Data transaction method and apparatus, electronic device, and storage medium

Abstract

The application discloses a data transaction method and device, electronic equipment and storage medium, and relates to the technical field of data transaction. The method comprises the following steps: in response to a data transaction request, determining an associated graph node corresponding to the data transaction request and data representation of the associated graph node based on a preset neural heterogeneous graph data space; generating a target delivery object based on the type of the data transaction request, the associated graph node and the data representation of the associated graph node; and delivering the target delivery object to a requester to complete the data transaction. The application accurately locates the multi-modal data node corresponding to the transaction request based on the neural heterogeneous graph data space, improves the accuracy of data discovery, adaptively generates the target delivery object according to the request type, meets diversified requirements, completes safe delivery under the premise of guaranteeing data privacy and sovereignty, realizes the intelligentization and automation of the whole process of data transaction, and effectively improves the transaction efficiency and user experience.

Description

Technical Field

[0001] This application relates to the field of data transaction technology, and in particular to a data transaction method, a data transaction device, a computer-readable storage medium, and an electronic device. Background Technology

[0002] Dataspaces technology aims to address the limitations of traditional data management when dealing with heterogeneous, large-scale data sources, enabling flexible data integration without strict schemas. However, current dataspace systems lack fine-grained privacy control mechanisms for multimodal data. Summary of the Invention

[0003] This application aims to at least partially address one of the technical problems in related technologies. To this end, the first objective of this application is to propose a data transaction method, comprising: responding to a data transaction request, determining the associated graph nodes and their data representations based on a pre-defined neural heterogeneous graph data space; generating a target delivery object based on the type of the data transaction request, the associated graph nodes, and their data representations; and delivering the target delivery object to the requester to complete the data transaction. This application, based on a pre-defined neural heterogeneous graph data space, accurately identifies the associated graph nodes and their data representations corresponding to the data transaction request, achieving efficient retrieval and location of multimodal data and significantly improving the accuracy of data discovery. Secondly, it adaptively generates corresponding target delivery objects according to the type of data transaction request (such as data transaction or model transaction), meeting the diverse needs of users in different scenarios such as raw data acquisition and model training. Finally, by securely delivering delivery objects such as augmented datasets or federated learning models to the requester, the transaction is completed under the premise of protecting data privacy and sovereignty, realizing intelligent and automated processing of the entire data transaction process, effectively improving the efficiency of data transactions and user experience.

[0004] The second objective of this application is to propose a data transaction device.

[0005] The third objective of this application is to provide a computer-readable storage medium.

[0006] The fourth objective of this application is to propose an electronic device.

[0007] To achieve the above objectives, the first aspect of this application proposes a data transaction method, comprising: responding to a data transaction request, determining the associated graph nodes corresponding to the data transaction request and the data representations of the associated graph nodes based on a preset neural heterogeneous graph data space; generating a target delivery object based on the type of the data transaction request, the associated graph nodes, and the data representations of the associated graph nodes; and delivering the target delivery object to the requester to complete the data transaction.

[0008] According to one embodiment of this application, the target delivery object includes a target document. Generating the target delivery object based on the type of the data transaction request, the association graph nodes, and the data representation of the association graph nodes includes: responding to the data transaction request being of the type of data transaction, performing a similarity search in a public data source based on the data representation of the association graph nodes to determine the corresponding first public data item; fusing the data representation of the association graph nodes with the data representation of the first public data item to generate a corresponding enhanced dataset; and generating the target document based on the descriptive information of the enhanced dataset.

[0009] According to one embodiment of this application, before performing similarity retrieval in a public data source based on the data representation of the association graph nodes, the process includes: determining the association graph nodes to be encrypted in the association graph nodes based on the data representation of the association graph nodes; encrypting the data representation of the association graph nodes to be encrypted to generate a privacy-protected data representation; performing a similarity retrieval in a public data source based on the data representation of the association graph nodes to be encrypted to determine the corresponding second public data item; and fusing the privacy-protected data representation with the data representation of the second public data item to generate a corresponding enhanced dataset.

[0010] According to one embodiment of this application, determining the nodes to be encrypted in a graph based on the data representation of the graph nodes includes: in the neural heterogeneous graph data space, determining the nearest neighboring graph nodes of the graph nodes centered on the data representation of the graph nodes; calculating the neighborhood density of the graph nodes based on the distance between the data representation of the graph nodes and the data representation of the corresponding nearest neighboring graph nodes; determining the information entropy value of the graph nodes based on the ratio between the neighborhood density of the graph nodes and the average density of the neural heterogeneous graph data space; determining the sensitivity level of the graph nodes based on the information entropy interval in which the information entropy of the graph nodes is located; and determining the graph nodes with a preset sensitivity level as the graph nodes to be encrypted.

[0011] According to one embodiment of this application, encrypting the data representation of the nodes in the graph to be encrypted to generate a privacy-protected data representation includes: obtaining a privacy label of the nodes in the graph to be encrypted, and determining a target encryption mechanism based on the privacy label; and using the target encryption mechanism to encrypt the data representation of the nodes in the graph to be encrypted to generate a privacy-protected data representation.

[0012] According to one embodiment of this application, the data representation of the nodes in the graph to be encrypted is encrypted using a target encryption mechanism to generate a privacy-protected data representation. This includes: in response to the target encryption mechanism being a differential privacy mechanism, generating calibrated Gaussian noise based on a sensitivity parameter and a noise multiplier, and adding the noise to the data representation of the nodes in the graph to be encrypted to generate a privacy-protected data representation, wherein the sensitivity parameter and the noise multiplier are determined according to the sensitivity level of the nodes in the graph to be encrypted; and in response to the target encryption mechanism being a homomorphic encryption mechanism, performing homomorphic encryption on the data representation of the nodes in the graph to be encrypted based on encryption security parameters to generate a privacy-protected data representation, wherein the encryption security parameters are determined according to the sensitivity level of the nodes in the graph to be encrypted.

[0013] According to one embodiment of this application, the target delivery object includes a target federated learning model. The target delivery object is generated based on the type of data transaction request, the association graph nodes, and the data representation of the association graph nodes. This includes: responding to the data transaction request type being a model transaction, using the original data holder corresponding to the association graph node as the federated client, and configuring differential privacy parameters for the corresponding federated client based on the sensitivity level of the association graph node; sending a first initial model to the federated client, wherein the federated client performs local training on the first initial model based on local private multimodal data, and prunes the model gradients generated during training and adds Gaussian noise according to the corresponding differential privacy parameters to generate a corresponding gradient parameter set; receiving the gradient parameter set fed back by each federated client, processing the gradient parameter set in an encrypted state through a secure aggregation protocol to generate corresponding model parameter ciphertext; and sending the model parameter ciphertext to the corresponding federated client for decryption and the next round of local training until a preset convergence condition is met to obtain the target federated learning model.

