Privacy-preserving distributed matching system using federated learning and secure computation

The system addresses legal and security issues in business matching by using federated learning and secure computation to protect sensitive data locally, enabling global matching with privacy preservation and compliance with data regulations.

KR103044791B1Active Publication Date: 2026-11-10TEAM ZERO CODE CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
KR1020260069654
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2026-04-17
Publication Date
2026-11-10
Estimated Expiration
2046-04-17

AI Technical Summary

Technical Problem

Existing business matching systems face legal, security, and participation challenges due to centralized data collection, which violates data sovereignty laws, exposes sensitive data to security risks, and is often rejected by companies with data transfer policies.

Method used

A privacy-preserving distributed matching system using federated learning and secure computation, where local nodes store sensitive data and generate statistical summaries, applying homomorphic encryption and differential privacy to protect data, and transmit encrypted updates for global model training without centralization.

Benefits of technology

Enables global business matching while ensuring data privacy, maintaining trade secrets, and allowing participation from countries with data export restrictions by using hybrid protection techniques and policy management.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112026047135518-PAT00001_ABST
    Figure 112026047135518-PAT00001_ABST
Patent Text Reader

Abstract

The present invention relates to a privacy-preserving distributed matching system utilizing federated learning and secure computation that realizes business matching between global enterprises while storing sensitive original data locally, characterized by comprising a plurality of local nodes each including a local data storage unit for storing sensitive original data, a local profile generation module for generating a local profile in the form of a statistical summary vector, a local matching feature calculation module for generating a feature vector for input to a matching algorithm, a local learning module for calculating parameter updates using a federated learning client method, a local encryption / noise module for selectively applying homomorphic encryption or differential privacy noise, and a local communication module for transmitting only encrypted values; a model management module for managing and redistributing global matching model parameters, an update aggregation module for aggregating encrypted updates, a secure computation module for calculating similarity without decryption, a matching score calculation module for calculating matching scores and providing them to corporate user terminals or administrator terminals, a policy / regulation management module for managing country-specific policy profiles and controlling the allowable range of each module, and a central coordination server including a central data storage unit and a central communication module.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] The present invention relates to business matching technology between companies or institutions, and more specifically, to a privacy-preserving distributed matching system using federated learning and secure computation that performs global business matching by combining federated learning and secure computation or differential privacy, wherein sensitive data distributed by country or company is not directly collected on a central server, but only local profile and model updates necessary for matching are shared while protecting sensitive data at each local node. Background Technology

[0003] Global business-to-business (B2B) matching platforms, supply chain networks, and investment matching systems function as core infrastructure in various economic activities of companies today, such as partner identification, procurement, and import / export transactions.

[0004] However, these existing platforms generally adopt a centralized data collection structure, employing a method where highly sensitive sales data—such as transaction history, unit prices, profit margins, financial information, inventory status, and supply and demand plans—is directly collected from various countries and companies on a central server, and then used to train matching algorithms or calculate matching scores.

[0005] This centralized structure contains three fundamental problems.

[0006] First is the legal risk associated with cross-border data transfer. In the process of physically transferring data from one country to a data center located in another to provide global services, there is a significant legal burden of complying with the European Union's General Data Protection Regulation (GDPR), national data sovereignty laws, and regulations on the export of data.

[0007] In particular, as many countries, including China, Russia, India, and Indonesia, have recently been tightening data localization regulations, it is practically impossible for companies or public institutions in these countries to participate in centralized matching platforms.

[0008] Second is the security risk of sensitive sales data. Since unit prices, profit margins, transaction terms, and financial indicators constitute core trade secrets, a security breach in the central system can result in irreversible commercial damage to the company.

[0009] Third, there is the issue of refusal to participate in platforms. Since many companies and public institutions have policies prohibiting the external transfer of sensitive data, situations frequently arise where the introduction of centralized matching systems is rejected at the very least.

[0011] Meanwhile, Federated Learning is a distributed learning method that keeps data from individual devices or institutions local and centrally collects only model parameter updates for training; it has already been proposed in specific fields such as smartphone keyboard prediction and medical data analysis.

[0012] However, there are very few cases proposing an integrated global matching system architecture that applies federated learning to the practical task of business matching between enterprises, incorporating local profile summary generation, federated learning-based matching model training, the integrated application of homomorphic encryption and differential privacy, and a policy-based control structure usable even in countries with data export regulations.

[0014] In this regard, examples of inventions include Registered Patent No. 10-2871200 (hereinafter Prior Art 1), Registered Patent No. 10-2789827 (hereinafter Prior Art 2), and Published Patent No. 10-2025-0154311 (hereinafter Prior Art 3).

[0016] Prior Art 1 discloses a technology that manages user profiles by dividing them into fragments based on a Multi-Party Computation (MPC) system and infers prediction labels for user profiles using a Gradient Boosting machine learning technique.

[0017] Prior Art 1 is significant in that it proposes a structure that enables individual computing systems to perform predictive inference without viewing the entire profile in a multi-party computing environment.

[0018] However, Prior Art 1 is a technology specialized in user profile inference in the advertising and recommendation system domain, and its application domain and system purpose are essentially different from the technical challenge of business matching between companies.

[0019] Furthermore, Prior Art 1 does not adopt a federated learning method and thus lacks a distributed learning structure in which multiple distributed enterprise nodes calculate gradients from their respective local data and share only those updates; it also has fundamental limitations in that it completely lacks the generation of business metric-based profile summaries at local nodes, integration with homomorphic encryption or differential privacy, and a policy-based control structure to address national regulations on data export.

[0021] Prior Art 2 discloses a privacy labeling technology that divides user profiles into multiple shares in a multi-party computing (MPC) system and identifies the labels of the clusters to which each profile belongs using a machine learning-based clustering model.

[0022] Prior Art 2 belongs to a similar technological lineage to Prior Art 1 in that it proposes a structure capable of label classification without fully exposing user profiles to individual computing systems.

[0023] However, Prior Art 2 is also a technology specialized in user labeling in the advertising and recommendation system domains, and is unrelated to business matching issues between companies.

[0024] Furthermore, Prior Art 2 has fundamental technical limitations as it does not adopt a federated learning architecture and does not include a local profile generation structure based on practical business indicators such as corporate unit price, margin rate, production capacity, and lead time, a policy management function for responding to data export regulations, or a calculation structure based on safe computation of matching scores.

