Trust management method based on ensemble learning in ocean cloud edge collaborative network

Abstract

The application discloses a kind of ocean cloud edge cooperative network based on integrated learning trust management method, comprising the following steps: (1) trust evidence collection;(2) trust value calculation and convergence;(3) trust value data processing;(4) trust evaluation;(5) trust update;The application first considers the dynamic characteristics of underwater channel and the subjectivity of node trust evidence collection, proposes a new trust evidence quantization method, then uses the method of integrated learning to complete trust decision by edge server in cooperation with cloud server, and updates the network structure according to the evaluation result.The application gives a new measurement method for communication trust and node trust, optimizes the trust value calculation method, and improves the accuracy of trust evaluation result.

Description

Technical fields:

[0001] This invention belongs to the field of radio network technology, specifically relating to a trust assessment and optimization method in a trust management mechanism for marine cloud-edge collaborative networks. Background technology:

[0002] Ocean cloud-edge collaborative networks are a new network paradigm that leverages edge computing and cloud computing technologies to enhance network processing capabilities, building upon Underwater Wireless Sensor Networks (UWSNs). They play a crucial role in marine exploration, monitoring, and protection, monitoring marine ecosystems and climate patterns, and providing information for resource exploration and disaster prevention. Traditional UWSNs consist of numerous sensor nodes deployed underwater, collecting and transmitting various data from the deep ocean via acoustic communication. However, the limited energy and resources of UWSNs restrict their development. By incorporating cloud-edge collaboration, complex computations are offloaded to higher-level devices with greater computing power, reducing resource consumption on lower-level sensor nodes and minimizing energy loss. Furthermore, the more efficient processing of tasks by more powerful devices leads to greater efficiency. However, because these networks operate in unattended environments, they face security challenges and are vulnerable to various attacks (such as data tampering and denial-of-service attacks). These issues can lead to degraded network performance, reduced data quality, and unmet needs. Therefore, effective network security mechanisms are essential.

[0003] Traditional encryption algorithms are unsuitable for resource-constrained sensor networks, and authentication schemes cannot defend against internal attacks. Trust management mechanisms, as a lightweight and efficient security mechanism, play a crucial role in network security research, characterized by low computational complexity and the ability to resist internal network attacks. Therefore, this invention utilizes a trust management mechanism to ensure network security.

[0004] However, most trust management mechanisms use relatively fixed trust value calculation methods (the coefficients and weights in their calculation formulas are pre-set and lack adaptability). Furthermore, most trust assessment methods define devices with a trust value score (normalized score) higher than 0.5 as trusted devices, but such fixed assessment mechanisms cannot adapt to changing and dynamic network environments. Therefore, this invention considers the characteristics of device interaction and the unique instability of underwater acoustic channels, proposing an optimized trust value calculation method; in addition, it proposes a trust assessment method based on ensemble learning to improve the accuracy of the assessment results.

[0005] The relevant research literature on trust management methods for underwater networks is as follows:

[0006] 1. In 2019, Guangjie Han et al. proposed a trust evaluation method based on support vector machine in "A Synergetic Trust Model Based on SVM in Underwater Acoustic Sensor Networks". This model collects the interaction records between nodes, the remaining energy of the nodes, and the accuracy of the data packets sent by the evaluated nodes to quantify the trust value of the nodes. Then, a hyperplane is used to distinguish between reliable and unreliable nodes.

[0007] 2. In 2020, Jinfang Jiang et al. proposed a dynamic trust evaluation method using C4.5 decision trees in their paper "A Dynamic Trust Evaluation and UpdateMechanism Based on C4.5 Decision Tree in Underwater Wireless Sensor Networks". This method collects trust evidence, discretizes the trust value according to fuzzy theory, and then calculates the information gain ratio between different categories to construct a C4.5 decision tree, thereby classifying the trustworthiness of different devices.

