Supply chain finance-oriented multi-party data collaboration and credit evaluation method and system

Abstract

The invention provides a supply chain finance-oriented multi-party data collaboration and credit evaluation method and system, and relates to the technical field of credit evaluation, and the method comprises the steps: creating a digital identity, constructing a distributed data collaboration mechanism, collecting supply chain finance data through a privacy calculation program, and constructing a supply chain credit evaluation network based on a business association relationship. And calculating credit conduction intensity to generate dynamic credit characteristics and calculating credit evaluation index values. According to the invention, secure collaborative sharing of multi-party data is realized, the privacy of original data is protected, and the accuracy and credibility of credit assessment of supply chain participants are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of credit evaluation, in particular to a multi-party data collaboration and credit evaluation method and system for supply chain finance. BACKGROUND

[0002] Supply chain finance is a financial service mode connecting upstream and downstream enterprises in the supply chain. Through the integrated management of fund flow, information flow and logistics in the supply chain, it provides financing, settlement and risk management services for supply chain participants. With the development of digital economy, supply chain finance has become an important way to solve the financing difficulties and high financing costs of small and medium-sized enterprises. Traditional supply chain finance mainly relies on the credit guarantee of core enterprises, and with the development of technology, multi-party data collaboration and credit evaluation methods based on blockchain, big data and artificial intelligence have gradually become a research hotspot in the industry.

[0003] At present, the multi-party data collaboration and credit evaluation in the field of supply chain finance has the following defects and deficiencies: There is serious information asymmetry between supply chain participants, and the data of each party is scattered and closed, making it difficult to form a comprehensive credit profile. Financial institutions cannot obtain complete supply chain transaction data, leading to inaccurate risk assessment and low efficiency of financing decision-making, which restricts the healthy development of supply chain finance business.

[0004] Existing data sharing mechanisms often face challenges in data security and privacy protection. Supply chain participants are reluctant to share raw data due to business secret protection considerations, and traditional data collaboration methods cannot effectively balance the contradiction between data value mining and privacy protection, hindering the deep collaboration of multi-party data. SUMMARY

[0005] The embodiment of the present application provides a multi-party data collaboration and credit evaluation method and system for supply chain finance, which can solve the problems in the prior art.

[0006] The first aspect of the embodiment of the present application provides a multi-party data collaboration and credit evaluation method for supply chain finance, comprising: receiving a data collaboration request submitted by a supply chain participant, and creating a digital identity for each supply chain participant in a blockchain network according to the subject information in the data collaboration request; building a distributed data collaboration mechanism based on the digital identity, connecting the local data nodes of the supply chain participants to the blockchain network, deploying smart contracts and privacy computing programs on each local data node, and the privacy computing program performs encrypted calculation on the local data through a federated learning model, and only submits the calculation result to the blockchain network, ensuring that the original data does not leave the local node; The privacy computing program collects supply chain financial data from each local data node, extracts business association relationships between supply chain participants based on the supply chain financial data, constructs a supply chain credit evaluation network according to the business association relationships, takes the supply chain participants as network nodes, takes the business association relationships as network connection edges, calculates the credit conduction strength between the network nodes based on a graph neural network, and the credit conduction strength is used to represent the credit influence degree between the supply chain participants. According to the credit conduction strength, a dynamic credit feature is generated for each network node, and the dynamic credit feature is used to calculate the credit evaluation index value of the supply chain participant by weighting.

[0007] Receiving a data collaboration request submitted by a supply chain participant, creating a digital identity for each supply chain participant in a blockchain network according to subject information in the data collaboration request includes: The data collaboration request includes enterprise registration information, business qualification information, legal person information, and data use range; The enterprise registration information is compared with the enterprise credit database, the business qualification information is verified with the administrative permission database, and the legal person information is checked with the identity authentication system; based on the verification result, the subject trustworthiness score is calculated; according to the subject trustworthiness score, the permission level of the supply chain participant is determined; Deploying an identity management smart contract in the blockchain network, generating an identity string containing the participant identifier, the permission level identifier, and the time limit identifier for the supply chain participant based on the permission level; The identity string is processed by fragmentation to obtain multiple identity fragments, and the identity fragments are sent to a preset number of trust nodes in the blockchain network; the trust nodes sign the received identity fragments; the signature results of the trust nodes are collected, and the signed identity fragments are combined to generate a digital identity certificate.

[0008] Based on the digital identity, a distributed data collaboration mechanism is constructed, the local data nodes of the supply chain participants are connected to the blockchain network, and smart contracts and privacy computing programs are deployed in each local data node, the privacy computing program performs encrypted calculation on the local data through a federated learning model, and only the calculation results are submitted to the blockchain network, including: Receiving a data node access request of a supply chain participant, assigning a node identifier to the data node based on the digital identity; Deploying a three-layer smart contract architecture in the data node, the three-layer smart contract architecture includes a data access layer contract for data access control, a business logic layer contract for feature extraction, and a collaborative computing layer contract for computing task distribution; the data access layer contract verifies the data operation authority according to the node identifier; The local data of the data node is preprocessed and feature extracted by the business logic layer contract to generate a training sample set; and a federated learning model is constructed based on the training sample set, and homomorphic encryption algorithm is used to encrypt the model parameters; The encrypted model parameters are distributed to each data node by the collaborative computing layer contract, each data node calculates the model gradient value based on the training sample set, and adds differential privacy noise based on the Laplace distribution to the model gradient value; and the intensity of the differential privacy noise is adaptively adjusted according to the calculation round; The model gradient value to which the differential privacy noise is added is aggregated in a weighted average manner to generate a new round of model parameters; the new round of model parameters are cross-validated among the data nodes through the data access layer contract, and the new round of model parameters are written into the blockchain network after the validation.

