Private lp-gp privacy matching and project due diligence risk control method and system based on federated learning

Abstract

This invention discloses a method and system for privacy matching and project due diligence risk control in private equity LP-GP based on federated learning, relating to the fields of artificial intelligence and fintech. The method utilizes a horizontal federated learning framework to achieve bidirectional compatibility calculation and accurate matching of original sensitive data between LPs and GPs without leaving their local storage. It also utilizes a vertical federated learning framework to achieve cross-entity joint modeling of original data from multiple heterogeneous data holders without leaving their domain, enabling the identification of financial fraud and project risk rating. Furthermore, it uses a federated coordination center to securely aggregate encrypted parameters and iteratively optimize the global model. This invention balances LP-GP matching accuracy and project due diligence risk control accuracy without disclosing core business data and personal privacy information throughout the process, meeting the compliance requirements of the Personal Information Protection Law and the Data Security Law. It solves the technical problem of balancing privacy protection and business effectiveness in private equity investment and financing scenarios in existing technologies, and can be widely applied to the entire process management of private equity funds.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and financial technology, specifically involving a private equity LP-GP privacy matching and project due diligence risk control method and system based on federated learning. Background Technology

[0002] The core business of the private equity fund industry is divided into two major segments: LP-GP matching on the fundraising side and project due diligence and risk control on the investment side. Currently, both segments have core technical pain points that make it difficult to balance business performance and data privacy and security.

[0003] On the fundraising side, the matching process between LPs and GPs suffers from severe information asymmetry and data privacy conflicts: LPs' core information, such as funding size, cash flow cycle, and risk appetite, constitutes highly sensitive business secrets and personal privacy information; GPs' historical performance details, actual profits and losses of managed projects, and management fee sharing rules are core business assets. Neither party is willing to submit raw data to a centralized third-party platform for matching calculations. Current technologies either rely solely on coarse-grained matching based on fuzzy labels, resulting in extremely low matching accuracy and failing to meet the true needs of both parties; or they employ centralized data aggregation models, posing serious data leakage risks and violating compliance requirements of the Personal Information Protection Law and the Data Security Law.

[0004] On the investment side, the core of pre-investment due diligence for private equity projects is the identification of financial fraud. Single-company financial statements are highly susceptible to manipulation, necessitating cross-verification of data from multiple sources, including bank statements, tax returns, supply chain transactions, and legal and public opinion data, to accurately identify risks. However, this multi-source data is held by different institutions and consists of highly sensitive operational data, making it impossible to directly aggregate it onto a centralized platform for joint modeling. This results in current technology being unable to achieve cross-verification of multi-source data across entities while maintaining compliance, leading to insufficient accuracy in identifying financial fraud and posing significant investment risks.

[0005] While existing federated learning technology has been applied in scenarios such as inclusive finance, consumer credit, and wealth management product recommendation, there is still no complete technical solution for the entire scenario of LP-GP matching and project due diligence and risk control in the private equity industry. This cannot resolve the core contradiction between data privacy and business effectiveness unique to the private equity industry. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a private equity LP-GP privacy matching and project due diligence risk control method and system based on federated learning. This solves the technical problem in existing private equity investment and financing scenarios that cannot simultaneously address data privacy and security, compliance requirements, and business accuracy. Under the premise that the original data of all participants does not leave their local area, it achieves accurate two-way matching between LPs and GPs and accurate identification of project financial fraud risks.

[0007] The core technical solution of this invention is as follows: A horizontal federated learning framework is used to address the privacy protection issue in LP-GP matching scenarios; a vertical federated learning framework is used to address the multi-source data joint modeling issue in project due diligence and risk control scenarios; and a unified federated coordination center completes the secure aggregation of encrypted parameters and model iteration, without touching or storing any participant's original data throughout the process, thus achieving the dual goals of privacy compliance and business effectiveness. Specific technical solutions are as described in claims 1-8, and will not be repeated here.

[0008] Compared with the prior art, the present invention has the following significant technical and business advantages:

[0009] 1. Completely solves data privacy and compliance pain points: All participants' original sensitive data never leaves their local devices / servers, and only model parameters are transmitted in encrypted form. This fundamentally avoids the risk of data leakage and fully complies with the compliance requirements of the Personal Information Protection Law and the Data Security Law for financial data, thus solving the core technical defects of existing centralized solutions.

[0010] 2. Significantly improves LP-GP matching accuracy: Based on joint modeling using complete private feature data from both parties, the accuracy of bidirectional matching is improved by more than 15% compared to traditional coarse labeling matching, while completely protecting the core business secrets of both parties, thus solving the pain point of information asymmetry in the private equity fundraising process.