[0008] 3. In 2022, Jiaxin Du et al. integrated edge computing technology into underwater acoustic sensor networks in their paper "n Edge-Computing-Enabled Trust Mechanism for Underwater Acoustic Sensor Networks". They used autonomous underwater vehicles (AUVs) as edge devices to provide low-latency trust modeling services. In addition, they proposed a trust model based on attention mechanism and long short-term memory (LSTM) to adaptively adjust the weighted combination weights between various trust evidences and dynamically evaluate the trustworthiness of the device. Summary of the Invention:

[0009] To address the aforementioned issues, this invention presents a trust management method based on ensemble learning in a marine cloud-edge collaborative network. Based on the characteristics of node interaction and the unique dynamic properties of the underwater acoustic channel, it proposes novel quantification methods for communication trust and energy trust. Furthermore, it evaluates the trustworthiness of nodes based on ensemble learning.

[0010] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0011] A trust management method based on ensemble learning in a marine cloud-edge collaborative network includes the following steps:

[0012] Step 1: Collection of Trust Evidence

[0013] Nodes collect and record trust evidence by listening to the behavior of their neighbors based on time slots, namely the results of communication between other nodes and themselves, the remaining energy of other nodes, and the data packets transmitted to them by other nodes; in each time slot, they record trust evidence for trust calculation, namely communication trust, node trust, and data trust.

[0014] Step Two: Trust Value Calculation and Aggregation

[0015] After collecting the three types of trust evidence—communication trust, node trust, and data trust—a node delivers it to the cluster head node of its cluster. The cluster head node then uploads this data to the edge server. The edge server quantifies this evidence and calculates the trust evaluations made by its lower-level nodes for other nodes. Each edge server is responsible for aggregating the evaluations of all nodes in its area of ​​responsibility, calculating the trust value of all nodes, and then uploading it to the cloud server. The cloud server then summarizes the trust values ​​of nodes in each network area.

[0016] Step 3: Trust Value Data Processing

[0017] After the cloud server aggregates the trust values ​​of all nodes in the network, it combines the evaluation values ​​from various trust assessment perspectives into a feature vector, then uses a clustering algorithm to add labels to the vector, and finally uses these labeled data as input for ensemble learning.

[0018] Step 4: Trust Assessment

[0019] After data preprocessing, the server uses an ensemble learning method to evaluate the trustworthiness of each device. The ensemble learning used in this invention consists of two layers. The first layer is a base learner composed of three intelligent classification methods, which is responsible for directly processing the data and providing a prediction of the classification result. The second layer is a meta-learner, which performs logistic regression on the results of the first layer and makes the final classification result after weighing the factors. After processing the trust values ​​of all nodes in the network in step three, the ensemble learning method is used to evaluate the trust of each node.

[0020] Step 5: Trust Update

[0021] After the cloud server assesses the trustworthiness of a node based on its trust value, it sends the assessment results to the edge server. The edge server then schedules its subordinate nodes based on the results and updates the network structure to eliminate abnormal nodes in order to maintain network security.

[0022] In step one above, the way a node collects trust evidence based on time slots is by listening to the communication behavior of other nodes and recording the interaction results with other nodes, including records of successful and failed communication, the remaining energy of other nodes, and the content of data packets transmitted by other nodes.

[0023] In step one above, the specific methods for setting communication trust, node trust, and data trust are as follows:

[0024] Communication trust is described using the number of failed interactions, the number of successful interactions, and the current channel state:

[0025]

[0026] Where s represents the number of successful interactions and f represents the number of failed interactions. Assuming node communication follows a beta distribution, when nodes i and j have no interaction, their communication records are empty. To avoid this, we added a constant to the calculation method; furthermore, ω... u ω represents the channel quality at time u. The better the channel quality, the higher the ω value. u The higher the value of u, the greater the weight of u in the network's time slot. Due to the varying characteristics of underwater acoustic channels, the channel quality of each time slot may differ from the previous one. Therefore, the interaction results of each time slot are weighted and summed based on the channel quality of that time slot, resulting in a higher weighting for communication results with good channel quality in the communication trust calculation. The purpose of this design is to give higher weight to communication results with good channel quality. Considering the irregular increase in noise in underwater communication environments, which interferes with the quality of communication links, it is difficult to distinguish between communication failures caused by node subjective factors or link interference when channel quality is poor. Communication failures with good channel quality, however, largely stem from the node itself.

