User credit risk assessment method, system and device based on multi-industry federal learning and medium

Abstract

The invention discloses a user credit risk assessment method, system, equipment and medium based on multi-industry federated learning in order to improve data security of federated learning, and belongs to the technical field of information. The method comprises the following steps: extracting operator data and bank financial data, and performing encryption preprocessing through a Hash mechanism; performing federated learning on the operator data and the bank financial data after encryption preprocessing, accessing a PSI module as a plug-in to an FATE core, converting RSA of an FATE algorithm into an elliptic curve, and performing secret sharing through a hexagonal grid cellular automaton to obtain a trained federated model; and evaluating the credit risk of the user by using the trained federal model. According to the method, the requirement of hiding a keyword of a queried object or a client ID is met, the hidden query service provides a matched query result, calculation can be carried out without going out data of both parties, and the method follows the data authentication and access process of multi-party security calculation and the overall business process. Data barriers among different industries are broken through, and deep data cooperation of multiple industries is realized.

Description

Technical Field

[0001] This invention relates to the field of information technology, and specifically to a user credit risk assessment method, system, device, and medium based on multi-industry federated learning. Background Technology

[0002] Federated learning is a distributed machine learning framework with privacy protection and secure encryption technology. Its purpose is to enable decentralized participants to collaborate on training machine learning models without disclosing their private data to other participants, so that data is available but not visible, and the model runs without the data running.

[0003] The FATE algorithm is an open-source project designed to provide a secure computing framework to support the federated AI ecosystem. Based on homomorphic encryption and multi-party computation (MPC) secure computing protocols, it supports secure computation of federated learning architectures and various machine learning algorithms. FATE's federated learning algorithms include horizontal federated learning, vertical federated learning, and transfer learning, covering mainstream industry algorithms and MPC secure computing protocols. Furthermore, FATE's underlying technical architecture includes TensorFlow / PyTorch (deep learning), EggRoll / Spark (distributed computing frameworks), and a multi-party federated communication network, while the upper layer consists of federated security protocols. It has built a federated algorithm library and provides functionalities such as federated blockchain, federated multi-cloud management, a federated model visualization platform, federated modeling pipeline scheduling, and federated online inference around practical scenarios. It aims to solve the balance between data privacy protection and data analysis computation by achieving "usable but invisible" data, thus protecting the security and privacy of raw data while completing data computation and analysis tasks.

[0004] The LightGBM (Light Gradient Boosting Machine) algorithm is a machine learning algorithm based on the gradient boosting framework, specifically designed to solve problems such as classification and regression, aiming to provide an efficient, fast, and accurate implementation of gradient boosting algorithms.

[0005] Privacy-preserving computation intersects with numerous disciplines such as cryptography, artificial intelligence, and computer hardware, forming major technical routes including secure multi-party computation, federated learning, and trusted execution environments (TEAs). Each branch of technology has a different security foundation. Secure multi-party computation, based on cryptography, performs joint computation on encrypted data, achieving provable security under a fixed security model. Federated learning, based on distributed machine learning and differential privacy, transmits only processed intermediate data during computation, protecting the privacy of all parties' data. TEAs rely on hardware and, through isolation techniques, create a secure and trusted execution space for users, providing confidentiality and integrity protection for code and data.

[0006] Some existing technologies start with two devices: each first member device has local data, and the second member device maintains the business model to be trained. The model correlation between the locally trained business model and the received business model is determined, and updates to the locally trained model are only provided to the second member device when the model correlation meets predetermined conditions. The second member device updates the business model based on the model updates received from the first member devices. Alternatively, based on the normalized gradient data and the cosine similarity between the previous aggregation result calculated interactively by the two servers, the aggregated gradient values ​​are returned to the client. The client calculates the total aggregated gradient value based on these aggregated gradient values. This interactive calculation between the two servers improves data security and data processing efficiency in existing federated learning systems.

[0007] Some existing technologies start from a central server and each client. Each client uses local data to train locally based on the global model parameters to obtain local model update parameters. Each client uses secret calculation to generate a mask and uploads the local model update parameters, which have been processed with their own mask for privacy protection, to the central server. The central server aggregates all global model update parameters.