[0011] 3. Significantly improve the accuracy of financial fraud detection: Through vertical federated learning, cross-validation of multi-source data across institutions is achieved. Compared with single data source identification solutions, the recall rate of financial fraud anomalies is improved by more than 22%, while meeting the data privacy protection needs of various data holders and solving the core risk control pain point of pre-investment due diligence in private equity.

[0012] 4. The solution is highly versatile and has a low deployment threshold: Each participant only needs to deploy a lightweight local computing module and can participate in federated modeling without modifying existing business systems. It is compatible with the technical capabilities of various participants in the private equity industry and has strong feasibility.

[0013] 5. High model iteration efficiency: The federated averaging algorithm is used to achieve efficient parameter aggregation, resulting in fast model convergence and low computational and communication overhead. It can quickly iterate and optimize the model based on changes in business data, adapting to the dynamic market environment of the private equity industry. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0015] Figure 1 This is a schematic diagram of the overall architecture of the private LP-GP privacy matching and project due diligence risk control system based on federated learning described in this invention. The diagram includes a federated coordination center module, an LP-side local computing module, a GP-side local computing module, and a multi-source due diligence data-side local computing module, showing the encrypted data interaction and collaborative working relationship between the modules.

[0016] Figure 2 This is a flowchart illustrating the method for implementing LP-GP bidirectional privacy matching using horizontal federated learning as described in this invention. The flowchart includes the complete process of model initialization, feature alignment, local model training, encrypted parameter uploading, secure aggregation and global model update, iterative convergence, and matching result output.

[0017] Figure 3 This is a schematic diagram of the method for implementing due diligence and risk control of private equity projects using vertical federated learning as described in this invention. The diagram includes a complete process of sample encryption and alignment, local feature calculation and model training, encrypted gradient uploading, secure parameter aggregation, model iterative optimization, and output of financial fraud risk rating. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the present invention will be further described in detail below with reference to specific embodiments. The specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0019] Example 1: LP-GP Bidirectional Matching Implementation Based on Horizontal Federated Learning

[0020] The application scenario of this embodiment is a private equity fund service platform that connects 120 qualified LPs and 80 GP institutions. The method described in this invention achieves two-way matching under privacy protection. The specific implementation steps are as follows:

[0021] 1. Initial Deployment: The platform deploys a federated coordination center, which deploys lightweight local computing modules for all connected LP and GP terminals. The center distributes the initial bidirectional matching model (using a fully connected neural network structure) and standardized feature templates to all terminals.

[0022] 2. Feature Alignment and Preprocessing: The LP side extracts 8 core features locally, such as fund size and lock-up period acceptance, while the GP side extracts 8 core features locally, such as historical management scale and historical annualized rate of return. Both parties complete data normalization processing locally to ensure feature dimension alignment. All raw data is stored locally throughout the process and is not transmitted externally.

[0023] 3. First round of local training: The LP and GP sides train the initial model based on local private data to obtain the weight update parameters of the local model. After encrypting the parameters using the Paillier homomorphic encryption algorithm, they are uploaded to the Federal Coordination Center.

[0024] 4. Parameter Security Aggregation: The Federal Coordination Center uses the FedAvg Federal Averaging algorithm to perform a weighted average of the encrypted parameters uploaded by all LP and GP ends. The weight is the proportion of the training sample size of each end, which is used to obtain the globally updated matching model and distribute it to all LP and GP ends.

[0025] 5. Iterative optimization to convergence: Repeat the process of local training - parameter upload - aggregation and distribution, set the maximum number of iterations to 50 rounds, and the convergence threshold is when the loss function decreases by less than 0.001. In this embodiment, the convergence condition is met in the 32nd round, and the iteration stops.

[0026] 6. Matching result output: The converged global model is distributed to each terminal. The LP terminal and GP terminal complete the bidirectional adaptation score locally. Only the encrypted score results are uploaded to the federated coordination center. The center finally outputs the Top 20 LP-GP bidirectional matching ranking list to complete the matching process.

[0027] In this embodiment, no original sensitive data of either party is leaked throughout the entire process, and the matching accuracy is improved by 18.7% compared with the traditional tag matching scheme, fully meeting compliance requirements and business needs.

[0028] Example 2: Project Due Diligence Financial Fraud Detection Based on Vertical Federated Learning

[0029] The application scenario of this embodiment is a GP institution conducting pre-investment due diligence on a proposed project. The participants include five entities: the GP institution (the initiating party), the company's bank, the tax authorities, the core enterprise in the supply chain, and a judicial data service provider. The method described in this invention is used to identify the risk of financial fraud. The specific implementation steps are as follows:

[0030] 1. Roles and Initialization: Define the GP institution as the active party, holding the project's historical default risk label; banks, tax authorities, supply chain agencies, and judicial institutions are the passive parties, holding cash flow, tax returns, transaction data, and litigation data respectively; the Federal Coordination Center issues the initial vertical XGBoost risk control model and feature template.

