KR20210065565A - Data Privacy-Preserving Distributed Knowledge Discovery system based on the Blockchain and Method thereof - Google Patents

Abstract

Disclosed is a distributed knowledge discovery system for providing data privacy based on blockchain. The system of the present invention comprises: a contract manager having an application logic to register a distributed data knowledge discovery contract and set initial parameters; an ML model builder having an application logic to subscribe to a blockchain group and register a model, as an ML model developer having its own data; an ML model data set which is not made public for privacy as the data owned by the ML model builder; an ML model developed by the ML model builder by using the ML model data set; and blockchain having a contract which is a software logic for registering an ML builder and managing an incentive system, as distributed ledger technology which cannot be modified once registered and whose record will last permanently. According to the present invention, knowledge can be discovered through a model shared in a distribution environment by third persons who do not trust each other by using blockchain.

Description

Data Privacy-Preserving Distributed Knowledge Discovery system based on the Blockchain and Method thereof

The present invention relates to a blockchain-based distributed knowledge discovery system and method for developing and sharing a model without disclosing training data and pursuing model improvement. A distributed knowledge discovery system that provides blockchain-based data privacy so that multiple parties can develop their own machine models with their own data and then share each other's learning models through blockchain and use them for model improvement; it's about how

If you want to develop an ML model with the prior art, a large amount of data is needed to develop a high-quality model. Data in the field of interest has potential value, so it is a very important asset and it is difficult to share it because it is reluctant to disclose it. Model developers at the same level need a shared way to improve their ML models, but it is difficult to share with third parties who do not trust each other.

Nevertheless, if their models and data are shared with each other for knowledge discovery in a distributed environment, builders who share first and do not share later may occur, and when ML builders who do not register a model occur, first share the model. Builders lose money, and honest builders lose if they all share a poorly performing model.

While solving these problems, it is intended to propose a method of discovering knowledge through a shared model in a distributed environment between third parties who do not trust each other using blockchain.

1. Republic of Korea Patent Publication No. 10-2019-0086301 (2019.07.22) 2. Republic of Korea Patent Publication No. 10-2019-0041784 (2019.04.23)

It is an object of the present invention to provide a distributed knowledge discovery system and method that provides blockchain-based data privacy so that knowledge can be discovered through a model shared between third parties who do not trust each other and in a distributed environment using blockchain. is to do

Another object of the present invention is to use the blockchain-based incentive group evaluation method or the deposit deposit method including the imposition of fines to solve various problems that are not based on trust between data owners. It is to provide a distributed knowledge discovery system and method that provides

In order to achieve the above object, a distributed knowledge discovery system that provides data privacy based on a block chain according to an embodiment of the present invention,

A contract manager for registering a distributed data knowledge discovery contract and setting initial parameters, the contract manager having an application logic; means an ML model developer with its own data, and an ML model builder with application logic that can join a group and register a model on the blockchain; an ML model data set that is not disclosed for privacy as its own data owned by the ML model builder; an ML model developed by the ML model builder using the ML model data set; and a blockchain having a contract, which is a distributed ledger technology that cannot be tampered with and records are permanently maintained once registered, a software logic that ML builder registers ML models, and manages incentive systems, etc. do.

A distributed knowledge discovery method that provides blockchain-based data privacy according to another embodiment of the present invention,

(a) an initialization step of a distributed knowledge discovery contract corresponding to an initialization task for performing distributed knowledge discovery; (b) forming a group of ML builders by recruiting ML builders; (c) ML model registration step in which the registered ML builder encrypts the developed ML model using its own data and registers it in the blockchain; (d) ML model recovery step of downloading and decrypting ML models registered by other ML builders belonging to the same group from the blockchain and recovering the ML models; (e) an ML model validation step of validating whether the models registered by the ML model builder are all the same model through a hash; (f) ML model performance reporting step of repairing the ML model generated by other ML builders to test the performance and report the result; and (g) ML model task completion step in which all ML performance reports are completed and deposit settlement is completed according to the corresponding procedure.

