Supply chain bill ecological interconnection system based on open API and smart contract

Abstract

The invention provides a supply chain bill ecological interconnection system based on an open API (Application Program Interface) and a smart contract, which relates to the field of electric digital data processing and comprises a self-adaptive open interface collaboration module, a cross-chain smart contract arrangement module, a multi-source credit intelligent calculation module and an end-to-end privacy security protection module, the self-adaptive open interface collaboration module is responsible for realizing intelligent access and authority management of a heterogeneous system, the cross-chain intelligent contract arrangement module is responsible for realizing unified contract management and automatic execution in a multi-block chain environment, and the multi-source credit intelligent calculation module is responsible for fusing multi-party data to construct a supply chain credit evaluation system. The end-to-end privacy security protection module is responsible for providing data encryption, privacy calculation and identity verification of the whole process; according to the system, the problems of data islands, unsmooth credit transmission, unbalanced privacy protection and data utility and the like in a supply chain bill system can be solved, and a complete technical solution is provided for digital ecology of supply chain finance.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing, and more specifically to a supply chain invoice ecosystem interconnection system based on open APIs and smart contracts. Background Technology

[0002] Supply chain bill financing is an important means to alleviate the financing difficulties of small and medium-sized enterprises (SMEs), but the traditional system has problems such as data silos, poor credit transmission, and contradictions between privacy protection and data sharing. Existing technologies mainly solve some problems through graph neural networks, federated learning, and blockchain technology, but a complete ecological interconnection solution has not yet been formed.

[0003] In terms of credit assessment algorithms, existing technologies mainly employ three methods. The first is based on traditional machine learning, using support vector machines or random forests to classify and predict corporate financial indicators. This method relies on manual feature engineering, struggles to capture complex dependencies within supply chain networks, is sensitive to data gaps, and has limited generalization ability. The second is based on graph attention networks, aggregating neighbor node information through multi-layer graph convolution to calculate credit scores. This method can model transaction relationships between companies, but suffers from oversmoothing, where the features of multi-hop neighbors tend to be consistent, leading to decreased node distinguishability. Furthermore, it does not consider the temporal characteristics and integrity constraints of the bill of exchange circulation path. The third is based on deep neural network embedding, using weighted adjacency matrices combined with deep network embedding algorithms to estimate the credit default distance between companies. This method performs well in capturing complex dependencies, but has high computational complexity, struggles to handle large-scale dynamic graph structures, and lacks modeling of supply chain hierarchical decay characteristics.

[0004] Regarding data privacy protection, existing federated learning techniques generally employ a uniform differential privacy budget, where all participants add noise of the same intensity to protect data privacy. This approach ignores differences in creditworthiness and data quality among participants. High-quality data providers may experience reduced model accuracy due to excessive noise, while low-credit participants may exploit the weaker privacy constraints to launch data poisoning attacks. Some studies have introduced blockchain technology to achieve decentralized coordination in federated learning, but they still employ fixed privacy budget allocation strategies, failing to dynamically adjust the strength of privacy protection based on participants' historical behavior and real-time performance. This makes it difficult to optimize the balance between data utility and privacy protection.

[0005] Regarding foreign patents, US Patent 20150379510A1 discloses a method for monetizing data changes in the data supply chain using blockchain infrastructure and smart contracts. This patent establishes a matching mechanism between data producers and data buyers, realizing micro-payments for data changes through smart contracts. The patent's shortcomings lie in its focus solely on the economic mechanism of data transactions, neglecting credit assessment issues in supply chain bill scenarios, failing to consider privacy protection requirements when training models through multi-party data collaboration, and lacking a dynamic management mechanism for credit differences among participants. This makes it difficult to apply to supply chain finance scenarios requiring credit transfer and risk assessment.

[0006] In terms of domestic patents, Chinese patent application CN113781215A discloses a supply chain finance value transfer system based on a consortium blockchain. This system designs an on-chain value transfer system and method based on the consortium blockchain, specifically including supporting tokens representing assets on the chain. The issuance, transfer, redemption, and settlement of tokens are carried out through the consortium blockchain based on cryptographic identity IDs and organizational member management. The system also anchors on-chain tokens to transaction documents, revenue receipts, warehouse receipts, inbound and outbound vouchers, logistics vouchers, etc. on the supply chain, and completes supply chain finance business through smart contracts, realizing the value transfer of the supply chain on the consortium blockchain. The patent's shortcomings lie in its use of a traditional single-core enterprise credit endorsement model for credit assessment, which only considers the credit transfer from the core enterprise to direct suppliers and fails to establish a multi-source credit assessment model based on the supply chain network topology. This results in inaccurate credit assessments of second-tier, third-tier, and other multi-level suppliers. Furthermore, while the patent uses blockchain to ensure data immutability, it fails to address data privacy protection issues during multi-party collaborative modeling. It lacks an adaptive privacy budget allocation mechanism for the credit differences among different participants, making it impossible to achieve high-quality federated learning model training while protecting data privacy. In addition, the patent does not consider the impact of the integrity and time efficiency of the bill circulation path on credit assessment and lacks quantitative analysis of the potential risk accumulation effects of bills during circulation. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings by proposing a supply chain invoice ecosystem interconnection system based on open APIs and smart contracts.

[0008] The present invention adopts the following technical solution:

[0009] A supply chain bill ecosystem interconnection system based on open APIs and smart contracts includes an adaptive open interface collaboration module, a cross-chain smart contract orchestration module, a multi-source credit intelligent computing module, and an end-to-end privacy and security protection module.

[0010] The adaptive open interface collaboration module is responsible for enabling intelligent access and permission management of heterogeneous systems; the cross-chain smart contract orchestration module is responsible for unified management and automated execution of contracts in a multi-blockchain environment; the multi-source credit intelligent computing module is responsible for integrating multi-party data to build a supply chain credit evaluation system; and the end-to-end privacy and security protection module is responsible for providing end-to-end data encryption, privacy computing, and identity verification.

[0011] The adaptive open interface collaboration module includes a multi-protocol intelligent adaptation unit, a hierarchical permission dynamic management unit, and an interface performance self-optimization unit. The multi-protocol intelligent adaptation unit is responsible for realizing intelligent protocol conversion and semantic mapping of heterogeneous systems. The hierarchical permission dynamic management unit is responsible for automatic permission adjustment and risk control based on behavior. The interface performance self-optimization unit is responsible for intelligent routing and load balancing self-learning.

[0012] The cross-chain smart contract orchestration module includes a multi-chain protocol adaptation unit, an event-driven contract engine unit, and a contract template library and configuration engine unit. The multi-chain protocol adaptation unit is responsible for achieving unified access and interoperability of heterogeneous blockchains. The event-driven contract engine unit is responsible for the automatic triggering and cascading execution of contracts based on business events. The contract template library and configuration engine unit is responsible for the rapid generation of visual contracts and industry customization.

[0013] The multi-source credit intelligent computing module includes a heterogeneous data fusion preprocessing unit, a graph neural network credit modeling unit, and a federated learning privacy protection unit. The heterogeneous data fusion preprocessing unit is responsible for cleaning, aligning, and extracting features from multi-source data. The graph neural network credit modeling unit is responsible for credit scoring based on supply chain network topology awareness. The federated learning privacy protection unit is responsible for data privacy protection under multi-party collaborative modeling.

[0014] The end-to-end privacy and security protection module includes a layered encrypted transmission unit, a cryptographic calculation and query unit, and a zero-knowledge proof verification unit. The layered encrypted transmission unit is responsible for multi-level encryption protection of data transmission across the entire link. The cryptographic calculation and query unit is responsible for data calculation and retrieval in encrypted state. The zero-knowledge proof verification unit is responsible for proving the authenticity of transactions without information leakage.

