A graph attention trust management method and model based on double trust encoding

Abstract

The application discloses a kind of graph attention trust management method and model based on double trust coding, the method includes: the original feature of input equipment to computer system is preprocessed, and obtains preprocessing result;Direct trust feature vector is obtained based on preprocessing result through two layers of fully connected neural network;Direct trust feature vector is obtained through two layers of graph attention network, and indirect trust feature vector is obtained, and pure neighbor feature is extracted after processing;Direct trust feature vector and pure neighbor feature are spliced to generate high-dimensional graph feature, and original feature is combined to form super feature vector;Based on super feature vector, discriminant is carried out using random forest classifier, and discriminant result is output.The method separates direct trust and indirect trust by modeling, introduces multi-head graph attention mechanism to learn neighbor collaborative mode, and realizes adaptive feature fusion using random forest, which improves the detection accuracy of malicious devices in complex environment and the adaptability to heterogeneous environment.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) device security technology, and in particular, to a graph attention trust management method and model based on dual-trust coding. Background Technology

[0002] Traditional trust management methods are based on probability and statistics theory, encompassing techniques such as Beta distribution, Bayesian inference, and DS evidence inference. While computationally simple, these methods lack self-learning capabilities and struggle to adapt to evolving malicious attack patterns. With the development of machine learning, methods like Support Vector Machines and Random Forests have been introduced into the IoT trust management field. While these methods have improved the accuracy of malicious node identification to some extent, they treat each device as an independent sample, ignoring the interaction relationships between devices and the characteristics of trust propagation. This results in insufficient detection rates when facing complex network threats.

[0003] In recent years, trust management methods based on Graph Neural Networks (GNNs) have gradually emerged. These methods leverage the network topology between devices to conduct trust assessments, achieving better detection performance compared to traditional methods. However, existing GNN-based trust management methods still have shortcomings. First, most existing methods construct static graphs based on feature similarity, ignoring the real interaction history and communication patterns between devices, resulting in a high false negative rate when dealing with temporal attacks. Second, existing methods typically combine direct trust assessment based on device behavior and indirect trust assessment based on neighbor recommendations in their modeling. These two types of information are highly correlated and have significant redundancy, limiting the accuracy of identifying Sybil attacks and hierarchical collaborative attacks. Finally, existing methods often employ a uniform fusion strategy to process all devices, failing to dynamically adjust the assessment method according to the characteristics of nodes under different attack scenarios, resulting in poor generalization performance when facing complex heterogeneous scenarios.

[0004] In summary, to address the aforementioned limitations, a graph attention trust management method and model based on dual-trust encoding is proposed. Summary of the Invention

[0005] The main objective of this invention is to provide a graph attention trust management method and model based on dual-trust coding, in order to address the shortcomings of existing IoT device trust management technologies in scenarios involving device malice detection, such as mutual interference between direct and indirect trust feature modeling, impure neighbor information extraction, neglect of network topology characteristics, poor generalization performance, and low accuracy in detecting malicious devices.

[0006] Based on a first key aspect of the present invention, a graph attention trust management method based on dual-trust encoding is provided, the method comprising the following steps:

[0007] Input the original features of the preset training set device into the computer system to perform preprocessing operations, adjust the distribution of the feature dimensions of the original features to a standard normal distribution with a mean of 0 and a standard deviation of 1, and obtain the preprocessing result;

[0008] Based on the preprocessing results, a direct trust feature vector is obtained through a two-layer fully connected neural network, including linear transformation, batch normalization, and ReLU activation.

[0009] Using the direct trust feature vector as input, the neighbor information is aggregated through a two-layer graph attention network to obtain the indirect trust feature vector, and then the pure neighbor features are extracted through feature difference and nonlinear transformation.

[0010] The direct trust feature vector is concatenated with the pure neighbor feature to generate a high-dimensional graph feature, and then the high-dimensional graph feature is combined with the original feature to form a super feature vector.

[0011] Based on the aforementioned super-feature vectors, a random forest classifier is used for adaptive fusion and final discrimination, and the discrimination result is output.

[0012] As a further preferred embodiment, in the aforementioned method, the preprocessing operation is performed in the following steps:

[0013] Calculate training set statistics: for each feature dimension of the original feature, calculate the mean and standard deviation of that dimension on the training set.

[0014] Based on the mean and standard deviation, a standardization transformation is performed to normalize all original features, thereby obtaining the preprocessing result and completing the preprocessing operation.

[0015] As a further preferred embodiment, in the aforementioned method, the two-layer fully connected neural network includes a first fully connected neural network and a second fully connected neural network, which respectively perform linear transformation, batch normalization and ReLU activation;

[0016] The execution steps of the linear transformation of the first fully connected neural network are as follows:

[0017] Based on the preprocessing results, a weighted combination of the preprocessing results is first obtained by performing a linear transformation using the weight matrix and the bias vector.

[0018] The batch normalization execution steps for the first fully connected neural network are as follows:

[0019] Batch normalization standardizes the output of each neuron in the current batch. First, it obtains the weighted combination of the preprocessing results corresponding to the neuron, and then combines the mean and variance of the neuron in the current batch to output the batch normalization result, thus completing the batch normalization.

[0020] The ReLU activation steps of the first fully connected neural network are as follows:

[0021] Based on the batch normalization result, ReLU activation is applied to set all negative values ​​to zero and keep positive values ​​unchanged, thus obtaining the final output of the first fully connected neural network.

[0022] Based on the final output of the first fully connected neural network, perform linear transformation, batch normalization, and ReLU activation on the second fully connected neural network;

[0023] The execution steps of linear transformation, batch normalization, and ReLU activation in the second fully connected neural network are the same as those in the first fully connected neural network.

[0024] Specifically, a Dropout operation is inserted between the output of the first fully connected neural network and the second fully connected neural network, randomly disabling 30% of the neurons during each training session.

[0025] As a further preferred embodiment, in the aforementioned method, the execution steps for aggregating neighbor information via a two-layer graph attention network are as follows:

[0026] The direct trust feature vector is input into the first layer graph attention network, and the first-order neighbors are combined through independent attention heads. The outputs of the attention heads are then concatenated to obtain the high-dimensional direct trust feature vector.