[0026] Prior Art 3 discloses an education and public policy resource allocation platform that integrates learner, provider, and policy data in the fields of education and public policy, calculates matching scores and rewards through graph neural networks (GNN) and reinforcement learning (RL), converts this into a binary variable optimization problem (QUBO), and searches for an optimal solution through a quantum approximation optimization algorithm.

[0027] Prior Art 3 reflects a certain interest in privacy protection in that it includes a configuration for injecting differential privacy noise during the data preprocessing stage.

[0028] However, the matching scope of Prior Art 3 is limited to resource allocation in the fields of education and public policy, and is essentially different from the technology structure for commercial business matching between global companies in all aspects, including application area, technological purpose, and system architecture.

[0029] In particular, prior art 3 adopts a structure that processes data centrally, so it cannot support global participation in countries with regulations on data export, and it lacks a distributed architecture in which multiple enterprise nodes store their respective sensitive data locally and jointly train a global matching model using a federated learning method.

[0030] Furthermore, the differential privacy of Prior Art 3 is limited to the input preprocessing stage, and is clearly distinguished in terms of technical depth and scope from a structure that integrates and applies homomorphic encryption, multi-party computation (MPC), and differential privacy to each stage of local profile generation, encryption, secure aggregation, and matching score calculation throughout the communication rounds of federated learning.

[0032] As such, existing prior art technologies fail to provide a distributed privacy preservation structure suitable for the practical task of business matching between enterprises. Specifically, a system integrating a local profile generation structure that stores sensitive original data in each company's local system while encrypting and providing only statistical summary profiles externally, a structure that trains high-quality global matching models without centralization through federated learning, a hybrid secure computation structure that integrates homomorphic encryption and differential privacy into the matching score calculation process, and a policy-based control structure that manages national data export regulations and differential privacy parameters has not yet been sufficiently proposed. Prior art literature

[0034] Registered Patent No. 10-2871200 Registered Patent No. 10-2789827 Published Patent No. 10-2025-0154311 The problem to be solved

[0035] The present invention was developed to improve upon the aforementioned problems and aims to provide a privacy-preserving distributed matching system utilizing federated learning and secure computation, which enables global inter-company business matching without exporting sensitive original data at all, by having each of multiple local nodes at the corporate or institutional level store sensitive original data, including transaction history, unit price, margin rate, financial information, inventory status, production capacity, quality grade, and lead time, only within their own systems; converting item-specific statistics regarding price range distribution, production capacity, quality indicators, and lead time distribution into local profiles in the form of statistical summary vectors instead of the original data; protecting these profiles by applying homomorphic encryption or differential privacy noise according to sensitivity levels; and transmitting them to a central coordination server.

[0036] Furthermore, the present invention aims to provide a privacy-preserving distributed matching system using federated learning and secure computation that enables the continuous improvement of a high-quality global business matching model reflecting the matching experiences of various regions and companies without centralizing the original data, by repeating the process over multiple communication rounds in which each of a plurality of local nodes receives global matching model parameters distributed from a central coordination server, calculates gradient or parameter updates using its own local data, encrypts or noiseizes and transmits them, and the update aggregation module of the central coordination server combines the received encrypted updates using a federated learning aggregation algorithm to calculate new global matching model parameters.

[0037] Furthermore, the present invention aims to provide a privacy-preserving distributed matching system using federated learning and secure computation, which enables accurate and reliable business matching results without exposing sensitive data of individual companies to the outside, by having a secure computation module of a central coordination server perform vector-to-vector inner product, cosine similarity, or Euclidean distance calculations on encrypted local profile vectors without decryption using homomorphic encryption-based computation or multi-party computation (MPC)-based distributed sum computation, and having a matching score calculation module calculate a matching score between pairs of companies by combining the secure computation results with global model parameters, and then applying differential privacy noise as a post-processing step to prevent back-inference.

[0038] Meanwhile, the present invention aims to provide a privacy-preserving distributed matching system using federated learning and secure computation that enables the realization of an optimal balance between the level of privacy protection and matching accuracy by performing field-specific hybrid protection processing for each data field of a local profile, wherein differential privacy noise is applied to the average value in the unit price and margin rate fields, the result of a homomorphic encryption-based average calculation is applied to the lead time field, and only low-, medium-, and high-priced range information is selectively applied to the categorical field including risk preference.

[0039] Furthermore, the present invention aims to provide a privacy-preserving distributed matching system using federated learning and secure computation, which enables companies or public institutions in countries with data export restrictions to participate in global business matching while storing sensitive data only on their domestic servers. This is achieved by having a policy / regulation management module integrate and store a policy profile including country-specific or institution-specific data export regulation information, data localization policies, and differential privacy parameters, and by integrally controlling the data processing allowance range and computation allowance level of the local encryption / noise module and local communication module within the local node, as well as the update aggregation module, secure computation module, and matching score calculation module within the central coordination server. means of solving the problem