[0027] Node trust is described using the difference in remaining energy between two nodes:

[0028]

[0029] Among them, E res (t,i) represents the remaining energy of the i-th node at time t, E res (t,j) represents the remaining energy of the j-th node at time t, and Eini represents the initial energy of the node;

[0030] This invention posits that, over time, nodes experiencing similar network processes will inevitably consume similar amounts of energy. However, nodes with significantly different remaining energy levels will participate in different network events within the same timeframe. This could be due to a node encountering a DoS attack or a node's reluctance to actively participate in network events due to a greedy strategy. Individuals tend to trust others who have similar experiences. Therefore, this invention optimizes traditional node (energy consumption) trust by incorporating the subjectivity of node evaluation, making the calculation of trust values ​​more reasonable.

[0031] Data trust is described using the content of data packets transmitted by nodes:

[0032]

[0033] In the formula, μ is the average value of this dataset, and v j The data content at target node j; the greater the difference between the data at node j and the average value, the lower its data trust value.

[0034] In step three above, the edge server calculates the trust evaluations of other nodes by the nodes in its assigned area, uploads these trust values ​​to the cloud server, and then the cloud server aggregates and updates the global trust values. After establishing the global trust values, it is necessary to calculate the expected trustworthiness evaluation for each device. The specific calculation method is as follows:

[0035]

[0036] In this invention, when calculating the expected value of the evaluated node, based on the small-world theory and the six degrees of separation theory, it is required that the number of evaluations received by the evaluated node is at least 6. If this condition is met, the calculation follows the first condition in the above formula; otherwise, a weighted calculation follows the second condition. Where T... n (j) represents the evaluation received by device j at the nth evaluation angle, T n (u,j) is the evaluation made by device u to device j from the nth evaluation perspective; the first formula is the average expected evaluation of device j that meets the conditions, and δ in the second formula... u These are weights, calculated as follows:

[0037]

[0038] In the formula T n (p,u) represents the evaluation (i.e. the calculated trust value) made by node p on node u in the nth trust dimension (communication trust, node trust, or data trust).

[0039] For node u that participated in evaluating j, if l nodes participated in evaluating u, then This indicates the credibility of u. The higher the credibility of u, the more worthwhile its evaluation of j is to be considered.

[0040] In step four above, after the cloud server processes the credibility evaluation of each node from various evaluation perspectives, it constructs input samples for the trust assessment. A clustering algorithm is then used to label each sample, constructing the complete trust assessment model input. Finally, an ensemble learning method is used to evaluate the credibility of all nodes. The ensemble learning method designed in this invention consists of two layers. The first layer comprises three base learners that directly learn from the input data and predict classification results. To broaden the learning perspective, this invention selects three models with different learning methods: support vector machines that use hyperplanes to classify samples, decision trees that use information gain ratios to classify continuous data, and multilayer perceptrons. Considering the scarcity of underwater data, this invention uses k-fold cross-validation to classify all samples to improve data utilization. After the first layer performs classification prediction on the credibility of the samples, all prediction results are concatenated and combined as the input to the meta-learner in the second layer. Although the meta-learner does not directly learn from the original data, it makes feature extraction beyond the sample features by weighing the prediction results of the base learners, i.e., "learning" of "learning". The meta-learning implemented in this invention specifically refers to logistic regression, which ultimately achieves trust assessment through two layers of learning and data flow.

[0041] In step five above, the trust update occurs after each trust assessment. The cloud server updates the stored node trustworthiness based on the most recent trust assessment result, and then sends the result to the edge server. The edge server, based on the received trust assessment result, notifies its subordinate cluster head nodes to perform operations. The trust assessment result can have three possible outcomes:

[0042] (i.) The new assessment results show no new malicious nodes compared to the previous results, which means that there is no need to update the network topology, the network security is not under new threat, and the edge server will not notify its subordinate cluster head nodes to perform network scheduling.

[0043] (ii.) If a new malicious node appears in the new assessment results, the network topology needs to be updated. The edge server will find the cluster head node of the cluster where the malicious node is located and notify the cluster head to remove the malicious node from its cluster network and prohibit it from participating in information transmission in the network.