In the (a) distributed knowledge discovery initialization step, when training in an ensemble approach after sharing a model as a parameter to initialize when registering a contract in the block chain, ‘builder’s deposit amount’, ‘bad model penalty amount’, ‘bad data penalty’ providing 'amount', 'minimum number of builders', 'maximum number of builders', 'builder recruitment period', and 'minimum performance score'; and providing as additional parameters 'type and architecture of the ML model' including the number of layers, the number of nodes, and the size of the kernel when training with the retraining approach.

The (b) ML builder group formation step includes: a builder who wants to participate during the 'builder recruitment period' sends the 'builder's guarantee amount' to the block chain as a deposit in cryptocurrency and receives an application; If the number of builders who have set a deposit is less than the 'minimum number of builders' during the 'builder recruitment period', the ML builder group formation is canceled and the received deposit is all returned to the applicant; And if the number of builders applied is more than the 'minimum number of builders', the ML builder group is successfully formed, and even if the 'builder recruitment period' has not expired, if the group's limited number of 'maximum number of builders' builders complete the application, ML builders It is characterized in that it includes; a step in which the group formation is successful and complete.

In the (c) ML model registration step, when the model formation is successful, the 'ML model type and architect' is checked and model training is started, and when the training model reaches the 'minimum performance score' or higher, the training is completed. completion stage; and registering a combination of the training model encrypted with the encryption key, the decryption key encrypted with the public key of the receiving builder, and a message including the hash value of the training model when the model training is completed. do it with

The (d) ML model recovery step includes the step of recovering the training model by decrypting the training model with the decryption key decrypted with the private key of the receiving builder in recovering the ML model by downloading the encrypted message registered in the block chain. characterized.

을 비교하여 동일 성을 확인하는 단계; 동일성이 확인되면 동일함을 그룹의 모든 빌더들에게 보고하고, 동일성이 확인되지 않으면 비동일함을 보고하는 단계; 그룹에서 비동일함을 보고한 빌더들이 한명이라도 있으면 모델의 개발자에게 재 등록을 요구하는 단계; 재 등록 하지 않거나, 재 등록된 모델을 재검사하여 동일성이 확인되지 않을 경우, 비동일 함을 최종 보고하는 단계; 및 최종 비동일함이 보고되는 빌더는 그룹에서 제외하고, 제외된 빌더의 보증금은 기부금으로 전환되는 단계;를 포함하는 것을 특징으로 한다.In the (e) ML model validation step, a hash value of the restored model is generated, and the hash value contained in the generated hash value and the message registered in the block chain (

verifying identity by comparing them; Reporting the same to all builders in the group if the identity is confirmed, and reporting non-identity if the identity is not confirmed; requesting re-registration from the developer of the model if there is at least one builder reporting non-identity in the group; Final reporting of non-identity if not re-registration or re-inspection of re-registered models to confirm identity; And the final non-identity reported builder is excluded from the group, and the deposit of the excluded builder is converted into a donation; characterized in that it comprises a.

In the (f) ML model performance reporting step, model performance checks are performed on models that have completed model validation, and for models that report performance that does not reach the 'minimum performance score', a 'poor model penalty' is given to the model developer. and granting a 'bad data penalty' to the builders who have checked the performance; There is a possibility that performance data may be poor when the performance does not come out properly, and there are cases where the performance data is false even when the performance is correct. Therefore, a fine is imposed on the performance inspector, but the amount of the 'bad data penalty' is a set less than the amount of '; and the step of including the 'incorrect data penalty' in the donation.

The (g) ML task completion step, when all the work is completed, the steps of first setting the amount of the builders who submitted the deposit; Returning the amount excluding the fine granted to each and including the amount collected as a penalty in the donation; and the step of entering the learning model improvement procedure through ensemble or retraining methods through model sharing through which high-quality models that have passed the performance test are shared.

According to the present invention, it is possible to improve the performance of one's own ML model through model sharing among untrusted third parties. In other words, more data is needed to develop a model with better performance, but through this model sharing method, a better machine learning model can be developed with limited own data. An honest model sharer has the advantage of not having to consider sharing a model with a third party who does not trust it, since it is possible to give back to society by giving a reasonable fine to the dishonest model sharer who only wants to take the model of other sharers.

