Deep clustering-based blockchain entity association relationship evaluation method, system and device

Abstract

The application provides a deep clustering-based blockchain entity association relationship evaluation method, system and device. First, block information is extracted, data describing entities is converted into a first matrix, and a pseudo label is randomly assigned to each entity. The first matrix is subjected to data enhancement and input into a feature extraction network to obtain feature information. The feature information is mapped to a subspace, contrast learning is performed, and the parameters of the feature extraction network are updated. The parameters of the feature extraction network are repeatedly updated, entities are classified according to the pseudo label, and the behavior similarity between entities is obtained. The deep clustering-based method can accurately and effectively classify the entities with complex relationships in the blockchain ecology and detect the behavior similarity. The training set does not need to be marked in advance, the blockchain entity clustering process can be completed in an unsupervised situation, and the construction of the three-dimensional integrated entity association relationship of accounts, transactions and chain groups is realized.

Description

Technical Field

[0001] This invention relates to the field of blockchain technology, and specifically to a method for evaluating the association relationship of blockchain entities based on deep clustering. Technical Background

[0002] In recent years, blockchain technology, characterized by decentralization and immutability, has developed rapidly and is widely used in many fields. Correspondingly, blockchain security and regulation have become a focus of attention both domestically and internationally.

[0003] In the blockchain ecosystem, the number of interconnected entities such as accounts, transactions, and blockchain clusters is vast and their relationships are complex, making type classification and behavioral similarity detection difficult. Currently, there is a lack of expert rules capable of accurate feature recognition. Furthermore, supervised learning methods, which are relatively effective in classification tasks, are difficult to apply. There is a lack of fully labeled training sets, and manually assigning data labels to blockchain entities is labor-intensive and inefficient. Therefore, there is an urgent need for an accurate and efficient method to construct relationships between blockchain entities. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method for evaluating blockchain entity relationships based on deep clustering. Deep clustering is a method that combines representation learning and clustering techniques. Specifically, it includes the following steps:

[0005] (1) Extract block information, convert the data describing entities into matrix form, and randomly assign a pseudo-label to each entity;

[0006] (2) After data augmentation of the matrix, input it into the feature extraction network and output feature information;

[0007] (3) Map the feature information to the subspace, perform comparative learning, and update the parameters of the feature extraction network;

[0008] (4) Calculate and update the pseudo-label for each entity, and update the parameters of the feature extraction network;

[0009] (5) Repeat the training process of steps (2)-(4), then classify entities according to pseudo-labels and calculate the behavioral similarity between entities.

[0010] Furthermore, the specific implementation of step (1) is as follows:

[0011] 1.1 Extract block information, including entity account, transaction, and chain group information. Each entity is represented as X. i ,i∈[n], where n represents the number of entities;

[0012] 1.2 Randomly assign a pseudo-label y to each entity. i∈[m], i∈[n], where m is the number of entity categories to be divided.

[0013] Furthermore, the feature extraction network in step (2) consists of convolutional layers and a transformer encoder, and its specific implementation is as follows:

[0014] 2.1 Using different data augmentation methods T a and T b Perform operations on X to generate x. a,o =T a (X i ),x b,i =T b (X i ), i∈[n];

[0015] 2.2 for x a ,x b The embedding operation is performed by transforming it into a matrix of dimension d using lightweight stacked convolutions. The convolutional layers use four 3×3 convolutions with a stride of 2. Each convolution is followed by batch normalization and a linear rectified function, and finally by a 1×1 convolution with a stride of 1.

[0016] 2.3 Input the matrix generated in step 2.2 into the transformer encoder and output the feature h. a ,h b .

[0017] Furthermore, the specific implementation of step (3) is as follows:

[0018] 3.1 For feature h a ,h b Perform two linear transformations, batch normalization, and linear rectification functions respectively, and then linearly map them to the subspace z. a ,z b ;

[0019] 3.2 Conduct comparative learning to improve the {z} generated by the same entity. a,i ,z b,i The similarity between} reduces the {z} generated by different entities. a,i ,z b,j The similarity between {i ≠ j}, where similarity is defined as... Calculate the loss function Then, Adam is used as the optimizer to update the parameters of the feature extraction network.