[0015] Furthermore, the heterogeneous data fusion preprocessing unit includes an entity recognition processor, a time-series alignment processor, and a missing value imputation processor. The entity recognition processor is used to extract entity objects such as enterprises, invoices, and transactions from multi-source heterogeneous data. It identifies the relationship types between entities, such as supply, financing, and guarantee, through named entity recognition and relationship extraction technologies, and constructs a supply chain knowledge graph containing enterprise nodes and relationship edges. The time-series alignment processor is used to unify the timestamp benchmark from different data sources, handle issues such as time zone differences and inconsistent sampling frequencies, and use interpolation or downsampling methods to align time-series data to a unified time grid, providing standardized input for subsequent time-series analysis. The missing value imputation processor is used to handle missing items generated during data acquisition. Based on the information propagation algorithm of graph neural networks, it uses known information of adjacent nodes in the supply chain network to infer the attribute values ​​of missing nodes, or uses time series interpolation methods to fill the gaps in time-series data, thereby improving data integrity.

[0016] The entity recognition processor calculates the path credibility of the ticket from enterprise i to enterprise j according to the following formula. :

[0017] ;

[0018] ;

[0019] Where k is an intermediate node in the path, e0 represents the starting enterprise i, and e k+1 Indicates the endpoint company j, N represents the reliability of the relationship between adjacent firms. ff N is the number of times a contract can be successfully fulfilled. total N represents the total number of transactions. od The number of overdue payments. d represents the overdue penalty coefficient. n This represents the time delay for the nth hop.

[0020] Furthermore, the graph neural network credit modeling unit includes a graph structure construction processor, a credit propagation processor, and a temporal evolution processor. The graph structure construction processor is used to construct a multi-layer heterogeneous graph structure for enterprises, bills, and transaction relationships in the supply chain, define node types and edge types, and assign an initial feature vector to each node, including attributes such as the enterprise's financial indicators, the bill's amount and term, and the transaction history. The credit propagation processor is used to aggregate the neighbor information of nodes through multi-layer graph convolution operations, calculate the contribution weight of different neighbors to the credit score of the central node, realize the layer-by-layer propagation and accumulation of credit in the supply chain network, and generate the node's credit embedding vector. The temporal evolution processor is used to introduce a time-aware mechanism to capture the dynamic changes in credit scores, jointly model the graph structure at historical moments and the current graph structure, learn the evolution pattern of credit over time through a temporal graph convolutional network, and predict credit trends and default risks in future periods.

[0021] The graph structure building processor calculates the credit potential of enterprise i according to the following formula. :

[0022] ;

[0023] Where Core represents the core set of enterprises, and S c This indicates the strength of company C's credit source. Indicates path length. Let Ngb(i) be the potential energy decay coefficient, and let V be the set of direct trading partners of firm i. bill (i, j) represents the total amount of bill transactions for companies i and j, V total (i) represents the total amount of all bill transactions for company i. This represents the credit potential value of company j in the previous iteration.

[0024] Furthermore, the credit propagation processor converts the credit potential energy into a credit score Cr(i) according to the following formula:

[0025] ;

[0026] in, This is the scaling factor for the rating. D represents the average credit potential of all firms. i This is the counterparty diversity index for firm i. This is the diversity reward coefficient.

[0027] Furthermore, the federated learning privacy protection unit includes a gradient encryption processor, a differential privacy processor, and a Byzantine fault-tolerant processor. The gradient encryption processor is used to encrypt the gradient information uploaded by each participant during the model training process of federated learning using a secure multi-party computation protocol, and complete the gradient aggregation operation in encrypted state to ensure that the local data of each participant is not obtained by other parties. The differential privacy processor is used to add carefully designed noise perturbation before gradient uploading, so that the addition or removal of a single sample will not significantly affect the model output result. The Byzantine fault-tolerant processor is used to detect and prevent malicious participant attacks in the federated learning process, and identify and remove maliciously uploaded gradient information.

[0028] The gradient encryption processor calculates the privacy budget of enterprise i in the t-th round of federated learning according to the following formula. :

[0029] ;

[0030] in, Based on privacy budget, Cr max The highest credit score among all businesses. For credit sensitivity index, Let Variance be the gradient volatility of firm i. This is the fluctuation penalty coefficient.

[0031] The beneficial effects achieved by this invention are:

[0032] This system proposes a reliable quantification mechanism for bill circulation paths, enabling accurate assessment of the complete bill circulation path. It effectively identifies weak links and high-risk paths in the supply chain. The proposed credit potential field model overcomes the limitations of traditional linear credit decay. Through the power-law decay transmission of core enterprise credit and the local potential diffusion of trading partners, it fully considers the topological characteristics of the supply chain network and the positional value of enterprises within the network. This allows second- and third-tier suppliers at the end of the supply chain to obtain more accurate credit scores based on their actual performance within the network, solving the problem of SMEs struggling to obtain reasonable credit in the traditional model. Regarding privacy protection, a credit-anchored differential privacy budget allocation mechanism is proposed. This mechanism dynamically adjusts the noise addition intensity based on the participants' supply chain credit scores and gradient fluctuation variance, allowing participants with high credit and good data quality to contribute more effective information, while participants with low credit or abnormal behavior are subject to stricter privacy constraints and impact suppression. Simultaneously, a real-time gradient fluctuation monitoring mechanism effectively prevents data poisoning attacks, significantly enhancing the model's robustness.

[0033] To further understand the features and technical content of the present invention, please refer to the following detailed description and drawings of the present invention. However, the drawings provided are for reference and illustration only and are not intended to limit the present invention. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall structural framework of the present invention;

[0035] Figure 2 This is a schematic diagram of the multi-source credit intelligent computing module of the present invention;

[0036] Figure 3 This is a schematic diagram of the bill path credibility calculation flowchart of the present invention;

[0037] Figure 4 This is a schematic diagram of the calculation flowchart for the credit potential energy field model of this invention;

[0038] Figure 5 This is a schematic diagram of the credit-anchored privacy budget allocation and application process of the present invention;

[0039] Figure 6 This is a diagram comparing the accuracy of credit assessment at different supply chain levels between the present invention and traditional methods;

[0040] Figure 7 This is a schematic diagram illustrating the effect comparison of the path credibility mechanism of the present invention. Detailed Implementation

[0041] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can understand the advantages and effects of the present invention from the content disclosed in this specification. The present invention can be implemented or applied through other different specific embodiments, and various details in this specification can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. Furthermore, the accompanying drawings of the present invention are for simple illustrative purposes only and are not depictions of actual dimensions; this is stated beforehand. The following embodiments will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0042] Example 1.

[0043] This embodiment provides a supply chain invoice ecosystem interconnection system based on open APIs and smart contracts, combined with... Figure 1 It includes an adaptive open interface collaboration module, a cross-chain smart contract orchestration module, a multi-source credit smart computing module, and an end-to-end privacy and security protection module;

[0044] The adaptive open interface collaboration module is responsible for enabling intelligent access and permission management of heterogeneous systems; the cross-chain smart contract orchestration module is responsible for unified management and automated execution of contracts in a multi-blockchain environment; the multi-source credit intelligent computing module is responsible for integrating multi-party data to build a supply chain credit evaluation system; and the end-to-end privacy and security protection module is responsible for providing end-to-end data encryption, privacy computing, and identity verification.

[0045] The adaptive open interface collaboration module includes a multi-protocol intelligent adaptation unit, a hierarchical permission dynamic management unit, and an interface performance self-optimization unit. The multi-protocol intelligent adaptation unit is responsible for realizing intelligent protocol conversion and semantic mapping of heterogeneous systems. The hierarchical permission dynamic management unit is responsible for automatic permission adjustment and risk control based on behavior. The interface performance self-optimization unit is responsible for intelligent routing and load balancing self-learning.

