An edge node-oriented dynamic trust modeling calculation method and system

Abstract

The application discloses an edge node-oriented dynamic trust modeling calculation method and system, and the method is to convert an industrial edge network into a dynamic heterogeneous graph; a dynamic trust calculation model composed of a snapshot input layer, a space aggregation layer, a time aggregation layer and a prediction layer connection is constructed for training, the trained dynamic trust calculation model is deployed on the edge network side, a current time network snapshot collected and a historical snapshot are input into the dynamic trust calculation model, and a security decision module is used to trigger a dynamic security strategy according to a preset trust threshold for a low trust degree node or an abnormal interaction relationship; the system comprises a data modeling module, a dynamic trust calculation model and a security decision module. Through quantitative modeling and real-time calculation on the trust relationship between nodes, the application solves the problem of how to accurately and real-timely quantitatively evaluate the complex and implicit trust relationship between nodes in a dynamic open and highly heterogeneous industrial edge environment.

Description

Technical Field

[0001] This invention relates to the fields of industrial Internet of Things and edge computing security technology, and in particular to a dynamic trust modeling calculation method and system for edge nodes. Background Technology

[0002] As internet and computing architectures evolve towards a cloud-edge-device collaborative model, edge nodes have become crucial hubs connecting massive numbers of terminal devices with core cloud intelligence. These nodes engage in high-frequency data exchange and task collaboration in open environments, exhibiting highly dynamic and heterogeneous network topology and node behavior. They also face security threats such as malicious internal nodes, data tampering, and collusive attacks. Against this backdrop, establishing a precise and adaptive dynamic trust mechanism between nodes is the core foundation for ensuring the secure and reliable operation of edge computing systems. This also presents a significant challenge to trust assessment technologies, requiring them to simultaneously address dynamism, heterogeneity, and complex dependencies.

[0003] Existing trust management technologies are ill-equipped to fully address the aforementioned challenges, with limitations manifesting at multiple levels. Traditional static trust models, such as those based on fixed rules or pre-defined whitelists, generate trust values ​​that remain constant over long periods. They fail to capture the temporal evolution of node behavior and struggle to perceive dynamically evolving trust relationships due to network topology changes. In edge environments with frequent node joining and leaving and rapidly changing connection states, they often lead to misjudgments due to evaluation lag. To improve adaptability, machine learning-based trust assessment methods have been extensively studied, typically relying on direct historical interaction data between nodes for modeling. However, these methods have fundamental limitations. At the information utilization level, their modeling is inherently local and shallow, primarily relying on the direct interaction history of paired nodes. They fail to explicitly represent and utilize multi-hop trust dependencies transmitted through the network topology, resulting in insufficient utilization of the rich information inherent in the network structure, a one-sided evaluation perspective, and weak ability to identify complex threat patterns requiring multi-step collaboration. At the data representation level, most models implicitly assume that the network is homogeneous, that is, all nodes and connections are regarded as the same type. This is seriously inconsistent with the real edge network, which is composed of heterogeneous devices, diverse communication protocols and different task flows. Forcibly compressing multidimensional heterogeneous information into a homogeneous model for processing will result in the loss of key semantic information, introduce representation bias from the data source and limit the upper limit of accuracy.

[0004] Furthermore, existing research on models designed to handle dynamics still faces challenges in architectural design. Many methods attempt to introduce temporal processing units to capture trust evolution, but typically couple the spatial topological dependencies of node behavior with the temporal evolution within the same nonlinear model. This coupled modeling makes it difficult for the model to clearly separate and fully utilize different patterns of spatiotemporal features, resulting in limited analytical and adaptive capabilities for edge network dynamics where instantaneous abrupt changes in topological connections and relatively gradual changes in node behavior are intertwined. In addition, regarding the data modeling format of the model input, even when existing works introduce graph structures, they are mostly limited to simple homogeneous graphs or only contain a few types of heterogeneous graphs. There is a lack of a complete and unified formal system to accurately describe the multidimensional heterogeneous elements in the network, such as devices, protocols, interaction relationships, and node attributes, and their dynamic evolution. This makes the model input itself a highly simplified and information-lossy image of the real network, fundamentally limiting the potential performance of the model.