In addition, since the model sharer with poor performance will be jointly verified with other participating builders, collaboration with high-quality ML models verified through this verification method will improve the performance and value of the model, which will contribute to the development of machine learning in the future. .

1 is a block diagram showing the overall configuration of a distributed knowledge discovery system that provides blockchain-based data privacy according to the present invention.
2 is a diagram schematically illustrating a distributed knowledge discovery method that provides data privacy based on a block chain according to the present invention.

Hereinafter, a distributed knowledge discovery system and method for providing blockchain-based data privacy according to the present invention will be described in detail with reference to the accompanying drawings.

The present invention relates to a method of managing through a blockchain to discover distributed knowledge without disclosing data, wherein a plurality of parties develop their own machine models with their own data and then each other through the blockchain It is about a method of sharing the learning model of the model and using it to improve the model.

At this time, in order to solve various problems that are not based on trust between data owners, it includes a method of using a block chain-based incentive group evaluation method and a deposit deposit method including the imposition of a fine. Incentive method that converts to final donation is applied.

Builders who can develop models in a distributed environment should be able to freely join groups and join model sharing regardless of mutual trust, and should be able to share models with the minimum number of parties while providing data privacy. Among them, it should be possible to limit dishonest builders who do not share their own models in a timely manner and want to take only the models shared by other builders, and even if the model is shared, do not share models that do not help other models due to poor performance. should be able to prevent

In addition, all actions are carried out using a block chain contract that lasts forever without tampering with the results, and in order to limit builders who can harm normal builders by sharing models with poor performance with builders who have done fraudulent actions, We propose a blockchain-based incentive mechanism that receives a deposit, imposes a fine, and returns the settled amount. A model sharing method using encryption and key distribution so that specific data can be shared only with specific targets as needed while providing privacy. to present

Among the English abbreviations below, ML stands for machine learning, MLSP stands for machine learning model developer, and MLC stands for machine learning model customer.

1 is a block diagram showing the overall configuration of a distributed knowledge discovery system that provides blockchain-based data privacy according to the present invention.

Referring to FIG. 1, the distributed knowledge discovery system that provides data privacy based on the block chain according to the present invention is a contract manager (1), ML model builder (2), ML model data set (3), ML model (4). ) and blockchain (5).

The contract manager 1 is a contract manager that registers a distributed data knowledge discovery contract and sets initial parameters, and includes application logic.

: 2)는 자체 데이터(

)를 가지고 있는 ML 모델 개발자를 의미하며, 블록체인에 그룹에 가입하고 모델을 등록할 수 있는 어플리케이션 로직을 포함한다.The ML model builder (

: 2) own data (

), which includes application logic that can join groups and register models on the blockchain.

: 3)은 ML 모델 빌더(

: 2)가 소유한 자체 데이터를 의미하며, 프라이버시를 위해 공개하지 않는다.The ML model data set (

: 3) is the ML Model Builder (

: 2) means its own data, which is not disclosed for privacy.

: 4)은 ML 모델 빌더(

가 상기 ML 모델 데이터 셋(

을 가지고 개발한 ML 모델 (

)을 의미한다.The ML model (

: 4) is the ML Model Builder (

is the ML model data set (

ML model developed with

) means

The blockchain (Blockchain: 5) is a distributed ledger technology that cannot be tampered with and records are permanently maintained once registered, and includes a contract that is a software logic for ML builders to register ML models and manage incentive systems.

2 is a diagram schematically illustrating a distributed knowledge discovery method that provides data privacy based on a block chain according to the present invention.