[0046] The cross-chain smart contract orchestration module includes a multi-chain protocol adaptation unit, an event-driven contract engine unit, and a contract template library and configuration engine unit. The multi-chain protocol adaptation unit is responsible for achieving unified access and interoperability of heterogeneous blockchains. The event-driven contract engine unit is responsible for the automatic triggering and cascading execution of contracts based on business events. The contract template library and configuration engine unit is responsible for the rapid generation of visual contracts and industry customization.

[0047] The multi-source credit intelligent computing module includes a heterogeneous data fusion preprocessing unit, a graph neural network credit modeling unit, and a federated learning privacy protection unit. The heterogeneous data fusion preprocessing unit is responsible for cleaning, aligning, and extracting features from multi-source data. The graph neural network credit modeling unit is responsible for credit scoring based on supply chain network topology awareness. The federated learning privacy protection unit is responsible for data privacy protection under multi-party collaborative modeling.

[0048] The end-to-end privacy and security protection module includes a layered encrypted transmission unit, a cryptographic calculation and query unit, and a zero-knowledge proof verification unit. The layered encrypted transmission unit is responsible for multi-level encryption protection of data transmission across the entire link. The cryptographic calculation and query unit is responsible for data calculation and retrieval in encrypted state. The zero-knowledge proof verification unit is responsible for proving the authenticity of transactions without information leakage.

[0049] The heterogeneous data fusion preprocessing unit includes an entity recognition processor, a time-series alignment processor, and a missing value imputation processor. The entity recognition processor is used to extract entity objects such as enterprises, invoices, and transactions from multi-source heterogeneous data. It identifies the relationship types between entities, such as supply, financing, and guarantee, through named entity recognition and relationship extraction technologies, and constructs a supply chain knowledge graph containing enterprise nodes and relationship edges. The time-series alignment processor is used to unify the timestamp benchmark from different data sources, handle issues such as time zone differences and inconsistent sampling frequencies, and use interpolation or downsampling methods to align time-series data to a unified time grid, providing standardized input for subsequent time-series analysis. The missing value imputation processor is used to handle missing items generated during data acquisition. Based on the information propagation algorithm of graph neural networks, it uses known information of adjacent nodes in the supply chain network to infer the attribute values ​​of missing nodes, or uses time series interpolation methods to fill the gaps in time-series data, thereby improving data integrity.

[0050] The entity recognition processor calculates the path credibility of the ticket from enterprise i to enterprise j according to the following formula. :

[0051] ;

[0052] ;

[0053] Where k is an intermediate node in the path, e0 represents the starting enterprise i, and e k+1 Indicates the endpoint company j, N represents the reliability of the relationship between adjacent firms. ff N is the number of times a contract can be successfully fulfilled. total N represents the total number of transactions. od The number of overdue payments. d represents the overdue penalty coefficient. n This represents the time delay for the nth hop.

[0054] The graph neural network credit modeling unit includes a graph structure building processor, a credit propagation processor, and a temporal evolution processor. The graph structure building processor is used to construct a multi-layer heterogeneous graph structure of enterprises, bills, and transaction relationships in the supply chain, define node types and edge types, and assign an initial feature vector to each node, including attributes such as the enterprise's financial indicators, the bill's amount and term, and the transaction history. The credit propagation processor is used to aggregate the neighbor information of nodes through multi-layer graph convolution operations, calculate the contribution weight of different neighbors to the credit score of the central node, realize the layer-by-layer propagation and accumulation of credit in the supply chain network, and generate the credit embedding vector of the node. The temporal evolution processor is used to introduce a time-aware mechanism to capture the dynamic changes of credit scores, jointly model the graph structure at historical moments and the current graph structure, learn the evolution pattern of credit over time through a temporal graph convolutional network, and predict credit trends and default risks in future periods.

[0055] The graph structure building processor calculates the credit potential of enterprise i according to the following formula. :

[0056] ;

[0057] Where Core represents the core set of enterprises, and S c This indicates the strength of company C's credit source. Indicates path length. Let Ngb(i) be the potential energy decay coefficient, and let V be the set of direct trading partners of firm i. bill (i, j) represents the total amount of bill transactions for companies i and j, V total (i) represents the total amount of all bill transactions for company i. This represents the credit potential value of company j in the previous iteration.

[0058] The credit transmission processor converts credit potential energy into a credit score Cr(i) according to the following formula:

[0059] ;

[0060] in, This is the scaling factor for the rating. D represents the average credit potential of all firms. i This is the counterparty diversity index for firm i. This is the diversity reward coefficient.

[0061] The federated learning privacy protection unit includes a gradient encryption processor, a differential privacy processor, and a Byzantine fault-tolerant processor. The gradient encryption processor is used to encrypt the gradient information uploaded by each participant during the model training process of federated learning using a secure multi-party computation protocol, and complete the gradient aggregation operation in encrypted state to ensure that the local data of each participant is not obtained by other parties. The differential privacy processor is used to add carefully designed noise perturbation before gradient uploading, so that the addition or removal of a single sample will not significantly affect the model output. The Byzantine fault-tolerant processor is used to detect and prevent malicious participant attacks in the federated learning process, and identify and remove maliciously uploaded gradient information.

[0062] The gradient encryption processor calculates the privacy budget of enterprise i in the t-th round of federated learning according to the following formula. :

[0063] ;

[0064] in, Based on privacy budget, Cr max The highest credit score among all businesses. For credit sensitivity index, Let Variance be the gradient volatility of firm i. This is the fluctuation penalty coefficient.

[0065] Example 2.

[0066] This embodiment includes all the contents of Embodiment 1, and provides a supply chain bill ecosystem interconnection system based on open APIs and smart contracts, including an adaptive open interface collaboration module, a cross-chain smart contract orchestration module, a multi-source credit intelligent calculation module, and an end-to-end privacy and security protection module;

[0067] The adaptive open interface collaboration module is responsible for enabling intelligent access and permission management of heterogeneous systems; the cross-chain smart contract orchestration module is responsible for unified management and automated execution of contracts in a multi-blockchain environment; the multi-source credit intelligent computing module is responsible for integrating multi-party data to build a supply chain credit evaluation system; and the end-to-end privacy and security protection module is responsible for providing end-to-end data encryption, privacy computing, and identity verification.

[0068] The adaptive open interface collaboration module includes a multi-protocol intelligent adaptation unit, a hierarchical permission dynamic management unit, and an interface performance self-optimization unit. The multi-protocol intelligent adaptation unit is responsible for realizing intelligent protocol conversion and semantic mapping of heterogeneous systems. The hierarchical permission dynamic management unit is responsible for automatic permission adjustment and risk control based on behavior. The interface performance self-optimization unit is responsible for intelligent routing and load balancing self-learning.

[0069] The cross-chain smart contract orchestration module includes a multi-chain protocol adaptation unit, an event-driven contract engine unit, and a contract template library and configuration engine unit. The multi-chain protocol adaptation unit is responsible for achieving unified access and interoperability of heterogeneous blockchains. The event-driven contract engine unit is responsible for the automatic triggering and cascading execution of contracts based on business events. The contract template library and configuration engine unit is responsible for the rapid generation of visual contracts and industry customization.

[0070] Combination Figure 2 The multi-source credit intelligent computing module includes a heterogeneous data fusion preprocessing unit, a graph neural network credit modeling unit, and a federated learning privacy protection unit. The heterogeneous data fusion preprocessing unit is responsible for cleaning, aligning, and extracting features from multi-source data. The graph neural network credit modeling unit is responsible for credit scoring based on supply chain network topology awareness. The federated learning privacy protection unit is responsible for data privacy protection under multi-party collaborative modeling.