[0014] To achieve the above objectives, a second aspect of this application provides a data transaction apparatus, comprising: a determining module, configured to, in response to a data transaction request, determine the associated graph nodes corresponding to the data transaction request and the data representations of the associated graph nodes based on a preset neural heterogeneous graph data space; a generating module, configured to generate a target delivery object based on the type of the data transaction request, the associated graph nodes, and the data representations of the associated graph nodes; and a delivery module, configured to deliver the target delivery object to the requester to complete the data transaction.

[0015] To achieve the above objectives, a third aspect of this application provides a computer-readable storage medium storing a data transaction program thereon, which, when executed by a processor, implements the aforementioned data transaction method.

[0016] To achieve the above objectives, a fourth aspect of this application provides an electronic device characterized in that it stores a data transaction program thereon, which, when executed by a processor, implements the aforementioned data transaction method.

[0017] According to the data transaction method, apparatus, electronic device, and storage medium of this application embodiments, in response to a data transaction request, based on a preset neural heterogeneous graph data space, the associated graph nodes corresponding to the data transaction request and their data representations are determined; based on the type of the data transaction request, the associated graph nodes, and their data representations, a target delivery object is generated; and the target delivery object is delivered to the requester to complete the data transaction. This application accurately locates the multimodal data nodes corresponding to the transaction request based on the neural heterogeneous graph data space, improving the accuracy of data discovery; adaptively generates target delivery objects according to the request type to meet diverse needs; and completes secure delivery while ensuring data privacy and sovereignty, realizing intelligent and automated data transaction throughout the entire process, effectively improving transaction efficiency and user experience. Attached Figure Description

[0018] Figure 1 Here is a flowchart of a data transaction method according to some embodiments of this application; Figure 2 This is a schematic diagram providing an overview of data transactions in a multi-agent system according to some embodiments of this application; Figure 3 Here is a flowchart of a data transaction method according to other embodiments of this application; Figure 4 This is a schematic diagram of the structure of a data transaction apparatus according to some embodiments of this application; Figure 5 This is a block diagram of an electronic device according to some embodiments of this application. Detailed Implementation

[0019] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0020] The data transaction method, apparatus, electronic device, and storage medium of this application are described in detail below with reference to the accompanying drawings.

[0021] The concept of dataspaces emerged to overcome the inherent limitations of traditional data management techniques in handling heterogeneous, large-scale data sources. Traditional methods typically require the pre-definition of strict schemas and the implementation of centralized integration, resulting in overly rigid systems that struggle to adapt to dynamically changing data environments. In contrast, dataspaces emphasize a data management paradigm that does not require strict schemas and supports flexible, incremental integration. In recent years, they have incorporated dynamic secure sharing mechanisms, combining encryption technology with fine-grained access control to address increasingly complex data exchange needs.

[0022] Driven by the International Data Spaces (IDS) framework, data sovereignty has achieved initial standardized practices, enabling data providers to retain control over their data while sharing it. However, the IDS framework is primarily designed for structured data, limiting its application scope. It cannot effectively support complex scenarios involving multimodal data such as text, images, and sensor streams, thus restricting its widespread deployment in modern AI-driven applications.

[0023] To address the challenges of integrating large-scale, multi-source data, existing research has made progress in several technological dimensions. For example, secure multi-attribute aggregation technology enables data summarization without disclosing original records; scalable multi-tenant serverless cloud database architectures provide the infrastructure for resilient data services; and unified genealogy systems enhance data credibility and traceability by tracing data origins and evolution processes and combining cryptographic integrity verification mechanisms to achieve verifiable audit logs. While these technologies have, to some extent, built the rudiments of an interoperable ecosystem, existing reviews point out that the lack of a data governance framework and sustainability challenges (such as energy consumption and storage costs) remain pressing issues to be addressed.

[0024] In terms of privacy protection, Data Space Technology draws on and integrates various classic and advanced privacy-preserving computation techniques. Differential Privacy (DP) maintains statistical utility of data while protecting individual privacy by adding carefully calibrated noise to query results. To prevent re-identification attacks, group anonymization techniques such as k-anonymity and l-diversity are used to ensure that individuals in the dataset cannot be easily distinguished. For collaborative computation across organizations and in low-trust environments, Secure Multi-Party Computation (SMPC) and Homomorphic Encryption (HE) provide mathematical guarantees for performing joint analysis without exposing the original plaintext data. Research shows that by integrating differential privacy and homomorphic encryption, privacy-preserving machine learning can be achieved in scalable multi-party scenarios; however, the efficiency of homomorphic encryption in mining encrypted data, especially its computational overhead when dealing with large-scale datasets, remains a key bottleneck restricting its widespread application. While these technologies support encrypted queries, federated analysis, and multimodal data sharing for the Internet of Things, the latency and computational burden they introduce become particularly prominent in real-time or near-real-time data pipelines.

[0025] In terms of cross-organizational data sharing and collaboration, Federated Learning (FL), as a distributed machine learning paradigm, enables participants to collaboratively train models without centralizing raw data, thus naturally aligning with the distributed nature of the data space. Verifiable auditing mechanisms can enhance the transparency and traceability of data transactions through cryptographic ledgers and automated access control policies. Data market mechanisms, represented by typical data market platforms, further explore token-based data assetization and monetization models. Although existing research reviews emphasize the importance of incentive mechanism design and point out paths to interoperability through standardization initiatives, a significant research-practice gap remains in building a sustainable ecosystem that includes supply chains and metadata structures. In particular, security threats such as gradient leakage attacks and membership inference attacks in federated learning expose the inadequacy of existing frameworks in providing formal privacy guarantees under heterogeneous, non-independent, and identically distributed data distributions, often sacrificing 5–10% of model utility for privacy protection.

[0026] Attribute-based encryption provides a solution for fine-grained access control in cloud-IoT environments, while solutions combining federated learning and distributed ledger technologies are being explored for policy-driven vehicle-to-everything (V2X) data exchange. Platforms such as Conclave demonstrate the feasibility of using secure multi-party computation for big data analytics without exposing raw data, while homomorphic encryption is used to protect incentive mechanisms in collaborative research. However, these solutions generally face performance bottlenecks and struggle to meet the low-latency, high-throughput requirements of industrial applications.