[0044] (iii.) A new, normal node appears in the new evaluation results compared to the previous one. This situation is related to the node's historical behavior. If the node's reliability assessment happens to be on the verge of being untrustworthy, it may be misjudged by the network. Since underwater acoustic networks have transmission delays, the cloud server may reassess the node as reliable after supplementary evidence is obtained. In this case, the new normal node needs to be examined (because it was mistakenly judged as unreliable and has been removed from its cluster). In order to rejoin the network (from the node's own perspective, it knows it is a normal node and not a malicious one), the node needs to provide detailed logs and evidence to higher-level devices (these logs and evidence can include communication records, interaction messages, verification results, etc.). After the higher-level devices verify that the node is indeed normal, it is allowed to rejoin the network.

[0045] The beneficial effects of this invention are:

[0046] The method of this invention first considers the dynamic characteristics of the underwater channel and the subjectivity of node trust evidence collection, and proposes a new trust evidence quantification method. Then, the edge server and the cloud server use an ensemble learning method to complete the trust decision and update the network structure based on the evaluation results.

[0047] First, based on the characteristics of node interaction and the instability of underwater acoustic channels, a novel metric for communication trust and node trust is presented, and the calculation method for trust values ​​is optimized. Furthermore, a trust management method based on ensemble learning is proposed, improving the accuracy of trust assessment results. Attached Figure Description

[0048] Figure 1 This is a flowchart of the present invention;

[0049] Figure 2 This is a schematic diagram of k-fold cross-validation;

[0050] Figure 3 This is a schematic diagram of trust assessment based on ensemble learning. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] The purpose of this invention is to optimize the trust management mechanism in traditional underwater sensor networks. Firstly, it introduces a collaborative mechanism of edge computing and cloud computing to optimize the trust value calculation method in underwater acoustic networks, and proposes a trust management method based on ensemble learning in marine cloud-edge collaborative networks. In this method, based on the characteristics of node interaction and the unstable nature of underwater acoustic channels, a novel measurement method for communication trust and node trust is presented. Furthermore, this invention proposes a trust management method based on ensemble learning in marine cloud-edge collaborative networks, such as... Figure 1 As shown, the specific steps include:

[0053] Step 1: Collection of Trust Evidence

[0054] Nodes collect and record trust evidence by listening to the behavior of their neighbors based on time slots, namely the results of communication between other nodes and themselves, the remaining energy of other nodes, and the data packets transmitted to them by other nodes; in each time slot, they record trust evidence for trust calculation, namely communication trust, node trust, and data trust.

[0055] Step Two: Trust Value Calculation and Aggregation

[0056] After collecting the three types of trust evidence—communication trust, node trust, and data trust—a node delivers it to the cluster head node of its cluster. The cluster head node then uploads this data to the edge server, which quantifies the evidence and calculates the trust evaluation made by the node. Each edge server is responsible for aggregating the evaluations of all nodes in its area of ​​responsibility, calculating the trust value of all nodes, and then uploading it to the cloud server. The cloud server then summarizes the trust values ​​of nodes in each network area.

[0057] Step 3: Trust Value Data Processing

[0058] After the cloud server aggregates the evaluations made by all nodes in the network, it needs to calculate the expected evaluation of each node (the trust value of other nodes in evaluating the node), combine the evaluation values ​​from various trust evaluation perspectives into a feature vector, and then use a clustering algorithm to add labels to the vector. Finally, these labeled data are used as input for ensemble learning.

[0059] Step 4: Trust Assessment

[0060] After data preprocessing, the server uses an ensemble learning method to evaluate the trustworthiness of each device. The ensemble learning used in this invention consists of two layers. The first layer is a base learner composed of three intelligent classification methods, which is responsible for directly processing the data and providing a prediction of the classification result. The second layer is a meta-learner, which performs logistic regression on the results of the first layer and makes the final classification result after weighing the factors. After processing the trust values ​​of all nodes in the network in step three, the ensemble learning method is used to evaluate the trust of each node.

[0061] Step 5: Trust Update

[0062] After the cloud server assesses the trustworthiness of a node based on its trust value, it sends the assessment results to the edge server. The edge server then schedules its subordinate nodes based on the results and updates the network structure to eliminate abnormal nodes in order to maintain network security.

