Multi-modal feature fusion method and device for cross-chain behavior analysis and medium

Abstract

The invention relates to a multi-modal feature fusion method and device for cross-chain behavior analysis, and a medium. The method comprises the following steps: obtaining various modal data on a heterogeneous chain, carrying out feature extraction by adopting a corresponding feature extractor, carrying out normalization processing, and converting original data into a uniform numerical feature vector; carrying out vector alignment on the basis of the converted numerical feature vectors, dynamically allocating a weight for each vector by utilizing an attention mechanism, correcting attention scoring on the basis of reliability, time sequence attenuation and structural sensitivity in dynamic weight allocation, and introducing a gating coefficient based on modal-context combination to suppress a noise modal; weighted summation is carried out on the numerical feature vectors based on the weights to obtain fusion vectors, and then on-chain behavior fingerprint vectors are generated; and the fingerprint vector is encrypted by using a homomorphic encryption technology and then is used for cross-chain behavior analysis. Compared with the prior art, the method has the advantages of being capable of processing cross-chain heterogeneous data, high in privacy and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology and data science, and particularly to a multi-modal feature fusion method for cross-chain behavior analysis, a device and a medium. BACKGROUND

[0002] With the rapid development of information technology, various data systems are emerging, forming a highly heterogeneous and complex environment. In the field of blockchain technology, the parallel development of public chains, consortium chains, and private chains has led to a multi-chain, multi-protocol coexistence ecosystem. This heterogeneity brings huge management and security challenges.

[0003] The existing data analysis and monitoring methods mainly have the following problems: 1. Data silos and single-modal analysis: Most existing technologies are limited to analyzing a single blockchain network and a single type of data (e.g., transaction data). This method cannot obtain a complete system overview, and it is even more difficult to discover complex security threats that arise from cross-chain, multi-dimensional coordination (e.g., by distributing malicious assets across different chains or coordinating attacks).

[0004] 2. Difficulty in processing heterogeneous data: In a multi-chain environment, the data structures, consensus mechanisms, and communication protocols of different chains differ greatly, resulting in highly heterogeneous data formats. Traditional methods are difficult to effectively integrate and process multiple data modalities (e.g., transaction data, text data, image data, etc.) from different chains, thus limiting the ability of in-depth analysis.

[0005] 3. Privacy and security risks: When conducting cross-chain data analysis, directly transmitting and sharing raw data can pose a serious risk of privacy leakage. Traditional data aggregation methods usually require centralized processing, which not only poses a single point of failure risk, but also cannot meet the compliance requirements of privacy protection.

[0006] Chinese patent CN120671191A discloses a privacy computing method and system based on cross-modal data fusion, the method comprising: obtaining visual, text and audio modal original data and extracting features thereof to obtain initial feature representation; performing dimension standardization and combination to generate a unified initial feature vector; homomorphic encryption and dimension reduction are performed on the unified initial feature vector to obtain an encrypted feature representation, and modal alignment optimization is completed to construct a feature space mapping relationship; identifying the correlation path and screening the intensity based on the mapping relationship to obtain the correlation feature path; completing weighted fusion based on the correlation feature path to generate a fusion weight distribution matrix; using the matrix and the encrypted feature representation to perform nonlinear mapping and noise processing to obtain a privacy-protected fusion output; finally, the accuracy is evaluated and the weight is optimized to output the final fusion result. Although this method realizes multi-modal data fusion and solves the problem of insufficient privacy protection in the fusion process, the method does not consider the characteristics of cross-chain multi-modal data in the fusion process, and the fusion result cannot accurately detect cross-chain abnormal behavior and attack events. SUMMARY

[0007] The purpose of the present application is to overcome the defects of the prior art and provide a multi-modal feature fusion method, device and medium for cross-chain behavior analysis, to solve the problem of insufficient data analysis capability and single processing method in a heterogeneous multi-source distributed environment such as a cross-chain blockchain cluster.

[0008] The purpose of the present application can be achieved by the following technical solutions: According to a first aspect of the present application, a multi-modal feature fusion method for cross-chain behavior analysis is provided, the method comprising the following steps: Data modalization and vectorization: obtaining multiple modal data on a heterogeneous chain, using a corresponding feature extractor to extract features, and performing normalization processing to convert the original data of each modal into a unified numerical feature vector; Cross-chain feature vector alignment and dynamic weight distribution: based on the converted numerical feature vector, the vectors are aligned, and the attention mechanism is used to dynamically distribute weights to each vector, wherein in the dynamic weight distribution, the attention score correction is based on reliability, time decay and structure sensitivity, and a modal-context joint based gating coefficient is introduced to suppress noise modal; Multi-modal fusion and fingerprint vector generation: based on the weights, the numerical feature vectors are weighted and summed to obtain a fusion vector, and based on the fusion vector, a chain behavior fingerprint vector is generated; Encryption processing: the fingerprint vector is encrypted using homomorphic encryption technology for cross-chain behavior analysis.