[0031] 2. Sample Encryption and Alignment: Using the Unified Social Credit Code of the invested enterprise as the unique primary key, each participant processes the primary key through HASH encryption and uses the Diffie-Hellman algorithm to find the intersection of privacy sets to complete the sample ID alignment. During the alignment process, no non-intersecting enterprise information is disclosed, thus protecting the data privacy of each participant.

[0032] 3. First round of local training: The GP (active party) completes the financial report feature extraction locally, and each passive party completes the feature extraction of the corresponding dimension locally. Each party trains the model based on local data. The active party calculates the loss function and gradient, and each passive party assists in completing the feature cross calculation. All participants use homomorphic encryption to encrypt the gradient parameters and then upload them to the Federal Coordination Center.

[0033] 4. Parameter aggregation and update: The Federal Coordination Center securely aggregates the encrypted gradient parameters and distributes the gradient update results to each participant, who then updates the model parameters locally.

[0034] 5. Iterative optimization to convergence: Repeat the process of local training-gradient upload-aggregation and distribution, and set the maximum number of iteration rounds to 40 rounds. In this embodiment, the convergence condition is met in the 28th round, and the iteration stops.

[0035] 6. Risk Rating Output: The trained global model is sent to the GP. The GP inputs the corresponding characteristics of the proposed investment project. The model identifies three core anomalies in the financial statement revenue, bank statements, and tax return data through cross-validation of multi-source data. Finally, it outputs a high-risk financial fraud risk rating for the project, providing support for the GP's investment decision.

[0036] In this embodiment, all institutions' original operating data never leaves their local area, and the recall rate of financial fraud anomalies is 24.3% higher than that of a single financial statement identification solution, while fully meeting the data compliance requirements of each institution.

[0037] Example 3: Private Equity LP-GP Matching and Due Diligence Risk Control System Example

[0038] This embodiment provides a private LP-GP privacy matching and project due diligence risk control system based on federated learning, used to execute the methods described in Embodiments 1 and 2. The system includes:

[0039] 1. Federal Coordination Center Module: Deployed on a cloud server, configured with a 24-core CPU, 48GB of memory, and a 500GB solid-state drive, it is used to distribute the initial model and standardized feature templates, complete the secure aggregation of encrypted parameters, global model updates and distribution, and output the final matching results and risk rating results;

[0040] 2. LP Local Computing Module: Deployed on the LP's mobile / PC device, configured with a 4-core CPU, 8GB of memory, and a 200GB solid-state drive, it is used for encrypted storage of LP's private data, feature extraction, local model training, and encrypted parameter uploading. Data access permission control is set to ensure that the original data never leaves the local device.

[0041] 3. GP Local Computing Module: Deployed on the GP organization's local server, configured with an 8-core CPU, 16GB of memory, and a 200GB solid-state drive, it is used for encrypted storage of GP management data and project data, feature extraction, local model training, and encrypted parameter uploading. It is set up with intranet isolation to ensure that the original data is not leaked to the outside world.

[0042] 4. Local computing module for multi-source due diligence data: Deployed on local servers of institutions such as banks, tax bureaus, and supply chains, it is used for storing data of corresponding dimensions, extracting features, training local models, and uploading encrypted parameters. It connects with the institution's existing business system through an intranet interface, and the raw data is not transmitted across networks.

[0043] This system achieves privacy-preserving computation for two core functions—LP-GP matching and project due diligence and risk control—through modular deployment, ensuring no leakage of raw data throughout the process and adapting to the compliance requirements and business needs of the private equity industry.

[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for privacy matching and project due diligence risk control of private equity LP-GPs based on federated learning, characterized in that, Includes the following steps: The S1 Federal Coordination Center distributes initial basic models and standardized feature templates to multiple local data holders to complete feature alignment and sample alignment for each holder; the local data holders include LP holders, GP holders, and multi-source due diligence data holders. S2 is designed for LP-GP bidirectional matching scenarios. Through a horizontal federated learning framework, the LP and GP ends can complete local model training without leaving the local device with the original data. The homomorphically encrypted model update parameters are then uploaded to the federated coordination center. S3 targets due diligence and risk control scenarios for private equity projects. Through a vertical federated learning framework, multiple multi-source due diligence data terminals complete cross-terminal feature cross-calculation and local model training without the original data leaving the local server. The homomorphically encrypted gradient parameters are then uploaded to the federated coordination center. The S4 Federal Coordination Center securely aggregates the received encrypted parameters, updates the global model, and distributes it to each local data holder. After a preset round of iterative optimization, it finally outputs the LP-GP bidirectional matching and ranking results and the risk rating results for financial fraud of private equity projects.