[0005] In summary, existing technologies fall short in adaptability, representation capabilities, architectural design, and data modeling foundations when dealing with dynamic, open, and highly heterogeneous edge network environments, making it difficult to achieve accurate, reliable, and adaptive node trust assessment. Therefore, there is an urgent need in this field for an innovative and systematic modeling and computation framework to fundamentally address these multi-layered problems. Summary of the Invention

[0006] Purpose of the Invention: Addressing the limitations of existing static trust models in effectively handling dynamic changes in network topology, the difficulty of using traditional machine learning methods to uniformly model the multidimensional heterogeneity of network devices and protocols, and the failure of existing technologies to fully explore and utilize multi-hop trust dependencies transmitted between nodes through intermediate nodes, this invention aims to propose a dynamic trust modeling and calculation method and system for industrial edge nodes. This method and system are applicable to dynamic, open, and heterogeneous edge computing environments. By quantitatively modeling and calculating the trust relationships between nodes in real time, it solves the problem of how to accurately and in real-time quantify and evaluate the complex and implicit trust relationships between nodes in dynamic, open, and highly heterogeneous industrial edge environments.

[0007] Technical solution: To achieve the above objectives, the dynamic trust modeling and calculation method for edge nodes of this invention includes the following steps:

[0008] 1) The core of this step is to abstract and transform the dynamic, heterogeneous industrial edge network in the physical world into a formal mathematical structure—a dynamic heterogeneous graph—that can be directly processed by the computational model. The specific transformation and modeling process is as follows: First, continuously collect raw interaction data from the network. Each data record includes the source node, target node, interaction evidence strength, timestamp, and the protocol or type followed by the interaction. Based on the raw interaction data, determine a dynamic heterogeneous graph. ;in, Represents the set of all edge nodes; Represents a set of edges of multiple types, edges Additional weights Used for quantizing node representation and The strength or quality of interaction under relation type r, where relation type r is control information flow, sensor data flow, or physical neighbor relationship; It represents a predefined set of relationship types, such as: control information flow, sensor data flow, physical proximity relationship, and collaborative relationship for the same task; It represents a set of node attributes, including multi-dimensional features such as device type, real-time computing resource status, and historical behavior summary.

[0009] Since the network state changes continuously over time, the time axis is discretized into continuous time slots. The network state observed in any specific time slot t is a snapshot of the dynamic graph at the current moment. .here, This represents the subset of nodes that are active in time slot t. This represents the set of edges actually observed within that time slot and their current weights. Through this modeling process, dynamic and heterogeneous network entities and relationships are transformed into structured graph data, laying a precise data foundation for subsequent hierarchical trust computation.

[0010] 2) Construct a hierarchical dynamic trust computation model: Construct a four-layer dynamic trust computation model consisting of a snapshot input layer, a spatial aggregation layer, a temporal aggregation layer, and a prediction layer connected in sequence.

[0011] 2.1) The snapshot input layer will be a sequence of T consecutive dynamic graph snapshots arranged in chronological order. As input to the model, the temporal evolution information of the network is preserved. Representing the earliest historical snapshot, This represents the most recent snapshot. Representing chronological order, in The next dynamic graph snapshot. This step preserves information about the evolution of the network topology and interactions over time, including changes in the node set, the addition or removal of the edge set, and changes in edge weights.

[0012] 2.2) Spatial aggregation layer in each independent snapshot Internally, computation is performed to generate a spatial embedding representation by aggregating multi-level neighbor information of the target node. The computation process explicitly distinguishes between the two roles of nodes in trust interactions: "trust initiator" and "trust receiver."