[0020] Furthermore, the specific implementation of step (4) is as follows:

[0021] 4.1 Input X into the updated feature extraction network to obtain feature information h;

[0022] 4.2 Perform two linear transformations, batch normalization, and linear rectification on feature h, and then pass it through the flexible maximum transfer function to obtain the pseudo-label y'. n ;

[0023] 4.3 To prevent the clustering process from grouping most entities into the same class, if a pseudo-label c exists in a certain step (1) in y' n If it doesn't exist, perform a reassignment operation. Select the label category with the most entities, keep half of the entities' labels unchanged, and change the labels of the other half of the entities to 'c'. If y' n If the tag includes all types, no reassignment operation will be performed;

[0024] 4.4 Calculate y obtained from the previous training round n and the recalculated y' n The cross-entropy loss function is used, and SGD is used as the optimizer to update the parameters of the feature extraction network.

[0025] Furthermore, in step (5), after obtaining the optimized feature extraction network through the training process of repeating steps (2)-(4), steps 4.1 and 4.2 are executed to obtain the labels corresponding to the entities. Entities corresponding to the same label are classified into the same type of entities. Selecting a small number of entities within the same type for specific analysis can make the features of that type of entity more interpretable. Feature extraction is performed on the entities, and the behavioral similarity between entities can be associated through steps 3.1 and 3.2 to construct the relationship between entities.

[0026] The significant benefit of this invention is that, based on deep clustering, it can accurately and effectively classify entities with complex relationships within the blockchain ecosystem and detect behavioral similarities. This method does not require a pre-labeled training set and can complete the blockchain entity clustering process in an unsupervised manner, achieving the construction of a three-dimensional entity relationship framework encompassing accounts, transactions, and blockchain clusters. Attached Figure Description

[0027] Figure 1 This is a flowchart of a blockchain entity association evaluation method in a specific embodiment of the present invention. Detailed Implementation

[0028] To explain the technical solutions and advantages of the present invention more clearly and accurately, the present invention will be further described in detail below with reference to the accompanying drawings in the embodiments of the present invention.

[0029] like Figure 1 As shown, a method for evaluating the association relationship of blockchain entities based on deep clustering mainly consists of five steps.

[0030] (1) Extract block information, including entity account, transaction, and chain group information, and transform the data describing the entity into a matrix form, with each entity represented as X. i ,i∈[n], and randomly assign a pseudo-label y to each entity. i ∈[m], i∈[n], where n is the number of entities and m is the number of entity categories to be divided;

[0031] (2) Using different data augmentation methods T a and T b Generate x a,i =T a (X i ),x b,i =T b (X i For i∈[n], pass in a lightweight stacked convolutional layer, specifically four 3×3 convolutions with a stride of 2. Each convolution is followed by Batch Norm and ReLU, and finally passed through a 1×1 convolution with a stride of 1. The resulting matrix is ​​then passed to the transformer encoder to output the feature h. a ,h b ;

[0032] (3) Conduct comparative learning on feature h a ,h b Perform two linear transformations, batch normalization, and linear rectification functions respectively, and then linearly map them to the subspace z. a ,z b , improve the {z} generated by the same entity a,i ,z b,i The similarity between} reduces the {z} generated by different entities. a,i ,z b,j The similarity between {i ≠ j}, where similarity is defined as... Using loss function The loss value is obtained, and Adam is used as the optimizer to update the parameters of the feature extraction network;

[0033] (4) Input X into the updated feature extraction network to obtain feature information h. Perform two linear transformations, batch normalization, and linear rectification functions on feature h, and then pass it through the flexible maximum transfer function to obtain the pseudo-label y'. n If a pseudo-label c exists in a certain step (1) in y' n If it doesn't exist, perform a reassignment operation. Select the label category with the most entities, keep half of the entities' labels unchanged, and change the labels of the other half of the entities to 'c'. If y' nIf the label includes all types, no reassignment operation is performed; instead, the y value obtained from the previous training round is calculated. n and the recalculated y' n The cross-entropy loss between them is used, and SGD is used as the optimizer to update the parameters of the feature extraction network;

[0034] (5) After obtaining the optimized feature extraction network through the training process of repeating steps (2)-(4), the matrix representing the entity is re-inputted. After passing through the feature extraction network and the clustering process, the labels corresponding to the entities are obtained. Entities with the same label are classified into the same type of entity. Features are extracted from the entities, and the behavioral similarity between entities is associated through step (3) to construct the relationship between entities.