[0071] The end-to-end privacy and security protection module includes a layered encrypted transmission unit, a cryptographic calculation and query unit, and a zero-knowledge proof verification unit. The layered encrypted transmission unit is responsible for multi-level encryption protection of the entire link data transmission. The cryptographic calculation and query unit is responsible for data calculation and retrieval in the encrypted state. The zero-knowledge proof verification unit is responsible for proving the authenticity of transactions without information leakage.

[0072] The multi-protocol intelligent adaptation unit includes a protocol identification processor, a semantic mapping processor, and a version compatibility processor. The protocol identification processor is used to automatically detect the data format type sent by the access party, identify the protocol standard by parsing the data packet structure features, and generate a protocol type identifier for subsequent processing. The semantic mapping processor is used to construct a semantic conversion relationship graph between different protocol standards, realize bidirectional conversion at the field level through a pre-set mapping rule base, accurately map the data elements of the source protocol to the corresponding elements of the target protocol, and ensure the semantic consistency of cross-protocol data exchange. The version compatibility processor is used to manage the coexistence of multiple versions of the same protocol standard, maintain a version difference comparison table, and automatically handle field additions, deletions, and structural changes between versions.

[0073] The hierarchical permission dynamic management unit includes a behavior profiler, a permission adjustment processor, and an anomaly detection processor. The behavior profiler is used to collect and analyze the API call history data of the access party, extract behavioral features such as call frequency, time distribution, and interface preferences, construct a multi-dimensional behavior profile model, and calculate a credit rating score based on historical performance. The permission adjustment processor is used to dynamically adjust the access permission range of the access party according to the credit rating output by the behavior profiler, realize the time window limit of permissions and the elastic scaling of concurrent requests, automatically expand permissions when the credit rating is improved, and shrink permissions when the credit rating is reduced. The anomaly detection processor is used to monitor API call patterns in real time, identify abnormal behaviors such as high-frequency calls, illegal access, and data crawling through set threshold rules and machine learning anomaly detection algorithms, and trigger automatic rate limiting, temporary blocking, or downgrade processing strategies.

[0074] The interface performance self-optimization unit includes a routing optimization processor, a request prediction processor, and a load balancing processor. The routing optimization processor is used to analyze the response time, success rate, and resource consumption data of historical requests, learn the optimal request routing strategy, and dynamically select the best service node and data transmission path to minimize latency. The request prediction processor is used to predict the API call volume and type distribution in future periods based on a time series analysis model, and pre-allocate computing and storage resources in advance. The load balancing processor is used to monitor the real-time load status of each service node and allocate new requests, and automatically switch to a backup node when a node fails.

[0075] The multi-chain protocol adaptation unit includes a cross-chain relay processor, an asset mapping processor, and a state synchronization processor. The cross-chain relay processor is used to establish communication bridges between different blockchain networks, monitor transaction events on the source chain, convert event information into a format recognizable by the target chain, and transmit it to the target chain through the relay network to execute corresponding operations, thereby realizing message interoperability between consortium chains, public chains, and private chains. The asset mapping processor is used to manage the mapping relationship and lock-and-release mechanism of cross-chain assets. When assets are locked on the source chain, corresponding mapping certificates are generated on the target chain. When certificates are destroyed on the target chain, the original assets are unlocked on the source chain, ensuring the atomicity and consistency of cross-chain asset transfer. The state synchronization processor is used to maintain the consistency of business states on multiple chains, verify the authenticity of on-chain data through light node verification or Merkle proof technology, and periodically synchronize the block headers and key state data of each chain.

[0076] The event-driven contract engine unit includes an event listener processor, a state machine execution processor, and a rollback compensation processor. The event listener processor is used to establish a unified listening mechanism for on-chain events on the blockchain and off-chain events of external business systems. It subscribes to business events such as bill issuance, transfer, and financing applications, captures event trigger signals, parses event parameters, and pushes event information to the state machine execution processor. The state machine execution processor is used to determine the current state and target state after receiving an event trigger signal according to predefined business process state transition rules, and automatically calls the corresponding smart contract methods to execute business logic. It supports complex process orchestration such as sequential execution, parallel execution, and conditional branching. The rollback compensation processor is used to handle abnormal situations during smart contract execution. When a transaction fails or business conditions are not met, a rollback operation is triggered to restore the contract state to the snapshot before execution, or to execute a predefined compensation transaction to correct the side effects that have occurred, ensuring the transaction consistency of the business process.

[0077] The contract template library and configuration engine unit includes a template manager, a visual configurator, and a security audit processor. The template manager stores and manages parameterized smart contract templates, covering typical business scenarios such as financing, splitting, endorsement, and redemption of supply chain bills. It provides template version control, categorized retrieval, and reuse mechanisms, and supports rapid contract instantiation based on business needs. The visual configurator provides a graphical contract configuration interface, allowing users to combine business process nodes, set contract parameters and trigger conditions through drag-and-drop operations, and automatically generate contract code or configuration files, reducing the technical threshold for contract development and enabling zero-code contract generation. The security audit processor performs automated security checks on the generated smart contract code, scanning for common security vulnerabilities such as reentrancy attacks, integer overflows, and access control defects. It outputs a security audit report and provides remediation suggestions to ensure the security of the contract before it is uploaded to the blockchain.

[0078] The heterogeneous data fusion preprocessing unit includes an entity recognition processor, a time-series alignment processor, and a missing value imputation processor. The entity recognition processor extracts entity objects such as enterprises, invoices, and transactions from multi-source heterogeneous data. It identifies the relationship types between entities, such as supply, financing, and guarantee, through named entity recognition and relationship extraction technologies, and constructs a supply chain knowledge graph containing enterprise nodes and relationship edges. The time-series alignment processor unifies the timestamp benchmark from different data sources, handles issues such as time zone differences and inconsistent sampling frequencies, and uses interpolation or downsampling methods to align time-series data to a unified time grid, providing standardized input for subsequent time-series analysis. The missing value imputation processor handles missing items generated during data acquisition. Based on graph neural network information propagation algorithms, it infers the attribute values ​​of missing nodes using known information from adjacent nodes in the supply chain network, or uses time-series interpolation methods to fill gaps in time-series data, improving data integrity.

[0079] Combination Figure 3 The entity recognition processor calculates the path reliability of the invoice from enterprise i to enterprise j according to the following formula. :

[0080] ;

[0081] ;

[0082] Where k is an intermediate node in the path, e0 represents the starting enterprise i, and e k+1 Indicates the endpoint company j, N represents the reliability of the relationship between adjacent firms. ff N is the number of times a contract can be successfully fulfilled. total N represents the total number of transactions. od The number of overdue payments. d represents the overdue penalty coefficient. n The time delay for the nth hop;

[0083] The entity recognition processor records all path credibility information in the supply chain knowledge graph;

[0084] The graph neural network credit modeling unit includes a graph structure building processor, a credit propagation processor, and a temporal evolution processor. The graph structure building processor is used to construct a multi-layered heterogeneous graph structure from enterprises, bills, and transaction relationships in the supply chain, define node types and edge types, and assign an initial feature vector to each node, including attributes such as the enterprise's financial indicators, the bill's amount and term, and the transaction history. The credit propagation processor is used to aggregate the neighbor information of nodes through multi-layer graph convolution operations, calculate the contribution weight of different neighbors to the credit score of the central node, realize the layer-by-layer propagation and accumulation of credit in the supply chain network, and generate the node's credit embedding vector. The temporal evolution processor is used to introduce a time-aware mechanism to capture the dynamic changes in credit scores, jointly model the graph structure at historical moments and the current graph structure, learn the evolution pattern of credit over time through a temporal graph convolutional network, and predict credit trends and default risks in future periods.

