A trust mining method and system for intrinsic security based on graph neural network

Abstract

The present invention provides an intrinsically secure trust mining method based on a graph neural network, comprising the following steps: constructing a data interaction graph between data owners and data, obtaining an initial data owner embedding vector and a data embedding vector based on the data interaction graph; constructing a trust relationship graph between data owners, obtaining a trust embedding vector based on the trust relationship graph; updating the trust relationship graph to obtain node weights; using the initial data owner embedding vector, data embedding vector, and trust embedding vector as inputs to the graph neural network, training the network based on the node weights to obtain a hidden layer data embedding vector and a final data owner embedding vector; and predicting a data owner's interest score in data owned by one or more other data owners based on the final data owner embedding vector and the hidden layer data embedding vector. The present invention can securely and reliably mine data of interest for a target data owner.

Description

Technical Field

[0001] The present invention relates to the field of data mining and security protection technology, and in particular to a trust mining method and system with intrinsic security based on graph neural networks. Background Art

[0002] In today's era of explosive data growth, data mining techniques are becoming increasingly important for extracting valuable information or knowledge from vast amounts of low-value data. However, the data mining process can be accompanied by various data security threats, such as data leaks during transmission due to technical vulnerabilities or human attacks; malicious data providers may intentionally generate large amounts of meaningless data to confuse the true data owner and even induce the data owner to make diametrically opposed decisions. Therefore, a key research focus in data mining is how to ensure the security and reliability of the data mining process.

[0003] In existing technologies, a common security approach is to encrypt data during transmission and transmit the key separately over a secure channel. Without the correct key, valid content cannot be decrypted. Another common security approach is to conduct external audits to promptly identify and resolve security risks. However, both approaches rely heavily on additional security components. Given that external components themselves are susceptible to security attacks, these approaches cannot fundamentally address the various security threats facing data mining. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the present invention proposes an intrinsically secure trust mining method and system based on graph neural networks, which can safely and reliably mine data of interest for the target data owner.

[0005] The technical solution adopted in the present invention is as follows:

[0006] In the first aspect, a method for trust mining based on intrinsic security of a graph neural network is provided, comprising the following steps:

[0007] Construct a data interaction graph between the data owner and the data, and obtain the initial data owner embedding vector and data embedding vector based on the data interaction graph;

[0008] Construct a trust relationship graph between data owners and obtain a trust embedding vector based on the trust relationship graph;

[0009] Update the trust relationship graph and obtain the node weights;

[0010] The initial data owner embedding vector, data embedding vector, and trust embedding vector are used as the input of the graph neural network. Training is performed based on the node weights to obtain the hidden layer data embedding vector and the final data owner embedding vector.

[0011] Based on the final data owner embedding vector and the hidden layer data embedding vector, predict the interest score of a data owner for the data owned by one or more other data owners.

[0012] Furthermore, obtaining an initial data owner embedding vector includes: generating a first one-hot vector using the data owner's identity number according to the data interaction graph, and using the first one-hot vector as the initial data owner embedding vector.

[0013] Furthermore, according to the data interaction graph, the identity identification number of the data that has an interactive relationship with the data owner is used to generate a corresponding second one-hot vector, and the second one-hot vector is used as the data embedding vector.

[0014] Furthermore, obtaining a trust embedding vector includes: obtaining a trust embedding vector by extracting trust features from a trust relationship graph.

[0015] Furthermore, the trust relationship graph includes a trust matrix. The update process of the trust matrix is as follows:

[0016]

[0017] In the above formula, α and β are parameters that control the trust decay process, and l represents the number of updates of the trust relationship.

[0018] Furthermore, the data embedding vector of the hidden layer and the final data owner embedding vector are obtained, including:

[0019] Aggregate the updated trust relationship graph to obtain the global data owner data interaction graph;

[0020] Get the node weight of each node in the global data owner data interaction graph;

[0021] The initial data owner embedding vector, data embedding vector, and trust embedding vector are used as the input of the graph neural network. The hidden layer features are updated according to the node weight of each node to obtain the hidden layer data owner embedding vector, hidden layer data embedding vector, and hidden layer trust vector.