[0027] Meanwhile, Multi-Agent Systems (MAS) have evolved from focusing on autonomous problem-solving to achieving complex coordinated decision-making through reinforcement learning, significantly enhancing interactive intelligence in data-sharing scenarios. Breakthroughs in Large Language Models (LLMs) have enabled agent systems to handle advanced tasks such as document understanding and natural language communication. Recent MAS-LLM fusion frameworks, such as typical multi-agent frameworks, have demonstrated significant potential through collaborative workflows and are being explored for emerging behavior simulation and scalable collaborative mechanisms in data ecosystems. Theoretically, LLM-driven agent societies can integrate real and synthetic data structures, optimizing hierarchical decision-making through neural orchestration. Tool-enhanced LLM agents further expand the boundaries of capabilities, assisting in causal discovery of multimodal data. The capabilities of LLMs have far exceeded simple document processing; their applications have expanded to network semantic data search, rule interpretation, and prediction or synthetic data generation through embedding fusion.

[0028] To address the multimodal nature of real-world data, multimodal fusion techniques aim to integrate various information sources such as text, images, audio, and structured data to improve decision-making accuracy in fields like medical diagnosis and intelligent education. Technical approaches include early fusion, late fusion, and hybrid fusion. Review studies further refine fusion strategies such as joint representation, collaborative representation, and hierarchical representation. Visual-language fusion models have been successfully applied in areas like robot planning, while intermediate fusion strategies have proven effective in bridging the semantic gap between heterogeneous data. Representative works include the Tongyi Qianwen visual-language model, which demonstrates multimodal understanding capabilities; CLIP (Contrastive Language-Image Pre-training), which facilitates the transfer of language abilities to visual tasks; and visual-language model-to-vector, which focuses on training large-scale multimodal embeddings. Neural graph technology is used to support fusion queries, and reviews on multimodal graph learning systematically summarize the latest advancements in this field. Community detection algorithms such as Louvain and Leiden have been optimized for identifying fusion communities, while the Hybrid-DIA method (a mass spectrometry analysis technique) has been applied to multimodal science fields such as proteomics. However, combining these multimodal fusion capabilities with a secure and controllable data space remains an open challenge.

[0029] In summary, despite significant progress in areas such as data space, privacy-preserving computation, federated learning, multi-agent systems, and multimodal fusion, these technical solutions reveal a series of fundamental flaws when constructing a secure exchange system for multimodal data that possesses real-time adaptability and fine-grained privacy guarantees: The inherent contradiction between privacy protection and real-time performance: While traditional data space systems excel at managing heterogeneous data sources, they generally lack fine-grained privacy control mechanisms and real-time adaptability. Integrating advanced privacy technologies (such as Differential Privacy (DP) and Homomorphic Encryption (HE)) into data pipelines introduces significant computational overhead. Specifically, applying differential privacy to high-dimensional embeddings can lead to a 10–15% decrease in retrieval accuracy, while homomorphic encryption can introduce latency overhead of up to 100 times. Existing systems cannot dynamically and optimally allocate privacy budgets across multiple data processing stages while maintaining sub-200 millisecond query latency and high throughput, especially when processing multimodal data with varying sensitivities and complex dimensions, where performance bottlenecks are particularly pronounced. Based on this, this application calculates the sensitivity level of each graph node based on a preset neural heterogeneous graph data space, and performs privacy processing only on highly sensitive nodes to avoid unnecessary computational overhead. Secondly, it dynamically selects between differential privacy (DP) and homomorphic encryption (HE) according to the application scenario. For scenarios that allow approximate results, DP with lower computational overhead is preferred, and HE is only enabled when precise calculation is required. At the same time, a multi-agent parallel architecture is adopted to distribute the privacy computation task to multiple agent nodes for execution, and dynamic batch processing and caching mechanisms are used to reduce redundant computation.

[0030] The disconnect between static centralized architecture and dynamic security needs: Most existing frameworks rely on static, centralized, or centralized architectures, which not only introduce single points of failure but also expose data to a wider attack surface. They cannot support truly fine-grained, attribute-based access policies, lack dynamic pattern evolution capabilities, and struggle to achieve effective data collaboration across complex organizational trust boundaries. Systems cannot automatically interpret and enforce policies based on the sensitivity of data content, nor can they adaptively route data requests, often leading to a dilemma in practice: either over-protection (such as over-encryption and over-anonymization) results in a severe decline in data utility; or insufficient protection leads to compliance violations and security vulnerabilities. Based on this, this application adopts a distributed multi-agent architecture, where a main coordinator dynamically orchestrates multiple specialized agents (such as privacy agents, fusion agents, and search agents) to work collaboratively, eliminating the risk of single points of failure and distributing data across a pre-defined heterogeneous neural graph data space, effectively reducing the attack surface. Secondly, based on the pre-defined heterogeneous neural graph data space, the information entropy value and sensitivity level of each graph node are calculated, enabling automatic interpretation and execution of fine-grained access policies according to the real-time sensitivity of the data content, rather than relying on static global rules. Simultaneously, data requests are adaptively routed according to sensitivity levels; highly sensitive data is guided to a homomorphic encrypted channel for precise calculation, while medium- and low-sensitivity data is routed to a differential privacy channel for lightweight processing, avoiding one-size-fits-all over-protection or under-protection. Furthermore, the pre-defined heterogeneous neural graph data space supports dynamic pattern evolution; newly accessed data sources can be integrated into the graph structure in real time and automatically establish relationships without requiring downtime to reconstruct the mode, thus achieving flexible data collaboration across organizational trust boundaries. In this way, this application, while ensuring data sovereignty, achieves sensitivity-based dynamic policy execution and adaptive request routing, effectively balancing privacy protection and data utility.

[0031] The disconnect between multimodal data processing and privacy guarantees: Existing standardized frameworks such as the International Data Space (IDS) were primarily designed for structured data and cannot meet the integration and exchange needs of multimodal data (such as images, videos, text, and sensor streams) prevalent in modern applications. Although the field of artificial intelligence (especially large language models and multimodal fusion technologies) has demonstrated the ability to process such complex data, these AI agent systems do not natively integrate advanced privacy protection mechanisms. In multi-party environments, they often fall short when faced with the enormous computational overhead of multimodal data and lack a systematic solution that combines differential privacy, homomorphic encryption, and other technologies. Based on this, this application constructs a pre-defined neural heterogeneous graph data space, which uniformly represents multimodal data such as text, images, videos, and audio as embedded nodes in the graph, thereby achieving unified management and association of cross-modal data. On this basis, differential privacy or homomorphic encryption mechanisms are dynamically selected according to the sensitivity level of each graph node, and privacy protection operations are automatically performed in the entire process of data ingestion and multimodal fusion. Furthermore, the computational overhead is effectively controlled through a multi-agent parallel architecture and the pre-defined neural heterogeneous graph data space optimization. Thus, while supporting the integration and exchange of complex multimodal data, a systematic advanced privacy protection scheme is provided, which solves the problem that existing frameworks are limited to structured data and lack native privacy integration.