[0009] The model gradient value is added with differential privacy noise based on the Laplace distribution; and the intensity of the differential privacy noise is adaptively adjusted according to the calculation round, including: The value range of each parameter in the model gradient value is calculated, the sensitivity of the model gradient value is determined based on the value range, and the sensitivity represents the maximum influence degree of a single sample on the model gradient value; An initial privacy budget and a budget decay coefficient are set, the initial privacy budget is divided by the sensitivity to obtain an initial scale parameter of the Laplace distribution, and a location parameter of the Laplace distribution is determined by the model gradient value; A probability density function of the Laplace distribution is constructed based on the initial scale parameter and the location parameter, differential privacy noise is sampled from the probability density function, and the differential privacy noise is superimposed on the model gradient value to generate a noisy gradient value that satisfies differential privacy protection; The initial privacy budget is multiplied by the training round power of the budget decay coefficient to obtain the privacy budget of the current round; and the privacy budget of the current round is divided by the sensitivity to obtain the scale parameter of the Laplace distribution of the current round; The updated Laplace distribution is used to protect the differential privacy of the model gradient value of the next round.

[0010] A supply chain credit evaluation network is constructed according to the business association relationship, a supply chain participant is taken as a network node, a business association relationship is taken as a network connection edge, and the credit conduction intensity between the network nodes is calculated based on a graph neural network, including: According to the business association relationship, a multi-layer information propagation structure of a supply chain credit transmission network is constructed, in each layer of information propagation, a message vector between a node pair is calculated based on a node feature vector corresponding to a network node, the message vector is obtained by linear transformation of the node feature vector and through a nonlinear activation function, and the message vector is multiplied by an association weight between the node pair to obtain a weighted message vector; An attention mechanism is introduced to selectively aggregate neighbor information of the network node, an attention weight between the network node and neighbor nodes thereof is calculated, the attention weight is multiplied by the weighted message vector and summed to obtain a propagation feature of the network node at a current layer; The propagation feature of the network node is iteratively updated in multiple layers, the message vector calculation and attention aggregation operations are repeatedly performed in each layer of information propagation to obtain a final feature representation of the network node; Based on the final feature representation, a credit transmission strength between the node pair is calculated, the credit transmission strength is calculated by similarity of the node features, and a final transmission probability is obtained through normalization processing.

[0011] According to the credit transmission strength, a dynamic credit feature of each network node is generated, and the dynamic credit feature is used to calculate a credit evaluation index value of a supply chain participant through weighted calculation, including: A time encoding vector is obtained by encoding time information using a trigonometric function of different frequencies, and the time encoding vector and historical credit records are input into a long short-term memory network to generate a memory state vector of the network node; The memory state vector and the credit transmission strength are fused to obtain a dynamic credit feature of the network node through nonlinear transformation; Based on the dynamic credit feature, a first-order influence of the network node is calculated, and the first-order influence is obtained by weighting the dynamic credit features of the direct neighbor nodes; A multi-hop influence of the network node is calculated, and the multi-hop influence is obtained by iterative calculation, in each hop propagation, the first-order influence of the last hop is combined with the dynamic credit feature of the current network node through weighting; The multi-hop influence is adaptively aggregated to obtain a final credit evaluation index value of the supply chain participant.

[0012] In a second aspect of the embodiment of the application, a multi-party data collaboration and credit evaluation system for supply chain finance is provided, including: A first unit is configured to receive a data collaboration request submitted by a supply chain participant, and create a digital identity for each supply chain participant in a blockchain network according to subject information in the data collaboration request; The second unit is configured to build a distributed data collaboration mechanism based on the digital identity, access local data nodes of the supply chain participants to a blockchain network, and deploy smart contracts and a privacy computing program at each local data node, wherein the privacy computing program performs encrypted calculation on the local data through a federated learning model, and only submits the calculation result to the blockchain network, so that the original data is ensured not to leave the local node. The third unit is configured to collect supply chain financial data from each local data node through the privacy computing program, extract a business association relationship between the supply chain participants based on the supply chain financial data, build a supply chain credit evaluation network according to the business association relationship, take the supply chain participants as network nodes and the business association relationship as network connection edges, and calculate a credit conduction strength between the network nodes based on a graph neural network, wherein the credit conduction strength is used to represent a credit influence degree between the supply chain participants. The fourth unit is configured to generate a dynamic credit feature for each network node according to the credit conduction strength, and calculate a credit evaluation index value of the supply chain participant by weighting using the dynamic credit feature.

[0013] In a third aspect, an electronic device is provided, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0014] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0015] The present application has the following advantages: By creating a blockchain digital identity for the supply chain participants and building a distributed data collaboration mechanism, the information island problem in the traditional supply chain finance is solved, and the secure sharing and collaborative analysis of multi-party data are realized, while the data privacy and security of each participant are ensured.

[0016] Based on the privacy computing and federated learning technology, the present method ensures that the original data does not leave the local node, and only the calculation result is shared, effectively balancing the relationship between data value mining and privacy protection, and enhancing the willingness and trust of the supply chain participants to share data.

[0017] By building a supply chain credit evaluation network and introducing a graph neural network to calculate the credit conduction strength, the present method can dynamically capture the credit association and conduction effect between the supply chain participants, generate a more comprehensive and accurate credit evaluation result, and provide a scientific basis for supply chain financial risk management, thereby effectively improving the efficiency and security of financial services. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 A flowchart of a multi-party data collaboration and credit evaluation method for supply chain finance of an embodiment of the present application is shown in Figure 2 A supply chain data collaboration mechanism architecture based on a blockchain and federated learning is shown in Figure 3 A multi-hop influence propagation process is shown in DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in connection with the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0020] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0021] Figure 1 A flowchart of a multi-party data collaboration and credit evaluation method for supply chain finance of an embodiment of the present application is shown in Figure 1 The method comprises: receiving a data collaboration request submitted by a supply chain participant, and creating a digital identity for each supply chain participant in a blockchain network according to subject information in the data collaboration request; building a distributed data collaboration mechanism based on the digital identity, connecting local data nodes of the supply chain participants to the blockchain network, deploying smart contracts and privacy computing programs on each local data node, and performing encrypted calculation on local data through a federated learning model, and only submitting calculation results to the blockchain network to ensure that original data does not leave the local node; collecting supply chain financial data from each local data node through the privacy computing program, extracting business association relationships between the supply chain participants based on the supply chain financial data, building a supply chain credit evaluation network according to the business association relationships, taking the supply chain participants as network nodes and the business association relationships as network connection edges, calculating credit conduction intensity between the network nodes based on a graph neural network, and using the credit conduction intensity to represent the credit influence degree between the supply chain participants; According to the credit conduction intensity, a dynamic credit feature is generated for each network node, and the dynamic credit feature is used to calculate a credit evaluation index value of the supply chain participant by weighting.