[0009] As a preferred technical solution, the feature extractor comprises: The structured feature extractor is used for transaction behaviors and block structures of the blockchain in the multi-modal data, input data is numerical, enumeration or time series data, the structured feature extractor is used for standardization processing of the input data, and a fixed dimension transaction / block feature vector is constructed; The unstructured feature extractor is used for text modal data in unstructured data in the multi-modal data, the unstructured feature extractor generates semantic embedding vectors through a BERT or Word2Vec model, and is used for multimedia modal data in the unstructured data, the unstructured feature extractor adopts a ResNet or 3D-CNN structure to extract features of video frames or images, and generates a fixed dimension feature vector through a pooling layer. The time series feature extractor is used for time series data in the multi-modal data, and the time series feature extractor calculates statistical features through a sliding window to construct a fixed dimension vector.

[0010] As a preferred technical solution, the normalization processing is specifically that the feature vector extracted by the feature extractor is normalized to the interval [0, 1], and the dimension is unified to 128.

[0011] As a preferred technical solution, the vector alignment is specifically that a fully connected neural network is used as a projection layer to map feature vectors of all modes to a unified low-dimensional vector space.

[0012] As a preferred technical solution, the fully connected neural network is trained by using a contrastive loss, a weighting mechanism based on chain / sample reliability, time correlation and topological similarity is introduced into the contrastive loss, and a graph structure regularization term is introduced to retain topological information of a graph on a chain, and is expressed as: wherein, the positive sample weight , respectively represent reliabilities of nodes and , represent a time correlation between nodes and , represent a structural similarity between nodes and , represent a structural similarity weight, is an index of a positive sample node, is an index of a node pair of the node with the index i ; the negative sample weight , is a negative sample index, represent nodes Reliability, Represents a node and nodes The time correlation between them Represents a node and nodes Structural similarity between them; For feature vectors The vector after projection; Temperature coefficient; For structural regularization terms; This is the adjustment coefficient; To compare the losses.

[0013] As a preferred technical solution, the correction based on reliability, temporal decay, and structural sensitivity for attention scoring specifically includes: in, The original similarity, The corrected similarity. Indicates modal reliability. Represents the time series factor. Indicates modal structure sensitivity. For trainable or configurable coefficients, To prevent The parameters.

[0014] As a preferred technical solution, the gating coefficient is expressed as follows: in, The gating coefficient, It is the Sigmoid activation function. For learnable weight vectors, For feature vectors The vector after projection For context vectors, This refers to modal structure sensitivity.

[0015] As a preferred technical solution, in the dynamic weight allocation, the corrected similarity Combined with gating coefficient Then, it is converted into attention weights through a Softmax layer: in, The weights are dynamically allocated.

[0016] According to a second aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0017] According to a third aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method described thereon.

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) Enhanced data analysis capabilities: Through multimodal data fusion, this invention can comprehensively capture and analyze complex correlation behaviors in heterogeneous and multi-source environments, greatly improving the depth and accuracy of anomaly detection and security monitoring.

[0019] (2) Privacy protection: Homomorphic encryption technology is adopted to enable collaborative analysis without leaking the original data, thus solving the contradiction between data sharing and security in traditional methods.

[0020] (3) Universality and flexibility: The methodology of this invention is universal and can be applied not only to the blockchain field, but also to other multi-source heterogeneous distributed environments (such as the Internet of Things), providing a novel and reliable solution for the security management of various complex systems.

[0021] (4) In the process of dynamic weight allocation, the present invention uses reliability, temporal decay and structural sensitivity as correction terms for attention scoring. It takes into account the characteristics of different modes / chains having different credibility, temporal importance and structural sensitivity, so that the fusion result can improve the accuracy of cross-chain abnormal behavior and attack event detection.

[0022] (5) In the process of dynamic weight allocation, the present invention introduces a mode-context joint gating to suppress noise modes, allowing noise modes (such as text garbage) to be forcibly shut down in extreme scenarios (such as anomaly detection), thus solving the problem of sparse and noisy unstructured data on the chain. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0025] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0026] This embodiment provides a multimodal feature fusion method for cross-chain behavior analysis, such as... Figure 1 As shown, the method includes the following steps: S1, Data Modalization and Vectorization: Obtain multiple modal data from heterogeneous chains, extract features using corresponding feature extractors, and perform normalization processing to convert the original data of each modality into a unified numerical feature vector.