[0032] Trust and utility are difficult to balance in distributed collaboration: In distributed collaboration frameworks such as federated learning, related technologies struggle to achieve a balance between providing formal privacy guarantees (e.g., against member inference and gradient leakage attacks) and maintaining model utility. Research shows that in heterogeneous data distributions with non-independent identically distributed (Non-IID) models, existing privacy protection measures typically result in a 5–10% loss in model utility. Furthermore, multi-party data transaction scenarios lack dynamic protection mechanisms capable of withstanding complex inference attacks, leaving the data sovereignty and privacy of participating parties under continuous threat during collaboration. To address this, this application dynamically configures differentiated differential privacy parameters for each federated client based on the sensitivity level of graph nodes. High-sensitivity nodes are configured with stricter noise multipliers to enhance privacy protection, while low-sensitivity nodes are configured with lower noise to retain more data utility, avoiding over-protection or under-protection due to a unified privacy budget. During the local training phase, each client trains the model based on private multimodal data and prunes the gradients and adds calibrated Gaussian noise according to the allocated differential privacy parameters, completing privacy protection processing before gradient uploading, effectively defending against member inference and gradient leakage attacks. During the aggregation phase, the distributed gradient parameter set is processed in an encrypted state through a secure aggregation protocol, ensuring that the gradient information of any individual client is not leaked even during the aggregation process. Through multiple rounds of iterative training until convergence, this application provides formal (ε,δ)-DP privacy guarantees while keeping the model utility loss within 5%, and maintains stable model performance even under non-independent and identically distributed data distributions, achieving a dynamic balance between privacy protection strength and data utility.

[0033] Figure 1 This is a flowchart of a data transaction method according to some embodiments of this application. (Refer to...) Figure 1 The data transaction method in this application embodiment may include the following steps: S110, in response to a data transaction request, determines the associated graph node and the data representation of the associated graph node corresponding to the data transaction request based on the preset neural heterogeneous graph data space.

[0034] S120 generates the target delivery object based on the type of data transaction request, the associated graph nodes, and the data representation of the associated graph nodes.

[0035] S130, deliver the target delivery object to the requester to complete the data transaction.

[0036] Specifically, the Preset Neural Heterogeneous Graph Data Space refers to a hybrid data indexing framework that integrates graph structure relationships and vector embedding representations for the efficient organization and retrieval of multimodal data. The Preset Neural Heterogeneous Graph Data Space uniformly models data from different sources and modalities (such as text, images, videos, audio, and tabular data) as graph nodes. Each graph node corresponds to a data entity, and the edges between graph nodes represent the relationships between the data. Simultaneously, each graph node is associated with an embedding vector generated using multimodal models such as CLIP and InternLM. This vector encodes the data representation of the original data; for example, semantic content for text and pixel distribution for images.

[0037] Upon receiving a data transaction request initiated by a user, the main coordinator can parse key information in the request (such as data type, domain scope, keywords, etc.) and use multimodal models such as CLIP to transform the data transaction request into a query vector. Subsequently, a hybrid retrieval is performed based on the pre-built neural heterogeneous graph data space. At the same time, the cosine similarity between the query vector and the embedding vector of each graph node is calculated, and the cosine similarity is sorted. Finally, the associated graph nodes with high semantic matching degree with the request and meeting the transaction conditions, as well as the data representation of the associated graph nodes, are located.

[0038] Upon receiving a data transaction request, the system will generate a target delivery object based on the type of the request (e.g., data transaction or model transaction), combined with the identified nodes in the relationship graph and their data representations, using an appropriate processing strategy. Specifically, if the data transaction request is for a data transaction, the generated target delivery object is a document, which is ultimately provided to the requester to complete the transaction; if the data transaction request is for a model transaction, a trained model is generated through federated learning, and the model is ultimately provided to the requester to complete the transaction.

[0039] This application, based on a pre-defined neural heterogeneous graph data space, accurately identifies the associated graph nodes and their data representations corresponding to data transaction requests, achieving efficient retrieval and location of multimodal data and significantly improving the accuracy of data discovery. Secondly, it adaptively generates corresponding target delivery objects according to the type of data transaction request (such as data transaction or model transaction), meeting the diverse needs of users in different scenarios such as raw data acquisition and model training. Finally, by securely delivering delivery objects such as augmented datasets or federated learning models to the requester, the transaction is completed under the premise of protecting data privacy and sovereignty, realizing intelligent and automated processing of the entire data transaction process, effectively improving the efficiency of data transactions and user experience.

[0040] In some embodiments, the target delivery object includes a target document. Generating the target delivery object based on the type of data transaction request, the association graph node, and the data representation of the association graph node includes: responding to the data transaction request being of the type of data transaction, performing a similarity search in a public data source based on the data representation of the association graph node to determine the corresponding first public data item; fusing the data representation of the association graph node with the data representation of the first public data item to generate a corresponding enhanced dataset; and generating the target document based on the descriptive information of the enhanced dataset.

[0041] Specifically, when the data transaction request type is data transaction, firstly, based on the data representation of the association graph nodes, similarity retrieval is performed in public data sources by calculating vector cosine similarity. Public data items that are semantically related to the association graph nodes and have similarity scores exceeding a preset threshold are selected as the first public data items. Subsequently, the data representations of the association graph nodes and the data representations of the first public data items are input into a multimodal fusion model and fused through attention mechanisms or feature concatenation to generate a fused augmented dataset. Finally, a large language model is invoked to automatically understand and summarize the descriptive information such as the content, source, and quality indicators of the augmented dataset, generating a target document containing data descriptions, terms of use, and compliance statements, which is delivered to the requesting party as the transaction result.

[0042] In some embodiments, before performing similarity retrieval in a public data source based on the data representation of the association graph nodes, the process includes: determining the association graph nodes to be encrypted in the association graph nodes based on the data representation of the association graph nodes; encrypting the data representation of the association graph nodes to be encrypted to generate a privacy-protected data representation; performing a similarity retrieval in a public data source based on the data representation of the association graph nodes to be encrypted to determine the corresponding second public data item; and fusing the privacy-protected data representation with the data representation of the second public data item to generate a corresponding enhanced dataset.