[0022] In an optional embodiment, a data collaboration request submitted by a supply chain participant is received, and a digital identity is created for each supply chain participant in a blockchain network according to subject information in the data collaboration request, including: The data collaboration request includes enterprise registration information, business qualification information, legal person information, and data use range; The enterprise registration information is compared with an enterprise credit database, the business qualification information is verified with an administrative license database, and the legal person information is checked with an identity authentication system; a subject trustworthiness score is calculated based on the verification result; and the permission level of the supply chain participant is determined according to the subject trustworthiness score; An identity management smart contract is deployed in the blockchain network, and an identity string containing a participant identifier, a permission level identifier, and a time limit identifier is generated for the supply chain participant based on the permission level; The identity string is processed by fragmentation to obtain multiple identity fragments, and the identity fragments are sent to a predetermined number of trust nodes in the blockchain network; the trust nodes sign the received identity fragments; the signature results of the trust nodes are collected, and the signed identity fragments are combined to generate a digital identity certificate.

[0023] In order to realize receiving a data collaboration request submitted by a supply chain participant, and creating a digital identity for each supply chain participant in a blockchain network according to the subject information in the request.

[0024] A data collaboration request submitted by a supply chain participant is received, and the request contains enterprise registration information, business qualification information, legal person information, and data use range. The enterprise registration information includes enterprise name, unified social credit code, registered address, registered capital, date of establishment, etc. The business qualification information includes business license, industry license, special business license, and other qualification documents. The legal person information includes the name of the legal person, the ID number, the contact information, etc. The data use range indicates the types and purposes of data that the enterprise hopes to access and share in supply chain collaboration.

[0025] The enterprise registration information is compared with the enterprise credit database to verify the authenticity of the enterprise entity. For example, the system checks whether the unified social credit code "91X30Y75Z12345678" matches the record in the enterprise credit database, and checks the consistency of the registered capital, date of establishment, etc. Then the business qualification information is verified with the administrative license database to confirm the authenticity and validity of the qualification certificate held by the enterprise. For example, the system verifies whether the business license number "X10Y20Z3040" in the administrative license database is "valid", and whether the permitted business scope conforms to the data collaboration type applied. At the same time, the system checks the information of the legal person with the identity authentication system to confirm the authenticity of the legal person's identity. The system verifies the authenticity of the legal person's "Zhang" identity card number "3101X2Y9Z80123456" through the identity authentication interface, and checks the name and photo information.

[0026] Based on the above verification results, the system calculates the subject credibility score. The calculation method considers the verification results of the three dimensions and gives different weights. The verification result of the enterprise registration information has a weight of 0.4, the verification result of the business qualification information has a weight of 0.35, and the verification result of the legal person information has a weight of 0.25. The verification result of each dimension is divided into three cases: complete match (100 points), partial match (60 points), and mismatch (0 points). Assuming that the verification results of a certain enterprise in the three dimensions are complete match, complete match, and partial match, respectively, the subject credibility score of the enterprise is 0.4x100+0.35x100+0.25x60=90 points.

[0027] According to the subject credibility score calculated, the system determines the permission level of the supply chain participants. The permission level is divided into four levels: 90-100 points for A level (highest permission), 70-89 points for B level (high permission), 50-69 points for C level (basic permission), and less than 50 points for D level (restricted permission). Each permission level corresponds to different data access and interaction permissions. For example, A-level participants can access and share all types of supply chain data, B-level participants can access most data but some sensitive data need to be approved, C-level participants can only access basic business data, and D-level participants can only access public data.

[0028] The system deploys an identity management smart contract in the blockchain network, which contains functions such as identity generation, verification, update, and revocation. Based on the determined permission level, the smart contract generates an identity string for the supply chain participant, which includes the participant identifier, permission level identifier, and time limit identifier. The participant identifier is composed of the enterprise unified social credit code and a random hash value, such as "91X30Y75Z12345678-8fd7e9a2c5b1". The permission level identifier directly corresponds to the permission level, such as "Auth-A" representing A-level permission. The time limit identifier includes the identity generation time and validity period, such as "202405010000-202505010000" representing a validity period of one year. An example of a complete identity string is "91X30Y75Z12345678-8fd7e9a2c5b1-Auth-A-202405010000-202505010000".

[0029] To improve security, the system performs fragmentation processing on the identity string, dividing it into multiple identity fragments. Using a threshold secret sharing algorithm, the identity string is divided into 5 fragments, and only any 3 fragments are needed to reconstruct the complete identity. Examples of fragments include fragment 1 "91X30Y75Z12345678-8fd7#AUTH-*#2024*", fragment 2 "*12345678#fd7e9a2c5b1-A*#*0000-2025*", etc. The system sends these identity fragments to a predetermined number of trust nodes in the blockchain network, such as 5 high-reputation nodes in the network.

[0030] After receiving the identity fragments, the trust nodes use their respective private keys to digitally sign the fragments. The signing process uses the ECDSA algorithm to ensure the integrity and source verifiability of the fragments. The system collects the signature results of all trust nodes, and at least needs to collect a predetermined threshold number (such as 3) of valid signatures. The system combines the signed identity fragments and adds metadata information (such as a list of signing nodes, a timestamp of signing, etc.), and finally generates a digital identity certificate. The generated certificate is in X.509 format, containing complete identity information, fragment reconstruction rules, signature verification information, etc., serving as a credential for supply chain participants to perform identity authentication and permission control in the blockchain network.