Claims

1. A method for evaluating the association relationship of blockchain entities based on deep clustering, characterized in that, Includes the following steps: Extract block information, transform the relevant data of the obtained entities into a first matrix, and randomly assign a pseudo-label to each entity. The entities include accounts, transaction accounts and chain groups on the blockchain network, and the relevant data of the entities include account information, transaction data and chain group information on the blockchain network. The first matrix is ​​subjected to data augmentation processing and then input into a feature extraction network to obtain feature information; The feature extraction network includes convolutional layers and Encoder; Based on different data augmentation methods and Perform operations on the first matrix to generate ,in, Represents an entity, For the number of entities, Index representing an entity, , These represent the different vectors formed after the same entity has undergone two different data augmentation methods. right The embedding operation is performed, and it is transformed into a second matrix using lightweight stacked convolution; Input the second matrix into The encoder obtains feature information. ; The feature information is mapped to a subspace, comparative learning is performed, and the parameters of the feature extraction network are updated. The pseudo-label of each entity is calculated and updated based on the feature vectors in the subspace, and the parameters of the feature extraction network are updated accordingly. Repeat the steps of updating the parameters of the feature extraction network until the parameters of the feature extraction network converge, and obtain the constructed association model; Entities are classified based on the pseudo-labels, and the behavioral similarity between entities is obtained.

2. The method for evaluating blockchain entity association relationships based on deep clustering according to claim 1, characterized in that, The lightweight stacked convolution uses 4 Size, step size The convolutions are followed by batch normalization and linear rectified functions after each convolution, and finally passed through a... Step size is The convolution.

3. The method for evaluating blockchain entity association relationships based on deep clustering according to claim 1, characterized in that, The process of mapping the feature information to a subspace, performing comparative learning, and updating the parameters of the feature extraction network includes the following steps: The feature information is linearly mapped to a subspace by performing two linear transformations, batch normalization, and linear rectification functions on the feature information. ; Perform comparative learning to enhance the learning generated by the same entity. The similarity between them reduces the generation of different entities The similarity between them, where similarity is defined as ,in, Representing two feature vectors with the same dimension, using the loss function The loss value is obtained as the optimization objective, and then used... The parameters of the feature extraction network are updated by the optimizer.

4. The method for evaluating blockchain entity association relationships based on deep clustering according to claim 1, characterized in that, The process of calculating and updating the pseudo-label for each entity based on the feature vectors in the subspace, and updating the parameters of the feature extraction network, includes the following steps: The matrix representing all entities is input into the updated feature extraction network to obtain feature information; The pseudo-tags are obtained by performing two linear transformations, batch normalization, and linear rectification functions on the feature information, followed by a flexible maximum transfer function. ; If the entity's initial pseudo-label Not included If the first half of the entities are in the first half, a redistribution operation is performed. The label with the largest number of entities is selected, and half of the entities in that category retain their original labels while the labels of the other half are changed. ,if If all types of pseudo-tags are included, no reallocation operation will be performed; Calculate the pseudo-labels obtained from the previous training round and The cross-entropy loss between them is used as the optimization objective, and... The parameters of the feature extraction network are updated by the optimizer.

5. The method for evaluating blockchain entity association relationships based on deep clustering according to claim 1, characterized in that, The process of classifying entities based on the pseudo-labels and obtaining behavioral similarity between entities includes the following steps: The matrix representing all entities is input into the updated feature extraction network to obtain feature information; The pseudo-tags are obtained by performing two linear transformations, batch normalization, and linear rectification functions on the feature information, followed by a flexible maximum transfer function. Entities corresponding to the same tag are classified into entities of the same type; Selecting a small number of entities within the same type for specific analysis makes the features more interpretable; Feature extraction is performed on the entity, and the feature information is subjected to two linear transformations, batch normalization, and linear rectification functions to linearly map it to a subspace. ; Comparative learning can improve the performance of products generated from the same entities. The similarity between them reduces the generation of different entities The similarity between them, where similarity is defined as Using loss function Obtain the loss value and use it. The parameters of the feature extraction network are used as optimizers to update the network, evaluate the behavioral similarity between entities, and construct the relationships between entities.

6. A blockchain entity association evaluation system based on deep clustering, characterized in that, It includes an information extraction module, a feature processing module, a first update module, a second update module, and a classification processing module; The information extraction module is used to extract block information, convert the data describing entities into a first matrix, and randomly assign a pseudo-label to each entity. The entity includes accounts, transaction accounts, and chain groups on the blockchain network, and the relevant data of the entity includes account information, transaction data, and chain group information on the blockchain network. The feature processing module performs data augmentation on the first matrix and inputs it into the feature extraction network to obtain feature information; The first update module is used to map the feature information to a subspace, perform comparative learning, and update the parameters of the feature extraction network; The second update module is used to calculate and update the pseudo-label of each entity based on the feature vectors in the subspace, and update the parameters of the feature extraction network. The classification processing module is used to repeatedly update the parameters of the feature extraction network, classify entities based on the pseudo-labels, and obtain the behavioral similarity between entities.

7. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 5.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 5.

Images