[0041] To achieve the above objectives, the present invention comprises a privacy-preserving distributed matching system using federated learning and secure computation, wherein the system includes: a plurality of local nodes, each operated at the level of an enterprise or institution and storing sensitive original data independently; and a central coordination server connected to the plurality of local nodes via a secure channel through a network communication unit, which oversees the learning of a global matching model and the calculation of matching scores between enterprises; wherein each of the plurality of local nodes comprises: a local data storage unit that stores sensitive original data including transaction history, unit price, margin rate, financial information, inventory status, production capacity, quality grade, and lead time; and a local profile generation module that calculates item-specific statistics from the sensitive original data stored in the local data storage unit, including average unit price, unit price standard deviation, and quantile regarding price range distribution; maximum production volume, average production volume, and production volume variability regarding production capacity; defect rate and certification level regarding quality indicators; and average delivery time and delay probability regarding lead time distribution, and converts these into a local profile in the form of a real value vector or a structured statistical summary vector including category characteristics. A local matching feature calculation module that generates a fixed-length feature vector or a structured feature set to be input into a matching algorithm by additionally combining derivative indicators, including risk preference categories and transaction stability indicators, with statistics received from the local profile generation module; and a local learning module of a federated learning client method that receives global matching model parameters distributed from the central coordination server and calculates a gradient or parameter update for the global matching model using past matching results, transaction success status, and satisfaction data stored in the local data storage unit.A local encryption / noise module that generates only encrypted or noise-processed values ​​through hybrid protection processing, wherein, for the local profiles and feature vectors generated by the local profile generation module and the local matching feature calculation module and the gradients or parameter updates calculated by the local learning module, homomorphic encryption or differential privacy noise including Laplace noise or Gaussian noise is selectively applied to high-sensitivity unit price, margin rate, and financial indicator fields, low, medium, and high-sensitivity bucketization processing is performed on medium-sensitivity fields, and masking or lightweight processing is performed on non-sensitive fields; and includes a local communication module that transmits only encrypted or noise-processed local profiles and parameter updates received from the local encryption / noise module to the central coordination server via the network communication unit, while not transmitting sensitive original data, individual contract conditions, and detailed unit price / margin information stored in the local data storage unit outside the local system at all; and the central coordination server includes a model management module that maintains and manages parameters of a global matching model including a graph embedding model or a matching score prediction model, updates global parameters based on aggregation results received from an update aggregation module, and repeatedly redistributes the updated global parameters to each of the plurality of local nodes over a plurality of communication rounds; and an update aggregation module that calculates new global matching model parameters by combining encrypted gradients or parameter updates received from the local communication modules of each of the plurality of local nodes according to an aggregation algorithm including a simple average (FedAvg), a weighted average proportional to the data size, or secure aggregation.A secure computation module that performs calculations of inner product, cosine similarity, or Euclidean distance between vectors without decryption for encrypted local profile vectors received from the plurality of local nodes, using homomorphic encryption-based computation or multi-party computation (MPC)-based distributed sum computation; and a matching score calculation module that calculates a matching score between pairs of companies by combining the similarity computation result received from the secure computation module with global matching model parameters of the model management module, protects individual companies' sensitive data from being reverse-inferred by applying additional differential privacy noise as post-processing to the calculated matching score or ranking, and provides a matching result including a top partner candidate list and similarity and risk indicators to a corporate user terminal or an administrator terminal. A privacy-preserving distributed matching system using federated learning and secure computation can be provided, characterized by comprising: a policy / regulation management module that stores and manages a policy profile including information on data export regulations by country and institution, data localization policies, differential privacy parameters, and encryption levels, controls the data processing allowance range and computation allowance level of local encryption / noise modules and local communication modules within the plurality of local nodes, and update aggregation modules, secure computation modules, and matching score calculation modules within the central coordination server, and records all computation execution history and parameter change history as an audit log; a central data storage unit that stores non-sensitive data including matching results received from the matching score calculation module, global model parameters, aggregated statistics, and policy setting values; and a central communication module that receives encrypted local profile and parameter updates from the local communication modules of each of the plurality of local nodes and transmits updated global matching model parameters to the plurality of local nodes according to the instructions of the model management module.

[0042] Here, the local profile generation module configures the local profile in the form of a real value vector having elements of average unit price, unit price volatility, maximum production volume, average lead time, lead time volatility, average margin rate, and quality defect rate, and the local encryption / noise module performs field-specific selective protection processing for each field of the local profile, providing an average value with differential privacy noise applied to the unit price and margin rate fields, applying a homomorphic encryption-based average calculation to the lead time field, and providing only low / medium / high price range information to the categorical field including risk preference.

[0043] At this time, the policy / regulation management module stores a policy profile for a local node corresponding to a specific country or institution, including a ban on data export abroad, exclusive exchange of encrypted profiles through a designated path, and country-specific upper limits for differential privacy parameters, and applies it to the local node; however, for a local node operated by a public institution, it only allows local learning on corporate data of the country and transmission of encrypted parameter updates, and restricts the data movement path and the scope of permitted operations so as to receive only matching scores with domestic companies and a list of top partner candidates from the matching score calculation module. Effects of the invention

[0045] According to the present invention with the above-described configuration, the following effects can be achieved.

[0046] First, the present invention has the distinctive advantage of enabling global business matching between companies while absolutely protecting core trade secrets such as transaction history, unit prices, profit margins, and financial information, by generating only encrypted or noise-processed values ​​through hybrid protection processing in which the local encryption / noise module of each local node selectively performs homomorphic encryption, differential privacy noise addition, and segmentation processing according to the sensitivity level, and by having the local communication module transmit only the protected values ​​while not exporting any sensitive original data outside the local system.

[0047] Furthermore, the present invention has the distinctive advantage of being able to secure a high-quality global business matching model that reflects the actual matching experiences of various regions and companies without any original data from companies being centralized, by adopting a federated learning structure that combines encrypted gradient or parameter updates of multiple local nodes over multiple communication rounds using a federated learning aggregation algorithm to continuously improve the global matching model.

[0048] Furthermore, the present invention has the distinctive advantage of providing accurate and reliable matching results while ensuring that even a central coordination server cannot reverse-engineer the exact unit price, margin rate, or financial figures of individual companies, by having a secure computation module perform similarity calculations on encrypted local profile vectors without decryption and a matching score calculation module apply differential privacy noise as post-processing to the calculated matching scores.

[0049] Furthermore, the present invention has the distinct advantage of enabling participation in international business matching even in conservative regulatory environments, as the policy / regulation management module integrates and manages country-specific data export regulations, differential privacy parameters, and data location fixing policies, and simultaneously controls the data processing allowance range and computational allowance level of each module within the local node and the central coordination server, thereby allowing companies or public institutions in countries with data export restrictions to participate in the platform by keeping sensitive data confined to their domestic servers and receiving only the matching results globally.

[0050] In addition, the present invention has the advantage of realizing an optimal balance of privacy and utility by selectively applying homomorphic encryption, differential privacy noise, segmentation, or masking to each data field of a local profile according to the sensitivity level, thereby maintaining a balance between theoretical complete security and the feasibility of actual service implementation, while preserving sufficient information for fields that have a critical impact on matching accuracy and strongly protecting precise numeric fields with a high risk of back-inference. Brief explanation of the drawing

[0052] FIG. 1 is a block diagram illustrating the overall configuration of a privacy-preserving distributed matching system using federated learning and secure computation according to an embodiment of the present invention. Also, FIG. 2 is a flowchart illustrating the local profile generation and encryption / noise processing process at the local node shown in FIG. 1. Also, FIG. 3 is a sequence diagram illustrating the training process of a federated learning-based global matching model between the central coordination server and a plurality of local nodes shown in FIG. 1. In addition, FIG. 4 is a flowchart illustrating the process of calculating a matching score based on safety calculation in the central coordination server shown in FIG. 1. In addition, FIG. 5 is a block diagram illustrating a regulation and policy-based control structure centered on the policy / regulation management module illustrated in FIG. 1. Specific details for implementing the invention

[0053] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described in detail below together with the accompanying drawings.