[0027] This step aims to transform the various forms of raw data on heterogeneous chains into a unified, computer-processable numerical vector.

[0028] S11, Feature Extraction.

[0029] This embodiment designs independent feature extractors based on different data modalities, and these extractors can be flexibly combined according to application requirements: (1) Structured feature extractor: For blockchain transaction behavior and block structure in multiple modal data, the input data includes transaction amount, transaction frequency, number of participating accounts, block size, block header information, etc., which are numerical, enumerated or time series data. The structured feature extractor standardizes the input data and constructs a fixed-dimensional transaction / block feature vector.

[0030] Specifically, the structured feature extractor uses a feature extraction function. Perform feature extraction: in, Represents structured feature vectors; It represents data such as transaction behavior and block structure.

[0031] (2) Unstructured feature extractor: For unstructured data in various modal data, such as text modality and multimedia modality, an unstructured feature extractor is used to extract features.

[0032] Among them, text modality processing supports the parsing of multilingual and smart contract terms, and generates semantic embedding vectors through BERT or Word2Vec models.

[0033] Multimedia modalities can use ResNet or 3D-CNN structures to extract features from video frames or images, and generate fixed-dimensional feature vectors through pooling layers.

[0034] Specifically, the unstructured feature extractor uses a feature extraction function. : in, Represents unstructured feature vectors; Represents text or multimodal metadata.

[0035] (3) Time series feature extractor: For time series data in various modal data, such as changes in transaction frequency, the time series feature extractor calculates statistical features through a sliding window to construct a vector of fixed dimensions.

[0036] The time series feature extractor uses a feature extraction function. : in, Represents the temporal feature vector; 、 This represents the time series feature extraction interval.

[0037] S12, normalization processing.

[0038] All modal feature vectors will be normalized to the [0,1] interval and have a unified dimension of 128 to ensure comparability of vectors in subsequent fusion.

[0039] The normalization process is represented as: in, express The first ind One portion, d The number of characteristic components.

[0040] Normalize each component: get: Unified Dimension: in, Represents a linear mapping matrix. ; This represents the bias vector. .

[0041] S2, Cross-chain feature vector alignment and dynamic weight allocation: Vector alignment is performed based on the transformed numerical feature vectors, and weights are dynamically allocated to each vector using an attention mechanism. In the dynamic weight allocation, attention scoring is corrected based on reliability, temporal decay and structural sensitivity, and a mode-context joint gating coefficient is introduced to suppress noisy modes.

[0042] Since the feature vectors generated by different modalities and different chains have different dimensions and scales, this invention uses the following method to unify them in preparation for subsequent fusion.

[0043] S21, Vector alignment.

[0044] A projection layer maps the feature vectors of all modalities to a unified low-dimensional vector space; this projection layer is implemented using a fully connected neural network. To overcome the differences in quality, temporal sequence, and topology among heterogeneous chain data, this invention introduces a weighted mechanism based on chain / sample reliability, temporal correlation, and topological similarity into the contrastive loss, and introduces a graph structure regularization term to preserve the topological information of the graph on the chain.

[0045] The modal eigenvectors after projection are: in, Represents the eigenvectors before projection. Indicates the projection layer. It is the projected feature vector. It is the projected dimension. The feature space after projection, i.e., dimension 1. The real vector space.

[0046] Projection maps feature vectors of different modalities to the same space, allowing for comparison and weighted calculation of feature vectors of different modalities after projection.

[0047] Define the contrastive loss, which includes the weighted InfoNCE loss function: Among them, positive sample weights , Representing nodes respectively and nodes Reliability metrics, such as historical node confirmation rate, total node lifetime, and node reputation. Represents a node and nodes The time correlation between them Represents a node and nodes Structural similarity between them Represents the structural similarity weight. The index of the positive sample node. For index i The positive sample node index of the node pairing; the negative sample weight. , For negative sample index, Represents a node Reliability, Represents a node and nodes The time correlation between them Represents a node and nodes Structural similarity between them; For feature vectors The vector after projection; Temperature coefficient; For structural regularization terms; This is the adjustment coefficient; To compare the losses.

[0048] A higher weight for positive samples indicates that the model is more inclined to bring the representation of these samples closer together. This is because on the chain, some nodes / events are more reliable, closer, and structurally more important than others. Therefore, the weight of positive samples should be linked to these features rather than treated equally.