[0008] Some existing technologies perform computation through a two-step encryption method: training the federated learning model using virtual datasets held by at least two virtual participants to obtain the first parameter for each virtual participant; aggregating and calculating the first parameters of each virtual participant to obtain the second parameter; encrypting the second parameter based on an encryption protocol to obtain the parameter ciphertext, and sending the parameter ciphertext to the collaborating party.

[0009] Some existing technologies involve dual-device serial computing, or a central device receiving and aggregating data parameters from surrounding clients, or performing multi-step encrypted aggregation.

[0010] Existing technologies rarely focus on improvements to federated learning technical frameworks, built-in operators, and model return values.

[0011] With increasingly stringent data privacy regulations, data sharing between parties is often impossible. For example, while "XX operator" possesses a wealth of user attribute and behavioral data, such as location tracking, browsing logs, search history, and app usage records, "XX bank" struggles to access this information, resulting in data fragmentation. This decline in data quality directly impacts the predictive accuracy of credit scores, particularly affecting new products and users, significantly reducing service quality. Therefore, the market urgently needs multi-industry federated learning modeling methods to break down data barriers between different sectors and achieve deep data collaboration across multiple industries. Summary of the Invention

[0012] This invention addresses the shortcomings of existing technologies by providing a user credit risk assessment method, system, device, and medium based on multi-industry federated learning.

[0013] To achieve the above objectives, the present invention adopts the following technical solution:

[0014] A user credit risk assessment method based on multi-industry federated learning includes the following steps:

[0015] Extract operator data and bank financial data, and perform encrypted preprocessing using a hash mechanism;

[0016] Federated learning is performed on encrypted preprocessed operator data and bank financial data to obtain a trained federated model.

[0017] Use a trained federated model to assess user credit risk.

[0018] To optimize the above technical solution, the specific measures also include:

[0019] Furthermore, the hash mechanism specifically uses a random key and asymmetric encryption.

[0020] Furthermore, the federated learning of the encrypted preprocessed operator data and banking financial data specifically involves using the FATE framework for vertical federated learning, as detailed below:

[0021] The encrypted preprocessed operator data and banking financial data are vertically segmented. The PSI module of the FATE framework is used to find the intersection of the user IDs of the encrypted preprocessed operator data and banking financial data. The data with the same user ID but different user characteristics are extracted as the training dataset for vertical federated learning. The LightGBM algorithm is used to build a federated model. The federated model performs vertical federated learning on the training dataset and secret sharing is carried out through a hexagonal grid cellular automaton.

[0022] Furthermore, the PSI module uses an elliptic curve cryptography algorithm to find the intersection of the user IDs in the encrypted preprocessed operator data and the bank financial data, specifically:

[0023] The operator and the bank negotiated an elliptic curve group;

[0024] Each party generates a random private key for each element in its own dataset, and then calculates the corresponding public key;

[0025] Each party maps the elements in the dataset to points on an elliptic curve, and then encrypts these points using a private key;

[0026] The two parties exchange encrypted point sets;

[0027] Both parties then encrypt the point set sent by the other party again using their private keys;

[0028] The two parties compare the two encrypted point sets and find the same points. The original elements corresponding to these same points are the intersection.

[0029] Furthermore, the federated model is a gradient boosting decision tree model. The federated model performs vertical federated learning on the training dataset, and secret sharing is specifically achieved through a hexagonal grid cellular automaton:

[0030] All participants are deployed on a hexagonal grid of nodes, each with 6 neighbors. Participants include operators and banks. One party is designated as the active party and the other as the passive party, as needed. Each node receives samples from the training dataset. The active party node holds the labels, and the passive party node holds the feature data. The active party node calculates the gradient g and the second-order partial derivative h of the sample. The gradient g and the second-order partial derivative h are homomorphically encrypted into encrypted gradient Enc(g) and encrypted second-order partial derivative Enc(h). The active party node sends the encrypted gradient Enc(g) and encrypted second-order partial derivative Enc(h) to its 6 direct neighbors. The neighboring nodes forward the data to their non-overlapping neighbors until it reaches the target passive party node.