[0022] The data owner embedding vector of the hidden layer and the trust embedding vector of the hidden layer are fused to obtain the final data owner embedding vector.

[0023] Furthermore, it also includes: determining whether to accept the data based on the interest score and a preset security threshold.

[0024] Furthermore, the calculation result of the interest score is optimized by the following method:

[0025] The existing data owner's known score ε for the data is used as the reference score Reference score The corresponding data owner and the predicted interest score of the data are used as the predicted score By reference score and predicted scores The calculation results are optimized by the root mean square error between them.

[0026] Secondly, we provide an intrinsically secure trust mining system based on graph neural networks, which includes the following modules:

[0027] The trust measurement module is used to construct a data interaction graph between data owners and data, and obtain the initial data owner embedding vector and data embedding vector based on the data interaction graph; it is also used to construct a trust relationship graph between data owners and obtain the trust embedding vector based on the trust relationship graph.

[0028] Trust propagation module, used to update the trust relationship graph;

[0029] The trust aggregation module is used to obtain node weights based on the updated trust relationship graph. It is also used to use the initial data owner embedding vector, data embedding vector, and trust embedding vector as inputs to the graph neural network, train it based on the node weights, and obtain the hidden layer data embedding vector and the final data owner embedding vector.

[0030] The trust optimization module is used to predict a data owner's interest score in the data owned by one or more other data owners based on the final data owner embedding vector and the hidden layer data embedding vector, and is also used to determine whether to accept the data based on the interest score and a preset security threshold.

[0031] According to a third aspect, an electronic device is provided, including:

[0032] One or more processors; a storage device for storing one or more programs;

[0033] When one or more programs are executed by one or more processors, the one or more processors implement any one of the intrinsically secure trust mining methods based on graph neural networks provided in the first aspect.

[0034] It can be seen from the above technical solution that the beneficial technical effects of the present invention are as follows:

[0035] 1. The trust mining method provided in this embodiment models trust relationships in a graph neural network, uses an attention mechanism to bind trust relationships to social networks, embeds the data owner's interest score in the data into the trust level, and roots "trust" in the data mining method, thus breaking away from the constraints of security protection passively following data mining and possessing endogenous security protection capabilities.

[0036] 2. In real life, direct trust relationships between data owners are rare and often buried in a large amount of irrelevant data. The trust mining method provided in this embodiment introduces graph neural network technology for data mining to extract sufficient trust relationships and expand the trust relationships between data owners, thereby solving the sparse trust problem. At the same time, by updating the trust relationship graph and trust matrix, it can well reflect the changing process of data owners' trust, and when mining trust relationships, it can enhance the credibility of the results while ensuring accuracy.

[0037] 3. When the technical solution of the present invention is implemented, the data owner can manually adjust the degree of trust in a specific data owner to control the direction of data mining, so that the mining results are more in line with the subjective wishes of the data owner. Therefore, the trust mining method provided by the present invention has good scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0039] Figure 1 A flow chart of a trust mining method according to an embodiment of the present invention;

[0040] Figure 2 A schematic diagram of the relationship between data owners and data in one embodiment of the present invention;

[0041] Figure 3 Schematic diagram of a trust propagation process according to an embodiment of the present invention;

[0042] Figure 4 A schematic diagram of a graph-based attention mechanism for predicting data owners' ratings of data in one embodiment of the present invention;

[0043] Figure 5 A schematic diagram of a system architecture in an embodiment of the present invention;

[0044] Figure 6 The figure is a schematic diagram of the internal structure of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following embodiments of the technical solution of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore only examples and are not intended to limit the scope of protection of the present invention.

[0046] It should be noted that, unless otherwise specified, the technical or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0047] Example

[0048] In the process of social interaction, data owners not only follow their personal preferences but also the opinions of the media and their friends. This means that data mining results are easily influenced by trust relationships, which are inherent and long-standing in social networks. Therefore, this embodiment inherently embeds the data owner's "trust" as a "security gene" into the data mining process.