[0031] In an optional implementation, a distributed data collaboration mechanism is constructed based on the digital identity, and local data nodes of supply chain participants are connected to the blockchain network. Smart contracts and privacy computing programs are deployed on each local data node, and the privacy computing program performs encrypted computation on local data through a federated learning model, and only submits the computation results to the blockchain network, including: receiving a data node access request of a supply chain participant, assigning a node identifier to the data node based on the digital identity; deploying a three-layer smart contract architecture in the data node, the three-layer smart contract architecture including a data access layer contract for data access control, a business logic layer contract for feature extraction, and a collaborative computing layer contract for computing task distribution; the data access layer contract verifies data operation authority according to the node identifier; preprocessing and feature extraction of local data of the data node through the business logic layer contract to generate a training sample set; constructing a federated learning model based on the training sample set, and encrypting model parameters using a homomorphic encryption algorithm; distributing the encrypted model parameters to each data node through the collaborative computing layer contract, each data node calculating model gradient values based on the training sample set and adding Laplace distribution-based differential privacy noise to the model gradient values; and adaptively adjusting the intensity of the differential privacy noise according to the calculation round; aggregating the model gradient values with added differential privacy noise using a weighted average method to generate new round model parameters; cross- verifying the new round model parameters between data nodes through the data access layer contract, and writing the new round model parameters to the blockchain network after verification.

[0032] The application provides a supply chain data collaboration mechanism based on blockchain and privacy computing, which realizes data security sharing and collaborative computing through digital identity and federated learning.

[0033] Figure 2 A supply chain data collaboration mechanism based on blockchain and federated learning is provided, in which a supply chain participant first needs to access a blockchain network. The data node access process includes node identity authentication and identifier assignment. The system receives a data node access request of a supply chain participant, verifies the validity of its digital identity certificate. The digital identity certificate contains the organization information, public key and digital signature of the participant. After verification, the system assigns a unique identifier to the data node based on the digital identity, in the format of "ORG-NODE-Random String", such as "MANUFACTURER-NODE-7B32F9E1", which will be used for subsequent permission control and data traceability.

[0034] The system deploys a three-layer smart contract architecture at each data node. The data access layer contract (DAC) is responsible for permission management and data access control. This contract maintains an access control list (ACL) that records the operation permissions of each node on data resources. When a data node requests access to data, the DAC queries the ACL based on the node identifier to verify the operation permissions. For example, for raw material quality inspection data, the manufacturer node can have read permission, while the supplier node has read and write permissions. The permission verification example code logic is to check if the request node identifier is in the ACL and determine if the request operation type is within the allowed operation set. If the verification is successful, access is allowed, otherwise it is denied and the access attempt is recorded.

[0035] The business logic layer contract (BLC) is responsible for data preprocessing and feature extraction. This contract performs cleaning, normalization, and feature extraction on raw data. For example, in supply chain quality prediction, the BLC extracts key features from raw production data, such as production batch, temperature, humidity, raw material grade, etc. The data preprocessing steps include: removing outliers (such as temperature records exceeding the normal range), filling missing values (using the average or median), and normalizing numerical features (normalizing to the 0-1 range). The feature extraction process generates a training sample set, each sample containing an input feature vector and a target label (such as product pass rate).

[0036] The collaborative computing layer contract (CCC) is responsible for distributing computing tasks and coordinating the federated learning process. Based on the extracted training sample set, the system builds a federated learning model. The initial model parameters are randomly generated, with parameter values ranging from -0.1 to 0.1. The system uses additive homomorphic encryption algorithm to encrypt the model parameters, using a 2048-bit key to encrypt the parameter values, ensuring that the original parameters are not directly exposed.

[0037] CCC distributes the encrypted model parameters to each data node. After receiving the encrypted parameters, each node uses the local training sample set to train the model and calculate the model gradient value. To enhance privacy protection, the system adds differential privacy noise based on the Laplace distribution to the gradient value. The noise generation process is: generate Laplace noise according to the sensitivity value (set to 0.5) and privacy budget ε (initial value set to 1.0), and add the generated noise to each dimension of the gradient value. The system adaptively adjusts the intensity of differential privacy noise according to the number of calculation rounds, with the formula: current round privacy budget = initial privacy budget × (1 + round decay coefficient × current round), where the round decay coefficient is set to 0.05. In this way, as the number of training rounds increases, the noise intensity gradually decreases, balancing privacy protection and model accuracy.

[0038] Each node uploads the gradient value added with differential privacy noise to the blockchain network. The system aggregates these gradient values using a weighted average method, with the weights being proportional to the data volume of each node. For example, if node A has 100 samples and node B has 200 samples, the weight of node A is 0.33 and the weight of node B is 0.67. The aggregated gradient value is used to update the model parameters to generate new round model parameters.

[0039] The new round model parameters are cross-validated among the data nodes through the data access layer contract. The validation process includes: each node calculates the model performance indicators (such as accuracy, F1 score, etc.) using the local validation set, and if the performance indicators meet the preset threshold (such as accuracy greater than 0.85), it is considered to pass the validation. The model parameters that pass the validation are written into the blockchain network, along with the transaction hash, timestamp and digital signature of each node, ensuring that the results are traceable and tamper-proof.

[0040] Through the above mechanism, the application realizes data collaboration and model sharing among supply chain participants under the premise of protecting the privacy of original data, providing technical support for supply chain risk prediction, quality control and other applications. In the whole process, the original data always remains local, only the encrypted model parameters and the gradient added with noise are transmitted in the network, effectively protecting the data privacy and security.

[0041] In an optional implementation, differential privacy noise based on Laplace distribution is added to the model gradient value; and the strength of the differential privacy noise is adaptively adjusted according to the calculation round, comprising: The value range of each parameter in the model gradient value is calculated, and the sensitivity of the model gradient value is determined based on the value range, the sensitivity representing the maximum influence degree of a single sample on the model gradient value; An initial privacy budget and a budget decay coefficient are set, the initial privacy budget is divided by the sensitivity to obtain an initial scale parameter of the Laplace distribution, and a position parameter of the Laplace distribution is determined by the model gradient value; A probability density function of the Laplace distribution is constructed based on the initial scale parameter and the position parameter, differential privacy noise is sampled from the probability density function, and the differential privacy noise is superimposed on the model gradient value to generate a noisy gradient value meeting the differential privacy protection; The initial privacy budget is multiplied by the training round power of the budget decay coefficient to obtain the privacy budget of the current round; and the privacy budget of the current round is divided by the sensitivity to obtain the scale parameter of the Laplace distribution of the current round; The updated Laplace distribution is used to perform differential privacy protection on the model gradient value of the next round.