[0054] However, the present invention is not limited to the embodiments disclosed below but will be implemented in various different forms.

[0055] The embodiments described in this specification are provided to ensure that the disclosure of the invention is complete and to fully inform those skilled in the art of the scope of the invention.

[0056] And the present invention is defined only by the scope of the claims.

[0057] Accordingly, in some embodiments, well-known components, well-known operations, and well-known techniques are not specifically described to avoid the invention being interpreted ambiguously.

[0058] Additionally, throughout the specification, the same reference numerals refer to the same components, and the terms used (mentioned) in this specification are for describing embodiments and are not intended to limit the invention.

[0059] In this specification, the singular form includes the plural form unless specifically stated otherwise in the text, and components and operations referred to as 'comprising (or comprising)' do not exclude the presence or addition of one or more other components and operations.

[0060] Unless otherwise defined, all terms used in this specification (including technical and scientific terms) may be used in a meaning that is commonly understood by those skilled in the art to which the present invention belongs.

[0061] Also, terms defined in commonly used dictionaries are not interpreted ideally or excessively unless otherwise defined.

[0062] When a part of a specification is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0063] Additionally, terms such as "...part," "...module," etc., as described in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.

[0064] The expression "at least one of a, b, and c" described throughout the specification may include 'a alone', 'b alone', 'c alone', 'a and b', 'a and c', 'b and c', or 'a, b, and c all'.

[0065] The "terminal" mentioned below can be implemented as a computer or portable terminal capable of connecting to a server or other terminals via a network.

[0066] Here, the computer includes, for example, a notebook, desktop, laptop, etc. equipped with a web browser, and the portable terminal is, for example, a wireless communication device that ensures portability and mobility, and may include all kinds of handheld-based wireless communication devices such as communication-based terminals such as IMT (International Mobile Telecommunication), CDMA (Code Division Multiple Access), W-CDMA (W-Code Division Multiple Access), LTE (Long Term Evolution), smartphones, tablet PCs, etc.

[0067] Below, embodiments of the present invention are described in detail with reference to the attached drawings so that those skilled in the art can easily implement the present invention.

[0068] However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein.

[0069] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings.

[0070] In describing the embodiments, technical details that are well known in the technical field to which the present invention belongs and are not directly related to the present invention are omitted.

[0071] This is intended to convey the essence of the invention more clearly without obscuring it by omitting unnecessary explanations.

[0072] For the same reason, some components in the attached drawings have been exaggerated, omitted, or schematically depicted.

[0073] In addition, the size of each component does not entirely reflect its actual size. Identical or corresponding components in each drawing have been assigned the same reference number.

[0074] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described in detail below together with the accompanying drawings.

[0075] However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms, and these embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims.

[0076] Throughout the specification, the same reference numerals refer to the same components.

[0077] At this point, it will be understood that each block of the process flow diagrams and combinations of the flow diagrams can be executed by computer program instructions.

[0078] Since these computer program instructions can be loaded into the processor of a general-purpose computer, a special-purpose computer, or other programmable data processing equipment, the instructions executed through the processor of the computer or other programmable data processing equipment create a means to perform the functions described in the flowchart block(s).

[0079] Since these computer program instructions may be stored in computer-available or computer-readable memory that can be directed toward a computer or other programmable data processing equipment to implement a function in a specific way, the instructions stored in said computer-available or computer-readable memory may also produce a manufactured item containing instruction means that performs the function described in the flowchart block(s).

[0080] Since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that perform a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in the flowchart block(s).

[0081] Additionally, each block may represent a module, segment, or part of code containing one or more executable instructions for executing a specific logical function(s).

[0082] Additionally, it should be noted that in some alternative execution examples, the functions mentioned in the blocks may occur out of order.

[0083] For example, two blocks that are depicted in succession can actually be performed substantially simultaneously, or the blocks can sometimes be performed in reverse order according to their corresponding functions.

[0085] First, FIG. 1 is a block diagram illustrating the overall configuration of a privacy-preserving distributed matching system using federated learning and secure computation according to one embodiment of the present invention.

[0086] And, FIG. 2 is a flowchart illustrating the local profile generation and encryption / noise processing process at the local node illustrated in FIG. 1.

[0087] And, FIG. 3 is a sequence diagram illustrating the learning process of a federated learning-based global matching model between a central coordination server and a plurality of local nodes as illustrated in FIG. 1.

[0088] In addition, Fig. 4 is a flowchart illustrating the process of calculating a matching score based on safety operations in the central coordination server shown in Fig. 1.

[0089] In addition, FIG. 5 is a block diagram illustrating a regulation and policy-based control structure centered on the policy / regulation management module illustrated in FIG. 1.

[0091] For reference, in FIG. 2, the local node internal processing steps (210 to 240, 260) are shown as solid lines, the encryption / noise processing step (250) as a dotted line, and the central transmission step (270) as a dotted line.

[0092] Additionally, the central coordination server shown as 300 in FIG. 3 is the same component as the central coordination server (120) in FIG. 1, and local node A shown as 310 and local node B shown as 320 correspond to local node (110a) and local node (110b) in FIG. 1, respectively.

[0093] In addition, the policy / regulation management module shown as 510 in FIG. 5 is the same component as the policy / regulation management module (125) in FIG. 1, and FIG. 5 separately illustrates the internal control structure of the module in detail.

[0095] The present invention can be applied as an embodiment to a privacy-preserving distributed matching system (100) using federated learning and secure computation, comprising a plurality of local nodes (110), a central coordination server (120), and a network communication unit (130) as shown in FIG. 1.

[0096] Multiple local nodes (110) are each operated at the level of a company or institution and store sensitive original data independently, and a central coordination server (120) is connected to the multiple local nodes (110) via a secure channel through a network communication unit (130) to oversee the training of the global matching model and the calculation of matching scores between companies.