[0013] When calculating first-order trust propagation, the neighbor information directly connected to the target node is aggregated. When the first node u acts as the trust receiver, the receiver role embedding of the first node u is calculated. First, check all in-degree neighbors of the first node. Trust score given Through a learnable rating transformation matrix ,have to: Then, construct the message from neighbor v. ,in Let v be the current feature vector of the neighbor v. This represents a vector concatenation operation. Finally, aggregate the messages from all in-degree neighbors: ,in This is an aggregation function. Similarly, the embedding of the first node u as the trust initiator is calculated. , represented as Finally, by fusing these two role embeddings through a fully connected layer, a unified spatial embedding of the first node u in the current snapshot is obtained. , represented as The weight matrix of the fully connected layer. Compress the concatenated vector into Compared to direct splicing, fully connected layers can extract high-order interaction features, avoiding redundant information and high-dimensional computational overhead.

[0014] When performing high-order trust propagation calculations, multi-hop neighbor information is aggregated by stacking multiple layers (let's call it L layers) of the above aggregation operations, thereby capturing the transitivity of trust relationships in the network. In the... layer In the vectorized representation, the trust level of node u is as follows: ,in It is the first The learnable transformation matrix of the layer, This is the original trust score. The message is constructed as follows: ,in Indicates that neighbor v is in the th order. Layer embedding, Representing the The layer passes messages from neighbor v to node u. The aggregated messages obtain the first node u at the [node position]. The layer is embedded as a trust receiver:

[0015] ;

[0016] Similarly, the embedding of the first node u as the trust initiator is calculated. :

[0017] ;

[0018] Finally, the output of this layer is obtained by fusing the dual-role embedding: ;

[0019] here, This represents the final spatial embedding of the first node u at the first layer of the spatial aggregation layer (step 2.2), which incorporates the dual role information of "trust receiver" and "trust initiator". Among them, Represents a non-linear activation function. Representing the Learnable weight matrix for layer fusion operation Representing the The bias vector for layer fusion operations.

[0020] 2.3) The temporal aggregation layer fuses the temporal features of the first node u across multiple consecutive time snapshots to capture the evolutionary patterns of its behavior. The temporal features of the first node u are derived from the first node u at different times. Spatial embedding constitutes a sequence. .in Indicates the time of the first node u. The spatial aggregation layer output embedding is used to capture the temporal evolution of node behavior, resulting in the final feature vector of the first node u that integrates spatiotemporal features. The final feature vector of the second node v The process is as follows: First, for each moment... Features Add time difference coding Obtain spatiotemporal context features ,in The reference time for the current moment, here, is the sequence. This serves as the temporal feature used in subsequent calculations; then, historical moments are calculated. The temporal characteristics of the target time attention weights :

[0021] ;

[0022] in, For learnable parameters, For activation functions; Representing time Features for predicting target time The importance weights are determined. Finally, the temporal features are weighted and aggregated, and max pooling is used to capture global temporal patterns, resulting in the final feature vector of the first node u that integrates spatiotemporal features. The final feature vector of the second node v :

[0023] ;

[0024] .

[0025] 2.4) The prediction layer will pair the target nodes. The final embedding of (i.e., the first node u and the second node v) and Input a multilayer perceptron, and the multilayer perceptron outputs a scalar score. After activation by the sigmoid function, the predicted trust relationship score between the first node u and the second node v is obtained. , The range of values The closer the value is to 1, the higher the level of trust predicted by the dynamic trust calculation model.

[0026] The dynamic trust calculation model employs a time-constrained binary cross-entropy loss function. Training is performed to minimize the predicted values. With genuine trust The difference between them. The loss function is defined as follows:

[0027] ;

[0028] in, This represents the total number of edges in the training set. The set of edges representing time t, This indicates the pair of nodes at time t. The predicted probability of a trust relationship existing. Indicates the negative sampling node The predicted probability, This represents the negative sampling quantity.