Referring to FIG. 2 , the distributed knowledge discovery method for providing data privacy based on the block chain according to the present invention includes an initialization step (a) of a distributed knowledge discovery contract corresponding to an initialization task for performing distributed knowledge discovery, and an ML builder. ML builder group formation step (b) of recruiting and forming a group, ML model registration step (c) in which the registered ML builder encrypts the ML model developed using its own data and registers it in the blockchain (c), other ML belonging to the same group ML model recovery step (d) to download and decrypt the ML model registered by the builder from the block chain and recover the ML model, ML model validation step to validate whether the models registered by the ML model builder are all the same model through hash (e), ML model performance reporting step (f), in which the ML model generated by other ML builders is restored to test the performance and report the result, and the ML model in which the ML performance reporting is completed and deposit settlement is completed according to the corresponding procedure task completion step (g).

The distributed knowledge discovery initialization step (a) is a step of starting the distributed knowledge discovery mechanism, and is a step in which the contract manager registers the distributed knowledge discovery contract in the block chain. In the case of the retraining approach, the initialization parameters are 'type and architecture of ML model', 'deposit amount of builders', 'amount of bad model fine', 'amount of bad data fine', 'minimum number of builders', 'maximum bill i.e. above Distributed knowledge discovery initialization step (a) is a parameter that is initialized when registering a contract in the block chain. When training in an ensemble approach after sharing the model, 'builder's deposit amount', 'bad model fine amount', 'bad data fine amount' ', 'minimum number of builders', 'maximum number of builders', 'builder recruitment period', and 'minimum performance score'; and providing 'type and architecture of the ML model', including layers, number of nodes, kernel size, and the like, as additional parameters when training with the retraining approach.

The ML builder group formation step (b) is a step in which distributed builders can freely participate in the group. Builders participating in the group must deposit an amount that matches the 'builder's deposit amount', and if builders who satisfy the 'minimum number of builders' participate during the 'builder recruitment period', a group is formed and the 'builder recruitment period' is not closed. Even if not, if the 'maximum number of builders' joins the group, the ML builders group is formed and recruitment ends. If the ML builder group formation fails because the 'minimum number of builders' is not satisfied within the 'builder recruitment period', the collected deposit will be returned as it is.

That is, the ML builder group formation step (b) includes: a builder who wants to participate during the 'builder recruitment period' sends the 'builder's guarantee amount' to the block chain as a deposit in cryptocurrency and receives an application; If the number of builders who have set a deposit is less than the 'minimum number of builders' during the 'builder recruitment period', the ML builder group formation is canceled and the received deposit is all returned to the applicant; And if the number of applied builders is greater than or equal to the 'minimum number of builders', the ML builder group is successfully formed, and even if the 'builder recruitment period' has not expired, if the 'maximum number of builders' builder has completed the application, the ML builder group and a step in which group formation is successful and complete.

으로 등록한다.In the ML model registration step (c), if the model formation is successful, 'type and architecture of the ML model' is checked and model training is started. When the performance of the trained model reaches more than the 'minimum performance score', the training is completed and registered in the blockchain for transmission to all builders. At the time of registration, a combination of the training model encrypted with the private key, the decryption key encrypted with the public key of the receiving builder, and a message containing the hash value of the training model (

register as

That is, in the ML model registration step (c), when the model formation is successful, the 'ML model type and architect' is checked and model training is started, and when the training model reaches the 'minimum performance score' or higher, training is completed. model training completion stage; And when the model training is completed, it is registered in the block chain, at this time registering a combination of a training model encrypted with an encryption key, a decryption key encrypted with the public key of the receiving builder, and a message including a hash value of the training model.

를 다운로드 하여 ML 모델을 다운로드하여 복호화 한다. 복화화 절차는 먼저 비밀키로 복호화키(s)를 복구한다.The ML model recovery step (d) is an encrypted message (

Download and decrypt the ML model. The decryption procedure first recovers the decryption key (s) with the secret key.

That is, the ML model recovery step (d) includes the step of restoring the training model by downloading the encrypted message registered in the block chain and restoring the training model with the decryption key decrypted with the private key of the receiving builder. .