[0097] In FIG. 1, for convenience of explanation, only the internal configuration of the local node (110a) is illustrated in detail, but the local node (110b), local node (110c), and all subsequent local nodes each have the same components.

[0099] Each of the multiple local nodes (110) includes a local data storage unit (111), a local profile generation module (112), a local matching feature calculation module (113), a local learning module (114), a local encryption / noise module (115), and a local communication module (116).

[0101] The local data storage unit (111) stores sensitive original data including transaction history, unit price, margin rate, financial information, inventory status, production capacity, quality grade, and lead time.

[0102] Since such sensitive original data is a key indicator of competitiveness between companies and can cause serious commercial damage if exposed externally, the present invention establishes as a core principle that sensitive original data stored in the local data storage unit (111) is not transmitted outside the local system in any form.

[0104] The local profile generation module (112) calculates item-specific statistics from sensitive original data stored in the local data storage unit (111), including average unit price, standard deviation of unit price, and quantile regarding price range distribution, maximum production volume, average production volume, and production volume variability regarding production capacity, defect rate and certification level regarding quality indicators, and average delivery time and delay probability regarding lead time distribution, and converts them into a local profile in the form of a real value vector or a structured statistical summary vector including category characteristics.

[0105] In this way, by using statistical summary vectors instead of the original figures, sufficient information for matching can be extracted without exposing individual transaction history or exact unit price figures.

[0107] The local matching feature calculation module (113) further combines derivative indicators, including risk preference categories and transaction stability indicators, with statistics received from the local profile generation module (112) to generate a fixed-length feature vector or a structured feature set to be input into the matching algorithm.

[0109] The local learning module (114) receives global matching model parameters distributed from the central coordination server (120) and calculates a gradient or parameter update (ΔW) for the global matching model using past matching results, transaction success status, and satisfaction data stored in the local data storage unit (111).

[0110] Since the local learning module (114) operates as a federated learning client, only the calculated gradient or parameter update values ​​are transmitted to the central coordination server (120), and the local data itself is not transmitted.

[0112] The local encryption / noise module (115) generates only encrypted or noise-processed values ​​through hybrid protection processing, which applies homomorphic encryption or differential privacy noise including Laplace noise or Gaussian noise to high-sensitivity unit price, margin rate, and financial indicator fields, performs low, medium, and high-sensitivity bucketization processing on medium-sensitivity fields, and performs masking or lightweight processing on non-sensitive fields, for local profiles and feature vectors generated by the local profile generation module (112) and the local matching feature calculation module (113), and gradients or parameter updates calculated by the local learning module (114).

[0113] This hybrid structure, which selectively applies different protection techniques based on sensitivity levels, achieves backtracking prevention at a practical level—even if not theoretically a perfect encryption level—while keeping the computational burden within a realistic range.

[0115] The local communication module (116) transmits only the encrypted or noise-processed local profile and parameter updates received from the local encryption / noise module (115) to the central coordination server (120) via the network communication unit (130), but does not transmit sensitive original data, individual contract conditions, and detailed unit price / margin information stored in the local data storage unit (111) outside the local system at all.

[0116] The network communication unit (130) provides a VPN or TLS-based secure channel between the local communication module (116) and the central communication module (127) described later.

[0118] Meanwhile, referring to FIG. 2, the local profile creation and protection processing process (200) at the local node (110) is illustrated in more detail.

[0119] In the original data loading step (210) illustrated by a solid line in FIG. 2, sensitive original data is read from the local data storage unit (111), in the summary statistics calculation step (220), statistics for each item are calculated, in the feature vector generation step (230), a feature vector for matching is generated, and in the sensitive field classification step (240), the sensitivity level of each field is determined.

[0120] Next, in the encryption / noise processing step (250) illustrated by a dotted line, hybrid protection processing according to the sensitivity level is performed, and in the local verification step (260) illustrated by a solid line, after verifying whether the processed profile meets the policy criteria of the policy / regulation management module (125), only the protected processed profile is transmitted to the central coordination server (120) in the central transmission step (270) illustrated by a dotted line.

[0122] Next, the central coordination server (120) includes a model management module (121), an update aggregation module (122), a safety calculation module (123), a matching score calculation module (124), a policy / regulation management module (125), a central data storage unit (126), and a central communication module (127).

[0124] The model management module (121) maintains and manages parameters of a global matching model including a graph embedding model or a matching score prediction model, updates the global parameters based on the aggregation results received from the update aggregation module (122), and repeatedly redistributes the updated global parameters to each of the multiple local nodes (110) over multiple communication rounds.

[0125] Through this multi-communication round iteration structure, the global-level matching model is progressively improved without the sensitive original data of each local node (110) ever being centralized.

[0127] The update aggregation module (122) combines encrypted gradient or parameter updates received from each local communication module (116) of a plurality of local nodes (110) according to an aggregation algorithm including a simple average (FedAvg), a weighted average proportional to the data size, or secure aggregation to produce new global matching model parameters.

[0128] When a secure summation method is applied, the summation operation is performed in an encrypted state, so the central coordination server (120) can aggregate without decrypting the update values ​​of individual local nodes.

[0130] Meanwhile, referring to Fig. 3, the learning process of a federated learning-based global matching model is illustrated in more detail.

[0131] When the model management module (121) distributes the current global matching model parameters to local nodes (110a, 110b) (global parameter distribution), each local node (110) calculates gradient or parameter updates using its local data (local learning), and then transmits the updates, which have been protected by the local encryption / noise module (115), to the central coordination server (120) (parameter update).

[0132] The update aggregation module (122) aggregates the received updates to produce updated global parameters (aggregation) and redistributes them to each local node (110) (redistribution of updated parameters).

[0133] By repeating this loop over multiple communication rounds, a high-quality global matching model is secured that reflects the matching experiences of various regions and companies without centralizing the original data.

[0135] The secure computation module (123) performs the calculation of the inner product, cosine similarity, or Euclidean distance between vectors without decryption on encrypted local profile vectors received from multiple local nodes (110) using a homomorphic encryption-based computation or a multi-party computation (MPC)-based distributed sum computation.

[0136] Homomorphic encryption-based operations utilize the characteristic that direct numerical operations are possible while the data is encrypted, and MPC-based distributed sum operations derive the operation result through a secret distribution method among multiple operation nodes.