2. The method according to claim 1, characterized in that, In step S1, the standardized feature template includes a feature set specific to the LP side, a feature set specific to the GP side, and a feature set specific to the due diligence data. The LP-specific feature set includes fund size, acceptance of fund lock-up period, cash flow repayment cycle, industry investment preference, risk tolerance threshold, tolerance for maximum historical investment drawdown, and expected annualized rate of return. The GP-specific feature set includes historical assets under management, historical annualized returns of projects, historical maximum drawdown, industry distribution of projects under management, fund duration, management fee structure, manager co-investment ratio, and historical IPO exit rate of projects. The specific feature set of due diligence data includes the investee company's financial statements, bank transaction data, tax return data, upstream and downstream supply chain transaction data, legal litigation data, and business public opinion data.

3. The method according to claim 1, characterized in that, In step S1, feature alignment and sample alignment specifically include: For LP-GP matching scenarios, based on standardized feature templates, feature dimensions of the LP side and GP side are aligned to ensure that the feature spaces of the training data of both sides are consistent. For project due diligence and risk control scenarios, the unified social credit code of the invested company is used as the unique primary key. Through HASH encryption and Diffie-Hellman key exchange algorithm, the sample IDs of various multi-source due diligence data terminals are aligned without disclosing any non-intersecting sample information during the alignment process.

4. The method according to claim 1, characterized in that, In step S2, the specific execution steps of horizontal federated learning are as follows: The S21 Federal Coordination Center distributes the initial bidirectional matching model and feature templates to all participating LP and GP ends. The S22 LP and GP ends complete feature extraction and data normalization preprocessing locally, train the initial model based on local private data, and obtain local model update parameters; The S23 LP and GP terminals use homomorphic encryption algorithms to encrypt the local model update parameters before uploading them to the Federal Coordination Center. The S24 Federal Coordination Center securely aggregates all encrypted parameters to obtain a globally updated matching model, which is then distributed to each LP and GP. S25 Repeat steps S22-S24 until the model converges. Each end completes the fitness scoring locally, and only the encrypted scoring results are uploaded to the center. Finally, the LP-GP bidirectional matching and sorting list is output.

5. The method according to claim 1, characterized in that, In step S3, the specific execution steps of vertical federated learning are as follows: The S31 Federal Coordination Center identifies the active and passive parties. The GP holding the project risk label is the active party, while the multi-source due diligence data holding characteristic data of various dimensions is the passive party. After S32 completes sample alignment, the active party and each passive party complete feature extraction and preprocessing locally, and perform local model training based on the vertical XGBoost algorithm. The active party is responsible for calculating the loss function and gradient, and the passive party assists in completing feature cross calculation. S33 The active and passive parties upload the encrypted gradient parameters to the federal coordination center, and the center distributes the gradient update results after completing parameter aggregation. S34 Repeat steps S32-S33 until the model converges. Use the trained model to perform cross-validation of multi-source data, identify anomalies in the financial data of the invested companies, and output the financial fraud risk rating results.

6. The method according to claim 1, characterized in that, In step S4, the secure aggregation uses the FedAvg algorithm to perform a weighted average of the encrypted model parameters, with the weights being the proportion of training samples held by each local data holder.

7. The method according to claim 1, characterized in that, In step S4, the convergence condition for model iteration is that the decrease in the loss function of the global model is lower than a preset threshold, or the number of iterations reaches a preset maximum number of rounds.

8. A private equity LP-GP privacy matching and project due diligence risk control system based on federated learning, characterized in that, include: The Federal Coordination Center module is used to distribute the initial model and standardized feature templates, complete the secure aggregation of encrypted parameters, update and distribute the global model, and output the final business results. The LP-side local computing module is deployed on the LP-side device and is used for storing LP private data, extracting features, training local models, and uploading encrypted parameters, without leaking the original data throughout the process. The GP-side local computing module is deployed on the GP-side server and is used for storing GP management data and project data, extracting features, training local models, and uploading encrypted parameters, without leaking the original data throughout the process. The local computing module for multi-source due diligence data is deployed on the servers of each data holder and is used for storage, feature extraction, local model training and encrypted parameter uploading of due diligence data for the corresponding dimension, without leaking the original data throughout the process. The system is used to execute the private LP-GP privacy matching and project due diligence risk control method based on federated learning as described in any one of claims 1-7.

Images