를 복호화하여 해시 값을 생성하고, 생성된 해시 값과 메시지(

로부터 복호화된 해시 값(

을 비교하여 동일성을 확인하고 동일함을 보고함. 모든 빌더가 해시값의 동일함을 확인하고 동일하지 않음이 보고된 빌더는 그룹에서 제외하고, 제외된 빌더의 보증금은 기부금에 포함한다.In the ML model validation step (e), the model (

decrypts to generate a hash value, and the generated hash value and message (

hash value decrypted from

by comparing them to confirm the identity and report the identity. All builders confirm that the hash value is the same, and builders reporting non-identity are excluded from the group, and the deposit of the excluded builders is included in the donation.

과 메시지(

를 비교하여 동일 성을 확인하는 단계; 동일성이 확인되면 동일함을 그룹의 모든 빌더들에게 보고하고, 동일성이 확인되지 않으면 비동일함을 보고하는 단계; 그룹에서 비동일함을 보고한 빌더들이 한명이라도 있으면 모델의 개발자에게 재 등록을 요구하는 단계; 재 등록 하지 않거나, 재 등록된 모델을 재검사하여 동일성이 확인되지 않을 경우, 비동일 함을 최종 보고하는 단계; 및 최종 비동일함이 보고되는 빌더는 그룹에서 제외하고, 제외된 빌더의 보증금은 기부금으로 전환되는 단계를 포함한다.That is, the ML model validation step (e) generates a hash value of the restored model and includes the generated hash value and the hash value (

and message (

verifying identity by comparing them; Reporting the same to all builders in the group if the identity is confirmed, and reporting non-identity if the identity is not confirmed; requesting re-registration from the developer of the model if there is at least one builder reporting non-identity in the group; Final reporting of non-identity if not re-registration or re-inspection of re-registered models to confirm identity; And the final non-identity is reported builder is excluded from the group, and the deposit of the excluded builder is converted into a donation.

In the ML model performance reporting step (f), the model performance check is performed on the model for which the model validation has been completed, and for the model reporting the performance that does not reach the 'minimum performance score', the model developer is subject to a 'poor model penalty'. , and a 'bad data penalty' is given to builders who have inspected their performance. There is a possibility that performance data may be poor when proper performance is not obtained, and there are cases when false performance is reported even when proper performance is obtained. Therefore, a fine is imposed on the performance inspector, but the amount of 'bad data fine' is 'bad model fine'. Set it to be less than the amount of ', and 'penalty for bad data' is included in the donation.

In the ML task completion step (g), when all the work is completed, first, the amount of the builders who submitted the deposit is settled. The amount excluding the fine imposed on each individual is returned, and the amount collected from the fine is included in the donation. High-quality models that have passed the performance test enter the learning model improvement process through ensemble or retraining methods through model sharing.

Although the invention made by the present inventor has been described in detail according to the above embodiment, the present invention is not limited to the above embodiment, and it is common knowledge in the art that various changes can be made without departing from the gist of the present invention. self-evident to those who have

1: Contract Manager
2: ML Model Builder
3: ML model dataset
4: ML model
5: Blockchain

Claims (9)

