CN113065143A - Block chain based secure sharing of industrial data - Google Patents

Abstract

The invention realizes a privacy protection sharing scheme of industrial big data based on a block chain, and establishes an alliance chain among system users. And considering a replacement scheme of cloud storage based on expandability and easiness in deployment. And establishing an incentive mechanism based on the evolutionary game theory, constructing an intelligent contract and encouraging the data sharer to share data. In addition, decentralized machine learning models are built based on federal learning to respect privacy concerns when sharing data, and blockchain-based architectures allow training data to be kept secret, distributed, and shared within the system. In this way, all participating users can observe a unified model while preserving privacy of sensitive input data. On the basis, an industrial data sharing platform based on the block chain is designed and realized.

Description

Block chain based secure sharing of industrial data

One, the technical field

The invention relates to the field of application industry, in particular to a privacy protection sharing scheme for realizing industrial big data based on a block chain.

Second, background Art

The industry is the foundation and the pillar of national economy and is also an important mark of economic strength and competitiveness of a country. With the development of industrial technology, various production devices and sensors of industrial enterprises generate data all the time. The industrial big data is a product of combination of the internet, the big data and the industrial industry and is the footfall of national strategies such as the industrial internet, the industrial 4.0 and the like in enterprises. In recent years, under the situation that industrial internet is vigorously developed in China, deep fusion of manufacturing industry and internet is promoted, and high-quality development of manufacturing industry is promoted, large data development of China presents a explosive growth situation. Research institute data shows that the market scale of the global big data is $ 454 billion in 2018, the industrial big data accounts for more than 50% of the total scale of the global big data, and the industrial big data is seen to become a main field for the development of the global big data industry.

The industrial big data has the characteristics of high real-time performance, large data volume, low density, strong data source isomerism and the like. The industrial process requires an industrial analysis model with high precision, high reliability and strong causal relationship, so that the requirement of daily industrial production can be met, and the pure data-driven data analysis means can not meet the requirement of an industrial scene. Therefore, a feature of industrial big data analysis is to emphasize the deep fusion of domain-of-expertise and data mining.

With the expansion of industrial scale, in order to realize data analysis and application in an industrial scene, the ship height of the data demand can rise, and how to establish a data security sharing platform becomes a key research trend. The key for realizing industrial big data sharing is to guarantee the privacy and safety of industrial enterprise data and establish an open data transaction platform. In a traditional mode, data such as operation, production and maintenance records of equipment are stored in a cloud storage mode, namely, a centralized server, but the mode has many hidden dangers, for example, a single node is failed to bring the risk of user data loss, and in addition, due to lack of strong data protection measures, the possibility of data file theft and tampering may occur in the data storage and transmission processes.

Blockchains are distributed ledger techniques that combine intelligent contracts and other ancillary techniques under network consensus. The core advantage of the block chain technology is decentralization, and by means of data encryption, timestamps, distributed consensus, economic incentive and the like, network nodes can realize trusted point-to-point transaction, coordination and cooperation in a distributed system without mutual trust, so that the problems of high cost, low efficiency, unsafe data storage and the like universally existing in a centralization mechanism are solved. When a blockchain is introduced into industrial big data sharing, the following advantages can be brought:

1) data between upstream and downstream of an industrial chain are linked up through a block chain, so that ecological sharing of core enterprises, mutual trust sharing between industrial enterprises and value sharing between industrial internet platforms are facilitated, and the barrier that data between enterprises are not mutually trusted is broken. The block chain technology is utilized to provide a corresponding solution for the challenges of production cooperation, industrial safety, information sharing, resource fusion, flexible supervision and the like in the promotion of industrial networked production.

2) The block chain has credibility, safety and non-tamper property. Any records written on the distributed ledger may not be deleted or changed. This means that the user cannot delete a record or insert a record in the blockchain, effectively guaranteeing the rights of all parties.

3) Based on the decentralized concept of the block chain technology, distributed control and storage are utilized, and through a shared economic mode, stock storage and computing resources on a live network are utilized, so that the data storage and operation and maintenance pressure of an industrial big data operator is effectively relieved, and data mining and value increment are further effectively realized.

Third, the invention

The invention realizes a privacy protection sharing scheme of industrial big data based on a block chain, and establishes an alliance chain among system users. And considering a replacement scheme of cloud storage based on expandability and easiness in deployment. And establishing an incentive mechanism based on the evolutionary game theory, constructing an intelligent contract and encouraging the data sharer to share data. In addition, decentralized machine learning models are built based on federal learning to respect privacy concerns when sharing data, and blockchain-based architectures allow training data to be kept secret, distributed, and shared the generated models within the system.

The technical scheme adopted by the invention is as follows:

the invention mainly provides a data sharing model and a data sharing excitation model based on a block chain and an evolutionary game theory.

The data sharing model based on the block chain is BCDSM, and the model ensures the openness, auditable tracking and tamper resistance in the system data sharing transaction process. The BCDSM model mainly comprises a system initialization stage, a data request processing stage, a model updating stage, a data sharing stage and the like.