[0042] In the implementation of adding Laplace distribution-based differential privacy noise to the model gradient value, by systematically processing the model gradient value and adaptively adjusting the privacy noise intensity, an effective differential privacy protection mechanism can be realized.

[0043] In the specific implementation process, it is first necessary to calculate the value range of each parameter in the model gradient value. For example, for a deep neural network model containing 1 million parameters, by analyzing the gradient value calculated based on the current batch of training data, the maximum and minimum values of each parameter are determined. Assuming that in a certain training process, the maximum value of the parameter gradient is 0.85 and the minimum value is -0.72, the value range is 1.57. Based on this value range, the sensitivity of the model gradient value can be determined as 1.57, which indicates that the maximum influence of a single sample on the model gradient does not exceed 1.57. The calculation of sensitivity is crucial for ensuring the effectiveness of differential privacy protection, as it directly affects the amount of noise that needs to be added.

[0044] Next, set the initial privacy budget and the budget decay coefficient. In actual applications, the initial privacy budget can be set to 1.0 and the budget decay coefficient to 0.95. The initial privacy budget divided by the sensitivity gives the initial scale parameter of the Laplace distribution. Taking the above sensitivity 1.57 as an example, the initial scale parameter is 1.0 / 1.57≈0.637. The location parameter of the Laplace distribution is determined by the model gradient value itself, i.e. for each gradient value of a model parameter, the corresponding location parameter of the Laplace distribution is that gradient value. For example, if the gradient value of a parameter is 0.3, the location parameter of the Laplace distribution corresponding to that parameter is 0.3.

[0045] Based on the initial scale parameter and the location parameter, the probability density function of the Laplace distribution is constructed. In actual implementation, a random number generation function provided by the programming language can be used to sample from the distribution. Assuming that for a Laplace distribution with a location parameter of 0.3 and a scale parameter of 0.637, the sampled noise value is 0.42. Adding this differential privacy noise 0.42 to the original gradient value 0.3 gives a noisy gradient value of 0.72. The same operation is performed for each parameter in the model to generate noisy gradient values that satisfy the differential privacy protection requirements.

[0046] In different rounds of the training process, the privacy budget needs to be dynamically adjusted to balance model performance and privacy protection. Assuming that the current training has reached the 10th round, multiply the initial privacy budget of 1.0 by the 10th power of the budget decay coefficient of 0.95 (about 0.599) to obtain the privacy budget of the current round as 0.599. Divide this value by the sensitivity of 1.57 to obtain the scale parameter of the Laplace distribution for the current round as 0.599 / 1.57≈0.382. This scale parameter is less than the initial value of 0.637, indicating that as the training round increases, the intensity of the added noise is gradually increasing, thereby providing stronger privacy protection.

[0047] The updated Laplace distribution is used to protect the differential privacy of the model gradient values in the next round. For example, in the 11th round of training, the gradient value of a certain parameter is 0.25, and the noise 0.58 is generated using the Laplace distribution with a scale parameter of 0.382, resulting in a noisy gradient value of 0.83. In this way, as the training round increases, the noise intensity gradually increases, providing stronger privacy protection for the later training data.

[0048] In the early stage of training, the noise is relatively small, which is beneficial for the model to quickly learn the main features of the data; while in the later stage, the noise gradually increases, preventing the model from overfitting to the details of the data that may leak privacy. For example, in the 50th round of training, the budget decays to about 7.7% of the initial value, and the corresponding noise scale parameter becomes about 0.049, at which point the added noise significantly increases, providing strong privacy protection for the training data.

[0049] By this method of adaptively adjusting the intensity of differential privacy noise, the privacy of the training data can be effectively protected while ensuring the effectiveness of the model training. Experiments show that the accuracy of the model trained using this method on test data is only 2.5% lower than that of the model without privacy protection, while providing ε=3.0 differential privacy guarantee, which is considered strong enough in most application scenarios.

[0050] In actual deployment, the initial privacy budget and the budget decay coefficient can be adjusted according to the privacy sensitivity requirements of specific application scenarios.

[0051] In an optional implementation, a supply chain credit evaluation network is constructed according to the business association relationship, the supply chain participants are taken as network nodes, the business association relationship is taken as network connection edges, and the credit conduction strength between network nodes is calculated based on a graph neural network, including: According to the business association relationship, a multi-layer information propagation structure of a supply chain credit conduction network is constructed, in each layer of information propagation, a message vector between a node pair is calculated based on a node feature vector corresponding to a network node, the message vector is obtained by linear transformation of the node feature vector and through a nonlinear activation function, and the message vector is multiplied by an association weight between the node pair to obtain a weighted message vector; An attention mechanism is introduced to selectively aggregate neighbor information of the network node, an attention weight between the network node and a neighbor node of the network node is calculated, the attention weight is multiplied by the weighted message vector and summed to obtain a propagation feature of the network node at a current layer; The propagation feature of the network node is iteratively updated in multiple layers, the message vector calculation and the attention aggregation operation are repeatedly performed in each layer of information propagation to obtain a final feature representation of the network node; Based on the final feature representation, a credit conduction strength between the node pair is calculated, the credit conduction strength is calculated by similarity of the node features, and a final conduction probability is obtained through normalization processing.

[0052] The specific implementation of constructing a supply chain credit evaluation network according to the business association relationship is as follows: In a specific implementation, the supply chain credit evaluation system first collects basic data of each participant, including enterprise basic information, historical transaction records, financial data, and performance, etc. The system constructs an initial supply chain credit conduction network by taking each participant in the supply chain as a network node and taking business association relationships such as transaction and cooperation projects as network connection edges. Each node is represented as a feature vector, and the initial feature vector contains attribute information of the node, such as enterprise scale of 50 million yuan, establishment time of 7 years, historical default rate of 0.5%, etc.