[0027] The high-dimensional direct trust feature vector is input into the second layer graph attention network and aggregated using single-head attention. The feature dimension of the high-dimensional direct trust feature vector is compressed to be consistent with the feature dimension of the direct trust feature vector, and finally the indirect trust feature vector is obtained.

[0028] The steps for extracting the pure neighbor features are as follows:

[0029] First, feature difference is performed by subtracting the indirect trust feature vector from the direct trust feature vector to obtain the feature difference result.

[0030] Then, a nonlinear transformation is performed on the feature difference results to linearly recombine the original difference features. Combining the transformation matrix and the bias matrix, the pure neighbor features are extracted by processing with the ReLU activation function.

[0031] As a further preferred option, in the aforementioned method, the specific process of adaptive fusion and final discrimination using a random forest classifier includes a training phase and a discrimination phase;

[0032] Specifically, the training phase employs a two-stage strategy:

[0033] The first stage involves supervised learning training of the graph neural network, incorporating orthogonal regularization constraints. The classification loss employs focus loss, as shown in the following equation:

[0034]

[0035] in, This indicates the probability of the model predicting the malice of sample i. This represents the true label of sample i. The weighting coefficients represent the factors that balance the differences in the number of positive and negative samples. Indicates the modulation factor;

[0036] The orthogonal regularization uses the absolute value of the inner product of the direct trust feature vector and the indirect trust feature vector as the loss term;

[0037] Specifically, as shown in the following formula:

[0038]

[0039] in, This represents the total number of training samples. Represents the indirect trust feature vector. Represents the direct trust feature vector;

[0040] The total loss function of a graph neural network is a weighted sum of the classification loss and the orthogonality loss, where the weight coefficients of the orthogonality loss are determined through grid search.

[0041] Once the graph neural network has been trained and its parameters are fixed, the second stage of training will begin:

[0042] A random forest classifier is trained using the super-feature vectors and the corresponding sample real labels as training data. During training, a random training subset is generated for each decision tree through bootstrapping, and the optimal features and split points are recursively selected to construct the decision tree.

[0043] As a further preferred embodiment, in the aforementioned method, the specific process of the discrimination stage is as follows:

[0044] The sample to be tested is input into the decision tree constructed by the trained random forest, and each decision tree outputs a class label;

[0045] The final category is determined by majority vote, and the probability of malice is calculated at the same time.

[0046] Based on the malicious probability, a binary classification decision is made through a preset threshold, and the discrimination result is output.

[0047] As a further preferred embodiment, in the aforementioned method, the binary classification decision specifically includes:

[0048] The result of judging the sample is determined by combining a preset threshold with the probability of malice. If the probability of malice is greater than the preset threshold, the sample to be tested is determined to be malicious.

[0049] If the probability of malice is less than or equal to the preset threshold, the sample to be tested is determined to be normal.

[0050] Based on a second key aspect of the present invention, a graph attention trust management model based on dual-trust encoding is provided, comprising:

[0051] The preprocessing module is used to receive the original features of the device, perform preprocessing operations on the original features, and output the preprocessing results.

[0052] The direct trust encoding module adopts a two-layer fully connected neural network structure, which sequentially includes a linear transformation unit, a batch normalization unit, and a ReLU activation unit, used to encode the features of the preprocessing results and output a direct trust feature vector.

[0053] The indirect trust coding module includes a first-layer graph attention network unit, a second-layer graph attention network unit, a feature difference unit, and a nonlinear transformation unit;

[0054] The first-layer graph attention network unit and the second-layer graph attention network unit are used to aggregate neighbor information with direct trust feature vector as input and output indirect trust feature vector;

[0055] The feature difference unit and nonlinear transformation unit are used to calculate the difference results of the direct trust feature vector and the indirect trust feature vector, and perform linear recombination and ReLU activation processing to output pure neighbor features.

[0056] The super-feature construction module is used to concatenate direct trust features and pure neighbor features to generate high-dimensional features, and then combine the preprocessed original features to output a super-feature vector.

[0057] The trust classification module uses a random forest classifier to adaptively fuse super-feature vectors for feature discrimination and malicious intent detection, and outputs the discrimination results.

[0058] According to a third key aspect of the present invention, an electronic device is provided, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0059] The memory stores a computer program that, when executed by the processor, causes the processor to perform the graph attention trust management method based on dual-trust coding as described above.

[0060] Based on a fourth key aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed, implements the graph attention trust management method based on dual-trust coding as described above.

[0061] Compared with existing technologies, this invention proposes a graph attention trust management method and model based on dual-trust coding. First, this invention adopts a dual-trust coding architecture. Direct trust coding is used to encode and extract the original features to reflect the device's own behavior, focusing only on the device's own historical interaction records, performance indicators, and trust metrics, without considering the influence of neighboring devices. Indirect trust coding takes the direct trust features as input, aggregates neighbor information through a graph attention network, and then removes the device's own information through feature difference and nonlinear transformation to obtain pure neighbor features. This solves the defects of existing technologies, such as mutual interference between direct and indirect trust feature modeling and impure neighbor information extraction, and provides a high-quality feature foundation for subsequent accurate discrimination.

[0062] Secondly, this invention introduces graph attention to mine the topological relationships between IoT devices in the network. Simultaneously, it introduces orthogonal regularization constraints during the graph neural network training phase, fusing focus loss and orthogonal loss through the total loss function. This enhances the information complementarity of direct and indirect trust features and improves the adaptability of this invention to imbalanced scenarios. Furthermore, it employs a two-stage training strategy, combining the strong feature learning capabilities of deep learning with the efficient fusion and discrimination advantages of random forests. This overcomes the shortcomings of existing methods, such as neglecting network topological characteristics and poor generalization performance in complex scenarios, thus improving the stability and robustness of malicious device detection.