[0085] Combination Figure 4 The graph structure construction processor calculates the credit potential of enterprise i according to the following formula. :

[0086] ;

[0087] Where Core represents the core set of enterprises, and S c This indicates the strength of company C's credit source. Indicates path length. Let Ngb(i) be the potential energy decay coefficient, and let V be the set of direct trading partners of firm i. bill (i, j) represents the total amount of bill transactions for companies i and j, V total (i) represents the total amount of all bill transactions for company i. This represents the credit potential value of firm j in the previous iteration;

[0088] The credit transmission processor converts credit potential energy into a credit score Cr(i) according to the following formula:

[0089] ;

[0090] in, This is the scaling factor for the rating. D represents the average credit potential of all firms. i This is the counterparty diversity index for firm i. For diversity reward coefficient;

[0091] The federated learning privacy protection unit includes a gradient encryption processor, a differential privacy processor, and a Byzantine fault-tolerant processor. The gradient encryption processor encrypts the gradient information uploaded by each participant during model training using a secure multi-party computation protocol, performing gradient aggregation in encrypted form to ensure that each participant's local data is not accessed by other parties. The differential privacy processor adds carefully designed noise perturbations before gradient uploading, ensuring that the addition or removal of a single sample does not significantly affect the model output, preventing attackers from inferring the privacy information of the training data through reverse engineering, and providing quantifiable privacy protection strength. The Byzantine fault-tolerant processor detects and prevents malicious attacks during the federated learning process, identifying and removing maliciously uploaded gradient information through majority voting, abnormal gradient detection, and robust aggregation algorithms, ensuring that the model can still converge to the correct result even when some participants exhibit malicious behavior.

[0092] Combination Figure 5 The gradient encryption processor calculates the privacy budget of enterprise i in the t-th round of federated learning according to the following formula. :

[0093] ;

[0094] in, Based on privacy budget, Cr max The highest credit score among all businesses. For credit sensitivity index, Let Variance be the gradient volatility of firm i. This is the fluctuation penalty coefficient;

[0095] Based on the calculated privacy budget, the system applies it to the gradient noise addition process in federated learning: After completing model training locally, enterprise i generates a noise vector, adds it element by element to the original gradient, and uploads the resulting noisy gradient to the central server for aggregation and update. The smaller the privacy budget, the greater the noise intensity that needs to be added, thus protecting its data privacy while also reducing the negative impact of its potential abnormal data on the global model. The larger the privacy budget, the smaller the added noise intensity, retaining more effective information and enabling its high-quality data to fully contribute to the improvement of model performance. In addition, the system restricts the data access scope of participants according to the privacy budget level, thereby achieving full-link linkage control from privacy protection to access control.

[0096] The layered encrypted transmission unit includes a key negotiation processor, a transmission encryption processor, and a quantum-safe processor. The key negotiation processor executes a key exchange protocol when the two communicating parties establish a connection, and uses the RSA asymmetric encryption algorithm to negotiate and generate a session key, ensuring that each communication uses an independent temporary key, thereby achieving forward security to prevent the cracking of historical communications. The transmission encryption processor implements multi-layered encryption protection for the data transmission channel, establishing an encrypted tunnel at the transmission layer and performing secondary encryption on sensitive fields at the application layer, forming a dual protection mechanism of transmission layer encryption and application layer encryption. The quantum-safe processor is used to address the threat of future quantum computing to the existing encryption system and to construct a quantum-resistant key exchange and digital signature system.

[0097] The encrypted computation and query unit includes a homomorphic encryption processor, a searchable encryption processor, and a secure computation coordinator. The homomorphic encryption processor supports direct mathematical operations in encrypted form, employing a fully homomorphic encryption algorithm to perform addition and multiplication operations on the encrypted data. This allows data owners to upload encrypted data to the computing platform, which performs statistical analysis or model inference without decryption. The computation results are then decrypted by the data owner. The searchable encryption processor enables keyword retrieval of the encrypted database, constructing an encrypted index structure. Users submit encrypted query keywords, and the system matches matching data records in the encrypted index and returns them. The entire process does not reveal the plaintext content of the query keywords and retrieval results. The secure computation coordinator coordinates joint computation tasks involving multiple data holders, decomposing the computation task into multiple sub-tasks and assigning them to each participant. Each party performs local computation and returns encrypted intermediate results. The coordinator aggregates these results to obtain the final computation output, enabling joint querying and statistical analysis where data is available but not visible.

[0098] The zero-knowledge proof verification unit includes a proof generation processor, a range proof processor, and a proof aggregation processor. The proof generation processor generates zero-knowledge proofs for transactions or identity information. The prover generates concise proof data based on the original data and the proof key. The verifier only needs the proof data and the verification key to verify the authenticity of the statement without needing to obtain the original data content. The range proof processor proves that a certain value is within a specific range without revealing the specific value. It is applied to the compliance verification of transaction amounts and the qualification verification of identity and age. Through the commitment scheme and the range proof protocol, the verifier is convinced that the transaction amount does not exceed the limit or the user's age meets the requirements, but the exact amount or age cannot be known. The proof aggregation processor merges multiple independent zero-knowledge proofs into a single aggregated proof, which significantly reduces the space overhead of storing proof data on the blockchain and the time cost of verification computation. It is particularly suitable for batch transaction scenarios. The verifier can verify the validity of multiple transactions simultaneously through a single aggregation verification operation.

[0099] The n and p mentioned above are ordinal numbers used to represent sequence numbers and have no actual meaning.

[0100] To verify the beneficial effects of the present invention, we conducted a comparative experiment based on real data from a provincial supply chain invoice platform.

[0101] The first part is the credit assessment performance verification experiment. First, the dataset was divided into training and test sets in an 8:2 ratio to ensure a balanced distribution of enterprises across all supply chain levels. Then, credit assessment models for four comparative methods were constructed: a traditional machine learning method based on support vector machines and random forests; a standard graph neural network method using graph attention networks; a deep network embedding method based on Node2Vec; and the method proposed in this patent that combines invoice path credibility with a credit potential field. For the traditional machine learning method, we extracted 15 financial indicators and 8 transaction behavior features of enterprises as inputs and used grid search for hyperparameter optimization. For the graph neural network method, we constructed a supply chain network graph containing all enterprise nodes and transaction relationship edges, trained it using a three-layer graph convolutional network with a learning rate of 0.001. For the method in this patent, we first calculated the credibility of all invoice circulation paths according to Formula 1, incorporated the path credibility as edge weights into the supply chain network graph, and then constructed a credit potential field model according to Formula 2, setting the potential decay index α to 2.0, iterating until convergence to obtain the credit score of each enterprise. Further analysis of the evaluation results across different supply chain levels revealed that for Tier 1 suppliers, the accuracy rate of this patented method reached 91%, a 9 percentage point improvement compared to the traditional method's 82%. For Tier 3 and Tier 4 suppliers, the accuracy rates reached 79% and 72%, respectively, representing improvements of 28 and 34 percentage points compared to the traditional methods' 51% and 38%, respectively. This fully validates the superiority of the credit potential field model in multi-level credit transfer scenarios. To verify the effectiveness of the path credibility mechanism, we designed an ablation experiment to test the credit evaluation results with and without path credibility. The results show that for a two-hop bill transfer, introducing path credibility improved the accuracy from 0.71 to 0.89. For longer paths of four and five hops, the accuracy improved from 0.52 and 0.41 to 0.77 and 0.71, respectively, representing improvements of 30% to 35%. This demonstrates that the path credibility quantification mechanism can effectively capture the risk accumulation effect during bill transfer.