[0029] 3) Deploy the trained dynamic trust calculation model on the edge network. The system collects network interaction data in real time, constructs a network snapshot of the current moment, and inputs this snapshot along with several previous historical snapshots into the dynamic trust calculation model. The dynamic trust calculation model outputs a real-time trust scoring matrix among all nodes in the network. A security decision module then uses preset trust thresholds... (For example For low-trust nodes or abnormal interactions, dynamic security policies are triggered. These dynamic security policies include: generating security alarm logs, isolating suspicious nodes in resource scheduling, or blocking specific access control commands initiated from low-trust nodes.

[0030] The decision-making logic of the security decision module is as follows: for any element in the trust scoring matrix... ,like If the interaction is deemed untrustworthy, the corresponding strategy will be triggered.

[0031] This invention presents a dynamic trust modeling and calculation system employing a dynamic trust modeling and calculation method oriented towards edge nodes. The system includes: a data modeling module for real-time data acquisition and construction of dynamic heterogeneous graph snapshots; a hierarchical dynamic trust calculation model for performing spatial aggregation, temporal aggregation, and trust score prediction; and a security decision module for executing security policies based on the trust score matrix output by the dynamic trust calculation model and preset thresholds.

[0032] The dynamic trust modeling and calculation method for edge nodes proposed in this invention aims to achieve three main objectives: First, to uniformly represent multi-dimensional heterogeneous elements in the network, such as device types, communication protocols, and interaction modes, overcoming the biases of traditional homogeneous modeling; second, to explicitly capture and calculate the dependence and influence of trust relationships transmitted through multi-hop paths in complex network topologies, rather than being limited to direct interactions; and third, to adaptively integrate the spatiotemporal evolution characteristics of node behavior, that is, to simultaneously analyze the topological dependence of trust in the spatial dimension and the dynamic changes in the temporal dimension, thereby achieving continuous and accurate calculation of the trust state of nodes.

[0033] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0034] 1) Powerful modeling capabilities and accurate heterogeneous representation. Step 1) formally defines the industrial edge network as a dynamic heterogeneous graph. This invention unifies multi-dimensional heterogeneous elements such as devices, protocols, and tasks. Compared with existing technologies that simplify networks into homogeneous graphs, this invention can more accurately represent real industrial scenarios, laying a more realistic theoretical and data foundation for subsequent calculations and fundamentally reducing computational errors caused by modeling biases.

[0035] 2) A complete computational framework with coordinated spatiotemporal features. The hierarchical dynamic trust computation model constructed in step 2)—"snapshot input - spatial aggregation - temporal aggregation - prediction"—systematically solves the dynamic trust computation problem. The spatial aggregation layer (step S2.2) specifically handles topological dependencies within a single snapshot, supporting multi-hop trust propagation and overcoming the limitations of existing methods that only utilize first-order neighbor information. The temporal aggregation layer (step S2.3) dynamically fuses historical sequence information through an attention mechanism, specifically handling the temporal evolution of node behavior. This spatiotemporally separated yet coordinated computational framework achieves a comprehensive and refined characterization of the spatial dependencies and temporal evolution of trust relationships.

[0036] 3) The computation direction is refined, and the accuracy is significantly improved. In the spatial aggregation layer (step S2.2), the aggregation computation of nodes is innovatively divided into two dual roles: "trust receiver" and "trust initiator," and message construction and aggregation are performed separately for each role. and The calculation of trust scores is more precise. This refined directional modeling allows the dynamic trust calculation model to more accurately depict the asymmetry of trust transfer. Compared with the aggregation methods in existing technologies that do not distinguish directions or simply average, it significantly improves the calculation accuracy and rationality of trust scores.