[0138] The matching score calculation module (124) calculates a matching score between pairs of companies by combining the similarity calculation result received from the safety calculation module (123) and the global matching model parameters of the model management module (121), and protects against reverse inference of sensitive data of individual companies by applying additional differential privacy noise as post-processing to the calculated matching score or ranking, and then provides the matching result, including a top partner candidate list and similarity and risk indicators, to the company user terminal (10) or manager terminal (20).

[0140] Meanwhile, referring to FIG. 4, the matching score calculation processing process (400) in the central coordination server (120) is illustrated in more detail.

[0141] In the local profile reception step (410), the central communication module (127) receives the encrypted profile of each local node (110), and in the decryption / verification step (420, optional), if only differential privacy is applied, the value range and policy compliance are verified, and if homomorphic encryption is applied, the operation can proceed in an encrypted state without decryption, so this step is performed optionally.

[0142] In the safe operation preparation step (430), protocols and sessions required for similarity calculation are prepared, and in the similarity / fitness calculation step (440), vector-to-vector operations are performed through the safe operation module (123).

[0143] In the differential privacy post-processing step (450, optional), additional noise is applied to the matching score or ranking, and in the matching score calculation and sorting step (460), the final score and the top partner candidate list are calculated, and in the result storage and provision step (470), the result is stored in the central data storage unit (126) and provided to the corporate user terminal (10) and the administrator terminal (20).

[0145] The policy / regulation management module (125) stores and manages policy profiles including country-specific and institution-specific data export regulation information, data localization policies, differential privacy parameters (ε), and encryption levels, controls the data processing allowance range and operation allowance level of local encryption / noise modules (115) and local communication modules (116) within multiple local nodes (110), and update aggregation modules (122), secure computation modules (123), and matching score calculation modules (124) within the central coordination server (120), and records all operation execution history and parameter change history as an audit log.

[0146] The central data storage unit (126) stores non-sensitive data including global model parameters, aggregated statistics, matching results, and policy setting values, and the central communication module (127) receives encrypted local profile and parameter updates from each local communication module (116) of a plurality of local nodes (110) and transmits the updated global matching model parameters to the plurality of local nodes (110) according to the instructions of the model management module (121).

[0148] Meanwhile, referring to FIG. 5, a regulation and policy-based control structure (500) centered on a policy / regulation management module (125) is illustrated more specifically.

[0149] The policy / regulation management module (510) is interconnected with the policy profile storage unit (520) to query and update policy profiles by country and institution, and applies the policy of the corresponding country and institution to the local encryption / noise module (115), local learning module (114), and local communication module (116) of the local node (110) through the local policy application interface (530).

[0150] In addition, the allowable operation range and encryption level of the update aggregation module (122), safety operation module (123), and matching score calculation module (124) of the central coordination server (120) are controlled through the central operation policy application interface (540).

[0151] Through this, for example, for data from EU countries subject to GDPR, a ban on export and strict differential privacy parameters (ε) may be enforced, and aggregation and computation related to such data may be restricted to being performed only on servers within a designated region.

[0153] The present invention is applicable to the above-described embodiments, and it goes without saying that it is applicable to various embodiments as follows.

[0155] First, the local profile generation module (112) can configure the local profile in the form of a real value vector having an average unit price, unit price variability, maximum production volume, average lead time, lead time variability, average margin rate, and quality defect rate as elements.

[0156] In addition, the local encryption / noise module (115) can perform field-specific selective protection processing for each field of the local profile, providing an average value with differential privacy noise applied to the unit price and margin rate fields, applying a homomorphic encryption-based average calculation to the lead time field, and providing only low / medium / high price range information to the categorical field including risk preference.

[0157] This field-specific differentiated protection method achieves an optimal balance between privacy protection levels and matching accuracy by applying lightweight processing to fields that critically impact matching accuracy to minimize information loss, and by intensively applying strong encryption and noise to precision numeric fields with a high risk of backfiring.

[0159] Meanwhile, the policy / regulation management module (125) can apply a policy profile to a local node (110) corresponding to a specific country or institution, including a ban on data export, exclusive exchange of encrypted profiles through a designated path, and a country-specific upper limit of a differential privacy parameter (ε).

[0160] In particular, for local nodes (110) operated by public institutions, only local learning of corporate data of the country and transmission of encrypted parameter updates are allowed, and the range of data movement paths and operations allowed is restricted so that only matching scores with domestic companies and a list of top partner candidates are provided from the matching score calculation module (124).

[0161] For example, if a specific country's Ministry of Industry or investment promotion agency acts as a local hub, the hub can store the sensitive data of domestic companies solely on its own servers while receiving only lists of suitable overseas partner candidates and matching scores from global platforms; this enables a Business-to-Government (B2G) cooperation structure that allows participation in international business matching while fully maintaining data sovereignty.

[0163] Meanwhile, as an example of a method performed by the system (100), the system (100) calculates item-specific statistics from the local data storage unit (111) of each of the plurality of local nodes (110) through the local profile generation module (112) to generate a local profile in the form of a statistical summary vector.

[0164] The system (100) generates a fixed-length feature vector by adding derived indicators to a local profile through a local matching feature calculation module (113), and generates a protection profile by selectively performing homomorphic encryption, differential privacy noise addition, or segmentation processing according to the sensitivity of each field through a local encryption / noise module (115).

[0165] The system (100) receives global matching model parameters distributed from the model management module (121) through the local learning module (114), calculates gradient or parameter updates using local data, performs protection processing through the local encryption / noise module (115), and transmits them to the central coordination server (120) through the local communication module (116), but does not transmit sensitive original data.

[0166] The system (100) aggregates encrypted updates received from multiple local nodes (110) through an update aggregation module (122) to calculate new global matching model parameters and redistributes them through a model management module (121), and performs similarity calculations between encrypted profile vectors without decryption through a secure computation module (123).

[0167] The system (100) calculates a matching score through a matching score calculation module (124), applies differential privacy noise after post-processing, and provides a list of top partner candidates to an enterprise user terminal (10) or an administrator terminal (20). This local learning and central aggregation process is repeated over multiple communication rounds, but is performed under the control of a policy / regulation management module (125) while complying with the country-specific data export regulations and differential privacy parameters (ε) of each local node (110).