[0043] Specifically, before performing public data source similarity retrieval based on the data representation of the association graph nodes, privacy preprocessing is required. This is because the association graph nodes may contain sensitive data, and directly using the original representation for retrieval would expose privacy information. Therefore, sensitive nodes must be encrypted and protected before proceeding with subsequent operations.

[0044] Therefore, it is necessary to determine the nodes in the graph to be encrypted. For example, obtain the attribute information of each node, including data source type, content domain, historical transaction records, access frequency, etc., and input these multi-dimensional attributes into a pre-trained privacy classification model to output the initial privacy probability score of each node. At the same time, in the neural heterogeneous graph data space, calculate the graph distance and association strength between each node and known sensitive nodes (such as nodes marked as high privacy risk) to obtain the association risk score of the node. The initial privacy probability score and the association risk score are weighted and fused to generate a comprehensive privacy risk value for each node. Based on the preset range in which the risk value is located, the nodes are divided into low-risk, medium-risk, and high-risk levels. The medium-risk and high-risk nodes are determined as the nodes in the graph to be encrypted.

[0045] Subsequently, the data representation of the nodes in the graph to be encrypted is encrypted. For example, the corresponding encryption mechanism can be selected according to the privacy label of the node. Specifically, for nodes with moderate sensitivity, a differential privacy mechanism is adopted, and the sensitivity parameter and noise multiplier are determined based on its sensitivity level. Calibrated Gaussian noise is generated and added to the data representation to form a privacy-protected representation. For highly sensitive nodes, a homomorphic encryption mechanism is adopted, and the encryption security parameter is determined according to the sensitivity level to perform homomorphic encryption processing on the data representation.

[0046] After encryption is completed, based on the encrypted data representation of the nodes in the graph to be encrypted, a similarity search is performed in the public data source. By calculating the vector cosine similarity, public data items that are semantically related and have a similarity exceeding a preset threshold are selected as the second public data items.

[0047] Finally, the privacy-protected data representation and the data representation of the second public data item are input into the multimodal fusion model, and fusion processing is performed through attention mechanisms or feature splicing to generate an enhanced dataset that integrates external public information while protecting the original sensitive data.

[0048] In some embodiments, determining the nodes to be encrypted in a graph based on the data representation of the graph nodes includes: in the neural heterogeneous graph data space, determining the nearest neighboring graph nodes of the graph nodes centered on the data representation of the graph nodes; calculating the neighborhood density of the graph nodes based on the distance between the data representation of the graph nodes and the data representation of the corresponding nearest neighboring graph nodes; determining the information entropy value of the graph nodes based on the ratio between the neighborhood density of the graph nodes and the average density of the neural heterogeneous graph data space; determining the sensitivity level of the graph nodes according to the information entropy interval in which the information entropy of the graph nodes is located; and determining the graph nodes with a preset sensitivity level as the graph nodes to be encrypted.

[0049] Specifically, the nodes to be encrypted can be determined by analyzing the local structural features of the nodes in the neural heterogeneous graph data space. First, taking the data representation of each node as the center, the k nearest nodes are identified as the node's nearest neighbors by calculating the Euclidean or cosine distance between the node and its surrounding node representations. Then, based on the distance values ​​between the node and each of its nearest neighbors, the mean or median of these distances is calculated as a measure of the neighborhood density of the node's region. The smaller the neighborhood density, the sparser the distribution around the node, and vice versa. Finally, the neighborhood density of the node is compared with the global average density of the entire neural heterogeneous graph data space. The ratio of the two values ​​is calculated, reflecting the sparseness or density of the region where the node is located. This ratio is used as the information entropy value of the node. The smaller the ratio, the more isolated the node is, and it may contain unique or sensitive information. The larger the ratio, the more dense the node is, and the more common the information is. Then, according to the preset information entropy value interval division rules, the different intervals in which the node's information entropy value falls are corresponding to different sensitivity levels. For example, the low entropy value interval corresponds to the high sensitivity level, and the high entropy value interval corresponds to the low sensitivity level. Finally, the association graph nodes whose sensitivity level reaches the preset threshold (such as the high sensitivity level) are identified as association graph nodes to be encrypted, so that the data representation of these nodes can be encrypted and protected in the future.

[0050] In some embodiments, the data representation of the nodes to be encrypted in the graph is encrypted to generate a privacy-protected data representation, including: obtaining a privacy label of the nodes to be encrypted in the graph, and determining a target encryption mechanism based on the privacy label; and using the target encryption mechanism to encrypt the data representation of the nodes to be encrypted in the graph to generate a privacy-protected data representation.

[0051] Specifically, the privacy label of each node in the graph to be encrypted is obtained. This privacy label is pre-set based on the node's sensitivity level and the data usage scenario. Then, the target encryption mechanism is determined based on the privacy label. Specifically, a differential privacy mechanism is selected for scenarios requiring statistical analysis and real-time processing and allowing approximate results, while a homomorphic encryption mechanism is selected for scenarios requiring precise calculation results or end-to-end data encryption. After determining the encryption mechanism, the corresponding encryption operation is performed to encrypt the data representation of the nodes in the graph to be encrypted, thereby generating a privacy-protected data representation.

[0052] In some embodiments, the data representation of the nodes in the graph to be encrypted is encrypted using a target encryption mechanism to generate a privacy-preserving data representation. This includes: in response to the target encryption mechanism being a differential privacy mechanism, generating calibrated Gaussian noise based on a sensitivity parameter and a noise multiplier, and adding the noise to the data representation of the nodes in the graph to be encrypted to generate a privacy-preserving data representation, wherein the sensitivity parameter and the noise multiplier are determined according to the sensitivity level of the nodes in the graph to be encrypted; and in response to the target encryption mechanism being a homomorphic encryption mechanism, performing homomorphic encryption on the data representation of the nodes in the graph to be encrypted based on encryption security parameters to generate a privacy-preserving data representation, wherein the encryption security parameters are determined according to the sensitivity level of the nodes in the graph to be encrypted.

[0053] Specifically, if the target encryption mechanism is a differential privacy mechanism, the sensitivity level, sensitivity parameter, and noise multiplier of the nodes in the graph to be encrypted can first be determined by querying a preset mapping table between sensitivity level and sensitivity parameter-noise multiplier. This preset mapping table includes multiple sensitivity levels and a corresponding sensitivity parameter-noise multiplier for each sensitivity level. The sensitivity parameter reflects the maximum impact of a single data point on the query result, while the noise multiplier controls the balance between privacy protection strength and data availability. Then, calibrated Gaussian noise is generated based on these two parameters. This noise follows a normal distribution with a mean of zero and a variance determined by both the sensitivity parameter and the noise multiplier. The sensitivity parameter and noise multiplier are input into a preset formula to calculate the variance. Finally, the generated Gaussian noise is directly added to the original data representation of the nodes to be encrypted, resulting in a privacy-protected data representation. This prevents attackers from accurately inferring the original data values ​​while preserving the statistical characteristics of the data for subsequent analysis.