[0063] Finally, this invention employs a random forest classifier to achieve adaptive feature fusion and final discrimination. Its training process requires no gradient optimization; it constructs a decision tree ensemble through self-adoption, resulting in high training efficiency. In scenarios with thousands to tens of thousands of IoT devices, training time is typically completed within tens of seconds, far faster than the hours required for deep neural networks. Furthermore, theoretical analysis and practical application verification demonstrate that this invention outperforms existing management solutions in terms of accuracy in malicious device detection, providing reliable technical support for the security protection of IoT devices. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0065] Figure 1The following is an execution flowchart of a graph attention trust management method based on dual-trust encoding in one embodiment of the present invention;

[0066] Figure 2 The diagram illustrates a system architecture of a graph attention trust management method based on dual-trust encoding, according to one embodiment of the present invention.

[0067] Figure 3 The diagram illustrates a graph attention trust management method and its model based on dual-trust encoding, as well as a comparative analysis of different model variants under three threat scenarios, in one embodiment of the present invention.

[0068] Figure 4 The diagram illustrates a graph attention trust management method and its model based on dual-trust encoding, as well as a comparison of ROC curves with different model variants under three threat scenarios, in one embodiment of the present invention.

[0069] Figure 5 The figure shows a comparison of the F1-Scores of four variants in three threat scenarios in one embodiment of the present invention;

[0070] Figure 6 This diagram illustrates the ranking of feature importance calculated by random forest in three threat scenarios according to one embodiment of the present invention.

[0071] Figure 7 This diagram illustrates a comparison of the GAT attention weight distribution between a normal device and a malicious device in one embodiment of the present invention.

[0072] Figure 8 This diagram illustrates the evolution of the trust value between a normal device and a malicious device with the number of interactions in one embodiment of the present invention.

[0073] Figure 9 The diagram illustrates the weight allocation of direct trust features, indirect trust features, and original features in a random forest under three threat scenarios in one embodiment of the present invention. Detailed Implementation

[0074] The preferred embodiments of the present invention will be described in detail below to provide a clearer understanding of the purpose, features, and advantages of the present invention. It should be understood that the following embodiments are not intended to limit the scope of the present invention, but are merely illustrative of the essential spirit of the technical solution of the present invention.

[0075] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known techniques associated with the invention may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0076] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0077] like Figure 1 As shown, in one embodiment, a graph attention trust management method based on dual-trust encoding according to the present invention includes the following steps:

[0078] S100, input the original features of the training device into the computer system to perform preprocessing operations, adjust the distribution of the feature dimensions of the original features to a standard normal distribution with a mean of 0 and a standard deviation of 1, and obtain the preprocessing result;

[0079] S200, based on the preprocessing results, a direct trust feature vector is obtained through a two-layer fully connected neural network, including linear transformation, batch normalization and ReLU activation;

[0080] S300, taking the direct trust feature vector as input, the neighbor information is aggregated through a two-layer graph attention network to obtain the indirect trust feature vector, and then the pure neighbor features are extracted through feature difference and nonlinear transformation.

[0081] S400, the direct trust feature vector is concatenated with the pure neighbor feature to generate a high-dimensional graph feature, and then the high-dimensional graph feature is combined with the original feature to form a super feature vector;

[0082] S500, based on the super feature vector, uses a random forest classifier for adaptive fusion and final discrimination, and outputs the discrimination result.

[0083] This method is based on a graph attention trust management model (DTEM) based on dual-trust encoding.

[0084] The core design concept of DTEM is dual-trust separation encoding: direct trust and indirect trust are modeled independently in different feature spaces to avoid mutual interference between their representations.

[0085] like Figure 2 As shown, the model consists of four main modules: direct trust layer, indirect trust layer, feature fusion layer, and trust classification layer.

[0086] Direct Trust Layer: This module is responsible for extracting representations reflecting the device's own behavior from the device's 10-dimensional raw features. It employs a two-layer fully connected neural network, focusing only on the device's own historical interaction records, performance metrics, and trust metrics, without considering the influence of neighboring devices.

[0087] The output of this layer is a 64-dimensional direct trust feature vector, represented by the symbol... This module is designed to capture the autonomous behavior patterns of devices, providing a basis for subsequent judgments.

[0088] Indirect Trust Layer: This module, based on a graph attention network, learns indirect trust representations from collaborative information from neighboring devices. It uses direct trust features... As input, information from neighboring nodes is adaptively aggregated through a multi-head attention mechanism.

[0089] Crucially, this module calculates the influence weight of each neighbor on the current device, enabling the model to focus on neighbors who interact closely with or behave similarly to the current device. The output is a 64-dimensional indirect trust feature vector. .

[0090] To extract pure neighbor collaboration information, this module will also... The process involves removing the influence of the nodes themselves to obtain pure neighbor features. .

[0091] Feature fusion layer: The module will directly trust the features. and pure neighbor characteristics The features are concatenated to form a 128-dimensional fused feature vector. .

[0092] The concatenation operation preserves complete information from both feature spaces, enabling the classifier to make judgments using both the device's own behavior and neighbor collaboration patterns.

[0093] Trust Classification Layer: This module is a multilayer perceptron (MLP) that maps 128-dimensional fused features to the final trust score. The classifier contains three hidden layers with dimensions of 128, 64, and 32 respectively, each followed by batch normalization and a ReLU activation function.

[0094] The final layer uses Sigmoid activation to ensure the output is in the [0,1] interval, which can be interpreted as the probability of device trustworthiness. To further improve performance, the model also introduces hybrid features. Enhancement strategy: The 14-dimensional original features are concatenated with the deep fusion features again and fed into a random forest for secondary classification.

[0095] In one feasible implementation, step S100 is described in detail:

[0096] Before feeding the 10-dimensional raw features into the neural network, preprocessing must be performed to eliminate the two negative impacts caused by the inconsistency of the dimensions of different features.

[0097] First, it prevents the model from learning only large numerical features while ignoring small numerical features; second, it prevents the optimization efficiency from being reduced due to different feature scales, which would ultimately lead to an increase in training time.

[0098] This invention employs a standardized method to preprocess features. Its core idea is to adjust the distribution of each feature dimension to a standard normal distribution with a mean of 0 and a standard deviation of 1.

[0099] The specific steps are as follows:

[0100] Calculate the training set statistics for each feature dimension. Calculate the mean of this dimension on the training set. and standard deviation .