[0169] Meanwhile, each of the multiple local nodes (110) may further include a sensitivity automatic evaluation module (118) that calculates a field-specific sensitivity score by analyzing the entropy, variance, and correlation with other fields for each data field of the local data storage unit (111), and automatically determines the level of protection to apply among homomorphic encryption, differential privacy noise addition, segmentation, or masking according to the interval of the sensitivity score.

[0170] The sensitivity automatic evaluation module (118) dynamically adjusts the sensitivity score threshold based on the country-specific differential privacy parameter (ε) received from the policy / regulation management module (125), thereby realizing an adaptive protection level determination that automatically responds to changes in the country-specific regulation level or changes in the data structure.

[0171] This reduces the operational burden of administrators having to manually set protection levels for each field, while enabling the rapid and consistent reflection of policies in the event of regulatory changes.

[0173] Meanwhile, the update aggregation module (122) calculates a confidence score based on cosine similarity, norm deviation, and an outlier detection algorithm for each gradient or parameter update received from multiple local nodes (110), and performs adaptive weighted aggregation that reflects the confidence score as a weight, while performing Byzantine Fault Tolerant filtering to exclude or attenuate updates from local nodes (110) whose confidence score falls below a preset threshold.

[0174] In a global B2B matching platform environment, there may be companies with uneven data quality or malicious participants who submit intentionally distorted updates; in such situations, Byzantine-resistant filtering functions as a key safety mechanism to protect the training quality of global matching models and maintain platform reliability.

[0176] Meanwhile, the matching score calculation module (124) can perform a privacy-preserving matching basis provision function by analyzing the contribution of each field, including price range, production capacity, lead time, and quality indicators that contributed to the matching score, generating an explainable matching basis that includes only the contribution ranking of each field and the category level of low / medium / high instead of individual figures, and applying differential privacy noise to the matching basis and providing it to the corporate user terminal (10) or administrator terminal (20).

[0177] Corporate users can verify the basis for a partner's high matching score at the category level without disclosing exact unit prices or profit margins, thereby enhancing trust and acceptance of the platform.

[0179] Meanwhile, the central coordination server (120) may further include a blockchain-based audit log module (128) that hashes and records the data access history, differential privacy parameter (ε) change history, participation history per federated learning communication round, and matching result provision history of multiple local nodes (110) in a distributed ledger, and automatically blocks system access of local nodes (110) where policy violations are detected through a smart contract.

[0180] Blockchain-based audit logs cannot be tampered with after the fact due to the immutability of the distributed ledger, and can be utilized as an audit tool by regulatory or public institutions to independently verify the platform's policy compliance history.

[0182] Meanwhile, the central coordination server (120) may further include an adaptive round scheduling module (129) that, prior to the start of each federated learning communication round, dynamically selects an optimal set of participating nodes by comprehensively evaluating the data update status, current regulatory status, network response delay, and previous round contribution of each of the multiple local nodes (110), and proceeds with the corresponding round when the number of selected participating nodes satisfies a preset minimum participation condition (quorum), but when receiving a regulatory change event of a specific country or institution from the policy / regulation management module (125), immediately excludes the local node (110) from the corresponding communication round and re-participates after the policy update is completed.

[0183] This simultaneously ensures the legal stability and operational continuity of the platform by responding in real-time to unexpected situations in actual global operating environments, such as amendments to the Personal Information Protection Act, changes in regulations on data transfer abroad, or network failures.

[0185] The operation and effects of the present invention are described below.

[0186] First, the present invention has the effect that sensitive original data, such as transaction conditions, unit prices, margins, and financial information, does not leave the local system at all, while enabling global business matching, by having the local encryption / noise module (115) perform hybrid protection processing according to sensitivity levels and the local communication module (116) transmit only the encrypted or noise-processed values.

[0187] In addition, the present invention has the effect of enabling participation in international business matching even in a conservative regulatory environment, as the policy / regulation management module (125) integrates and manages country-specific data export regulations, differential privacy parameters (ε), and data location fixing policies, and simultaneously controls the allowable range of each module within the local node (110) and the central coordination server (120), thereby allowing countries or public institutions with data export restrictions to participate in the platform with a structure in which data is kept locally and only matching results are provided globally.

[0188] In addition, the present invention has the effect of securing a high-quality global matching model that reflects actual matching experiences of various regions and companies without centralizing original data, through a federated learning structure in which an update aggregation module (122) aggregates the encrypted gradients of a plurality of local nodes (110) over a plurality of communication rounds.

[0189] Furthermore, the present invention has the effect of securing strong technical barriers to entry and competitive edge in regulatory compliance while maintaining a balance between theoretical complete security and the feasibility of actual service implementation through a hybrid structure that combines multiple security techniques, such as homomorphic encryption, multi-party computation (MPC), and differential privacy, at a practical level in each stage of matching profile generation, model update, and score calculation.

[0191] As described above, the fundamental technical concept of the present invention is to provide a privacy-preserving distributed matching system using federated learning and secure computation, which enables global inter-company business matching without exporting sensitive original data at all, by having each of the multiple local nodes at the corporate or institutional level store sensitive original data, including transaction history, unit price, margin rate, financial information, inventory status, production capacity, quality grade, and lead time, only within their own systems; converting item-specific statistics regarding price range distribution, production capacity, quality indicators, and lead time distribution into local profiles in the form of statistical summary vectors instead of the original data; protecting these profiles by applying homomorphic encryption or differential privacy noise according to sensitivity levels; and transmitting them to a central coordination server.