[0054] If the target encryption mechanism is homomorphic encryption, the encryption security parameters of the nodes in the graph to be encrypted can first be determined by querying a preset mapping table between sensitivity levels and encryption security parameters. This preset mapping table includes multiple sensitivity levels and the corresponding encryption security parameters for each sensitivity level. The encryption security parameters determine the key length and computational complexity of the encryption algorithm; higher sensitivity levels require longer keys and higher security strength. Then, based on the encryption security parameters, homomorphic encryption is performed on the original data representation of the nodes to be encrypted, converting plaintext data into ciphertext data and generating a privacy-protected data representation. This allows specific computational operations to be performed directly on the ciphertext data without decryption. The decrypted computation result is consistent with the result of performing the same computation on the plaintext, thus protecting data privacy while supporting subsequent complex operations such as joint analysis and model aggregation.

[0055] In this way, differentiated privacy protection strategies can be flexibly selected according to the sensitivity level of the node to be encrypted and the application scenario requirements, ensuring that the data will not leak the original sensitive information during subsequent similarity retrieval, multimodal fusion and other processing, while maintaining the required computing functions.

[0056] In some embodiments, the target delivery object includes a target federated learning model. The target delivery object is generated based on the type of data transaction request, the association graph nodes, and the data representation of the association graph nodes. This includes: responding to the data transaction request type being a model transaction, using the original data holder corresponding to the association graph node as the federated client, and configuring differential privacy parameters for the corresponding federated client based on the sensitivity level of the association graph node; sending a first initial model to the federated client, wherein the federated client performs local training on the first initial model based on local private multimodal data, and prunes the model gradients generated during training and adds Gaussian noise according to the corresponding differential privacy parameters to generate a corresponding gradient parameter set; receiving the gradient parameter set fed back by each federated client, processing the gradient parameter set in an encrypted state through a secure aggregation protocol to generate corresponding model parameter ciphertext; and sending the model parameter ciphertext to the corresponding federated client for decryption and the next round of local training until a preset convergence condition is met to obtain the target federated learning model.

[0057] Specifically, the original data corresponding to the nodes in the association graph refers to the private multimodal data stored locally by each client that has not been anonymized or desensitized, including but not limited to electronic medical record texts from medical institutions, medical images (such as CT and MRI), laboratory test forms, transaction records from financial institutions, customer identity information, risk control labels, product images of enterprises, sales records, and sensor time series data. This original data cannot be directly shared or traded because it contains personal privacy or trade secrets.

[0058] The original data holders corresponding to the identified graph nodes are used as federated learning clients, and differential privacy parameters are configured for the corresponding clients based on the sensitivity level of each graph node. For example, the corresponding differential privacy parameters can be determined by querying a preset relationship mapping table between sensitivity levels and differential privacy parameters. The preset relationship mapping table includes multiple sensitivity levels and the differential privacy parameters corresponding to each sensitivity level.

[0059] The differential privacy parameters include privacy budget, privacy failure probability, maximum gradient norm, and noise multiplier. The privacy budget measures the strength of privacy protection; a smaller budget indicates stronger protection. The privacy failure probability represents the tolerance probability of privacy guarantees failing. The maximum gradient norm limits the maximum impact of a single sample on model updates. The noise multiplier controls the intensity of Gaussian noise added to the gradient. Specifically, clients with higher sensitivity levels have smaller privacy budgets and higher privacy failure probabilities.

[0060] Subsequently, the initial model is distributed to each client. Each client trains the initial model locally using its private multimodal data. The training process is as follows: the client inputs local data into the model according to a preset batch size, calculates the loss function through forward propagation, and then calculates the gradient of the weight parameters of each network layer relative to the loss function through backpropagation, forming the original gradient set. The gradient parameter set contains the gradient tensor of each network layer. After each batch of training is completed, the original gradient set is privacy-enhanced according to the configured differential privacy parameters. Specifically, firstly, the L2 norm of the gradient of each sample is calculated. If the norm exceeds the maximum gradient norm C, the gradient of that sample is scaled proportionally to the norm C, thus pruning the gradient magnitude and ensuring that a single sample does not excessively affect the model update. Then, a Gaussian distribution N(0,σ) is added to the pruned gradient. 2 C 2 The random noise is generated by adding σ, where σ represents the noise multiplier, to produce a set of noisy gradient parameters that satisfy differential privacy. Then, the client accumulates and averages the noisy gradients of all batches to obtain the final gradient parameter set of this round of local training. This parameter set contains the gradient information of all weighted layers and bias terms of the model, but has achieved formal differential privacy protection through pruning and adding noise, so that even if an attacker obtains this gradient parameter set, they cannot reverse engineer the original data of any single sample.

[0061] After receiving the privacy-protected gradient parameter sets uploaded by all clients, the gradient parameter sets are processed in an encrypted state through a secure aggregation protocol. Specifically, homomorphic encryption or secure multi-party computation technology is used to enable the aggregation server to directly perform a weighted average calculation on the gradient parameter sets in the encrypted state without being able to decrypt the gradients of each client. The aggregation weights are usually determined based on the amount of local data of each client, and a new encrypted model parameter is generated.

[0062] The generated encrypted model parameters are sent to the corresponding federated clients. Each client decrypts the encrypted model parameters locally to restore the updated model for the current round, and uses this as the basis for the next round of local training. The complete process of "local training - gradient upload - encrypted aggregation - encrypted distribution - decryption update" is executed iteratively until the model meets the preset convergence conditions (such as reaching the maximum number of training rounds, the change in the loss function being lower than the threshold, or the model accuracy no longer improving). Finally, the global model obtained at this point is determined as the target federated learning model. This enables cross-organizational collaborative modeling and secure transactions under the premise that the original data of each data holder does not leave their local machine and the gradient transmission process is protected by both encryption and differential privacy.