[0031] The passive node aggregates the encryption gradient Enc(g) and the encryption second-order partial derivative Enc(h) to obtain the encryption gradient histogram Enc(G) and the encryption second-order partial derivative histogram Enc(H);

[0032] The passive node passes the encrypted gradient histogram Enc(G) and the encrypted second-order partial derivative histogram Enc(H) in reverse to its neighbors, and finally to the active node.

[0033] The active node decrypts the histogram using its private key and calculates the feature split gain for each candidate split node based on the decrypted histogram. The candidate split node with the largest feature split gain is the optimal split node. The optimal split node divides the sample into left and right child nodes according to the sample feature values ​​of the node, and grows the gradient boosting decision tree according to the leaf-wise strategy until the set depth is reached.

[0034] This invention also proposes a user credit risk assessment system based on multi-industry federated learning, comprising:

[0035] The data preprocessing module is used to extract operator data and bank financial data and perform encrypted preprocessing through a hash mechanism;

[0036] The federated learning module is used to perform federated learning on encrypted preprocessed operator data and banking financial data to obtain a trained federated model.

[0037] The risk assessment module is used to assess user credit risk using a trained federated model.

[0038] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the user credit risk assessment method based on multi-industry federated learning as described above.

[0039] The present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute the user credit risk assessment method based on multi-industry federated learning as described above.

[0040] The beneficial effects of this invention are:

[0041] By integrating the high-performance PSI module from the open-source project Hidden Language computation core into the Fate core, a new federated learning platform framework was designed, significantly improving the overall performance of the Fate algorithm in federated learning. This approach combines the strengths of both the Fate algorithm and the open-source project Hidden Language. The Fate algorithm excels in federated learning, while Hidden Language excels in secure intersection calculations; their integration allows for a close synergy of their technical strengths.

[0042] When performing secure intersection using the Fate algorithm, the RSA encryption operator is replaced with an elliptic curve, which speeds up the computational process.

[0043] Based on cryptographic technology, the data service provider retains control over the data resources, and the querying party no longer uses plaintext queries. Instead, a random key is added to the query input parameters to ensure that no query traces are left and only the query results are obtained.

[0044] The secret sharing technique based on the principle of two-dimensional hexagonal grid cellular automata protects intermediate and transformation results. It does not use triangular or quadrilateral grid cellular automata, nor does it utilize polynomial-based secret sharing techniques such as homogeneous linear regression, making the model more realistic and natural.

[0045] The results not only return the PMML model file of the algorithm, but also the features and their quantities, as well as the probability value distribution file, which facilitates development and testing. Model developers can quickly develop the required features based on the number of features and the probability value distribution, and flexibly adjust the threshold by combining the number of probability values ​​and the distribution of the evaluation function. They are no longer limited to the probability value distribution of the test data, and can also broadly compare the evaluation function distribution of the training data.

[0046] It streamlines the entire process of federated learning, privacy-preserving computation, and mathematical modeling, with performance improvements at each stage. It not only focuses on the technical architecture of federated learning and privacy-preserving computation but also emphasizes the ease of use of mathematical modeling return values ​​from a practical application perspective. Attached Figure Description

[0047] Figure 1 This is a flowchart of the user credit risk assessment method based on multi-industry federated learning proposed in this invention.

[0048] Figure 2 A schematic diagram of secret sharing for a hexagonal grid cellular automaton. Detailed Implementation

[0049] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0050] Example 1

[0051] This invention proposes a user credit risk assessment method based on multi-industry federated learning. The overall process of this method is as follows: Figure 1 As shown, it includes the following steps:

[0052] Data from telecom operators and banks is extracted and pre-processed using a hash mechanism; the hash mechanism specifically uses a random key and asymmetric encryption.

[0053] The types of operator data characteristics include basic information, consumption behavior, social networks, and device information. Basic information includes, for example, network access duration, package type, and real-name authentication status; consumption behavior includes, for example, average monthly call charges, fluctuations in data usage, and number of international roaming calls; social networks include, for example, stability of the call circle (number of frequently contacted individuals) and abnormal call behavior; and device information includes, for example, frequency of mobile phone model changes and multiple device logins.