[0037] 4) Strong dynamic adaptability and good environmental adaptability. The attention mechanism used in the time aggregation layer (step S2.3) can adaptively assign weights to information from different historical moments. This allows the dynamic trust computation model to automatically focus on historical behavioral patterns most relevant to the current moment. Combined with direct processing of dynamic graph snapshot sequences, this invention can automatically adapt to dynamic changes in network topology and non-stationary evolution of node behavior, overcoming the limitations of static or fixed-time-window models. It exhibits stronger robustness and practicality in dynamic, open industrial edge environments. Attached Figure Description

[0038] Figure 1 This is an overall framework diagram of the dynamic trust modeling and computing system for industrial edge nodes provided in this embodiment of the invention;

[0039] Figure 2 This is a schematic diagram of the dynamic trust calculation model in an embodiment of the present invention;

[0040] Figure 3 shows the indicator values ​​on different trust datasets in the embodiments of the present invention;

[0041] Figure 3(a) shows the verification on the trust dataset OTC;

[0042] Figure 3 (b) shows the verification on the trust dataset Alpha. Detailed Implementation

[0043] The dynamic trust modeling and calculation method for edge nodes adopted in this invention includes the following steps:

[0044] Step 1): System initialization and construction of dynamic heterogeneous graph.

[0045] 1.1) Deploy the dynamic trust modeling system for edge nodes of this invention on an industrial edge computing platform to continuously monitor the network and collect data in the form of triples: <source node ID, target node ID, interaction evidence, timestamp, protocol type>. The interaction evidence can be converted into a trust score according to the specific application. In this embodiment, a successful interaction is +1, and a failure or anomaly is -1.

[0046] 1.2) Define a set of relation types R. Based on the "protocol type" field in the data, define three typical network edge protocol relations, namely R={'MODBUS_TCP', 'OPC_UA', 'MQTT'}.

[0047] 1.3) Using a fixed duration (e.g., 5 minutes) as a time window, slice the historical interaction data to generate a series of dynamic graph snapshots. Each snapshot In the process, if nodes i and j interact within a time window t, then an edge is established between them, and the weight of the edge is... This represents the average value of the interactive evidence within that time window.

[0048] Step 2): Construct and train a hierarchical dynamic trust computation model.

[0049] 2.1) Configuration of Hierarchical Dynamic Trust Computation Model: Number of Spatial Aggregation Layers This is to capture trust propagation within second-order neighbors. Initial node embedding dimension. The time aggregation layer uses the most recent The historical snapshot embeddings at each time step. The prediction layer MLP contains a 64-dimensional hidden layer. The optimizer uses Adam with an initial learning rate of 0.005. A preset trust threshold is set before training. For use in subsequent security decisions.

[0050] 2.2) Training Data Preparation: Training samples are constructed using a sliding window method. For each time point, using... Four consecutive snapshots are used as input to the dynamic trust computation model, in order to The edges that actually exist in the snapshot are used as positive samples, and a time-constrained negative sampling strategy is used to generate negative samples, which are used together to predict the trust relationship at time t+1.

[0051] 2.3) Training the dynamic trust computation model: On a historical dataset, minimize the time-constrained binary cross-entropy loss. Train the model for the target until the loss function converges on the validation set.

[0052] Step 3): Online assessment and security response.

[0053] 3.1) Deploy the trained dynamic trust calculation model on an edge server.

[0054] 3.2) When running online, the system maintains a sliding time window to build the latest network snapshot in real time. .

[0055] 3.3) will Along with the three previous historical snapshots, the dynamic trust calculation model is input into the already deployed model. The dynamic trust calculation model completes the calculation within milliseconds and outputs a real-time trust score matrix among all nodes in the network. , of which elements This represents the current level of trust that node i has in node j.

[0056] (4) Security policy execution. The security decision module scans the trust scoring matrix in real time. For any pair of nodes ,like If the request is deemed untrustworthy, the dynamic trust calculation system will automatically execute at least one of the following strategies: record a detailed security log containing the node ID, time, and trust score, and report an alarm; temporarily isolate node j in the resource scheduler and not assign it any new tasks; block specific critical control commands (such as emergency stop or parameter reset) sent from node i to node j and request manual confirmation.

[0057] Figure 1 The overall architecture of the dynamic trust computing system of this invention is illustrated. From left to right, the architecture includes: an industrial edge network perception layer (for real-time acquisition of node interaction data), a dynamic trust computing core layer (containing a dynamic heterogeneous graph modeling module, a hierarchical dynamic trust computing model, and a trust database), and a security decision application layer. Arrows indicate the data flow direction. Raw data, after modeling, is input into the dynamic trust computing model. The trust scoring matrix output by the dynamic trust computing model is sent to the security decision module, ultimately triggering specific response strategies.