[0049] A high negative sample weight indicates that the negative sample is difficult to distinguish, and there is greater pressure to differentiate it. A low negative sample weight indicates that the negative sample is unimportant or irrelevant.

[0050] By introducing positive and negative sample weights, the contrast loss not only depends on the similarity of the vectors themselves, but also incorporates features in areas such as chain reliability, time, and structure, thereby improving the accuracy and robustness of heterogeneous chain alignment.

[0051] S22, Dynamic weight allocation.

[0052] The process is performed in the projected vector space, introducing an attention mechanism module. Based on features of a uniform scale, weights are assigned to each modality according to the current security context, highlighting important modalities and suppressing secondary modalities. In the dynamic weight allocation (attention), reliability, temporal decay, and structural sensitivity are used as correction terms for attention scoring, and a modality-context joint gating is introduced to suppress noisy modalities, thereby achieving synergistic optimization of alignment and weighting.

[0053] First, the attention mechanism generates queries, keys, and values: in, Represents the security context vector Mapped to a query vector; This indicates that the modal eigenvectors are mapped to key vectors; This indicates that the modal eigenvectors are mapped to value vectors; Represents the mapping matrix; This indicates the dimensions of the query vector and the key vector. Indicates the dimension of the value vector.

[0054] Within a unified space, the original similarity is calculated using the query-key dot product. : in, Indicates the similarity between the query and the modality. In blockchain scenarios, different modalities / chains possess varying degrees of trustworthiness, temporal importance, and structural sensitivity; simple dot products are insufficient to represent the security context. Therefore, a correction is implemented based on attention scoring considering reliability, temporal decay, and structural sensitivity. in, The original similarity, The corrected similarity. Indicates modal reliability. The time series factor is represented by the fact that more recent events have a larger time series factor. Indicates modal structure sensitivity. For trainable or configurable coefficients, To prevent The parameters.

[0055] The gating coefficient is expressed as: in, The gating coefficient, It is the Sigmoid activation function. For learnable weight vectors, For feature vectors The vector after projection For context vectors, This refers to modal structure sensitivity.

[0056] Introducing gating coefficient It allows for the forced shutdown of noisy modalities (such as text garbage) in extreme scenarios (such as anomaly detection), solving the problem of sparse and noisy unstructured data on the chain.

[0057] Corrected similarity Combined with gating coefficient Then, it is converted into attention weights through a Softmax layer: in, The weights are dynamically allocated.

[0058] S3, Multimodal fusion and fingerprint vector generation: The numerical feature vectors are weighted and summed based on the weights to obtain a fusion vector, and an on-chain behavioral fingerprint vector is generated based on the fusion vector.

[0059] This step aims to deeply fuse aligned vectors from different modalities to form a unified "behavioral fingerprint".

[0060] After vector alignment and weight allocation, the feature vectors will enter the fusion layer for multimodal fusion. The fused vector is represented as: This represents the fused vector, which integrates features from all modalities.

[0061] The fused vectors are input into a multilayer feedforward neural network, which learns higher-order correlations and modal dependencies through nonlinear transformations.

[0062] The high-dimensional on-chain behavioral fingerprint vector output by the feedforward neural network can adaptively adjust its dimension, which can be adjusted according to the complexity of the event to cover the features of complex cross-chain events.

[0063] Generate behavioral fingerprint vector : in, It can adaptively adjust, adding or removing dimensions based on the complexity of the event. This represents a feedforward neural network function that can learn higher-order correlations between modes through multi-layer nonlinear transformations.

[0064] S4, Encryption Processing: The fingerprint vector is encrypted using homomorphic encryption technology and then used for cross-chain behavior analysis.

[0065] After the fingerprint vector is generated, it is immediately encrypted using a homomorphic encryption algorithm. Subsequent anomaly detection and aggregation calculations are performed in the encrypted state, fundamentally ensuring the privacy and security of on-chain data.

[0066] The generated fingerprint vector is immediately encrypted: In encrypted form, it can participate in subsequent anomaly detection or aggregation calculations.

[0067] This invention achieves unified analysis of cross-chain, multimodal data, overcoming the limitations of traditional single-chain or single-modal methods. Through multimodal feature fusion and dynamic weight allocation, it can more accurately detect cross-chain abnormal behavior and attack events, and complete aggregate calculations while ensuring on-chain data privacy. It is applicable to scenarios such as cross-chain asset transfer monitoring, smart contract security auditing, and on-chain abnormal behavior detection.