[0101] Specifically,

[0102]

[0103]

[0104] in, This indicates the number of devices in the training set. This represents the original value of the i-th device in the j-th feature dimension.

[0105] The standardization transformation uses the calculated mean and standard deviation to normalize all device features in the training and test sets.

[0106]

[0107] in, Represents the standardized eigenvalues. This represents a constant that prevents division by zero.

[0108] After this transformation, the mean of each feature dimension is approximately 0, and the standard deviation is approximately 1, thus bringing different features to the same numerical scale.

[0109] In one feasible implementation, step S200 is described in detail:

[0110] After preprocessing the input features, the standardized 14-dimensional features are... Mapped to a 64-dimensional direct trust representation .

[0111] This invention employs a two-layer fully connected neural network design, with each layer containing three operations: linear transformation, batch normalization, and ReLU activation.

[0112] The calculation process of the two-layer network is described in detail below:

[0113] The first layer's task is to map the standardized 10-dimensional input features to a 64-dimensional hidden space. This mapping process consists of three consecutive steps:

[0114] (1) Linear transformation:

[0115] Given standardized feature vectors First, through the weight matrix and bias vector Perform a linear transformation.

[0116]

[0117] in, For learnable parameters, each row of the matrix corresponds to a hidden neuron, and each column corresponds to an input feature dimension; Provide an adjustable baseline activation value for each neuron; linear transformation Weighted combination of input features.

[0118] (2) Batch normalization: Stabilize the training process through standardization operations. In deep neural networks, the input distribution of each layer changes continuously with the update of the parameters of the previous layer. This phenomenon is called internal covariate shift, which can lead to gradient vanishing or exploding.

[0119] Batch normalization standardizes the output of each neuron in the current batch.

[0120]

[0121] in, This represents the linear transformation output of the j-th neuron; and These are the mean and variance of the neuron in the current batch, respectively.

[0122] After completing batch normalization, The distribution has been adjusted to a suitable range.

[0123] Next, we need to introduce nonlinear transformations to enable the network to learn the complex interactions between features; this is the role of the ReLU activation function.

[0124] (3) ReLU activation

[0125] By setting all negative values ​​to zero and leaving positive values ​​unchanged, this simple nonlinear transformation enables the network to learn nonlinear relationships.

[0126]

[0127] The batch-normalized output is activated using ReLU to obtain the final output of the first layer:

[0128]

[0129] After these three steps, the first layer successfully maps the 10-dimensional normalized input to a 64-dimensional hidden layer representation. Combining the above three steps, the complete calculation formula for the first layer is:

[0130]

[0131] Output of the first layer It already contains a preliminary abstract representation of the input features, but this representation is still relatively coarse and insufficient to finely distinguish between normal and malicious devices.

[0132] The role of the second layer is to further refine these hidden features so that the final output can better serve the trust assessment task.

[0133] The computation process of the second layer is exactly the same as that of the first layer, consisting of three steps: linear transformation, batch normalization, and ReLU activation. The only difference is that the input changes from 14 dimensions to 64 dimensions, while the output remains a 64-dimensional direct trust feature vector.

[0134]

[0135] in, and These are the weight matrix and bias vector of the second layer.

[0136] In one embodiment, to prevent overfitting, a regularization mechanism needs to be introduced to constrain the complexity of the model.

[0137] This invention employs the Dropout regularization strategy. Its core idea is to randomly discard a portion of neurons during training, that is, to set their outputs to zero.

[0138] Specifically, in the first layer output A Dropout operation is inserted between the second layer, with a dropout rate set to p=0.3, meaning that 30% of the neurons are randomly masked during each training session.

[0139]

[0140] It is important to note that Dropout behaves completely differently during the training and testing phases, and these need to be explained separately.

[0141] During the training phase: during the second forward propagation, Each element in the array has a 30% probability of being set to zero.

[0142] This means that the effective structure of the network is different each time it is trained, which is equivalent to training multiple different sub-networks. This randomness brings two benefits. First, it prevents the network from becoming overly reliant on certain specific neurons, forcing the network to learn distributed and robust feature representations; second, it is equivalent to ensemble learning, and the final model can be viewed as the average of multiple sub-networks.

[0143] During the testing phase: Dropout is completely disabled, and all neurons participate in the computation. However, to maintain the expected output, the neuron's output needs to be scaled, i.e., multiplied by [a certain factor]. .

[0144] In one feasible implementation, step S300 is described in detail:

[0145] The indirect trust layer is the second core module of DTEM, and its goal is to extract collaborative trust information from the neighbor interaction patterns of devices. Unlike the direct trust layer, which only focuses on the device itself, this module uses a graph attention network to learn the influence of neighboring devices on the current device.

[0146] This invention solves this problem through an adaptive attention mechanism. An attention weight is calculated for each edge, reflecting the importance of neighboring nodes to the current node.

[0147] This adaptability enables the present invention to automatically focus on the neighbors most valuable for trust assessment and ignore noisy neighbors. Furthermore, to extract pure neighbor collaboration information, the present invention introduces a pure feature extraction mechanism, which uses feature difference operations to remove the influence of the nodes themselves, ensuring that the indirect trust representation contains only information from neighbors.

[0148] In one embodiment, the graph attention mechanism is first explained:

[0149] The core idea of ​​Graph Attention Networks (GAT) is to learn an attention weight for each edge in the graph, so that important neighbors receive higher weights. For node i and its set of neighbors... GAT requires calculating the attention weight of each neighbor to node i. .

[0150] This weight reflects the relative importance of neighbor node j to node i. If it is close to 1, it means that neighbor j has a very large influence on node i; if it is close to 0, it means that the influence of neighbor j is negligible.

[0151] The calculation of attention weights consists of three steps: first, the original attention score is calculated; then the score is normalized to ensure that the sum of the weights of all neighbors is 1; finally, the normalized weights are used to perform weighted aggregation of the neighbor features.

[0152] Calculate the attention score, given the features of node i. Characteristics of neighbor j (The feature here is the direct trust in the encoder's output), so the correlation score between them needs to be calculated first. .

[0153] Correlation score The larger the value, the more similar the two nodes are.