[0053] In constructing a multi-layer information propagation structure, the system designs a three-layer information propagation mechanism. In the first layer of information propagation, for node A and node B, the system calculates a message vector from A to B based on the feature vector of A. Specifically, the system linearly transforms the feature vector [5000, 7, 0.5] of A through a weight matrix to obtain an intermediate result [25, 14, -1.5], and then processes it through a ReLU activation function to obtain a message vector [25, 14, 0]. At the same time, the system determines the association weight between A and B according to the business association relationship, such as transaction frequency of 2 times per month, average transaction amount of 1 million yuan, cooperation time of 3 years, etc. The association weight 0.8 is calculated by comprehensively considering these factors. Multiply the message vector by the association weight to obtain the weighted message vector [20, 11.2, 0].

[0054] To achieve selective aggregation of neighbor information, the system introduces an attention mechanism. Take node C as an example, its neighbor nodes include A, B, and D. The system calculates the attention weight between C and each neighbor node, specifically through the matching degree of node features. For example, the feature matching degree between C and A is 0.4, the feature matching degree between C and B is 0.3, and the feature matching degree between C and D is 0.3. After normalization processing, the attention weights of C to A, B, and D are 0.4, 0.3, and 0.3 respectively. The system multiplies these attention weights with the corresponding weighted message vectors and sums them up. For example, the weighted message vector transmitted by A to C is [20, 11.2, 0], the weighted message vector transmitted by B to C is [15, 8, 1], and the weighted message vector transmitted by D to C is [18, 9.5, 0.5]. After attention weighted summation, the propagation feature of C in the first layer is [17.9, 9.7, 0.35].

[0055] In the second layer information propagation, the system uses the propagation feature obtained in the first layer as the new feature vector of the node, and repeats the message vector calculation and attention aggregation operation. Take node C as an example, using its first layer propagation feature [17.9, 9.7, 0.35] as input, after linear transformation, [35.8, 19.4, -1.05] is obtained, and after activation function processing, the new message vector [35.8, 19.4, 0] is obtained. Considering that the network structure may change, the system recalculates the association weight and attention weight between nodes. For example, the association weight between C and its neighbor nodes E and F is 0.7 and 0.6 respectively, and the attention weight is 0.6 and 0.4 respectively. After weighted aggregation, the propagation feature of C in the second layer is [25.1, 13.6, 0.2].

[0056] Similarly, in the third layer information propagation, the system continues to use the propagation feature of the second layer for iterative calculation. Finally, the feature representation of node C is [30.2, 15.8, 0.1]. The system performs the same iterative process on all nodes to obtain the final feature representation of all nodes in the network.

[0057] Based on the final feature representation, the system calculates the credit transmission strength between node pairs. Take nodes A and G as an example, the system calculates the cosine similarity between the feature representations of A and G. Assuming that the final feature representation of A is [32.5, 16.3, 0.05] and the final feature representation of G is [28.9, 14.7, 0.12], the calculated similarity is 0.998. The system normalizes the similarity value through the Softmax function to obtain the credit transmission probability of A to G as 0.85, indicating that the influence degree of A's credit status on G is 85%.

[0058] In an actual application scenario, an electronic product supply chain includes multiple participants such as raw material suppliers, component manufacturers, assembly plants, and sellers. The credit transmission network constructed by the system shows that the credit transmission strength between the core manufacturer M and the upstream supplier S is 0.82, indicating that the credit change of S will significantly affect the credit status of M. When S delays delivery, the system predicts that the credit score of M will decrease by 12 points (out of 100). Through the above implementation, the supply chain credit evaluation system can accurately capture the credit transmission relationship between participants, provide data support for supply chain financial risk management, and improve the stability and resilience of the overall supply chain.

[0059] In an alternative embodiment, a dynamic credit feature is generated for each network node based on the credit transmission strength, and the dynamic credit feature is used to calculate the credit evaluation index value of the supply chain participant by weighting. The time information is encoded by using a triangular function with different frequencies to obtain a time encoding vector, and the time encoding vector and historical credit records are input into a long short-term memory network to generate a memory state vector of the network node; The memory state vector and the credit transmission strength are fused to obtain a dynamic credit feature of the network node through nonlinear transformation; Based on the dynamic credit feature, a first-order influence of the network node is calculated, and the first-order influence is obtained by weighting the dynamic credit features of the direct neighbor nodes; The multi-hop influence of the network node is calculated, and the multi-hop influence is obtained by iterative calculation. In each hop propagation, the first-order influence of the last hop and the dynamic credit feature of the current network node are weighted and combined; The multi-hop influence is adaptively aggregated to obtain the final credit evaluation index value of the supply chain participant.

[0060] The present application provides a credit evaluation method based on a supply chain network, which calculates the credit transmission strength and the dynamic credit feature to generate accurate credit evaluation indicators for supply chain participants.

[0061] In this embodiment, the time information is encoded. The time encoding uses trigonometric functions with different frequencies. Specifically, for a given time point t, a time encoding vector with dimension d is generated. The sine and cosine functions are arranged alternately, and the frequencies increase from low to high. For example, when d = 4, the first element of the encoding vector is sin(t / 10000), the second element is cos(t / 10000), the third element is sin(t / 10000^{2 / d}), and the fourth element is cos(t / 10000^{2 / d}). This encoding method can capture the periodic characteristics of time, so that the encoding vectors of similar time points are close in high-dimensional space. In practical applications, d = 128 can be selected as the dimension of the time encoding vector.

[0062] The time encoding vector and the historical credit record are input into a long short-term memory network (LSTM) for processing. The LSTM network includes three control units: an input gate, a forget gate, and an output gate, which can effectively capture long-term dependencies. For each node in the supply chain network, its historical credit record is represented as time series data, including the credit status of the past n time points. In this embodiment, n = 24, representing the credit record of the past 24 months. The hidden layer dimension of the LSTM network is set to 64, and the final memory state vector h is calculated by forward propagation, also with a dimension of 64.

[0063] The memory state vector h needs to be fused with the credit conduction strength. The credit conduction strength matrix S describes the influence relationship between nodes in the supply chain network, and the matrix element S_ij represents the influence degree of node j on node i. For node i, the row vector S_i related to it is extracted from the matrix S as its conduction strength feature. The feature fusion uses the concatenation operation to concatenate the memory state vector h and the conduction strength feature S_i into a higher-dimensional vector [h; S_i]. Then, a multi-layer perceptron is used for nonlinear transformation to map the fused features to a dynamic credit feature space. The multi-layer perceptron includes two fully connected layers with ReLU activation functions in between, and the output layer has a dimension of 32, obtaining the dynamic credit feature z of the node.