[0049] In practice, the information generated during data mining can be used to construct two subgraphs. These graphs include a social graph representing the relationships between data owners and a data owner-data graph representing the interactions between data owners and their data. Data owners participating in both graphs can be connected. This natural approach, based on interpersonal recommendations, incorporates social network information into the underlying factors of learning data owners and data. Learning representations for data owners and data is key to data mining.

[0050] In machine learning, graph neural networks (GNNs), a deep neural network architecture, are used to learn meaningful representations of graph data. Their key idea is to use neural networks to iteratively aggregate feature information from local graph neighborhoods. Furthermore, node information can be transformed and aggregated after propagating through the graph. Therefore, GNNs naturally integrate node information and topological structure, demonstrating powerful learning capabilities for data mining. Trust mining methods based on GNNs can naturally integrate node information and topological structure, effectively addressing the sparse trust problem by learning embedding vectors to predict the data owner's potential interest in the data.

[0051] Based on the above analysis, this embodiment provides a method for trust mining with intrinsic security based on a graph neural network, including the following steps:

[0052] S11. Build a data interaction diagram between data owners and data

[0053] In the information-overloaded internet, there are a large number of interaction records between data owners and data. By treating data owners and data as nodes and the interactions between them as edges, we can construct a data interaction graph between data owners and data.

[0054] In a specific implementation, the set of data owners is denoted as The data set is denoted as The data owner's interest score set for the data is denoted as The data interaction graph between the data owner and the data is represented as G(U,I,ε). The sets U, I, and ε can all be obtained from existing information or databases.

[0055] S12. Obtain the initial data owner embedding vector and data embedding vector based on the data interaction graph

[0056] In the data interaction graph, the ID (identification number) of the data owner u corresponds to the data owner one-to-one and is unique. Use the ID (identification number) of the data owner u to generate the corresponding one-hot vector, and use this one-hot vector as the initial data owner embedding vector e u

[0057] Use the ID of data i that has an interactive relationship with the data owner u to generate the corresponding one-hot vector, and use this one-hot vector as the embedding vector e of data i that has an interactive relationship with the data owner u i .

[0058] e u and e i This is used for graph neural network training in subsequent steps.

[0059] S13. Build a trust relationship graph based on the trust relationship between data owners

[0060] In real life, data owners’ choices of new and needed data not only follow their personal preferences, but are also often influenced by their friends, and different friends have different influences on data owners. Specifically, Figure 2 As shown, data owner 1 and data owner 4 are both friends of data owner 3; data owner 3 trusts data owner 1 very much, but data owner 3 does not trust data owner 4 that much; thus, it can be concluded that data owner 3 is more likely to trust data 1 held by data owner 1 and believes that data 1 is safe and reliable.

[0061] In a specific implementation, the social relationship data of each data owner in real life is collected, and the trust relationship graph constructed based on the social relationship is represented as G uv (G u ,G v, ), where u represents the trustor, v represents the trusted person, The trust relationship between the trustor u and the trusted person v is represented by the trust matrix All data owners can be either trustors or trusted parties. Dividing data owners into trustors and trusted parties can better measure the trust relationship between data owners.

[0062] S14. Obtain trust embedding vector based on trust relationship graph

[0063] In a specific manner, the trust embedding vector is obtained by extracting the trust features in the trust relationship graph. The trust embedding vector e t It can be expressed as follows:

[0064] e t =e u ⊕e v (1)

[0065] In the above formula (1), e u represents the trust embedding vector of the trustor, e v represents the embedding vector of the trusted person, and ⊕ represents the concatenation operation of two embeddings.

[0066] S21. Update the trust relationship diagram

[0067] like Figure 2 As shown, data owner 2 is a friend of data owner 4, data owner 4 is a friend of data owner 3, and data owner 3 is a friend of data owner 1. Through the above social relationships, it can be inferred that data owner 2 is a friend of data owner 1, and data owner 2 may believe that data 1 held by data owner 1 is safe and reliable.