[0063] As a concrete example, an overview of data transactions in a multi-agent system is as follows: Figure 2 As shown, in the context of inter-enterprise data transaction space, the full process of multi-agent data fusion with enterprise A as the data demander and enterprise B as the data provider is as follows: The left side of the diagram first defines the basic rules of this space, that is, the circular nodes A to G represent different participating enterprises, and the arrow "B→A" clarifies the core transaction relationship of B selling data to A; In the core process on the right side of the diagram, enterprise A uses its own private dataset (A') as the basis, introduces enterprise B's private dataset (B'), and combines it with the public dataset (P). These three types of data are input into the core multi-agent data fusion system QUBE for processing. The entire fusion process strictly follows two key constraints: Condition 1: Protect data sensitivity and Condition 2: Ensure the data performance of C. Finally, a target fused dataset that meets both privacy and security requirements and has high availability is generated for enterprise A.

[0064] Reference Figure 3 The data transaction method in this application embodiment may further include the following steps: Step (1) The customer sends a purchase request: When a customer initiates a data purchase request through the system interface, it marks the start of the transaction process.

[0065] Step (2) QUBE interactive interface: The system's front-end interaction entry point is responsible for receiving customer requests and forwarding them to the back-end processing module.

[0066] Step (3) Multi-agent system coordinator: The multi-agent system coordinator parses client requests, determines the transaction type (e.g., whether it's a data transaction or a model transaction), schedules various specialized agents to work together, and manages data flow and processing.

[0067] Step (4) HMGI Basic Data Space: A pre-defined neural heterogeneous graph data space organizes multimodal data nodes in a graph structure and stores the data representation (such as embedding vectors) of each node. Based on this data space, the coordinator determines the graph nodes and their data representations associated with the client request.

[0068] Step (5) Data Ingestion Agent: It is responsible for extracting the raw data and representations of the nodes in the association graph from the HMGI basic data space, so as to prepare raw materials for subsequent processing.

[0069] Step (6) Privacy Agent: The work of the privacy agent includes: based on the data representation of the association graph nodes, determining the association graph nodes to be encrypted in the association graph nodes, selecting differential privacy or homomorphic encryption according to the privacy label, and adding Gaussian noise (differential privacy) or performing homomorphic encryption on the data representation of sensitive nodes.

[0070] Step (7) Web search agent: The web search agent retrieves highly similar public data items from public data sources (such as web pages and public datasets) based on the data representation of the nodes in the relational graph (which may be encrypted and privacy-preserving representations) for subsequent data augmentation.

[0071] Step (8) Document Agent: The document agent generates the target document based on the descriptive information of the augmented dataset.

[0072] Step (9) Data Fusion Agent: The data fusion agent merges the representations of private data (which may be encrypted) with the representations of public data items to generate an enhanced dataset.

[0073] Step (10) Merged dataset: This dataset contains the core value of the original private data, incorporates supplementary information from publicly available data, and has undergone privacy protection processing.

[0074] Step (11) The data was successfully delivered to the customer.

[0075] Corresponding to the above embodiments, this application also proposes a data transaction device.

[0076] Reference Figure 4 The data transaction device 400 includes: a determination module 410, a generation module 420, and a delivery module 430.

[0077] The determination module 410, in response to a data transaction request, determines the associated graph nodes and their data representations based on a preset neural heterogeneous graph data space. The generation module 420 generates a target delivery object based on the type of the data transaction request, the associated graph nodes, and their data representations. The delivery module 430 delivers the target delivery object to the requester to complete the data transaction.

[0078] According to one embodiment of this application, the target delivery object includes a target document. The generation module 420 is specifically configured to, in response to a data transaction request of type data transaction, perform a similarity search in a public data source based on the data representation of the association graph nodes to determine the corresponding first public data item; fuse the data representation of the association graph nodes with the data representation of the first public data item to generate a corresponding enhanced dataset; and generate a target document based on the descriptive information of the enhanced dataset.

[0079] According to one embodiment of this application, before performing similarity retrieval in a public data source based on the data representation of the association graph nodes, the generation module 420 is further configured to: determine the association graph nodes to be encrypted in the association graph nodes based on the data representation of the association graph nodes; encrypt the data representation of the association graph nodes to be encrypted to generate a privacy-protected data representation; perform a similarity retrieval in a public data source based on the data representation of the association graph nodes to be encrypted to determine the corresponding second public data item; and fuse the privacy-protected data representation with the data representation of the second public data item to generate a corresponding enhanced dataset.

[0080] According to one embodiment of this application, the generation module 420 is further configured to: determine the nearest neighboring graph nodes of the associated graph nodes in the neural heterogeneous graph data space, with the data representation of the associated graph node as the center; calculate the neighborhood density of the associated graph node based on the distance between the data representation of the associated graph node and the data representation of the corresponding nearest neighboring graph node; determine the information entropy value of the associated graph node based on the ratio between the neighborhood density of the associated graph node and the average density of the neural heterogeneous graph data space; determine the sensitivity level of the associated graph node according to the information entropy interval in which the information entropy of the associated graph node is located; and determine the associated graph nodes with a preset sensitivity level as the associated graph nodes to be encrypted.

[0081] According to one embodiment of this application, the generation module 420 is further configured to: obtain the privacy label of the node to be encrypted in the graph, and determine the target encryption mechanism based on the privacy label; and use the target encryption mechanism to encrypt the data representation of the node to be encrypted in the graph to generate a privacy-protected data representation.

[0082] According to one embodiment of this application, the generation module 420 is further configured to: in response to the target encryption mechanism being a differential privacy mechanism, generate calibrated Gaussian noise based on a sensitivity parameter and a noise multiplier, and add the noise to the data representation of the nodes in the graph to be encrypted to generate a privacy-protected data representation, wherein the sensitivity parameter and the noise multiplier are determined according to the sensitivity level of the nodes in the graph to be encrypted; and in response to the target encryption mechanism being a homomorphic encryption mechanism, perform homomorphic encryption on the data representation of the nodes in the graph to be encrypted based on encryption security parameters to generate a privacy-protected data representation, wherein the encryption security parameters are determined according to the sensitivity level of the nodes in the graph to be encrypted.

[0083] According to one embodiment of this application, the target delivery object includes a target federated learning model. The generation module 420 is specifically configured to: respond to a data transaction request of type model transaction; use the original data holder corresponding to the association graph node as the federated client; and configure differential privacy parameters for the corresponding federated client based on the sensitivity level of the association graph node; send a first initial model to the federated client, wherein the federated client performs local training on the first initial model based on local private multimodal data, and prunes the model gradient generated during training and adds Gaussian noise according to the corresponding differential privacy parameters to generate a corresponding gradient parameter set; receive the gradient parameter set fed back by each federated client, process the gradient parameter set in an encrypted state through a secure aggregation protocol to generate the corresponding model parameter ciphertext; and send the model parameter ciphertext to the corresponding federated client for decryption and the next round of local training until a preset convergence condition is met to obtain the target federated learning model.