[0064] Based on the dynamic credit feature, the first-order influence of the network node is calculated. The first-order influence reflects the influence of the node on its neighbor nodes through direct connection. For node i, its first-order influence is calculated by weighting and summing the dynamic credit features z_j of all its direct neighbor nodes j, with the weight being the normalized conduction strength S_ij. For example, if node i has three neighbor nodes with conduction strengths of 0.3, 0.5, and 0.2, and the corresponding dynamic credit features are three 32-dimensional vectors, then the first-order influence of node i is the weighted average of these three vectors with weight proportions of 3:5:2.

[0065] Figure 3 A multi-hop influence propagation process diagram is shown for further calculating the multi-hop influence of the network node, capturing the influence of the indirect connection relationship. The multi-hop influence is calculated by iteration, and the initial state is the dynamic credit feature of the node. In k-hop propagation, the influence of the last hop (k-1) is combined with the dynamic credit feature of the current node by weighting. The weight parameter a controls the attenuation rate of information propagation, and the value range is between 0 and 1. In this embodiment, a is set to 0.85, and the maximum number of hops K is set to 3. Taking three-hop propagation as an example, the first hop obtains the first-order influence, the second hop combines the first-order influence with the dynamic credit feature according to the proportion of 0.85:0.15, and the third hop continues to combine the second hop result with the dynamic credit feature according to the same proportion.

[0066] The multi-hop influence is adaptively aggregated to obtain the final credit evaluation index value of the supply chain participant. The adaptive weight is calculated by the attention mechanism, and the weight is dynamically allocated according to the importance of the influence of different hops. The attention score is calculated by a single-layer neural network, and the input is the influence vector of each hop, and the output is the corresponding weight coefficient. For example, for three-hop influence, the weight distribution may be [0.4, 0.35, 0.25], indicating that the importance of one-hop, two-hop and three-hop influence decreases in turn. The weights are weighted and summed with the corresponding influence vector to obtain the final credit evaluation index vector. The vector can be further mapped to a single credit score value by a fully connected layer, and the range is between 0 and 100, reflecting the comprehensive credit level of the supply chain participant.

[0067] Through the above method, the application can effectively fuse time information, historical credit records and network structure features, generate accurate dynamic credit features, capture complex credit relationships in the supply chain network through multi-hop information propagation, and finally realize accurate credit evaluation of the supply chain participants.

[0068] The multi-party data collaboration and credit evaluation system for supply chain finance in the embodiment of the application comprises: A first unit is configured to receive a data collaboration request submitted by a supply chain participant, and create a digital identity for each supply chain participant in a blockchain network according to subject information in the data collaboration request; A second unit is configured to construct a distributed data collaboration mechanism based on the digital identity, connect local data nodes of the supply chain participants to the blockchain network, and deploy smart contracts and privacy computing programs on the local data nodes. The privacy computing program encrypts the local data through a federated learning model, only submits the calculation result to the blockchain network, and ensures that the original data does not leave the local node. The third unit is configured to collect supply chain financial data from each local data node by the privacy computing program, extract a business association relationship between supply chain participants based on the supply chain financial data, construct a supply chain credit evaluation network according to the business association relationship, take the supply chain participants as network nodes, take the business association relationship as a network connection edge, calculate a credit conduction strength between the network nodes based on a graph neural network, and use the credit conduction strength to represent a credit influence degree between the supply chain participants. The fourth unit is configured to generate a dynamic credit feature for each network node according to the credit conduction strength, and use the dynamic credit feature to calculate a credit evaluation index value of the supply chain participant by weighting.

[0069] In a third aspect, an electronic device is provided, including: a processor; a memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0070] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions. When the computer program instructions are executed by a processor, the method described above is implemented.

[0071] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer-readable storage medium having computer-readable program instructions loaded thereon, which are used to execute various aspects of the present application.

[0072] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-party data collaboration and credit evaluation method for supply chain finance, characterized in that, The application comprises the following steps: Receiving a data collaboration request submitted by a supply chain participant, creating a digital identity for each supply chain participant in a blockchain network according to the subject information in the data collaboration request; Building a distributed data collaboration mechanism based on the digital identity, connecting the local data nodes of the supply chain participants to the blockchain network, deploying smart contracts and privacy computing programs on each local data node, and performing encrypted calculation on the local data through a federated learning model, and only submitting the calculation results to the blockchain network to ensure that the original data does not leave the local node; Collecting supply chain financial data from each local data node through the privacy computing program, extracting business association relationships between supply chain participants based on the supply chain financial data, constructing a supply chain credit evaluation network according to the business association relationships, taking the supply chain participants as network nodes and the business association relationships as network connection edges, calculating the credit conduction intensity between the network nodes based on a graph neural network, and using the credit conduction intensity to represent the credit influence degree between the supply chain participants; Generating a dynamic credit feature for each network node according to the credit conduction intensity, and using the dynamic credit feature to calculate the credit evaluation index value of the supply chain participant through weighted calculation.

2. The method of claim 1, wherein, Receiving a data collaboration request submitted by a supply chain participant, creating a digital identity for each supply chain participant in a blockchain network according to the subject information in the data collaboration request comprises the following steps: The data collaboration request includes enterprise registration information, business qualification information, legal person information and data usage range; The enterprise registration information is compared with the enterprise credit database, the business qualification information is verified with the administrative permission database, and the legal person information is checked with the identity authentication system; the subject credibility score is calculated based on the verification result; the permission level of the supply chain participant is determined according to the subject credibility score; Deploying an identity management smart contract in the blockchain network, generating an identity string containing the participant identifier, the permission level identifier and the time limit identifier for the supply chain participant based on the permission level; The identity string is processed by fragmentation to obtain multiple identity fragments, and the identity fragments are sent to a predetermined number of trust nodes in the blockchain network; the trust nodes sign the received identity fragments; the signature results of the trust nodes are collected, and the signed identity fragments are combined to generate a digital identity certificate.