[0154]

[0155] in, It is a non-linear activation function that allows negative values ​​to pass through with a small slope; This is the weight matrix, used for linear transformation to project the original features into a space more suitable for calculating similarity. This is a vector concatenation operation that joins the two transformed feature vectors end to end, preserving the complete information of the two nodes. This is a learnable attention vector.

[0156] To ensure that the attention weights sum to 1 across all neighbors, the attention scores are normalized using the softmax function.

[0157]

[0158] in, This represents the final attention weights after normalization.

[0159] Weighted aggregation of neighbor information, utilizing normalized attention weights The features of all neighbors are weighted and summed to obtain a new representation of node i. The physical meaning is that the transformed features of all neighbors are weighted and averaged according to attention weights; neighbors with larger weights contribute more to the aggregation result.

[0160] After the aggregation operation, the new representation of node i incorporates information from all its neighbors, but the contribution of each neighbor is dynamically adjusted according to its importance.

[0161]

[0162] in, It is the feature representation of the l-th layer neighbor j. This refers to the shared weight matrix used to calculate the attention score. This indicates that it is an activation function; in this embodiment, the ReLU function is used. This indicates that node i is in the th order. A new representation of layers.

[0163] Single-head attention can only learn one neighbor relationship pattern, resulting in limited representational capabilities. To enhance the model's representational power and stability, a multi-head attention mechanism is needed, allowing the model to learn multiple different neighbor relationship patterns simultaneously.

[0164] In one embodiment, the multi-head attention mechanism is illustrated as follows:

[0165] The core idea of ​​multi-head attention mechanism is to use K independent attention heads to compute in parallel. Each head has its own weight matrix and attention vector, and can learn different neighbor relationship patterns.

[0166] The outputs of the K heads are concatenated to form a richer node representation. Specifically, the attention weights of the k-th head are calculated in the same way as those of a single head, but with independent parameters.

[0167]

[0168] in, This represents a vector concatenation operation. Indicates the index of the attention head. This represents the attention weight calculated for the k-th head. It is the weight matrix of the k-th head.

[0169] Based on the above description, the indirect trust layer of the present invention adopts a two-layer GAT network to gradually aggregate neighbor information.

[0170] A single-layer GAT can only aggregate information from first-order neighbors (i.e., directly connected neighbors), while a two-layer GAT can aggregate information from second-order neighbors (i.e., neighbors of neighbors), thus broadening the scope of information propagation. This layer-by-layer design of expanding the receptive field ensures that the final representation of each node not only contains information from its direct neighbors but also indirectly contains information from more distant nodes, which helps to capture global patterns in the network.

[0171] The first layer uses multi-head attention to enhance representation capabilities, while the second layer uses single-head attention to reduce dimensionality and computational cost.

[0172] The first layer of GAT: The role of the first layer is to initially capture the collaborative patterns of neighbors and learn neighbor relationships from multiple perspectives through a multi-head mechanism.

[0173] Its input is a 64-dimensional feature vector output from the direct trust layer. The first-order neighbors are aggregated through four independent attention heads, each outputting a 64-dimensional feature vector, which is then concatenated to obtain a 256-dimensional feature vector.

[0174]

[0175] in, This represents the attention weight of the k-th head in the first layer. Let represent the direct trust vector of device j.

[0176] Second-layer GAT: The second layer outputs the 256-dimensional feature vector from the first layer. As input, single-head attention is used for aggregation, compressing the feature dimension back to 64 dimensions. The attention weight calculation formula is the same as above.

[0177]

[0178] Specifically, this step aggregates neighbor information using GAT, but the output... It is not simply a neighborly characteristic.

[0179] The reason lies in the fact that some design mechanisms of graph attention networks can cause node-specific information to be mixed into the aggregation result. This is detrimental to handling group-based coordinated attacks. This model uses the feature difference method to achieve this goal. Its core idea is to use the output of the indirect trust layer... Information minus the output of the direct trust layer The information will only include information about the neighbors.

[0180]

[0181] Although the result of direct subtraction Without the influence of the nodes themselves, simple linear differences cannot capture complex nonlinear relationships.

[0182] Therefore, it is necessary to perform a nonlinear transformation on the difference results and linearly recombine the original difference features to make them more suitable for subsequent classification tasks.

[0183]

[0184] in, Represents the transformation matrix. This represents the bias matrix.

[0185] In one feasible implementation, step S400 is described in detail:

[0186] First, directly trust the feature vector. and indirect trust feature vector splicing into high-dimensional graph features :

[0187]

[0188] Then, the graph features are combined with the original device behavior features to form a super feature vector, which contains the original features and graph structure features, providing rich discriminative information for the random forest classifier.

[0189]

[0190] In one feasible implementation, step S500 is described in detail.

[0191] The training objective in this invention is to learn feature representations that can accurately identify malicious devices.

[0192] This invention employs a two-stage training strategy. In the first stage, the graph neural network part is trained through supervised learning to learn high-quality feature representations.

[0193] In the second stage, the parameters of the graph neural network are fixed, and a random forest classifier is trained using the extracted features to learn adaptive feature fusion weights. This staged training strategy combines the powerful representation learning capabilities of deep learning with the robustness of ensemble learning, enabling the model to maintain excellent detection performance under various attack scenarios.

[0194] Given the true label of device i In this context, 1 represents a malicious device and 0 represents a normal device.

[0195] The model's predicted probability of malice is also given:

[0196]

[0197] in, This indicates the probability of the model predicting the malice of sample i. This represents the true label of sample i. The weighting coefficients represent the factors that balance the differences in the number of positive and negative samples. This represents the modulation factor.

[0198] To enhance the complementarity between direct and indirect trust features, this invention also introduces orthogonal regularization constraints during the training process.

[0199]

[0200] in, This represents the total number of training samples. Represents the indirect trust feature vector. Represents the direct trust feature vector; This represents the vector dot product.

[0201] This loss term encourages the model to learn mutually independent direct and indirect trust representations, thereby maximizing the information complementarity of the two branches.

[0202] Combining classification loss and orthogonality loss, the total loss function of a graph neural network is:

[0203]

[0204] in, The weighting coefficients representing the orthogonal loss are determined through a grid search.