[0068] According to the above analysis, in this step, the trust relationship graph can be updated by filling in the target data owner’s trust in the data owner who did not have a trust score before based on the transitivity of trust. The updated trust relationship graph can be expressed as Where l∈(1,2,…,L) is the number of updates of the trust relationship.

[0069] In some embodiments, when the trust relationship graph is updated and continuously expanded, it will eventually become one or more fully connected graphs, which may cause overfitting. To ensure the effectiveness of the update, the influence of friends should be set to become weaker as the trust relationship expands. In other words, Figure 2As shown in the figure, in the trust propagation chain of "data owner 2-data owner 3-data owner 4-data owner 1", data owner 4 has the strongest influence on data owner 2, the influence of data owner 3 on data owner 2 weakens, and data owner 1 has the weakest influence on data owner 2.

[0070] In the trust relationship diagram, the trust relationship between the trustor u and the trusted person v is represented by the trust matrix The trust matrix can be obtained based on the trust relationship between data owners in a known set of similar data owners. It can be expressed as follows:

[0071]

[0072] In the above formula (2), is a randomly generated vector, v represents the trusted person, U represents the data owner set, e t represents the trust embedding vector.

[0073] The updating process of the trust matrix is as follows: Figure 3 As shown, the logistic regression function can be used to represent this trust decay process, that is, the update process of the trust matrix can be expressed as follows:

[0074]

[0075] In the above formula (3), α and β are parameters that control the trust decay process, and l represents the number of updates of the trust relationship.

[0076] This step is a trust propagation process. After the trust propagation, the trust relationship graph between each pair of trustors and trusted persons is updated, and the trust relationship graph is represented by the trust matrix. The trust strength is stable at

[0077] S31. Aggregate the updated trust relationship graph to obtain the global data owner data interaction graph

[0078] In a specific embodiment, all trust relationship graphs obtained from step S21 can be aggregated to obtain the global data owner data interaction graph at this time:

[0079] G ′ (U,I,ε ′ )=Agg G (G ′ uv |u,v∈U)(4)

[0080] In the above formula (4), G ′ uvRepresents the updated trust relationship graph, u represents the trustor, v represents the trusted, Agg G Represents the aggregation function of the graph. In a specific embodiment, Agg G The selection is not limited and can be implemented in any feasible manner in the prior art.

[0081] S32. Obtain the node weight of each node in the global data owner data interaction graph

[0082] In a specific implementation, since each node contributes differently to the score, the attention mechanism is used to learn the node weight of each node in the global data owner data interaction graph. The process is shown as follows: Figure 4 The node weight of each node can be expressed as:

[0083]

[0084] In the above formula (5), σ(·) is the LeakyReLU activation function, k∈(u,i,t), p is the element in k; the data owner weight, data weight, and trust weight are represented by γ u , γ i and γ t .

[0085] S33. Use the initial data owner embedding vector, data embedding vector, and trust embedding vector as the input of the graph neural network, update the hidden layer features according to the node weight of each node, and obtain the hidden layer data owner embedding vector, hidden layer data embedding vector, and hidden layer trust vector.

[0086] In the process of initializing the embedding representation, the initial data owner obtained in step S11 is embedded in the vector e u , the data embedding vector e obtained in step S32 i and the trust embedding vector e obtained in step S13 t As the input of the graph neural network, the first hidden layer features of the data owner are obtained First data features and first trust characteristics In the subsequent embedding representation process, the hidden layer features are updated according to the following formula (6):

[0087]

[0088] In the above formula (6), q∈(1,2,…,Q) represents the number of hidden layers, γ k represents the node weight of each node, k∈(u,i,t), and They represent the p-th hidden layer features of the data owner embedding vector, the p-th hidden layer features of the data embedding vector, and the p-th hidden layer features of the trust embedding vector.