3. The method of claim 1, wherein, Based on the digital identity, a distributed data collaboration mechanism is built, the local data nodes of the supply chain participants are connected to the blockchain network, and smart contracts and privacy computing programs are deployed on each local data node; the privacy computing program performs encrypted calculation on the local data through a federated learning model, and only submits the calculation results to the blockchain network, which comprises the following steps: Receiving a data node access request of a supply chain participant, assigning a node identifier to the data node based on the digital identity; A three-layer smart contract architecture is deployed in the data node, including a data access layer contract for data access control, a business logic layer contract for feature extraction, and a collaborative computing layer contract for computing task distribution; the data access layer contract verifies data operation authority according to the node identifier; The local data of the data node is preprocessed and feature-extracted through the business logic layer contract to generate a training sample set; a federated learning model is constructed based on the training sample set, and homomorphic encryption algorithm is used to encrypt the model parameters; The encrypted model parameters are distributed to each data node through the collaborative computing layer contract, each data node calculates the model gradient value based on the training sample set, and adds differential privacy noise based on Laplace distribution to the model gradient value; the intensity of the differential privacy noise is adaptively adjusted according to the calculation round; The model gradient value to which the differential privacy noise is added is aggregated in a weighted average manner to generate a new round of model parameters; the new round of model parameters are cross-validated among the data nodes through the data access layer contract, and the new round of model parameters are written into the blockchain network after validation.

4. The method of claim 1, wherein, The model gradient value is added with differential privacy noise based on Laplace distribution; Adaptive adjustment of the intensity of the differential privacy noise according to the calculation round includes: The value range of each parameter in the model gradient value is calculated, the sensitivity of the model gradient value is determined based on the value range, and the sensitivity represents the maximum influence degree of a single sample on the model gradient value; An initial privacy budget and a budget decay coefficient are set, the initial privacy budget is divided by the sensitivity to obtain an initial scale parameter of the Laplace distribution, and the location parameter of the Laplace distribution is determined by the model gradient value; The probability density function of the Laplace distribution is constructed based on the initial scale parameter and the location parameter, differential privacy noise is sampled from the probability density function, and the differential privacy noise is superimposed on the model gradient value to generate a noisy gradient value that satisfies differential privacy protection; The initial privacy budget is multiplied by the training round power of the budget decay coefficient to obtain the privacy budget of the current round; the privacy budget of the current round is divided by the sensitivity to obtain the scale parameter of the Laplace distribution of the current round; The updated Laplace distribution is used to protect the differential privacy of the model gradient value of the next round.

5. The method of claim 1, wherein, A supply chain credit evaluation network is constructed according to the business association relationship, a supply chain participant is taken as a network node, a business association relationship is taken as a network connection edge, and the credit conduction intensity between network nodes is calculated based on a graph neural network, including: A multi-layer information propagation structure of the supply chain credit conduction network is constructed according to the business association relationship, in each layer of information propagation, a message vector between a node pair is calculated based on a node feature vector corresponding to the network node, the message vector is obtained by linearly transforming the node feature vector and passing through a nonlinear activation function, and a weighted message vector is obtained by multiplying the message vector and an association weight between the node pair; The attention mechanism is introduced to selectively aggregate the neighbor information of the network node, calculate the attention weight between the network node and its neighbor nodes, multiply the attention weight with the weighted message vector, and sum up to obtain the propagation feature of the network node in the current layer; The propagation feature of the network node is iteratively updated in multiple layers, and the message vector calculation and attention aggregation operation are repeatedly performed in each layer of information propagation to obtain the final feature representation of the network node; Based on the final feature representation, the credit conduction strength between node pairs is calculated, and the credit conduction strength is calculated by the similarity of node features and is normalized to obtain the final conduction probability.

6. The method of claim 1, wherein, According to the credit conduction strength, a dynamic credit feature is generated for each network node, and the dynamic credit feature is used to calculate the credit evaluation index value of the supply chain participant by weighting. The time information is encoded by using a triangular function with different frequencies to obtain a time encoding vector, and the time encoding vector and the historical credit record are input into a long short-term memory network to generate a memory state vector of the network node; The memory state vector and the credit conduction strength are fused to obtain the dynamic credit feature of the network node through nonlinear transformation; Based on the dynamic credit feature, the first-order influence of the network node is calculated, and the first-order influence is obtained by weighting the dynamic credit features of the direct neighbor nodes; The multi-hop influence of the network node is calculated, and the multi-hop influence is obtained by iterative calculation, and in each hop propagation, the first-order influence of the last hop and the dynamic credit feature of the current network node are combined by weighting; The multi-hop influence is adaptively aggregated to obtain the final credit evaluation index value of the supply chain participant.

7. A multi-party data collaboration and credit assessment system for supply chain finance, configured to implement the method of any one of claims 1-6, characterized in that, It comprises: The first unit is used for receiving the data collaboration request submitted by the supply chain participant, and creating a digital identity for each supply chain participant in the blockchain network according to the subject information in the data collaboration request; The second unit is used for constructing a distributed data collaboration mechanism based on the digital identity, connecting the local data nodes of the supply chain participants to the blockchain network, deploying smart contracts and privacy computing programs on each local data node, and encrypting the local data through a federated learning model to ensure that the original data does not leave the local node; The third unit is used for collecting supply chain financial data from each local data node through the privacy computing program, extracting business association relationships between supply chain participants based on the supply chain financial data, constructing a supply chain credit evaluation network according to the business association relationships, taking the supply chain participants as network nodes and the business association relationships as network connection edges, and calculating the credit conduction strength between network nodes based on a graph neural network, wherein the credit conduction strength represents the credit influence degree between the supply chain participants; The fourth unit is used for generating a dynamic credit feature for each network node according to the credit conduction strength, and calculating the credit evaluation index value of the supply chain participant by weighting the dynamic credit feature.

8. An electronic device, comprising: It comprises: A processor; A memory for storing processor-executable instructions; The processor is configured to invoke the instructions stored in the memory to execute the method in any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method in any one of claims 1 to 6.

Images