[0054] Bank financial data characteristics include account behavior, credit history, consumption habits, and risk labels. Account behavior includes monthly average balance, transaction frequency, and proportion of large transfers; credit history includes the number of overdue payments, loan limit utilization rate, and repayment timeliness; consumption habits include credit card spending categories (dining / luxury goods) and the proportion of online payments; and risk labels include historical default records.

[0055] Federated learning is performed on the encrypted preprocessed carrier data and banking financial data to obtain a trained federated model. The specific process of federated learning on the encrypted preprocessed carrier data and banking financial data is as follows: Vertical federated learning is performed using the FATE framework, as detailed below:

[0056] Because there is significant overlap between users across different industries but minimal overlap in user features, the encrypted preprocessed operator data and banking / financial data are vertically segmented. The PSI (Private Set Intersection) module of the FATE framework is used to find the intersection of user IDs from the encrypted preprocessed operator data and banking / financial data. Data with the same user IDs but different user features from both sides is extracted as the training dataset for vertical federated learning. In scenarios with large datasets, FATE's secure intersection performance is poor, resulting in very high overall processing time. To improve the overall performance of the FATE algorithm in federated learning, a new framework is designed. The high-performance PSI module of the open-source project Hidden Language Computation is extracted and integrated as a plugin into the FATE core. The effect is that FATE first activates Hidden Language's high-performance PSI core to complete the data intersection when performing intersection calculations, thus forming an innovative high-performance federated learning platform framework.

[0057] The FATE framework provides a multi-party secure sample alignment algorithm based on RSA encryption and hash functions. However, RSA encryption is too performance-intensive. Replacing finite field mathematical operations (RSA) with elliptic curve cryptography (ECC) can improve computational efficiency and performance, while still allowing secure feature crossing and model training. The PSI module uses elliptic curve cryptography to find the intersection of user IDs from encrypted preprocessed carrier data and banking financial data, specifically:

[0058] The operator and the bank negotiated an elliptic curve group;

[0059] Each party generates a random private key for each element in its own dataset, and then calculates the corresponding public key;

[0060] Each party maps the elements in the dataset to points on an elliptic curve, and then encrypts these points using a private key;

[0061] The two parties exchange encrypted point sets;

[0062] Both parties then encrypt the point set sent by the other party again using their private keys;

[0063] The two parties compare the two encrypted point sets and find the same points. The original elements corresponding to these same points are the intersection.

[0064] A federated model is constructed using the LightGBM algorithm. The federated model is a gradient boosting decision tree model. The federated model performs longitudinal federated learning on the training dataset and uses a hexagonal grid cellular automaton for secret sharing.

[0065] In the training phase of the federated model, instead of using polynomial-based secret sharing techniques such as homogeneous linear regression, a trustless multi-secret sharing scheme based on cellular automata principles is employed. Unlike typical cellular automata-based multi-secret sharing schemes, this scheme uses a two-dimensional hexagonal grid to construct the cellular automata. Hexagonal grid cellular automata are a grid-based simulation method and discrete model that can simulate various physical, chemical, and biological phenomena. While the model is somewhat complex and difficult to represent, it can better simulate isotropic phenomena, resulting in a more natural and realistic representation. In the cellular space, the changes in each cell according to the state update rules are synchronous, making it particularly suitable for parallel computing. Furthermore, the state changes of each cell are independent and do not affect each other. The parallel computing capabilities of hexagonal grid cellular automata accelerate the computational tasks in the secret sharing process. The hexagonal grid cellular automata-based method eliminates the need for a distributor in the distribution of secret shares, enabling the distribution of secret shares even without a distributor. From the perspective of the cellular automata method itself, it has higher computational efficiency than the current mainstream polynomial-based methods, which makes this approach more widely used.