[0063] In step one above, the way a node collects trust evidence based on time slots is by listening to the communication behavior of other nodes and recording the interaction results with other nodes, including records of successful and failed communication, the remaining energy of other nodes, and the content of data packets transmitted by other nodes.

[0064] In step one above, the specific methods for setting communication trust, node trust, and data trust are as follows:

[0065] Communication trust is described using the number of failed interactions, the number of successful interactions, and the current channel state:

[0066]

[0067] Where s represents the number of successful interactions and f represents the number of failed interactions. Assuming node communication follows a beta distribution, when nodes i and j have no interaction, their communication records are empty. To avoid this, we added a constant to the calculation method; furthermore, ω... u ω represents the channel quality at time u. The better the channel quality, the higher the ω value. u The higher the value of u, the greater the weight of u in the network's time slot. Due to the varying characteristics of underwater acoustic channels, the channel quality of each time slot may differ from the previous one. Therefore, the interaction results of each time slot are weighted and summed based on the channel quality of that time slot, resulting in a higher weighting for communication results with good channel quality in the communication trust calculation. The purpose of this design is to give higher weight to communication results with good channel quality. Considering the irregular increase in noise in underwater communication environments, which interferes with the quality of communication links, it is difficult to distinguish between communication failures caused by node subjective factors or link interference when channel quality is poor. Communication failures with good channel quality, however, largely stem from the node itself.

[0068] Node trust is described using the difference in remaining energy between two nodes:

[0069]

[0070] Among them, E res (t,i) represents the remaining energy of the i-th node at time t, E res (t,j) represents the remaining energy of the j-th node at time t, and Eini represents the initial energy of the node;

[0071] This invention posits that, over time, nodes experiencing similar network processes will inevitably consume similar amounts of energy. However, nodes with significantly different remaining energy levels will participate in different network events within the same timeframe. This could be due to a node encountering a DoS attack or a node's reluctance to actively participate in network events due to a greedy strategy. Individuals tend to trust others who have similar experiences. Therefore, this invention optimizes traditional node (energy consumption) trust by incorporating the subjectivity of node evaluation, making the calculation of trust values ​​more reasonable.

[0072] Data trust is described using the content of data packets transmitted by nodes:

[0073]

[0074] In the formula, μ is the average value of this dataset, and v j The data content at target node j; the greater the difference between the data at node j and the average value, the lower its data trust value.

[0075] In step three above, the edge server calculates the trust evaluations of other nodes by the nodes in its assigned area, uploads these trust values ​​to the cloud server, and then the cloud server aggregates and updates the global trust values. After establishing the global trust values, it is necessary to calculate the expected trustworthiness evaluation for each device. The specific calculation method is as follows:

[0076]

[0077] In this invention, when calculating the expected value of the evaluated node, based on the small-world theory and the six degrees of separation theory, it is required that the number of evaluations received by the evaluated node is at least 6. If this condition is met, the calculation follows the first condition in the above formula; otherwise, a weighted calculation follows the second condition. Where T... n (j) represents the evaluation received by device j at the nth evaluation angle, T n (u,j) is the evaluation made by device u to device j from the nth evaluation perspective; the first formula is the average expected evaluation of device j that meets the conditions, and δ in the second formula... u These are weights, calculated as follows:

[0078]

[0079] In the formula T n (p,u) represents the evaluation (i.e. the calculated trust value) made by node p on node u in the nth trust dimension (communication trust, node trust, or data trust).

[0080] For node u that participated in evaluating j, if l nodes participated in evaluating u, then This indicates the credibility of u. The higher the credibility of u, the more worthwhile its evaluation of j is to be considered.