[0102] The data was organized and obtained Figure 6 and Figure 7 .

[0103] Example 3.

[0104] This embodiment provides a specific implementation plan for a supply chain invoice ecosystem interconnection system based on open APIs and smart contracts. The system is deployed on the supply chain finance platform of a large manufacturing group, serving one core manufacturing enterprise, 127 Tier 1 suppliers, 438 Tier 2 suppliers, and 892 Tier 3 suppliers. The system adopts a hybrid cloud architecture, with the core computing nodes configured as Alibaba Cloud ECS compute-optimized c7 instances. Each node is equipped with a 64-core Intel Xeon Platinum 8369B processor with a main frequency of 2.7GHz, 256GB of DDR4 memory, and ESSD PL3 cloud disks providing a performance guarantee of 2 million IOPS. The database layer uses Alibaba Cloud PolarDB distributed version 5.0, configured with a primary / standby dual-node setup plus 5 read-only nodes, providing a query processing capacity of 120,000 queries per second.

[0105] The multi-protocol intelligent adaptation unit of the adaptive open interface collaboration module deploys a protocol recognition engine based on the BERT-Base Chinese pre-trained model. This model includes a 12-layer Transformer encoder, 768 hidden layer dimensions, 12 attention heads, and approximately 110 million parameters. The protocol recognition engine was fine-tuned on a training set containing 15 message types (including CAMT, PACS, and PAIN) from the ISO 20022 standard, 23 protocol formats (GS1 EPCIS 2.0 standard, and custom JSON and XML formats), with a training set of 580,000 real message samples and a validation set of 120,000. Training used the AdamW optimizer with a learning rate of 2e-5 and a batch size of 32. After 15 epochs, it achieved a recognition accuracy of 98.7% on the test set, a 22.4 percentage point improvement compared to the 76.3% accuracy of traditional rule-based recognition methods. As a technical variant, the protocol recognition engine can also use the XLNet-Base model instead of the BERT model. This model employs permutation language modeling and a two-stream attention mechanism, which can further improve the accuracy to 99.2% in long text protocol recognition scenarios. However, the inference time increases from an average of 32 milliseconds to 48 milliseconds, making it suitable for scenarios with lower real-time requirements but extremely high accuracy requirements. The semantic mapping processor maintains a rule base containing 8732 mapping rules, stored in Neo4j graph database version 4.4.12. Each rule is represented as a triplet structure of source field node, target field node, and transformation function edge, with a query response time consistently within 5 milliseconds. The version-compatible processor supports parallel processing of up to 5 historical versions of the same protocol standard. The version difference lookup table uses differential storage, saving 67% of storage space compared to full storage.

[0106] The behavior profile processor of the hierarchical permission dynamic management unit uses the time-series database InfluxDB 2.6 to record API call logs. The data retention strategy is set to keep the most recent 30 days of hot data in memory, migrate the 31 to 180 days of warm data to SSD, and archive the data older than 180 days to object storage. Behavioral feature extraction adopts a sliding time window mechanism with a window size of 7 days and a step size of 1 day. Each window extracts 18-dimensional features, including call frequency, peak QPS, interface type distribution, and time period distribution. The credit rating score uses the XGBoost 1.7.3 model with the following parameters: maximum tree depth of 8, learning rate of 0.05, subsampling rate of 0.8, and L2 regularization coefficient of 1.0. After training on a training set containing historical data of 23,000 access parties, the credit rating prediction accuracy reached 91.2%. The permission adjustment processor implements a five-level permission system, including a restricted level with a quota of 100 API calls per minute, a basic level with 500 calls, a standard level with 2000 calls, a priority level with 8000 calls, and a VIP level with no limit. The permission level is automatically adjusted every 24 hours based on the latest credit score and anomaly detection results. The anomaly detection processor employs a hybrid detection scheme combining the Isolation Forest algorithm and an LSTM (Long Short-Term Memory) network. The Isolation Forest is used to detect statistical anomalies such as sudden high-frequency calls, while the LSTM network is used to detect temporal pattern anomalies such as regular calls outside of working hours. The detection results of the two algorithms are weighted and fused, with the Isolation Forest weight at 0.6 and the LSTM weight at 0.4. The fused anomaly detection accuracy reaches 94.3%, with a false positive rate controlled below 2.1%. As an alternative, anomaly detection can also employ an unsupervised anomaly detection method based on an autoencoder. This method trains a neural network containing a 3-layer encoder and a 3-layer decoder, with hidden layer dimensions of 128, 64, and 32 respectively. Anomalies are identified through reconstruction errors. This method achieves a detection accuracy of 93.8% without labeled data and is suitable for scenarios where anomaly samples are scarce.

[0107] The cross-chain smart contract orchestration module is deployed in a hybrid chain environment containing three heterogeneous blockchains. The consortium blockchain uses Hyperledger Fabric version 2.5.4, configured with three organizations each running two peer nodes and one orderer node. The orderer uses the Raft consensus algorithm, with a block size limit of 2MB and a block generation interval of 500 milliseconds. The ledger database uses CouchDB 3.3.2 to provide rich query capabilities. The public blockchain uses Ethereum Geth client version 1.13.8 to connect to the Sepolia testnet. Contracts are written in Solidity 0.8.19 and compiled and deployed using Hardhat 2.19.2. The private blockchain uses FISCO BCOS version 3.6.0, configured with four consensus nodes and two observer nodes, using the PBFT consensus algorithm, with an average consensus latency consistently below one second. The multi-chain protocol adaptation unit implements a unified cross-chain message format. The message header includes eight mandatory fields such as source chain identifier, target chain identifier, message type, and timestamp, as well as business extension fields. The message body uses Protocol Buffers 3.21.9 encoding, reducing the amount of transmitted data by 42% compared to JSON format. The cross-chain relay processor is deployed on an independent relay service cluster, using Kafka 3.5.1 as the message queue to buffer cross-chain messages. Kafka is configured with three partitions and two replication factors to ensure no message loss. The asset mapping processor deploys locking contracts and mapping contracts on the source and target chains respectively. The locking contract uses a multi-signature mechanism, requiring at least three of the five verification nodes to sign and confirm before asset locking can be executed. The mapping contract automatically generates an equivalent amount of mapped assets based on the locking credentials, with a mapping ratio of 1:1 and an average mapping confirmation time of 3.2 seconds. The state synchronization processor performs a state synchronization check every 10 seconds, employing a Merkle tree verification method. It writes the latest block header hash and key state root hash of each chain to the synchronization log, achieving 100% accuracy in detecting state inconsistencies. Upon detection of an inconsistency, it immediately triggers an automatic rollback and resynchronization process. As a technological variation, cross-chain relay can also adopt a decentralized relay solution based on smart contracts. This involves deploying relay contracts on each chain to achieve peer-to-peer message passing. This solution eliminates the single point of failure risk of centralized relay nodes, but the cross-chain confirmation time increases to 5 to 8 seconds.