[0066] The federated model performs longitudinal federated learning on the training dataset, using a hexagonal grid cellular automaton for secret sharing, such as... Figure 2 Specifically:

[0067] All participants are deployed on a hexagonal grid of nodes, each with 6 neighbors. Participants include operators and banks. One party is designated as the active party and the other as the passive party, as needed. Each node receives samples from the training dataset. The active party node holds the labels, and the passive party node holds the feature data. The active party node calculates the gradient g and the second-order partial derivative h of the sample. The gradient g and the second-order partial derivative h are homomorphically encrypted into encrypted gradient Enc(g) and encrypted second-order partial derivative Enc(h). The active party node sends the encrypted gradient Enc(g) and encrypted second-order partial derivative Enc(h) to its 6 direct neighbors. The neighboring nodes forward the data to their non-overlapping neighbors until it reaches the target passive party node.

[0068] The passive node aggregates the encryption gradient Enc(g) and the encryption second-order partial derivative Enc(h) to obtain the encryption gradient histogram Enc(G) and the encryption second-order partial derivative histogram Enc(H);

[0069] The passive node passes the encrypted gradient histogram Enc(G) and the encrypted second-order partial derivative histogram Enc(H) in reverse to its neighbors, and finally to the active node.

[0070] The active node decrypts the histogram using its private key and calculates the feature split gain for each candidate split node based on the decrypted histogram. The candidate split node with the largest feature split gain is the optimal split node. The optimal split node divides the sample into left and right child nodes according to the sample feature values ​​of the node, and grows the gradient boosting decision tree according to the leaf-wise strategy until the set depth is reached.

[0071] In practical engineering, after building a federated model using the LightGBM algorithm, a PMML file, a TXT file showing the final features used by the model, their quantities, and the probability distribution of the predicted results will be generated. Because the number of leaf nodes (num_leaves) and the maximum depth (max_depth) of the gradient boosting decision tree model can vary, the final number of features used by the model will also differ. Therefore, after generating the model, a TXT file showing the final number of features used by the model, sorted by feature importance, is also generated. Features with an importance of 0 are not displayed in the results, allowing subsequent developers and testers to intuitively understand which features the model ultimately used and their importance.

[0072] The probability distribution of the predicted results includes the probability values ​​of the model's predictions, the corresponding number of predictions, and the distributions of F1, precision, and recall for different probability thresholds. A limited probability distribution can help developers quickly identify model underfitting. The distributions of the F1, precision, and recall evaluation functions for different probability thresholds facilitate subsequent selection of the result thresholds.

[0073] Table 1 shows examples of the features and their number used in the final model:

[0074] Table 1

[0075]

[0076]

[0077] The probability distribution of the prediction results is shown in Table 2:

[0078] Table 2

[0079]

[0080]

[0081] Table 3 shows examples of the distributions of F1, precision, and recall under different probability values:

[0082] Evaluation function Evaluation function corresponding to probability values ​​of 0-0.9 F1 f1[0.003,0.039,0.041,0.042,0.042,0.043,0.086,0.0,0.0,0.0] accuracy pr[0.001,0.02,0.021,0.022,0.022,0.022,0.047,0.0,0.0,0.0] Recall rate rc[1.0,0.986,0.972,0.972,0.972,0.944,0.493,0.0,0.0,0.0]

[0083] Use a trained federated model to assess user credit risk.

[0084] Operators can use the risk assessment results from the federal model to determine the user's risk of non-payment, churn, fraud, etc., while banks can refer to the risk assessment results from the federal model to decide whether to grant loans or other services to users.

[0085] Example 2

[0086] This invention proposes a user credit risk assessment system based on multi-industry federated learning, corresponding to the method in Embodiment 1, comprising:

[0087] The data preprocessing module is used to extract operator data and bank financial data and perform encrypted preprocessing through a hash mechanism;

[0088] The federated learning module is used to perform federated learning on encrypted preprocessed operator data and banking financial data to obtain a trained federated model.

[0089] The risk assessment module is used to assess user credit risk using a trained federated model.

[0090] The implementation methods of each module and its function in the system are completely consistent with the steps of the method in Implementation Example 1, so they will not be repeated here.

[0091] Example 3

[0092] This invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the user credit risk assessment method based on multi-industry federated learning as described in Embodiment 1.

[0093] Example 4

[0094] This invention proposes a computer-readable storage medium storing a computer program that causes a computer to execute a user credit risk assessment method based on multi-industry federated learning as described in Embodiment 1.