[0089] Through this step, we can obtain the p-th hidden layer features of the user vector, the p-th hidden layer features of the item vector, and the p-th hidden layer features of the trust vector. After completing Q updates, the hidden layer data owner embedding vector is expressed as The data owner embedding vector of the hidden layer is expressed as The trust vector of the hidden layer is expressed as

[0090] S34, merge the hidden layer data owner embedding vector and the hidden layer trust embedding vector to obtain the final data owner embedding vector

[0091] After L updates to the trust relationship graph in step S21 above, the trust matrix stabilizes at According to the following formula (7), the data owner embedding vector of the hidden layer and the trust embedding vector of the hidden layer are fused to obtain the final data owner embedding vector:

[0092]

[0093] In the above formula (7), represents the final data owner embedding vector, represents the data owner embedding vector of the hidden layer, represents the trust vector of the hidden layer, denoted as the trust matrix after updating L times, v denotes the trusted party, and U denotes the set of data owners.

[0094] S41. Predict a data owner's interest score for the data owned by one or more other data owners based on the final data owner embedding vector and the hidden layer data embedding vector.

[0095] In a specific embodiment, let the data embedding vector Equal to the data embedding vector of the hidden layer Embed the data owner obtained after training into the vector and data embedding vector According to the following formula, we can calculate and predict the interest score of a data owner in the data owned by one or more other data owners by multiplying:

[0096]

[0097] In some embodiments, the calculation result of the predicted interest score can be optimized by the following method: in the training phase of the method, the known scores ε of the data by the existing small number of data owners in step S11 are used as reference scores Reference score The scores of the corresponding data owner u and data i obtained in step S41 are used as the predicted scores By reference score and predicted scores The root mean square error between them is optimized according to formula (9):

[0098]

[0099] In the above formula (9), Represents the number of predicted ratings; the optimization goal is to use the Adam optimizer of the graph neural network to make the predicted ratings As close to the actual reference score as possible

[0100] In some embodiments, the decision on whether to accept the data may be based on the interest score and a preset security threshold. The highest score can be recorded as C ui The data owner can set a security threshold ζ based on his or her own security needs. If the score is , it can be determined that the data meets the security settings and is acceptable; otherwise, , deny the data.

[0101] The intrinsically secure trust mining method based on graph neural networks proposed in this paper can be used for data mining in various formats, such as articles, music, and videos. The following uses researchers as data owners and papers as data to illustrate the intrinsically secure trust mining process:

[0102] Assuming there are 4 researchers and 6 papers, the data interaction graph G is constructed according to step S11 as follows:

[0103]

[0104] The rows of the data interaction graph G represent researchers, and the columns represent papers. The values in G represent researchers' interest ratings for the papers, with 1 indicating interest and 0 indicating no rating at this time. As can be seen from the data interaction graph G, researcher 2, in the second row, has no data on any papers, making him the primary target for mining.

[0105] Then, we construct a trust relationship graph and a trust matrix T. To simplify the graph neural network part, we assume that the product of the random vector and the trust embedding vector is infinitely close to 1, and we can get a 4×4 trust matrix T:

[0106]

[0107] The trust matrix T represents the mutual trust scores of the four researchers, and the default trust level of the researchers in themselves is 1.

[0108] Next, for a trust propagation chain of “scientific researcher 2 - scientific researcher 3 - scientific researcher 4 - scientific researcher 1”, the trust relationship graph and the trust matrix are updated according to step S21. Assume that the parameters controlling the trust decay process are α = 1.1 and β = 0.6; after the first update, since researcher 2 trusts researcher 3, researcher 3's interest in the paper will be reflected in researcher 2, and researcher 2's trust score in the researchers changes to [0.0, 1.0, 0.16388802, 0.60938186]; after the second update, since researcher 3 trusts researcher 4, researcher 4's interest in the paper will be reflected in researcher 2, and researcher 2's trust score in the researchers changes to [0.12431152, 1.0, 0.53523427, 0.64281438]; after the third update, researcher 2's trust score in the researchers changes to [0.35060319, 1.0, 0.61098746, 0.64281438].