[0058] Figure 2 The detailed four-layer structure of the hierarchical dynamic trust computation model is shown. The bottom layer is the snapshot input layer, and the input is a time series snapshot sequence. Above the snapshot input layer is the spatial aggregation layer, which contains two sub-modules: first-order neighbor aggregation and higher-order neighbor aggregation. This demonstrates how nodes aggregate neighbor information along different edge types. Above the spatial aggregation layer is the temporal aggregation layer, showing how node embedding sequences are weighted and fused using an attention weight calculation module. The top layer is the prediction layer, demonstrating how node embeddings are processed by an MLP classifier to output the final trust relationship prediction probability.

[0059] Experimental verification:

[0060] To verify the effectiveness of this modeling and computation framework, a simulation experiment was conducted on the publicly available dynamic trust dataset Bitcoin-OTC. Transaction users were simulated as industrial edge nodes, and "trust" and "distrust" scores were used as interaction evidence. The experimental setup was consistent with step 2) of the implementation example. The dynamic trust computation model under this framework was compared with classic benchmark models (such as Guardian and GATrust).

[0061] As shown in Figures (3a) and (3b) of Figure 3, the verification process and key results are as follows: Under the same data partitioning and evaluation metrics (AUC, F1-score), the dynamic trust calculation model trained based on the framework of this invention achieves an AUC value of 0.85, a 5% improvement over the best baseline model; and an F1-score of 0.82, a 7% improvement over the best baseline model. This result demonstrates that the dynamic heterogeneous graph modeling and hierarchical dynamic trust calculation model proposed in this invention can effectively improve the accuracy of node trust relationship calculation in dynamic open environments. Specifically, this dynamic trust calculation model can better identify complex trust relationships generated by the temporal evolution of behavioral patterns or through multi-hop node transmission, thereby making more reliable assessments in dynamically changing network environments.

Claims

1. A dynamic trust modeling and calculation method for edge nodes, characterized in that: Includes the following steps: 1) Transform industrial edge networks into dynamic heterogeneous graphs Then, the time axis is discretized into continuous time slots; where, It is the set of all edge nodes; Represents a set of edges of multiple types, edges Additional weights Used for quantizing node representation and The strength or quality of interaction under relation type r; Represents a set of relation types; Represents a collection of node attributes; 2) Construct a dynamic trust computation model consisting of a snapshot input layer, a spatial aggregation layer, a temporal aggregation layer, and a prediction layer. The process is as follows: 2.1) The snapshot input layer will display a sequence of T dynamic graph snapshots arranged in time. As input, the changes in the node set, the addition or removal of the edge set, and the changes in edge weights in the network are preserved; where Representing the earliest historical snapshot, In order to be in The next animated snapshot, Represents the most recent snapshot; 2.2) Spatial aggregation layer in snapshot The computation is performed internally, generating spatial embedding by aggregating multi-order neighbor information of the target node; during first-order trust propagation computation, neighbor information directly connected to the target node is aggregated; when the first node u acts as a trust receiver, the receiver role embedding is calculated. First, check all in-degree neighbors. Trust score given Through the rating transformation matrix ,have to: Then construct the message from neighbor v. ,in Let v be the current feature vector of the neighbor v. This represents a vector concatenation operation; it aggregates messages from all in-degree neighbors. ,in The aggregation function is used to derive the embedding of the first node u as the trust initiator. : Finally, the spatial embedding of the first node u in the current snapshot is obtained through a fully connected layer. : ; 2.3) The time aggregation layer fuses the temporal features of the first node u across multiple consecutive time snapshots, and the temporal features are denoted as sequences. This yields the final feature vector of the first node u, which incorporates spatiotemporal features. The final feature vector of the second node v ;in Indicates the time of the first node u. The spatial aggregation layer output embedding; the process is as follows: First, for each time step... Features Add time difference coding Obtain spatiotemporal context features ,in Use the current time as a reference; then, calculate the time. For the target time attention weights : ; in, For learnable parameters, For activation functions; Representing time Features for predicting target time The importance weights are determined; finally, the temporal features are weighted and aggregated, and max pooling is used to capture global temporal patterns, resulting in the final feature vector of the first node u that integrates spatiotemporal features. The final feature vector of the second node v ; ； ； 2.4) The prediction layer will and Inputting data into a multilayer perceptron, the trust relationship score between the first node u and the second node v is obtained. ,in The scalar score for the multilayer perceptron. The function used is the sigmoid function; a time-constrained binary cross-entropy loss function is employed. Train the dynamic trust calculation model; 3) Deploy the trained dynamic trust calculation model on the edge network. Input the current network snapshot and historical snapshots into the dynamic trust calculation model, and output a real-time trust scoring matrix among all nodes in the network. Use the security decision module to determine the trust threshold. For nodes with low trust or abnormal interactions, generate security alert logs, isolate suspicious nodes in resource scheduling, or block access control commands initiated from nodes with low trust.