[0192] And, it goes without saying that many other variations and applications are also possible for those skilled in the art within the scope of the basic technical concept of the present invention. Explanation of the symbols

[0194] 10...Enterprise User Terminal 20...Administrator terminal 100... Privacy-preserving distributed matching system using federated learning and secure computation 110...Local Node 111...Local data storage 112...Local Profile Creation Module 113...Local Matching Feature Calculation Module 114...Local learning module 115...Local encryption / noise module 116...Local communication module 120...Central Coordination Server 121...Model Management Module 122...Update Aggregation Module 123...Safety calculation module 124...Matching score calculation module 125...Policy / Regulation Management Module 126...Central Data Storage Unit 127...Central Communication Module 130...Network Communications Department 200... Local profile creation and protection processing 210...Load original data 220...Summary statistics calculation 230...Generate feature vector 240...Sensitive field classification 250...Encryption / Noise processing 260...Local verification 270...Central transmission 300...Central Coordination Server 310...Local Node A 320...Local Node B 400...Matching score calculation processing 410...Receive local profile 420...Decryption / Verification 430...Prepare for safe operation 440... Similarity / Fitness Calculation 450... Differential privacy post-processing 460...Matching score calculation and sorting 470...Save and provide results 500...Regulatory and policy control system 510...Policy / Regulation Management Module 520...Policy Profile Storage 530...Local Policy Enforcement Interface 540...Central Computation Policy Application Interface

Claims

Claim 1 A privacy-preserving distributed matching system utilizing federated learning and secure computation comprises: a plurality of local nodes, each operating at the corporate or institutional level and storing sensitive original data independently; and a central coordination server connected to the plurality of local nodes via a secure channel through a network communication unit, which oversees the learning of a global matching model and the calculation of inter-corporate matching scores. Each of the plurality of local nodes includes: a local data storage unit that stores sensitive original data including transaction history, unit price, margin rate, financial information, inventory status, production capacity, quality grade, and lead time; and a local profile generation module that calculates item-specific statistics from the sensitive original data stored in the local data storage unit, including average unit price, unit price standard deviation, and quantile regarding price range distribution; maximum production volume, average production volume, and production volume variability regarding production capacity; defect rate and certification level regarding quality indicators; and average delivery time and delay probability regarding lead time distribution, and converts these into a local profile in the form of a real value vector or a structured statistical summary vector including category characteristics. A local matching feature calculation module that generates a fixed-length feature vector or a structured feature set to be input into a matching algorithm by additionally combining derivative indicators, including risk preference categories and transaction stability indicators, with statistics received from the local profile generation module; and a local learning module of a federated learning client method that receives global matching model parameters distributed from the central coordination server and calculates a gradient or parameter update for the global matching model using past matching results, transaction success status, and satisfaction data stored in the local data storage unit.A local encryption / noise module that generates only encrypted or noise-processed values ​​through hybrid protection processing, wherein, for the local profiles and feature vectors generated by the local profile generation module and the local matching feature calculation module and the gradients or parameter updates calculated by the local learning module, homomorphic encryption or differential privacy noise including Laplace noise or Gaussian noise is selectively applied to high-sensitivity unit price, margin rate, and financial indicator fields, low, medium, and high-sensitivity bucketization processing is performed on medium-sensitivity fields, and masking or lightweight processing is performed on non-sensitive fields; and includes a local communication module that transmits only encrypted or noise-processed local profiles and parameter updates received from the local encryption / noise module to the central coordination server via the network communication unit, while not transmitting sensitive original data, individual contract conditions, and detailed unit price / margin information stored in the local data storage unit outside the local system at all; and the central coordination server includes a model management module that maintains and manages parameters of a global matching model including a graph embedding model or a matching score prediction model, updates global parameters based on aggregation results received from an update aggregation module, and repeatedly redistributes the updated global parameters to each of the plurality of local nodes over a plurality of communication rounds; and an update aggregation module that calculates new global matching model parameters by combining encrypted gradients or parameter updates received from the local communication modules of each of the plurality of local nodes according to an aggregation algorithm including a simple average (FedAvg), a weighted average proportional to the data size, or secure aggregation.A secure computation module that performs calculations of inner product, cosine similarity, or Euclidean distance between vectors without decryption for encrypted local profile vectors received from the plurality of local nodes, using homomorphic encryption-based computation or multi-party computation (MPC)-based distributed sum computation; and a matching score calculation module that calculates a matching score between pairs of companies by combining the similarity computation result received from the secure computation module with global matching model parameters of the model management module, protects individual companies' sensitive data from being reverse-inferred by applying additional differential privacy noise as post-processing to the calculated matching score or ranking, and provides a matching result including a top partner candidate list and similarity and risk indicators to a corporate user terminal or an administrator terminal. A privacy-preserving distributed matching system using federated learning and secure computation, characterized by comprising: a policy / regulation management module that stores and manages a policy profile including information on data export regulations by country and institution, data localization policies, differential privacy parameters, and encryption levels, controls the data processing allowance range and computation allowance level of local encryption / noise modules and local communication modules within the plurality of local nodes, and update aggregation modules, secure computation modules, and matching score calculation modules within the central coordination server, and records all computation execution history and parameter change history as an audit log; a central data storage unit that stores non-sensitive data including matching results received from the matching score calculation module, global model parameters, aggregated statistics, and policy setting values; and a central communication module that receives encrypted local profile and parameter updates from the local communication modules of each of the plurality of local nodes and transmits updated global matching model parameters to the plurality of local nodes according to the instructions of the model management module. Claim 2 A privacy-preserving distributed matching system using federated learning and secure computation according to claim 1, wherein the local profile generation module configures the local profile in the form of a real value vector having elements of average unit price, unit price volatility, maximum production volume, average lead time, lead time volatility, average margin rate, and quality defect rate, and the local encryption / noise module performs field-specific selective protection processing for each field of the local profile, wherein for each field of the local profile, it provides an average value with differential privacy noise applied to the unit price and margin rate fields, applies a homomorphic encryption-based average calculation to the lead time field, and provides only low-price / medium-price / high-price range information to the categorical field including risk preference. Claim 3 A privacy-preserving distributed matching system using federated learning and secure computation according to claim 1, wherein the policy / regulation management module stores a policy profile for a local node corresponding to a specific country or institution, including a ban on data export abroad, exclusive exchange of encrypted profiles through a designated path, and country-specific upper limits for differential privacy parameters, and applies it to the local node, wherein for a local node operated by a public institution, only local learning on corporate data of the country and transmission of encrypted parameter updates are permitted, and the data movement path and the scope of permitted computation are restricted so as to receive only matching scores with domestic companies and a list of top partner candidates from the matching score calculation module.

Citation Information

Patent Citations

  • System and methods for validating and performing operations on homomorphically encrypted data

    KR1020180127506A

  • Manufacturing collaboration artificial intelligence matching service system based on corporate information

    KR1020250028837A

  • Method and system for protecting privacy by homomorphically encrypting SQL statements

    KR1020250055164A

  • Compliance requirements analysis and inspection automation device and method for controlling the same

    KR1020250085618A