[0109] After trust aggregation in steps S31-S35, a researcher's trust in other researchers will be reflected in the researcher's embedding vector.

[0110] Then, in step S41, the interest scores of researcher 2 for the five papers (Paper 1, Paper 2, Paper 3, Paper 4, Paper 6) can be obtained. Assuming that researcher 2's interest score for paper 4 is 0.673 (the highest score is 1), and the safety threshold set by researcher 2 is 0.6, then researcher 2 can accept paper 4.

[0111] The trust mining method provided in this embodiment models the trust relationship in a graph neural network, uses the attention mechanism to bind the trust relationship to the social network, embeds the data owner's interest score in the data into the trust level, and roots "trust" in the data mining method, thereby breaking away from the constraints of security protection passively following data mining and having endogenous security protection capabilities.

[0112] In real life, direct trust relationships between data owners are rare and are often buried in a large amount of irrelevant data. The trust mining method provided in this embodiment introduces graph neural network technology for data mining to extract sufficient trust relationships and expand the trust relationships between data owners, thereby solving the sparse trust problem. At the same time, by updating the trust relationship graph and trust matrix, it can well reflect the changing process of data owners' trust, and when mining trust relationships, it can enhance the credibility of the results while ensuring accuracy.

[0113] When the technical solution of the present invention is implemented, the data owner can control the data mining direction by manually adjusting the degree of trust in a specific data owner, so that the mining results are more in line with the subjective wishes of the data owner. Therefore, the trust mining method provided by the present invention has good scalability.

[0114] In one embodiment, a trust mining system with intrinsic security based on graph neural network is provided, including the following modules:

[0115] The trust measurement module is used to construct a data interaction graph between data owners and data, and obtain the initial data owner embedding vector and data embedding vector based on the data interaction graph; it is also used to construct a trust relationship graph between data owners and obtain the trust embedding vector based on the trust relationship graph.

[0116] Trust propagation module, used to update the trust relationship graph;

[0117] The trust aggregation module is used to obtain node weights based on the updated trust relationship graph. It is also used to use the initial data owner embedding vector, data embedding vector, and trust embedding vector as inputs to the graph neural network, train it based on the node weights, and obtain the hidden layer data embedding vector and the final data owner embedding vector.

[0118] The trust optimization module is used to predict a data owner's interest score in the data owned by one or more other data owners based on the final data owner embedding vector and the hidden layer data embedding vector, and is also used to determine whether to accept the data based on the interest score and a preset security threshold.

[0119] In one embodiment, an electronic device is also provided. The electronic device may be a computer, and its internal structure diagram may be as shown in FIG. Figure 6As shown. The electronic device includes a processor, a memory, a network interface and a database connected via a system bus. The processor of the electronic device is used to provide computing and control capabilities. The memory of the electronic device includes a non-volatile storage medium and an internal memory; the non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the electronic device is used to store configuration templates and can also be used to store target data. The network interface of the device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a trust mining method with intrinsic security based on a graph neural network.