[0084] It should be noted that the above explanations of the embodiments and beneficial effects of the data transaction method also apply to the data transaction apparatus of the embodiments of this application. To avoid redundancy, they will not be elaborated in detail here.

[0085] Corresponding to the above embodiments, this application also proposes a computer-readable storage medium.

[0086] The present application provides a computer-readable storage medium storing a data transaction program thereon, which, when executed by a processor, implements the aforementioned data transaction method.

[0087] It should be noted that the above explanations of the embodiments and beneficial effects of the data transaction method also apply to the computer-readable storage medium of the embodiments of this application. To avoid redundancy, they will not be elaborated in detail here.

[0088] Corresponding to the above embodiments, this application also proposes an electronic device.

[0089] See Figure 5As shown, the electronic device 500 of this application includes a memory 510, a processor 520, and a data transaction program stored in the memory 510 and capable of running on the processor 520. When the processor executes the data transaction program, it implements the aforementioned data transaction method.

[0090] It should be noted that the above explanations of the embodiments and beneficial effects of the data transaction method also apply to the electronic devices in the embodiments of this application. To avoid redundancy, they will not be elaborated in detail here.

[0091] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0092] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0093] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0094] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0095] In this application, unless otherwise expressly specified and limited, the terms "installation," "connection," "joining," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0096] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A data transaction method, characterized in that, include: In response to a data transaction request, based on a preset neural heterogeneous graph data space, the associated graph node corresponding to the data transaction request and the data representation of the associated graph node are determined; Based on the type of the data transaction request, the association graph node, and the data representation of the association graph node, a target delivery object is generated; The target delivery object is delivered to the requester to complete the data transaction.

2. The data transaction method according to claim 1, characterized in that, The target delivery object includes a target document. Generating the target delivery object based on the type of the data transaction request, the association graph nodes, and the data representation of the association graph nodes includes: In response to the data transaction request being of the type of data transaction, a similarity search is performed in a public data source based on the data representation of the association graph node to determine the corresponding first public data item; The data representations of the association graph nodes are fused with the data representations of the first public data item to generate a corresponding enhanced dataset; The target document is generated based on the descriptive information of the augmented dataset.

3. The data transaction method according to claim 2, characterized in that, Before performing similarity retrieval in a public data source based on the data representation of the association graph nodes, the process includes: Based on the data representation of the association graph nodes, determine the association graph nodes to be encrypted in the association graph nodes; The data representation of the nodes in the graph to be encrypted is encrypted to generate a privacy-protected data representation. Based on the data representation of the nodes in the graph to be encrypted, a similarity search is performed in the public data source to determine the corresponding second public data item; The privacy-protected data representation is fused with the data representation of the second public data item to generate a corresponding enhanced dataset.

4. The data transaction method according to claim 3, characterized in that, The step of determining the association graph nodes to be encrypted based on the data representation of the association graph nodes includes: In the neural heterogeneous graph data space, the nearest neighboring graph nodes of the associated graph nodes are determined with the data representation of the associated graph nodes as the center; The neighborhood density of the association graph node is calculated based on the distance between the data representation of the association graph node and the data representation of its corresponding nearest neighbor association graph node. The information entropy value of the association graph node is determined based on the ratio between the neighborhood density of the association graph node and the average density of the neural heterogeneous graph data space. The sensitivity level of the association graph node is determined based on the information entropy range in which the information entropy of the association graph node is located; The graph nodes with a preset sensitivity level are identified as the graph nodes to be encrypted.

5. The data transaction method according to claim 3, characterized in that, The encryption process for the data representation of the nodes in the graph to be encrypted, in order to generate a privacy-preserving data representation, includes: Obtain the privacy tags of the nodes in the graph to be encrypted, and determine the target encryption mechanism based on the privacy tags; The data representation of the nodes in the graph to be encrypted is encrypted using the target encryption mechanism to generate a privacy-protected data representation.

6. The data transaction method according to claim 5, characterized in that, The step of encrypting the data representation of the nodes in the graph to be encrypted using the target encryption mechanism to generate a privacy-preserving data representation includes: In response to the target encryption mechanism being a differential privacy mechanism, calibrated Gaussian noise is generated based on a sensitivity parameter and a noise multiplier, and the noise is added to the data representation of the nodes in the graph to be encrypted to generate a privacy-protected data representation, wherein the sensitivity parameter and the noise multiplier are determined according to the sensitivity level of the nodes in the graph to be encrypted; In response to the target encryption mechanism being a homomorphic encryption mechanism, the data representation of the nodes in the graph to be encrypted is homomorphically encrypted based on encryption security parameters to generate a privacy-protected data representation, wherein the encryption security parameters are determined according to the sensitivity level of the nodes in the graph to be encrypted.

7. The data transaction method according to claim 1, characterized in that, The target delivery object includes a target federated learning model. Generating the target delivery object based on the type of the data transaction request, the association graph nodes, and the data representation of the association graph nodes includes: In response to the data transaction request being of the type of model transaction, the original data holder corresponding to the association graph node is used as the federated client, and differential privacy parameters are configured for the corresponding federated client based on the sensitivity level of the association graph node. The first initial model is sent to the federated client, wherein the federated client trains the first initial model locally based on local private multimodal data, and prunes the model gradient generated during training and adds Gaussian noise according to the corresponding differential privacy parameters to generate the corresponding gradient parameter set. The gradient parameter set fed back by each of the federated clients is received, and the gradient parameter set is processed in an encrypted state through a secure aggregation protocol to generate the corresponding model parameter ciphertext. The encrypted model parameters are sent to the corresponding federated clients for decryption and the next round of local training until the preset convergence condition is met, thus obtaining the target federated learning model.

8. A data transaction device, characterized in that, include: The determination module is used to respond to a data transaction request and, based on a preset neural heterogeneous graph data space, determine the associated graph node corresponding to the data transaction request and the data representation of the associated graph node. The generation module is used to generate a target delivery object based on the type of the data transaction request, the association graph node, and the data representation of the association graph node; The delivery module is used to deliver the target delivery object to the requester to complete the data transaction.

9. A computer-readable storage medium, characterized in that, It stores a data transaction program, which, when executed by a processor, implements the data transaction method according to any one of claims 1-7.

10. An electronic device, characterized in that, It stores a data transaction program, which, when executed by a processor, implements the data transaction method according to any one of claims 1-7.

Images