[0205] After the graph neural network is trained, its parameters are fixed, and the super feature vectors and corresponding labels are extracted using the training set to train a random forest classifier.

[0206] The training process of random forests does not require gradient optimization; decision tree ensembles can be quickly built through bootstrapping and random feature selection.

[0207] Random Forest Training Process:

[0208] Tree construction phase: Bootstrap sampling is performed on the training set to generate a random training subset for each tree, and then the optimal features and split points are recursively selected to construct the decision tree.

[0209] Voting phase: Test samples The input is fed into all T decision trees, and each tree outputs a class label.

[0210] The type label outputs 0 or 1, where 0 indicates normal and 1 indicates malicious.

[0211] The final prediction is determined by majority vote, and the probability of malice is output (i.e., the proportion of trees that voted malice).

[0212]

[0213] in, This represents the total number of decision trees. This represents an exponential function.

[0214] Ultimately, a binary classification decision is made based on a threshold.

[0215]

[0216] The training time complexity of random forest is O(n log n). , Indicates the number of features.

[0217] In scenarios involving thousands to tens of thousands of IoT devices, training time is typically completed within tens of seconds, much faster than the hours required for training deep neural networks.

[0218] In one feasible implementation, the technical effects of the present invention are illustrated by the following embodiments:

[0219] In one embodiment, an overall analysis is performed first:

[0220] Four representative methods were selected as baseline comparisons, covering traditional trust models, machine learning methods, temporal models, and graph neural network methods. These are:

[0221] Bayesian Trust Model: A classic Bayesian trust model that uses a Beta distribution to model the uncertainty of device behavior and calculates trust values ​​based solely on the success rate of interactions.

[0222] Random Forest: A traditional machine learning method that directly trains 150 decision trees using 10-dimensional raw features without using graph structure information.

[0223] LSTM: A temporal neural network method that inputs the sequence of the device’s 50 most recent interactions into a two-layer LSTM (64 hidden dimensions) to capture temporal dependencies.

[0224] GCN: Graph Neural Network Baseline, using a two-layer GCN (64 hidden dimensions) for node classification, without distinguishing between direct and indirect trust.

[0225] All deep learning methods use the same training and test sets, the same data preprocessing procedures, the same optimizer (Adam), and the same early stopping strategy to ensure fairness in the comparison.

[0226] This embodiment constructs three typical IoT trust attack scenarios:

[0227] Threat 1 - Temporal Cooperative Attack: Simulates interactions between 13,181 real devices, including 1,920 single-controller devices, 1,608 multi-controller devices, 1,608 hierarchical networks, and 1,264 feature obfuscation devices. Attackers control 3-8 devices to form a group, initiating a cooperative attack after a 200-interaction camouflage period. The graph structure is updated after 720,000 interactions, with a training set clustering coefficient of 0.028.

[0228] Threat 2 - Sybil Attack: Simulates 13,181 real devices, including 1,801 in a cyclical pattern, 1,449 in a loose alliance, 1,441 in a star-shaped defamation pattern, 1,440 in a hierarchical defamation pattern, and 1,069 in a dynamic switching pattern. The attacker creates 12-18 fake identities, with a controller degree of 15-25. Training involves 720,000 interactions, resulting in a training set clustering coefficient of 0.004.

[0229] Threat 3 - Malicious Defamation Attack: Simulates 13,181 real devices, including 8,800 attackers, 3,520 victims, and 587 attack groups, with 800,000 training interactions.

[0230] In this embodiment, the following two metrics are evaluated.

[0231] F1-Score is used as the primary evaluation metric. It can simultaneously consider both false positive and false negative rates, preventing the model from obtaining artificially high accuracy by biasing towards the majority class.

[0232] ROC-AUC (Area Under the ROC Curve) measures the overall performance of a model under different classification thresholds. The ROC curve is plotted with the false positive rate on the horizontal axis and the true positive rate on the vertical axis. The closer the AUC is to 1, the better the performance. The advantage of ROC-AUC is that it is not affected by the classification threshold and can comprehensively reflect the model's discriminative ability.

[0233] Combination Figure 3 and Figure 4 As shown, the DTEM of the present invention performs best in all three threat scenarios, thanks to the dual-path trust representation being able to simultaneously capture the device’s own behavior and the influence of its neighbors.

[0234] The ROC curves were closest to the top left corner in all three scenarios, with AUCs of 0.990, 1.000, and 1.000, respectively, indicating that extremely high true positive rates and extremely low false positive rates could be maintained under different classification thresholds.

[0235] In one embodiment, ablation experiments and module contribution analysis are performed.

[0236] This implementation designed four variant models for comparison. The basic idea of ​​the ablation experiment is to gradually remove or simplify certain modules from the complete model, observe the degree of performance degradation, and thus quantify the contribution of that module.

[0237] Baseline: Train a random forest classifier using only 10-dimensional raw features, without using any graph structure information or GNN.

[0238] Direct Trust: Extract the device's own 64-dimensional behavioral features using direct trust, concatenate them with the original 10-dimensional features, and then input them into the random forest.

[0239] Indirect Trust: Only 64-dimensional indirect trust features are extracted using GAT, and then concatenated with the 10-dimensional original features before being input into a random forest.

[0240] DTEM: The complete model, which includes dual-path trust representation (64-dimensional direct + 64-dimensional indirect), orthogonal constraints and random forest adaptive fusion, with a total feature dimension of 138.

[0241] All variants use the same training set, test set, and random forest parameters (150 trees, maximum depth 10), the only difference being the source and dimension of the input features.

[0242] like Figure 5 As shown, the indirect trust layer makes the most significant contribution. After removing the indirect trust layer, the F1-Score of threat 1 decreased by 18%, threat 2 by 15%, and threat 3 even decreased by 20%. This is because the indirect trust layer captures the influence of neighboring nodes through GAT, effectively identifying the graph structure characteristics of coordinated attack groups. Removing the direct trust layer further leads to even greater performance loss, with threat 1 decreasing by 25.00%, threat 2 by 22.00%, and threat 3 by 28.00%, indicating that the device's own behavioral characteristics are also indispensable.