2. The dynamic trust modeling and calculation method for edge nodes according to claim 1, characterized in that: In step 2.2), during the calculation of higher-order trust propagation, multi-hop neighbor information is aggregated by stacking L-layer aggregation operations, thereby capturing the transitivity of trust relationships in the network. In the layer In the vectorization, the trust level of the first node u is: ,in It is the first The learnable transformation matrix of the layer, It is the original trust score; Message constructed as ,in Indicates that neighbor v is in the th order. Layer embedding, Representing the The message passed from neighbor v to node u; the aggregated message obtains the message from the first node u at the 1st rank. The layer is embedded as a trust receiver: ; Similarly, the embedding of the first node u as the trust initiator is calculated. : ； Finally, we get the number. Layer output: ; in, Represents a non-linear activation function. Representing the Learnable weight matrix for layer fusion operation Representing the The bias vector for layer fusion operations; This represents the final spatial embedding of the first node u in the first layer of the spatial aggregation layer.

3. The dynamic trust modeling and calculation method for edge nodes according to claim 1, characterized in that: In step 2.4), the binary cross-entropy loss function for: ； in, This represents the total number of edges in the training set. The set of edges representing time t, This indicates the pair of nodes at time t. The predicted probability of a trust relationship existing. Indicates the negative sampling node The predicted probability, This represents the negative sampling quantity.

4. The dynamic trust modeling and calculation method for edge nodes according to claim 1, characterized in that: In step 3), the elements in the trust rating matrix are... ,like If so, the interaction is deemed untrustworthy.

5. The dynamic trust modeling and calculation method for edge nodes according to claim 1, characterized in that: In step 1), the relation type r is a control information flow, a sensor data flow, or a physical neighbor relation.

6. The dynamic trust modeling and calculation method for edge nodes according to claim 1, characterized in that: In step 1), This includes control information flow, sensor data flow, physical proximity relationships, and collaborative relationships within the same task.

7. The dynamic trust modeling and calculation method for edge nodes according to claim 1, characterized in that: In step 1), It includes device type, real-time computing resource status, and historical behavior summary.

8. The dynamic trust modeling and calculation method for edge nodes according to claim 1, characterized in that: In step 2), the spatial aggregation layer includes a first-order neighbor aggregation submodule and a higher-order neighbor aggregation submodule.

9. A system employing the dynamic trust modeling and calculation method for edge nodes as described in claim 1, characterized in that: It includes a data modeling module for real-time data acquisition and construction of dynamic heterogeneous graph snapshots, a dynamic trust calculation model for performing spatial aggregation, temporal aggregation and trust score prediction, and a security decision module.

10. The system of the dynamic trust modeling and calculation method for edge nodes according to claim 9, characterized in that: The security decision module executes security policies based on the trust score matrix output by the dynamic trust calculation model and preset thresholds.

Images