[0068] In one embodiment, the electronic device includes a computing unit that can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) or a computer program loaded from a storage unit into a random access memory (RAM). The RAM may also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0069] Multiple components in an electronic device are connected to an I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0070] The computing unit can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of computing units include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit performs the various methods and processes described above, such as methods S1-S4. For example, in some embodiments, methods S1-S4 may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the computing unit, one or more steps of methods S1-S4 described above may be performed. Alternatively, in other embodiments, the computing unit may be configured to perform methods S1-S4 by any other suitable means (e.g., by means of firmware).

[0071] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0072] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0073] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0074] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multimodal feature fusion method for cross-chain behavior analysis, characterized in that, The method includes the following steps: Data Modalization and Vectorization: Obtain multiple modal data from heterogeneous chains, extract features using corresponding feature extractors, and perform normalization processing to convert the raw data of each modality into a unified numerical feature vector; Cross-chain feature vector alignment and dynamic weight allocation: Vector alignment is performed based on the transformed numerical feature vectors, and weights are dynamically allocated to each vector using an attention mechanism. In the dynamic weight allocation, attention scoring is corrected based on reliability, temporal decay and structural sensitivity, and a mode-context joint gating coefficient is introduced to suppress noisy modes. Multimodal fusion and fingerprint vector generation: The numerical feature vectors are weighted and summed based on the weights to obtain a fusion vector, and an on-chain behavioral fingerprint vector is generated based on the fusion vector; Encryption processing: The fingerprint vector is encrypted using homomorphic encryption technology and then used for cross-chain behavior analysis.

2. The multimodal feature fusion method for cross-chain behavior analysis according to claim 1, characterized in that, The feature extractor includes: Structured Feature Extractor: For blockchain transaction behavior and block structure in various modal data, the input data can be numerical, enumerated, or time series data. The structured feature extractor standardizes the input data and constructs a fixed-dimensional transaction / block feature vector. Unstructured Feature Extractor: For text modal data in unstructured data of various modalities, the unstructured feature extractor generates semantic embedding vectors through BERT or Word2Vec models; for multimedia modal data in unstructured data, the unstructured feature extractor uses ResNet or 3D-CNN structure to extract features from video frames or images, and generates fixed-dimensional feature vectors through pooling layers. Time series feature extractor: For time series data in various modalities, the time series feature extractor calculates statistical features through a sliding window to construct a fixed-dimensional vector.

3. The multimodal feature fusion method for cross-chain behavior analysis according to claim 1, characterized in that, The normalization process specifically involves normalizing the feature vectors extracted by the feature extractor to the [0,1] interval and unifying the dimension to 128.

4. The multimodal feature fusion method for cross-chain behavior analysis according to claim 1, characterized in that, The vector alignment specifically involves using a fully connected neural network as the projection layer to map the feature vectors of all modalities to a unified low-dimensional vector space.

5. A multimodal feature fusion method for cross-chain behavior analysis according to claim 4, characterized in that, The fully connected neural network is trained using contrastive loss. The contrastive loss incorporates a weighting mechanism based on chain / sample reliability, temporal relevance, and topological similarity, and a graph structure regularization term is introduced to preserve the topological information of the graph on the chain. This is expressed as: Among them, positive sample weights , Representing nodes respectively and nodes Reliability, Represents a node and nodes The time correlation between them Represents a node and nodes Structural similarity between them Represents the structural similarity weight. The index of the positive sample node. For index i The positive sample node index of the node pairing; the negative sample weight. , For negative sample index, Represents a node Reliability, Represents a node and nodes The time correlation between them Represents a node and nodes Structural similarity between them; For feature vectors The vector after projection; Temperature coefficient; For structural regularization terms; This is the adjustment coefficient; To compare the losses.

6. The multimodal feature fusion method for cross-chain behavior analysis according to claim 1, characterized in that, The correction based on reliability, temporal decay, and structural sensitivity for attention scoring specifically refers to: in, The original similarity, The corrected similarity. Indicates modal reliability. Represents the time series factor. Indicates modal structure sensitivity. For trainable or configurable coefficients, To prevent The parameters.

7. The multimodal feature fusion method for cross-chain behavior analysis according to claim 1, characterized in that, The gating coefficient is expressed as: in, The gating coefficient, It is the Sigmoid activation function. For learnable weight vectors, For feature vectors The vector after projection For context vectors, This refers to modal structure sensitivity.

8. The multimodal feature fusion method for cross-chain behavior analysis according to claim 1, characterized in that, In the aforementioned dynamic weight allocation, the corrected similarity Combined with gating coefficient Then, it is converted into attention weights through a Softmax layer: in, The weights are dynamically allocated.

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

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images