[0243] Removing graph structure information resulted in a decrease in F1 score across all three scenarios, validating the importance of graph topology in trust assessment. In summary, direct trust, indirect trust, and graph structure are all indispensable; only a complete dual-path architecture can achieve optimal performance.

[0244] In one embodiment, feature importance analysis is performed:

[0245] like Figure 6 As shown, the cumulative weight of graph topology features exceeds 40% in all scenarios, further validating that neighbor influence is the core of trust assessment. The cumulative weight of direct trust features is between 20% and 30%, indicating that the device's own historical behavior is also an important criterion for judgment.

[0246] The cumulative weight of the original network features is only 10%-15%, indicating that the underlying indicators also contribute to trust assessment, which verifies the necessity of the dual-path feature extraction proposed in this invention.

[0247] In one embodiment, GAT attention weight distribution analysis is performed:

[0248] Combination Figure 7 As shown, the attention weight of malicious devices is significantly higher than that of normal devices.

[0249] In Threat 1, the average attention weight of malicious devices is 0.559, while that of normal devices is only 0.284, reflecting the high weight of attention paid to other members of the group by members of the coordinated attack group.

[0250] In Threat 2, the average attention weight of the witch node reaches 0.612 because the witch node needs to frequently exchange information with the multiple identities it controls, forming dense attention connections.

[0251] In Threat 3, the victim being defamed has an unusually high attention weight towards the attacker. This is because the attacker frequently gives negative evaluations, causing GAT to allocate more attention to these neighbors. GAT's attention mechanism can automatically capture abnormal neighbor interaction patterns, thus providing key clues for trust assessment.

[0252] In one embodiment, a time-series evolution analysis of trust values ​​is performed:

[0253] Combination Figure 8 As shown, the trust value evolution of malicious devices exhibits a clear two-stage pattern: the camouflage phase and the attack phase. During the camouflage phase, the trust value of malicious devices is similar to that of normal devices, indicating that the attacker has successfully mimicked the behavior of normal devices. Upon entering the attack phase, the trust value of malicious devices rapidly decreases and remains at a low level, while the trust value of normal devices remains stable. This demonstrates that DTEM can capture the dynamic changes in attack behavior through temporal features, thereby achieving early detection.

[0254] In one embodiment, an adaptive weight allocation analysis of a random forest is performed.

[0255] Combination Figure 9 As shown, the weight of indirect trust features is the highest in all scenarios, further validating that neighbor influence is the core of trust assessment. Significant differences exist in weight allocation across different threat scenarios: in threat 1, the indirect trust weight is 60%; in threat 2, the indirect trust weight is 65%; while in threat 3, the direct trust weight is relatively high because the allegedly defamed victim actually performs well, making direct trust a valid criterion. The weight of original features remains around 15% in all scenarios, indicating that the contribution of underlying network metrics to the final decision is relatively limited. This adaptive weight allocation mechanism allows DTEM to dynamically adjust feature importance based on the characteristics of different threats, thereby maintaining high detection performance across diverse attack scenarios.

[0256] In summary, systematic experiments have validated the effectiveness of DTEM in detecting malicious IoT devices. Overall performance comparisons show that DTEM significantly outperforms baseline methods across three threat scenarios. Ablation experiments confirm the necessity of the dual-path trust architecture, with the indirect trust layer contributing the most, followed by the direct trust layer, thus validating the effectiveness of the separation modeling strategy. Feature importance analysis reveals the core role of graph topology features in trust assessment, while original features and direct trust features also play important complementary roles. Analysis of the model's internal mechanisms further demonstrates that the GAT attention mechanism can automatically capture anomalous collaborative patterns. The trust value evolution curve clearly reveals the two-stage behavior pattern of malicious devices, and the random forest weight allocation reflects the model's adaptability to different attack scenarios.

[0257] This fully demonstrates the robustness and interpretability of the invention in complex IoT environments.

[0258] The technical terms, principles, or means related to the technical solutions of the present invention mentioned in the above embodiments, which are not described in detail above, are all well-known technologies or common practices that are known to those skilled in the art.

[0259] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A graph attention trust management method based on dual-trust encoding, characterized in that, The method includes the following steps: The original features of devices in a preset training set are input into the Internet of Things (IoT) to perform preprocessing operations, adjusting the distribution of the feature dimensions of the original features to a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby obtaining the preprocessing result; Based on the preprocessing results, a direct trust feature vector is obtained through a two-layer fully connected neural network, including linear transformation, batch normalization, and ReLU activation. The fully connected neural network focuses only on the device's own historical interaction records, performance metrics, and trust metrics; Using the direct trust feature vector as input, the neighbor information is aggregated through a two-layer graph attention network to obtain the indirect trust feature vector, and then the pure neighbor features are extracted through feature difference and nonlinear transformation. The direct trust feature vector is concatenated with the pure neighbor feature to generate a high-dimensional graph feature, and then the high-dimensional graph feature is combined with the original feature to form a super feature vector. Based on the aforementioned hyperfeature vectors, a random forest classifier is used for adaptive fusion and final discrimination, and the discrimination result is output. The steps for aggregating neighbor information via a two-layer graph attention network are as follows: The direct trust feature vector is input into the first layer graph attention network, and the first-order neighbors are combined through independent attention heads. The outputs of the attention heads are then concatenated to obtain the high-dimensional direct trust feature vector. The high-dimensional direct trust feature vector is input into the second layer graph attention network and aggregated using single-head attention. The feature dimension of the high-dimensional direct trust feature vector is compressed to be consistent with the feature dimension of the direct trust feature vector, and finally the indirect trust feature vector is obtained. The steps for extracting the pure neighbor features are as follows: First, feature difference is performed by subtracting the indirect trust feature vector from the direct trust feature vector to obtain the feature difference result. Then, a nonlinear transformation is performed on the feature difference results to linearly recombine the original difference features. Combining the transformation matrix and the bias matrix, the pure neighbor features are extracted by processing with the ReLU activation function.