[0108] The event-driven contract engine unit's event listener processor subscribes to blockchain events using a WebSocket long-connection method, establishing two redundant connections for each chain to ensure high availability, with event capture latency controlled within 300 milliseconds. Off-chain business events are accessed through RabbitMQ 3.12.8 message queues, configured with persistent mode and mirrored queue strategies. Message confirmation employs a publisher confirmation mechanism to ensure reliable delivery. The state machine execution processor is implemented based on the finite state machine framework XState 4.38.2. The ticket issuance process defines seven state nodes, including application submission, document review, quota approval, contract generation, on-chain registration, notification confirmation, and issuance completion, as well as 12 state transition edges. Each transition edge is associated with a specific smart contract method and precondition checks. Parallel execution uses a coroutine pool approach, with the coroutine pool size dynamically adjusted from a minimum of 10 to a maximum of 200, automatically scaling based on the current number of pending events and system load, achieving a parallel processing capacity of 1500 events per second. The rollback compensation processor implements a checkpoint-based snapshot mechanism, saving a snapshot before each critical state transition. The snapshot includes contract state, database records, and external system call records. Snapshots use copy-on-write technology, recording only the changed parts. The average snapshot size is 3.7KB, and the average rollback operation time is 45 milliseconds. As an alternative, rollback compensation can also adopt an event-based sourcing compensation model, recording all state changes as an event sequence. During rollback, state recovery is achieved by reversing the execution of compensation events. This solution provides a complete operation audit trail, but increases storage overhead by approximately 1.8 times.

[0109] The heterogeneous data fusion preprocessing unit of the multi-source credit intelligent computing module connects to five types of data sources, including overdue announcement data from the Shanghai Commercial Paper Exchange, enterprise credit reports from the People's Bank of China Credit Reference Center, transaction data from core enterprise ERP systems, tax records from the tax system, and enterprise registration information from the industrial and commercial system. The entity recognition processor uses a named entity recognition model based on the BERT-BiLSTM-CRF architecture. The BERT layer uses the Chinese pre-trained model chinese-bert-wwm-ext, the BiLSTM layer contains two bidirectional LSTM layers with a hidden layer dimension of 256, and the CRF layer is used for sequence labeling. The model is trained on a self-built supply chain entity dataset containing 32,000 labeled samples, achieving an entity recognition F1 score of 94.6%. Relation extraction uses a rule-based method based on dependency parsing, with an accuracy of 89.3%. The time-series alignment processor handles differences in timestamp formats from various data sources, supporting automatic recognition and unified conversion of eight time formats, including ISO 8601, Unix timestamps, and China Standard Time. Time zone conversion is based on the IANA time zone database version 2024a. When handling data with inconsistent sampling frequencies, linear interpolation is used to upsample sparse data, and averaging is used to downsample dense data, aligning to a daily granularity. The missing value imputation processor employs different strategies for different types of missing values. For completely random missing values, mean imputation is used. For random missing values, the K-nearest neighbor algorithm is used to find corresponding values ​​from similar companies for imputation, with K set to 5. For non-random missing values, a graph neural network propagation algorithm is used to infer missing values ​​by aggregating known information from neighboring nodes. This algorithm achieves an accuracy of 87.2% after imputation on a test dataset containing a 30% missing value rate.

[0110] The graph structure construction processor of the graph neural network credit modeling unit extracted 15,734 enterprise nodes, 1.26 million bill nodes, and 2.73 million transaction relationship edges from the database. The constructed supply chain graph has an average node degree of 3.8, a graph diameter of 12 hops, and the largest connected component contains 99.2% of the nodes. The graph structure is stored in sparse matrix CSR format, saving 98.7% of memory space compared to dense matrix. The credit source strength calculation integrates seven indicators, including enterprise registered capital, credit rating, market capitalization, and revenue scale. The weights of each indicator are determined through principal component analysis, with the first principal component explaining 61.3% of the variance. The calculation of remote potential energy transmission and local potential energy diffusion is implemented using a message-passing neural network framework. Iterative calculations are performed on a server configured with NVIDIA A100 GPUs, with each iteration taking 1.2 seconds. Convergence typically occurs after 7 to 9 iterations, and the convergence threshold is set to the L2 norm of the potential energy changes of all nodes in two consecutive rounds being less than 0.01. The credit propagation processor uses graph sampling technology to accelerate computation, aggregating up to 50 neighbors for each node. The sampling strategy is importance sampling weighted by transaction amount, which improves computation speed by 6.8 times compared to full neighbor aggregation, while only decreasing accuracy by 0.3 percentage points. The diversity index is calculated based on the Shannon entropy formula. For a firm with only one trading counterparty, the diversity index is 0; for a firm with 10 trading counterparties and evenly distributed transactions, the diversity index is 2.3. The diversity reward coefficient β is set to 0.2, allowing the credit score of highly diverse firms to improve by up to 20%. As a technical variation, credit propagation can also use a graph attention network (GAT) instead of a message-passing neural network. GAT automatically learns the contribution weights of neighbors through an attention mechanism, eliminating the need for manually setting aggregation functions. It generalizes better when transferred from small-scale graphs to large-scale graphs after training, but the training time increases by about 2.3 times.

[0111] The federated learning privacy protection unit deployed a federated learning environment with 50 participants. The central server adopted the FedML 0.8.4 federated learning framework, supporting various aggregation algorithms such as FedAvg, FedProx, and FedOpt. This embodiment uses the FedAvg algorithm, and the global model after each round of aggregation is distributed to all participants. The gradient encryption processor uses the Paillier homomorphic encryption algorithm to encrypt gradients, with a key length of 2048 bits. Encrypting a gradient vector takes an average of 1.8 milliseconds, and decryption takes 2.3 milliseconds. The Laplacian noise scale added by the differential privacy processor is dynamically adjusted according to the privacy budget. The basic privacy budget ε is set to 1.0, the credit sensitivity index γ is set to 1.5, and the volatility penalty coefficient δ is set to 5.0. The gradient volatility variance is calculated using a sliding window method, with the window size being the last 5 training rounds. The Byzantine Fault Tolerance Processor employs the Krum aggregation algorithm. This algorithm calculates the Euclidean distance between each participant's uploaded gradient and all other gradients, selecting the k gradients with the smallest sum of distances for aggregation. The value of k is set to two-thirds of the total number of participants, allowing it to tolerate up to one-third of malicious participants. In a simulated data poisoning attack experiment, five malicious participants uploaded tampered gradients. The Krum algorithm successfully identified and excluded 4.6 averages, resulting in a model accuracy drop of only 1.2 percentage points. In contrast, the accuracy drop of the ordinary average aggregation method without Byzantine Fault Tolerance reached 8.7 percentage points. As an alternative, Byzantine Fault Tolerance can also use a median-based aggregation method, calculating the median of all uploaded values ​​for each gradient parameter as the aggregation result. This method has lower computational complexity but is slightly weaker in defending against multi-dimensional collaborative attacks.

[0112] The layered encrypted transmission unit of the end-to-end privacy and security protection module adopts the TLS 1.3 protocol at the transport layer, using the TLS_AES_256_GCM_SHA384 cipher suite, providing 256-bit AES-GCM encryption and SHA-384 hashing. Key exchange employs the elliptic curve Diffie-Hellman algorithm with the curve parameter secp384r1, providing 192-bit security strength. Application layer encryption uses the Chinese national standard SM4 algorithm with a 128-bit key length and CBC mode, performing secondary encryption on 12 sensitive fields such as the invoice amount, company name, and counterparty. The quantum-safe processor implements the lattice-based key exchange algorithm Kyber-768, which has been selected by NIST as a post-quantum cryptography standard. It provides approximately 192 bits of classical and quantum security strength. The key exchange process uses a public key size of 1184 bytes and a ciphertext size of 1088 bytes, approximately three times larger than the 384-byte public key of RSA-3072, but still resistant to quantum computing attacks. The encrypted computation and query unit uses the SEAL 4.1.1 homomorphic encryption library, with the BFV encryption scheme selected. The polynomial modulus degree is 8192, the coefficient modulus is 218 bits, and the plaintext modulus is 40 bits. This configuration supports approximately 20 levels of homomorphic operations, enabling complex calculations such as weighted summation and polynomial regression. The searchable encryption uses the symmetric searchable encryption scheme SSE, with the index optimized using a Bloom filter. The Bloom filter size is 1MB, the number of hash functions is 3, and the false positive rate is controlled within 0.1%. The zero-knowledge proof verification unit uses the Groth16 proof system. This system generates proofs of a fixed size of 128 bytes, with a fixed verification time of 2.3 milliseconds, which does not increase with circuit size, making it suitable for on-chain verification scenarios. The bill amount range proof uses the Bulletproofs protocol. The proof size for a 64-bit value range is 672 bytes, the generation time is 45 milliseconds, and the verification time is 5 milliseconds. Compared with the disadvantage of Groth16 requiring a trusted setting, Bulletproofs does not require a trusted setting and supports aggregation, but the proof size increases logarithmically with the number of bits in the range.