[0120] Those skilled in the art will understand that Figure 6 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the device to which the solution of the present application is applied. The specific device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0121] Obviously, those skilled in the art should understand that the modules or steps of the present invention described above can be implemented using a general-purpose computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices. Alternatively, they can be implemented using program code executable by the computing device, so that they can be stored in a computer storage medium (ROM / RAM, magnetic disk, optical disk) and executed by the computing device. In some cases, the steps shown or described can be performed in a different order than herein, or they can be made into individual integrated circuit modules, or multiple modules or steps can be made into a single integrated circuit module for implementation. Therefore, the present invention is not limited to any specific combination of hardware and software.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A trust mining method based on graph neural network with intrinsic security, characterized by: The following steps are involved: Construct a data interaction graph between the data owner and the data, and obtain the initial data owner embedding vector and data embedding vector based on the data interaction graph; Construct a trust relationship graph between data owners and obtain a trust embedding vector based on the trust relationship graph; Update the trust relationship graph to obtain node weights; the trust relationship graph includes a trust matrix, and the update process of the trust matrix is as follows: In the above formula, 、 is the parameter that controls the trust decay process, Indicates the number of updates of the trust relationship; The initial data owner embedding vector, the data embedding vector, and the trust embedding vector are used as inputs of the graph neural network, and training is performed according to the node weights to obtain the data embedding vector of the hidden layer and the final data owner embedding vector, including: aggregating the updated trust relationship graph to obtain a global data owner data interaction graph; obtaining the node weight of each node in the global data owner data interaction graph; using the initial data owner embedding vector, the data embedding vector, and the trust embedding vector as inputs of the graph neural network, updating the hidden layer features according to the node weight of each node to obtain the data owner embedding vector of the hidden layer, the data embedding vector of the hidden layer, and the trust vector of the hidden layer; fusing the data owner embedding vector of the hidden layer and the trust embedding vector of the hidden layer to obtain the final data owner embedding vector; Based on the final data owner embedding vector and the hidden layer data embedding vector, predict the interest score of a data owner for the data owned by one or more other data owners.

2. The trust mining method according to claim 1, characterized in that: Obtaining an initial data owner embedding vector includes: generating a first one-hot vector using the data owner's identity number according to the data interaction graph, and using the first one-hot vector as the initial data owner embedding vector.

3. The trust mining method according to claim 1, characterized in that: Obtaining the data embedding vector includes: generating a corresponding second one-hot vector using an identity identification number of data that has an interactive relationship with a data owner according to the data interaction graph, and using the second one-hot vector as the data embedding vector.

4. The trust mining method according to claim 1, characterized in that: Obtaining a trust embedding vector includes: obtaining a trust embedding vector by extracting trust features from a trust relationship graph.

5. The trust mining method according to claim 1, characterized in that: Also includes: Determine whether to accept the data based on the interest score and preset security thresholds.

6. The trust mining method according to claim 1, characterized in that: It also includes optimizing the calculation results of interest scores in the following ways: The known ratings of the data by the existing data owners are taken as reference ratings, and the interest ratings of the data owners and data corresponding to the reference ratings are taken as predicted ratings. The calculation results are optimized by the root mean square error between the reference ratings and the predicted ratings.

7. A trust mining system with intrinsic security based on graph neural network, characterized by: Includes the following modules: The trust measurement module is used to construct a data interaction graph between data owners and data, and obtain the initial data owner embedding vector and data embedding vector based on the data interaction graph. It is also used to construct a trust relationship graph between data owners and obtain the trust embedding vector based on the trust relationship graph. The trust propagation module is used to update the trust relationship graph; the trust relationship graph includes a trust matrix, and the updating process of the trust matrix is as follows: In the above formula, 、 is the parameter that controls the trust decay process, Indicates the number of updates of the trust relationship; A trust aggregation module is used to obtain node weights based on the updated trust relationship graph, and is also used to use the initial data owner embedding vector, data embedding vector, and trust embedding vector as inputs to the graph neural network, perform training based on the node weights, and obtain the hidden layer data embedding vector and the final data owner embedding vector, including: aggregating the updated trust relationship graph to obtain a global data owner data interaction graph; obtaining the node weight of each node in the global data owner data interaction graph; using the initial data owner embedding vector, the data embedding vector, and the trust embedding vector as inputs to the graph neural network, updating the hidden layer features based on the node weights of each node to obtain the hidden layer data owner embedding vector, the hidden layer data embedding vector, and the hidden layer trust vector; fusing the hidden layer data owner embedding vector and the hidden layer trust embedding vector to obtain the final data owner embedding vector; The trust optimization module is used to predict a data owner's interest score in the data owned by one or more other data owners based on the final data owner embedding vector and the hidden layer data embedding vector, and is also used to determine whether to accept the data based on the interest score and a preset security threshold.

8. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the intrinsically secure trust mining method based on graph neural network according to any one of claims 1 to 6.

Images