[0081] In step four above, after the cloud server processes the credibility evaluation of each node from various evaluation perspectives, it constructs input samples for trust assessment. A clustering algorithm is used to label each sample, constructing a complete trust assessment model input. Then, an ensemble learning method is used to evaluate the credibility of all nodes. The ensemble learning method designed in this invention consists of two layers (see...). Figure 3 The first layer consists of three base learners that directly learn from the input data and predict classification results. To broaden the learning perspective, this invention selects three models with different learning methods: a support vector machine that uses a hyperplane to classify samples, a decision tree that uses information gain ratio to classify continuous data, and a multilayer perceptron. Considering the scarcity of underwater data, to improve data utilization, this invention uses k-fold cross-validation (see...). Figure 2 The first layer classifies all samples; after the first layer performs classification prediction on the credibility of the samples, it concatenates and combines all the prediction results as the input of the meta-learner in the second layer; although the meta-learner does not directly learn from the original data, it makes feature extraction beyond the sample features by weighing the prediction results of the base learner, that is, the "learning" of "learning"; the meta-learning implemented in this invention specifically refers to logistic regression, which ultimately achieves trust assessment through two layers of learning and data flow.

[0082] In step five above, the trust update occurs after each trust assessment. The cloud server updates the stored node trustworthiness based on the most recent trust assessment result, and then sends the result to the edge server. The edge server, based on the received trust assessment result, notifies its subordinate cluster head nodes to perform operations. The trust assessment result can have three possible outcomes:

[0083] (i.) The new assessment results show no new malicious nodes compared to the previous results, which means that there is no need to update the network topology, the network security is not under new threat, and the edge server will not notify its subordinate cluster head nodes to perform network scheduling.

[0084] (ii.) If a new malicious node appears in the new assessment results, the network topology needs to be updated. The edge server will find the cluster head node of the cluster where the malicious node is located and notify the cluster head to remove the malicious node from its cluster network and prohibit it from participating in information transmission in the network.

[0085] (iii.) A new, normal node appears in the new evaluation results compared to the previous one. This situation is related to the node's historical behavior. If the node's reliability assessment happens to be on the verge of being untrustworthy, it may be misjudged by the network. Since underwater acoustic networks have transmission delays, the cloud server may reassess the node as reliable after supplementary evidence is obtained. In this case, the new normal node needs to be examined (because it was mistakenly judged as unreliable and has been removed from its cluster). In order to rejoin the network (from the node's own perspective, it knows it is a normal node and not a malicious one), the node needs to provide detailed logs and evidence to higher-level devices (these logs and evidence can include communication records, interaction messages, verification results, etc.). After the higher-level devices verify that the node is indeed normal, it is allowed to rejoin the network.

[0086] (iii.) A new normal node appears in the new evaluation results compared to the previous one. This situation is related to the node's historical behavior. If the node's reliability evaluation happens to be on the edge of being untrustworthy, it may be misjudged by the network. In this case, the new normal node needs to be examined. In order to rejoin the network, the node needs to provide detailed logs and evidence to the higher-level devices (these logs and evidence may include communication records, interaction messages, verification results, etc.). After the higher-level devices confirm that the node is indeed normal, they allow it to rejoin the network.

[0087] In summary:

[0088] This invention discloses a trust management method based on ensemble learning in marine cloud-edge collaborative networks, which optimizes the trust management mechanism in traditional underwater sensor networks. Firstly, based on the characteristics of node interactions and the instability of underwater acoustic channels, this method presents a novel metric for communication trust and node trust, optimizes the calculation method of trust values, and improves the accuracy of trust assessment results.

[0089] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A trust management method based on ensemble learning in a marine cloud-edge collaborative network. Its features include the following steps: (1) Collection of Trust Evidence Nodes collect trust evidence by listening to the behavior of their neighbors based on time slots; in each time slot, trust evidence used for trust calculation is recorded, namely communication trust, node trust, and data trust. (2) Trust value calculation and aggregation After collecting the three types of trust evidence—communication trust, node trust, and data trust—a node delivers it to the cluster head node of its cluster. The cluster head node then uploads this data to the edge server, which then uses the evidence to quantify the trust value. Each edge server is responsible for aggregating the evaluations of all nodes within its assigned area, calculating the trust value of all nodes, and then uploading it to the cloud server. The cloud server then aggregates the trust values ​​of nodes in each network area. (3) Trust value data processing After aggregating the evaluations made by all nodes in the network, the cloud server needs to calculate the expected evaluation of each node, then use a clustering algorithm to add labels to these expected values, and finally use these labeled data as input for ensemble learning. (4) Trust assessment The cloud server integrates a variety of intelligent classification methods. Using these methods as base learners, each base learner will learn from the processed trust values ​​and labels. During learning, k-fold cross-validation is used. For the results of the first layer of learning, a comprehensive second learning is adopted. After two rounds of learning, the final evaluation result is given. After processing the node trust values ​​of the entire network in step (3), the ensemble learning method is used to evaluate the trust of each node. (5) Trust update After the cloud server assesses the trustworthiness of a node based on its trust value, it sends the assessment results to the edge server. The edge server then schedules its subordinate nodes based on the results and updates the network structure to eliminate abnormal nodes in order to maintain network security.