[0113] The content disclosed above is only a preferred and feasible embodiment of the present invention, and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent technical changes made based on the content of the present invention specification and drawings are included within the scope of protection of the present invention. Furthermore, the elements therein can be updated as technology develops.

Claims

1. An open API and smart contract based supply chain invoice eco-interaction system, characterized in that, The adaptive open interface collaboration module, the cross-chain smart contract arrangement module, the multi-source credit intelligent computing module, and the end-to-end privacy security protection module are comprised. The adaptive open interface collaboration module is responsible for realizing intelligent access and permission management of heterogeneous systems, the cross-chain smart contract arrangement module is responsible for realizing unified management and automatic execution of contracts in a multi-blockchain environment, the multi-source credit intelligent computing module is responsible for fusing multi-party data to construct a supply chain credit evaluation system, and the end-to-end privacy security protection module is responsible for providing data encryption, privacy computing, and identity verification throughout the whole process. The adaptive open interface collaboration module comprises a multi-protocol intelligent adaptation unit, a hierarchical permission dynamic management unit, and an interface performance self-optimization unit, the multi-protocol intelligent adaptation unit is responsible for realizing intelligent protocol conversion and semantic mapping of heterogeneous systems, the hierarchical permission dynamic management unit is responsible for behavior-based automatic adjustment of permissions and risk control, and the interface performance self-optimization unit is responsible for self-learning of intelligent routing and load balancing. The cross-chain smart contract arrangement module comprises a multi-chain protocol adaptation unit, an event-driven contract engine unit, and a contract template library and configuration engine unit, the multi-chain protocol adaptation unit is responsible for realizing unified access and interoperation of heterogeneous blockchains, the event-driven contract engine unit is responsible for automatic triggering and cascading execution of contracts based on business events, and the contract template library and configuration engine unit is responsible for visual contract rapid generation and industry customization. The multi-source credit intelligent computing module comprises a heterogeneous data fusion preprocessing unit, a graph neural network credit modeling unit, and a federated learning privacy protection unit, the heterogeneous data fusion preprocessing unit is responsible for cleaning, alignment, and feature extraction of multi-source data, the graph neural network credit modeling unit is responsible for supply chain network topology-aware credit scoring, and the federated learning privacy protection unit is responsible for data privacy protection under multi-party collaborative modeling. The end-to-end privacy security protection module comprises a hierarchical encryption transmission unit, a ciphertext computing and query unit, and a zero-knowledge proof verification unit, the hierarchical encryption transmission unit is responsible for multi-level encryption protection of full-link data transmission, the ciphertext computing and query unit is responsible for data computing and retrieval in a ciphertext state, and the zero-knowledge proof verification unit is responsible for transaction authenticity proof without information leakage.

2. The open API and smart contract based supply chain invoice eco-interaction system of claim 1, wherein, The heterogeneous data fusion preprocessing unit comprises an entity recognition processor, a time alignment processor and a missing value filling processor. The entity recognition processor is configured to extract enterprise, bill and transaction entity objects from the multi-source heterogeneous data, identify supply, financing, guarantee and other relationship types between entities through named entity recognition and relationship extraction technology, construct a supply chain knowledge graph comprising enterprise nodes and relationship edges, and the time alignment processor is configured to unify the timestamp reference from different data sources, process time zone differences, inconsistent sampling frequencies and other problems, align the time series data to a unified time grid by using interpolation or down-sampling methods, and provide standardized input for subsequent time series analysis. The missing value filling processor is configured to process missing items generated in the data collection process, infer attribute values of missing nodes using known information of adjacent nodes in the supply chain network based on a graph neural network information propagation algorithm, or fill in the gaps of time series data using a time series interpolation method, and improve data integrity. The entity recognition processor calculates the path credibility of the bill flowing from enterprise i to enterprise j according to the following formula : ； ； where k is the intermediate node in the path, e0represents the starting enterprise i, e k+1 represents the end enterprise j, represents the relationship reliability between adjacent enterprises, N ff is the number of successful transactions, N total is the total number of transactions, N od is the number of overdue times, is the overdue penalty coefficient, d n is the time delay of the nth hop.

3. The open API and smart contract based supply chain invoice ecosystem interworking system of claim 2, wherein, The graph neural network credit modeling unit comprises a graph structure construction processor, a credit propagation processor and a time evolution processor. The graph structure construction processor is configured to construct enterprises, bills and transaction relationships in the supply chain into a multi-layer heterogeneous graph structure, define node types and edge types, and assign an initial feature vector to each node, including financial indicators of enterprises, amounts and maturities of bills, and historical records of transactions. The credit propagation processor is configured to aggregate neighbor information of nodes through multi-layer graph convolution operations, calculate contribution weights of different neighbors to the credit score of the center node, realize layer-by-layer propagation and accumulation of credit in the supply chain network, and generate a credit embedding vector of the node. The time evolution processor is configured to introduce a time perception mechanism to capture dynamic changes in the credit score, jointly model the graph structure at the historical time and the current graph structure, learn the evolution pattern of the credit over time through a time series graph convolution network, and predict the credit trend and default risk in the future period. The graph structure construction processor calculates the credit potential of the enterprise i according to the following formula : ； where Core denotes the set of core enterprises, S c denotes the credit source strength of enterprise c, denotes the path length, is the potential energy decay coefficient, Ngb(i) denotes the set of direct transaction partners of enterprise i, V bill (i,j) denotes the total amount of bill transaction between enterprise i and enterprise j, V total (i) denotes the total amount of all bill transactions of enterprise i, denotes the credit potential value of enterprise j in the last iteration.

4. The open API and smart contract based supply chain bill ecosystem interworking system of claim 3, wherein, The credit propagation processor converts the credit potential into the credit score Cr(i) according to the following formula: ； wherein, is a score scaling factor, is the average credit potential value for all enterprises, D i is the counterparty diversity index for enterprise i, is a diversity reward factor.

5. The open API and smart contract based supply chain invoice ecosystem interworking system of claim 4, wherein, The federated learning privacy protection unit comprises a gradient encryption processor, a differential privacy processor and a Byzantine fault tolerance processor. The gradient encryption processor is configured to encrypt the gradient information uploaded by each participant in the model training process of federated learning using a secure multi-party computation protocol, complete the gradient aggregation operation in the ciphertext state, and ensure that the local data of each participant is not obtained by other parties. The differential privacy processor is configured to add carefully designed noise disturbance before uploading the gradient, so that the addition or removal of a single sample does not significantly affect the model output result. The Byzantine fault tolerance processor is configured to detect and prevent malicious participant attacks in the federated learning process, and identify and exclude malicious uploaded gradient information. The gradient encryption processor calculates the privacy budget of the enterprise i in the tth round of federated learning according to the following formula : ； wherein, is the base privacy budget, Cr max is the highest credit score for all businesses, is the credit sensitivity index, is the gradient variance for business i, is the volatility penalty coefficient.

Images