[0095] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0096] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principle of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A user credit risk assessment method based on multi-industry federated learning, characterized in that, Includes the following steps: Extract operator data and bank financial data, and perform encrypted preprocessing using a hash mechanism; Federated learning is performed on encrypted preprocessed operator data and bank financial data to obtain a trained federated model. Use a trained federated model to assess user credit risk.

2. The user credit risk assessment method based on multi-industry federated learning as described in claim 1, characterized in that, The hash mechanism specifically uses a random key and asymmetric encryption.

3. The user credit risk assessment method based on multi-industry federated learning as described in claim 1, characterized in that, The federated learning of the encrypted preprocessed operator data and banking financial data specifically involves using the FATE framework for vertical federated learning, as detailed below: The encrypted preprocessed operator data and banking financial data are vertically segmented. The PSI module of the FATE framework is used to find the intersection of the user IDs of the encrypted preprocessed operator data and banking financial data. The data with the same user ID but different user characteristics are extracted as the training dataset for vertical federated learning. The LightGBM algorithm is used to build a federated model. The federated model performs vertical federated learning on the training dataset and secret sharing is carried out through a hexagonal grid cellular automaton.

4. The user credit risk assessment method based on multi-industry federated learning as described in claim 3, characterized in that, The PSI module uses an elliptic curve cryptography algorithm to find the intersection of the user IDs in the encrypted preprocessed operator data and the bank financial data, specifically: The operator and the bank negotiated an elliptic curve group; Each party generates a random private key for each element in its own dataset, and then calculates the corresponding public key; Each party maps the elements in the dataset to points on an elliptic curve, and then encrypts these points using a private key; The two parties exchange encrypted point sets; Both parties then encrypt the point set sent by the other party again using their private keys; The two parties compare the two encrypted point sets and find the same points. The original elements corresponding to these same points are the intersection.

5. The user credit risk assessment method based on multi-industry federated learning as described in claim 3, characterized in that, The federated model is a gradient boosting decision tree model. The federated model performs vertical federated learning on the training dataset, and secret sharing is specifically achieved through a hexagonal grid cellular automata: All participants are deployed on a hexagonal grid of nodes, each with 6 neighbors. Participants include operators and banks. One party is designated as the active party and the other as the passive party, as needed. Each node receives samples from the training dataset. The active party node holds the labels, and the passive party node holds the feature data. The active party node calculates the gradient g and the second-order partial derivative h of the sample. The gradient g and the second-order partial derivative h are homomorphically encrypted into encrypted gradient Enc(g) and encrypted second-order partial derivative Enc(h). The active party node sends the encrypted gradient Enc(g) and encrypted second-order partial derivative Enc(h) to its 6 direct neighbors. The neighboring nodes forward the data to their non-overlapping neighbors until it reaches the target passive party node. The passive node aggregates the encryption gradient Enc(g) and the encryption second-order partial derivative Enc(h) to obtain the encryption gradient histogram Enc(G) and the encryption second-order partial derivative histogram Enc(H); The passive node passes the encrypted gradient histogram Enc(G) and the encrypted second-order partial derivative histogram Enc(H) in reverse to its neighbors, and finally to the active node. The active node decrypts the histogram using its private key and calculates the feature split gain for each candidate split node based on the decrypted histogram. The candidate split node with the largest feature split gain is the optimal split node. The optimal split node divides the sample into left and right child nodes according to the sample feature values ​​of the node, and grows the gradient boosting decision tree according to the leaf-wise strategy until the set depth is reached.

6. A user credit risk assessment system based on multi-industry federated learning, characterized in that, include: The data preprocessing module is used to extract operator data and bank financial data and perform encrypted preprocessing through a hash mechanism; The federated learning module is used to perform federated learning on encrypted preprocessed operator data and banking financial data to obtain a trained federated model. The risk assessment module is used to assess user credit risk using a trained federated model.

7. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the user credit risk assessment method based on multi-industry federated learning as described in any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to execute the user credit risk assessment method based on multi-industry federated learning as described in any one of claims 1-5.

Images