2. The trust management method based on ensemble learning in the marine cloud-edge collaborative network according to claim 1, characterized in that... In step (1), the trust evidence collected by the node based on time slots by listening to the communication behavior of other nodes includes: the result of each communication between the node and other nodes, the remaining energy of other nodes, and the content of data packets transmitted by other nodes.

3. The trust management method based on ensemble learning in the marine cloud-edge collaborative network according to claim 1, characterized in that... The specific methods for setting communication trust, node trust, and data trust in step (1) are as follows: The communication trust is described using the number of failed interactions, the number of successful interactions, and the current channel state. The communication trust between node i and node j is calculated as follows: In the formula, s represents the number of successful interactions and f represents the number of failed interactions; This represents the channel quality at each moment; the better the channel quality, the better. The higher the u, the better; u represents each time slot of the network. Due to the changing characteristics of the underwater acoustic channel, the channel quality of each time slot may be different from the previous time slot. Therefore, the interaction results of each time slot are combined with the channel quality of the time slot and weighted and summed, so that the communication results when the channel quality is good account for a higher proportion in the calculation of communication trust. t represents the total number of time slots from network deployment to this trust calculation; The node trust is described using the difference in remaining energy between two nodes: in, Let represent the remaining energy of the i-th node at time t. This represents the remaining energy of the j-th node at time t. Represents the initial energy of the node; The data trust is: In the formula, f(v) is the probability density function of the content transmitted by the node; μ is the average value of this dataset. The data content at target node j; the greater the difference between the data at node j and the average value, the lower its data trust value.

4. The trust management method based on ensemble learning in the marine cloud-edge collaborative network according to claim 1, characterized in that... The expected evaluation method for each node in step (3) is as follows: In the formula, m represents the number of nodes that evaluate node j. This refers to the expected evaluation of the j-th node from the n-th trust perspective. The evaluation made by the u-th node to the j-th node; weight The calculation method is as follows: In the formula This represents the evaluation that node p makes of node u on the nth trust dimension. To evaluate the number of nodes in node u.

5. The trust management method based on ensemble learning in the marine cloud-edge collaborative network according to claim 1, characterized in that... In step (4), when trust evaluation is achieved by ensemble learning, k-fold cross-validation is required on the input data to improve the efficiency of data utilization. Each time, select 1 / k of the data, use the model to learn the remaining k-1 / k data, and then validate the model on the selected k data. The trust assessment scheme consists of two layers of intelligent learning. The first layer uses three different intelligent classification methods to learn the features of the data, including support vector machines, decision trees, and multilayer perceptrons, and makes classification predictions based on its learning results. Then, the meta-learner in the second layer learns from the classification predictions made by the base learner in the first layer and makes the final classification result after weighing the results.

6. The trust management method based on ensemble learning in the marine cloud-edge collaborative network according to claim 1, characterized in that... The issuance of the evaluation results in step (5) will result in three scenarios: (i.) If the new assessment results do not show any new malicious nodes compared to the previous results, there is no need to update the network topology, and the edge server will not notify its subordinate cluster head nodes to perform network scheduling. (ii.) The new assessment results show that there are new malicious nodes compared to the previous one. At this time, the network topology needs to be updated. The edge server will find the cluster head node of its cluster and notify the cluster head to remove the malicious node from the network and prohibit it from participating in the information transmission in the network. (iii.) If a new normal node appears in the new evaluation results, it is necessary to examine the new normal node and require it to provide detailed logs and evidence to the higher-level devices. After the higher-level devices confirm that the node is indeed normal, it can be allowed to rejoin the network.

Images