As a distributed knowledge discovery system that provides blockchain-based data privacy,
A contract manager for registering a distributed data knowledge discovery contract and setting initial parameters, the contract manager having an application logic;
means an ML model developer with its own data, and an ML model builder with application logic that can join a group and register a model on the blockchain;
an ML model data set that is not disclosed for privacy as its own data owned by the ML model builder;
an ML model developed by the ML model builder using the ML model data set; and
It is a distributed ledger technology that cannot be tampered with and records are permanently maintained once registered, and a block chain having a contract, a software logic that ML builders registers ML models, and manages incentive systems, etc. A distributed knowledge discovery system that provides blockchain-based data privacy. As a distributed knowledge discovery method that provides blockchain-based data privacy,
(a) a distributed knowledge discovery contract initialization step corresponding to an initialization task for performing distributed knowledge discovery;
(b) forming a group of ML builders by recruiting ML builders;
(c) ML model registration step in which the registered ML builder encrypts the developed ML model using its own data and registers it in the blockchain;
(d) ML model recovery step of downloading and decrypting ML models registered by other ML builders belonging to the same group from the blockchain and recovering the ML models;
(e) an ML model validation step of validating through a hash whether the models registered by the ML model builder are all the same model;
(f) ML model performance reporting step of repairing the ML model generated by other ML builders to test the performance and report the result; and
(g) ML model task completion step in which all ML performance reporting is completed and deposit settlement is completed according to the corresponding procedure; Distributed knowledge discovery method providing block chain-based data privacy, comprising: a. 3. The method of claim 2,
In the (a) distributed knowledge discovery initialization step,
It is a parameter to initialize when registering a contract on the blockchain. When training in an ensemble approach after sharing a model, 'Amount of Builder's Deposit', 'Amount of Bad Model Fine', 'Amount of Bad Data Fine', 'Minimum Number of Builders', 'Maximum Number of Builders' providing a 'number of builders', a 'builder recruitment period', and a 'minimum performance score'; and providing 'type and architecture of the ML model' including layers, number of nodes, and kernel size as additional parameters when training with the retraining approach; providing data privacy based on block chain comprising: distributed knowledge discovery method. 3. The method of claim 2,
The (b) ML builder group formation step,
Builders who want to participate during the 'builder recruitment period' send the 'builder's guarantee amount' to the block chain as a deposit in cryptocurrency and receive an application deposit setting step;
If the number of builders who have set a deposit is less than the 'minimum number of builders' during the 'builder recruitment period', the ML builder group formation is canceled and the received deposit is all returned to the applicant. and
If the number of builders applied is greater than or equal to the 'minimum number of builders', the ML builder group is successfully formed, and even if the 'builder recruitment period' has not expired, if the 'maximum number of builders', the group's limited number of builders, complete the application, the ML builder group A distributed knowledge discovery method that provides data privacy based on a block chain, characterized in that it comprises a step of successful formation and completion. 3. The method of claim 2,
The (c) ML model registration step is,
When the model formation is successful, the 'ML model type and architect' is checked and model training is started, and when the training model reaches the 'minimum performance score' or higher, the training is completed; and
When the model training is completed, it is registered in the block chain, and at this time, registering a combination of the training model encrypted with the encryption key, the decryption key encrypted with the public key of the receiving builder, and a message including the hash value of the training model; A distributed knowledge discovery method that provides blockchain-based data privacy. 3. The method of claim 2,
The (d) ML model recovery step is,
In restoring the ML model by downloading the encrypted message registered in the block chain, it comprises the step of decrypting and restoring the training model with the decryption key decrypted with the receiving builder's private key. distributed knowledge discovery method. 3. The method of claim 2,
The (e) ML model validation step is,
Generate a hash value of the recovered model, and generate hash value and message (

The hash value contained in the message registered in the blockchain from

confirming identity by comparing them;
Reporting the same to all builders in the group if the identity is confirmed, and reporting non-identity if the identity is not confirmed;
requesting re-registration from the developer of the model if there is at least one builder reporting non-identity in the group;
Final reporting of non-identity if not re-registration or re-inspection of re-registered models to confirm identity; and
A distributed knowledge discovery method that provides data privacy based on a block chain, comprising: excluding builders reporting final non-identity from the group, and converting the excluded builders' deposits into donations. 3. The method of claim 2,
In the (f) ML model performance reporting step, a model performance check is performed on a model that has completed model validation, and a 'poor model penalty' for a model reporting a performance that does not reach the 'minimum performance score' and granting a 'bad data penalty' to the builders who have checked the performance;
There is a possibility that performance data may be poor when the performance is not good, and there are cases where the performance data is false even when the performance is correct, so a fine is imposed on the performance inspector, but the amount of the 'bad data fine' is the 'bad model fine' set less than the amount of '; and
The distributed knowledge discovery method for providing data privacy based on a block chain, characterized in that it includes; including the 'fine data penalty' in the donation. 3. The method of claim 2,
The (g) ML task completion step is,
When all the work is completed, first settling the amount of the builders who submitted the deposit;
Returning the amount excluding the fine granted to each and including the amount collected as a penalty in the donation; and
A distributed knowledge discovery method that provides data privacy based on a block chain, characterized in that the high-quality models that have passed the performance test enter the learning model improvement procedure through an ensemble or retraining method through model sharing.

Images