash equilibrium is solved to obtain the cluster head competition probability. Cluster head election and secure clustering are accomplished based on cluster head competition probability, thereby achieving fuzzy trust enhancement.

2. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 1, characterized in that, The node interaction dataset specifically refers to: after the wireless sensor network is deployed, setting a fixed observation period for each node. For all its neighboring nodes Real-time interactive data collection is performed, and the collected content includes: nodes. To the node Total number of data packets sent, nodes Successfully received node Number of data packets sent, nodes To the node Actual transmission delay, node With nodes The number of effective interactions within the observation period T; the node is calculated based on packet transmission and reception data. To the node Forwarding packet loss rate (DPR); based on nodes To the node The latency ratio (DLR) is calculated based on the baseline latency and the actual transmission latency; the evidence credibility index (EVI) is calculated based on the number of effective interactions and a preset minimum effective interaction threshold; after normalizing the three indices, the data for each node is then processed. Normalized As a set of three-dimensional feature vectors, the feature vectors of all node pairs are integrated to form a complete node interaction dataset.

3. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 2, characterized in that, The network-level trust matrix is ​​specifically defined as: for the input The three feature parameters are defined into interval type II fuzzy sets by the interval type II fuzzy algorithm IT2-FLS. Each feature parameter is divided into three fuzzy semantic levels: low, medium, and high. Each fuzzy semantic level corresponds to a set of interval type II membership functions, forming upper and lower membership intervals. An interval-type fuzzy rule base is constructed based on the fuzzy semantic level combination of three input feature parameters. Each rule corresponds to an interval-type fuzzy trust output set. Interval-type fuzzy inference is performed based on the constructed fuzzy rule base. The upper and lower membership information of the three input feature parameters is fused to calculate the global interval-type fuzzy output of the trust degree. Based on the type-II fuzzy output set, Karnik-Mendel type reduction is performed to transform the interval type-II fuzzy set into a type-I fuzzy set. Then, defuzzification is completed using the centroid method to obtain the nodes. For nodes scalar trust value ; Traverse all node pairs in the network to generate a network-level trust matrix.

4. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 3, characterized in that, The Karnik-Mendel type reduction process is as follows: based on the type II fuzzy output set, the uncertainty redundancy of the interval type II fuzzy set is gradually eliminated by iteratively solving the switching point until the preset convergence condition is met, and the interval type II fuzzy trust output set is transformed into a type I fuzzy set to obtain the reduced type I fuzzy membership function.

5. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 4, characterized in that, The defuzzification process using the centroid method specifically involves: applying the centroid method to the type-I fuzzy set after KM type reduction, transforming the fuzzy trust semantics into scalar trust values ​​in the [0,1] interval. ; Traverse all node pairs in the wireless sensor network ,in For the node that initiates the interaction, For nodes that receive interactions, each node is paired with... Corresponding scalar trust value As matrix elements, construct a network-level trust matrix.

6. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 1, characterized in that, The MoCo contrastive learning model specifically includes a query encoder, a key encoder, and a queue memory. Input a network-level trust matrix, and the framework will process each node in the matrix. The corresponding row vector serves as the initial trust relationship feature vector for that node. The initial trust relationship feature vector of each node is subjected to multi-view augmentation processing to generate two different view features for the same node, which are used as inputs to the query encoder and key encoder respectively, and the output is the query feature. Bond features ; Key features Store in queue memory The queue memory maintains a fixed capacity and calculates query characteristics. With all key features in the queue memory Similarity, based on key features of the same node As positive samples, the key features of other nodes in the queue As negative samples, the InfoNCE loss function is used to optimize the parameters of the query encoder and key encoder. The optimization is iterative to maximize the similarity of the features encoded by different views of the same node and minimize the similarity of the features encoded by different nodes. Input the feature vector of the initial trust relationship between nodes into the trained query encoder, and output the node. Trust relationship representation vector Calculate the self-consistency outliers and memory outliers in the representation vector of each node, and fuse the two indices to obtain the node's self-consistency outliers. Suspicion rating.

7. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 1, characterized in that, The volunteer dilemma game model is as follows: after normalizing the node suspicion score, remaining energy, and communication cost, input them to construct the non-cooperative game utility function for each node; based on the constructed utility function, with the existence of cluster head nodes and the goal of them being low-suspicion, high-energy nodes, establish the payoff equilibrium equation for node strategy selection; by solving the equation, the mixed strategy Nash equilibrium solution for each node is obtained, and this equilibrium solution is the cluster head competition probability of the node.

8. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 7, characterized in that, The non-cooperative game utility function is specifically defined as follows: The local cluster head selection of node i is considered a two-strategy game, where each participating node can choose to become the cluster head, denoted as... Or it may not become a cluster head, denoted as ; When node selection At that time, the cluster structure will inevitably be successfully formed, and its effect is: When a node is selected And the rest At least one node is selected. Earn income If the rest Select all nodes Then suffer losses ; Each node has a probability ∈[0,1] selection With probability choose Then for the choice The nodes, the rest Select all nodes The probability is Therefore, choose The expected utility is: in, For nodes The total number of nodes in the local competitive domain. The probability of a node electing a cluster head. When no node is the cluster head in this round, the node... The losses suffered The public benefits resulting from the successful formation of the cluster structure, The expected utility of a node not running for cluster head; choose The expected utility is independent of other nodes and is always 0. , For nodes The additional energy cost of serving as the cluster head.

9. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 8, characterized in that, The specific benefit equilibrium equation is: Nodes select competing cluster heads. At times, higher energy consumption and communication costs are required, so choosing a non-competitive cluster head is preferable. At this time, it does not need to bear this cost, but it needs to rely on other nodes to become cluster heads to ensure the normal operation of the network. The condition for the hybrid strategy Nash equilibrium is: nodes Choose strategy With strategy Their utility is equal, that is Substituting the utility function to solve for the cluster head competition probability, we derive the equilibrium solution formula: Let the equilibrium probability be Substituting into Thus, the nodes are obtained. The equilibrium probability of choosing to become the cluster head is: Substituting the energy and security mappings, we obtain the closed form: in, To balance the probability, The additional energy cost of the cluster head relative to its members corresponds to the cost in the game theory. , To achieve cluster head-free loss by combining the number of local nodes with the security posture, For nodes The security gain / risk index of the area; When the cluster head has relatively higher energy consumption, that is As the size increases, the cost of the cluster head rises. The tendency is to decrease; when the security situation worsens, that is... Increase or local scale No cluster head loss when enlarged. Rise, thereby driving Increasing the size makes the network more prone to generating cluster heads, thus ensuring cluster structure formation and improving robustness.

10. The fuzzy trust enhancement method for wireless sensor networks based on game theory and contrastive learning according to claim 1, characterized in that, The cluster head election and secure clustering process specifically involves: comparing the cluster head competition probability of each node with a preset threshold to filter candidate cluster head nodes; for nodes in the candidate cluster head set, combining their cluster head competition probability with the trust value in the network-level trust matrix, and using a weighted sorting method to complete the cluster head election; employing a clustering strategy combining distance priority and trust priority to assign all ordinary member nodes to the corresponding cluster heads, thus completing secure clustering; after clustering is completed, based on the network-level trust matrix, constructing an intra-cluster trust submatrix to strengthen trust supervision between intra-cluster nodes.

Images