2. The graph attention trust management method based on dual-trust encoding according to claim 1, characterized in that, The preprocessing operation is performed in the following steps: Calculate training set statistics: for each feature dimension of the original feature, calculate the mean and standard deviation of that dimension on the training set. Based on the mean and standard deviation, a standardization transformation is performed to normalize all original features, thereby obtaining the preprocessing result and completing the preprocessing operation.

3. The graph attention trust management method based on dual-trust encoding according to claim 1, characterized in that, The two-layer fully connected neural network includes a first fully connected neural network and a second fully connected neural network, which respectively perform linear transformation, batch normalization and ReLU activation; The execution steps of the linear transformation of the first fully connected neural network are as follows: Based on the preprocessing results, a weighted combination of the preprocessing results is first obtained by performing a linear transformation using the weight matrix and the bias vector. The batch normalization process for the first fully connected neural network layer is as follows: Batch normalization standardizes the output of each neuron in the current batch. First, it obtains the weighted combination of the preprocessing results corresponding to the neuron, and then combines the mean and variance of the neuron in the current batch to output the batch normalization result, thus completing the batch normalization. The ReLU activation steps of the first fully connected neural network are as follows: Based on the batch normalization result, ReLU activation is applied to set all negative values ​​to zero and keep positive values ​​unchanged, thus obtaining the final output of the first fully connected neural network. Based on the final output of the first fully connected neural network, perform linear transformation, batch normalization, and ReLU activation on the second fully connected neural network; The execution steps of linear transformation, batch normalization, and ReLU activation in the second fully connected neural network are the same as those in the first fully connected neural network. Specifically, a Dropout operation is inserted between the output of the first fully connected neural network and the second fully connected neural network, randomly disabling 30% of the neurons during each training session.

4. The graph attention trust management method based on dual-trust encoding according to claim 1, characterized in that, The specific process of adaptive fusion and final discrimination using a random forest classifier includes a training phase and a discrimination phase. Specifically, the training phase employs a two-stage strategy: The first stage involves supervised learning training of the graph neural network, incorporating orthogonal regularization constraints. The classification loss employs focus loss, as shown in the following equation: in, This indicates the probability of the model predicting the malice of sample i. This represents the true label of sample i. The weighting coefficients represent the factors that balance the differences in the number of positive and negative samples. Indicates the modulation factor; The orthogonal regularization uses the absolute value of the inner product of the direct trust feature vector and the indirect trust feature vector as the loss term; Specifically, as shown in the following formula: in, This represents the total number of training samples. Represents the indirect trust feature vector. Represents the direct trust feature vector; The total loss function of a graph neural network is a weighted sum of the classification loss and the orthogonality loss, where the weight coefficients of the orthogonality loss are determined through grid search. Once the graph neural network has been trained and its parameters are fixed, the second stage of training will begin: A random forest classifier is trained using the super-feature vectors and the corresponding sample real labels as training data. During training, a random training subset is generated for each decision tree through bootstrapping, and the optimal features and split points are recursively selected to construct the decision tree.

5. The graph attention trust management method based on dual-trust encoding according to claim 4, characterized in that, The specific process of the discrimination stage is as follows: The sample to be tested is input into the decision tree constructed by the trained random forest, and each decision tree outputs a class label; The final category is determined by majority vote, and the probability of malice is calculated at the same time. Based on the malicious probability, a binary classification decision is made through a preset threshold, and the discrimination result is output.

6. The graph attention trust management method based on dual-trust coding according to claim 5, characterized in that, The binary classification decision specifically includes: The result of judging the sample is determined by combining a preset threshold with the probability of malice. If the probability of malice is greater than the preset threshold, the sample to be tested is determined to be malicious. If the probability of malice is less than or equal to the preset threshold, the sample to be tested is determined to be normal.

7. A graph attention trust management system based on dual-trust coding, characterized in that, include: The preprocessing module is used to receive the original features of the device, perform preprocessing operations on the original features, and output the preprocessing results. The direct trust encoding module adopts a two-layer fully connected neural network structure, which sequentially includes a linear transformation unit, a batch normalization unit, and a ReLU activation unit, used to encode the features of the preprocessing results and output a direct trust feature vector. The fully connected neural network structure focuses only on the device's own historical interaction records, performance indicators, and trust metrics; The indirect trust coding module includes a first-layer graph attention network unit, a second-layer graph attention network unit, a feature difference unit, and a nonlinear transformation unit; The first-layer graph attention network unit and the second-layer graph attention network unit are used to aggregate neighbor information with direct trust feature vector as input and output indirect trust feature vector; The feature difference unit and nonlinear transformation unit are used to calculate the difference results of the direct trust feature vector and the indirect trust feature vector, and perform linear recombination and ReLU activation processing to output pure neighbor features. The super-feature construction module is used to concatenate direct trust features and pure neighbor features to generate high-dimensional features, and then combine the preprocessed original features to output a super-feature vector. The trust classification module has a built-in random forest classifier, which is used to achieve adaptive feature fusion and maliciousness discrimination of super feature vectors and output the discrimination results. The steps for aggregating neighbor information are as follows: The direct trust feature vector is input into the first layer graph attention network unit, and the first-order neighbors are combined through independent attention heads. The outputs of the attention heads are then concatenated to obtain the high-dimensional direct trust feature vector. The high-dimensional direct trust feature vector is input into the second layer graph attention network unit, and single-head attention is used to aggregate it, compressing the feature dimension of the high-dimensional direct trust feature vector to be consistent with the feature dimension of the direct trust feature vector, and finally obtaining the indirect trust feature vector. The steps for outputting the pure neighbor features are as follows: First, feature difference is performed by subtracting the indirect trust feature vector from the direct trust feature vector to obtain the feature difference result. Then, a nonlinear transformation is performed on the feature difference results to linearly recombine the original difference features. Combining the transformation matrix and the bias matrix, the ReLU activation function is then applied to output the pure neighbor features.

8. An electronic device, characterized in that, include: The processor, communication interface, memory, and communication bus are connected, with the processor, communication interface, and memory communicating with each other via the communication bus. The memory stores a computer program that, when executed by the processor, causes the processor to perform the graph attention trust management method based on dual-trust coding as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the graph attention trust management method based on dual-trust encoding as described in any one of claims 1 to 6.

Images