And in the system initialization stage, a decentralized network of a block chain is constructed, data providers Pi upload data summary information of the data providers and public keys PKi of the data providers, the system distributes unique Identifiers (ID) of the data providers for registration, and meanwhile, initial credit coins are distributed to each data provider to serve as transaction currency in the data sharing process.

In the data request stage, a data requester R issues a data request Req to a blockchain system, requested data description information is attached to the request to indicate the type, size and the like of data required by the data requester R, and meanwhile, the local original model M of the data requester R is used₀Also uploaded into the blockchain.

In the data request processing stage, the system executes a retrieval mechanism of block data information based on a similarity matching algorithm according to the data request Req, queries and locates related data providers P { P ═ P₁,P₂,…,P_n}. Forwarding the data request Req to the phases by task distributionData off provides P ═ { P₁,P₂,…,P_n}。

In the model updating phase, relevant data providers P ═ P of the data sharing request Req are received₁,P₂,…，P_nAnd (4) forming a federated learning community, and running a federated learning algorithm by members in the community according to the data request Req to iteratively train the same machine learning model until the model is fitted.

In the data sharing stage, community members execute a consensus mechanism to select a certain data provider P_iThe data provider P_iPackaging the data sharing transaction information into blocks, adding the blocks into a block chain system, broadcasting the whole network for block information verification, and using the public key PK of the data requester R_rAnd encrypting the trained model M and sending the encrypted model M to a data requester R. The data requestor R accepts the final machine learning model M and the data sharing transaction is completed.

The data sharing excitation model based on the block chain and the evolutionary game theory is BC-EGT, and the specific steps are as follows:

the system distributes an initial credit currency to the newly added user; for a certain user initiating a data request Req, the user can request for announcement in a block chain system and explains the credit currency which can be earned; after the system receives the Req, the provider { P of the related data₁,P₂,…,P_nSending the request Req; data provider P_iCalculating the income generated under different decisions by combining the evolutionary game theory; calculating P during model training for Federal learning_iContribution value C_i(ii) a After the transaction is completed, according to C_iTo P_iThe resulting credit is assigned.

Description of the drawings

Fig. 1 BCDSM model overall framework diagram.

Fig. 2 is a block diagram of a BCDSM data block.

Fig. 3 data request and sharing process.

FIG. 4 shows a specific process of FL-SMPC-DPP.

Fifth, detailed description of the invention

The following describes embodiments of the present invention in further detail with reference to the accompanying drawings.

As shown in fig. one, the BCDSM model of the present description mainly consists of five stages, i.e., a system initialization stage, a data request processing stage, a model update stage, and a data sharing stage. The blockchain type used in the model is first introduced and the block structure is designed in detail. A data sharing process of the system, a data retrieval process of model design and a similarity matching algorithm are introduced. The BCDSM model relates to distributed machine learning in a data sharing phase, and provides a data privacy protection algorithm FL-SMPC-DPP based on federated learning and SMPC aiming at problems in the data sharing phase. The method specifically comprises the following steps:

step 1, in order to better guarantee system security, in combination with the particularity of an application scene, a block chain type used by a model is an alliance block chain, the disclosure degree of the block chain is between a public chain and a private chain, the block chain is only used by specific alliances, a plurality of specified organizations or organizations participate in management together, and transaction data are recorded together; in order to effectively guarantee data privacy and safety, the block chain of the alliance is responsible for retrieving related data and recording transaction information, and users do not need to upload own original data information; in the BCDSM model, aiming at a user who firstly joins in a block chain of the alliance, a system distributes a unique Identity (ID) for the user to register, and meanwhile, the user needs to upload data summary information of the user, such as data type, data size and the like; each chunk in the federation chain is responsible for recording data digest information, request records, and shared records, as shown in fig. 2.

Step 2, the data request sharing process of the system is as shown in fig. 3, and for a specific data sharing request Req, a request is issued from a data requester to a system processing request, and the final data sharing is completed, which is specifically divided into the following steps:

(1) a data requester R issues a data sharing request Req on a block chain;

(2) performing a specific data retrieval algorithm on the data summary information on the blockchain according to the data request record, locating the data provider associated with the data request ReqSet P ═ P₁,P₂,…,P_n}；

(3) The system forwards the data sharing request Req to P ═ { P ═ P₁,P₂，…,P_nEach member of (1);

(4) each data provider P_iResponding to the data sharing request Req to form a federated learning community C;

(5) each member in the community C executes a federal learning algorithm, trains the same machine learning model until the model fits or the iteration times are reached;

(6) after model training is completed, at data provider P ═ P₁,P₂,…,P_nExecuting a consensus algorithm, uploading the data sharing transaction to a block chain by a specific node SN, and broadcasting the data sharing transaction to the whole network for verification so as to achieve consensus;

(7) the SN uses the public key PK of the data requestor R_rEncrypting the final model M after the completion of the learning training of the Nippon;

(8) the SN transmits the encrypted machine learning model to a data requester R;

(9) the data requester R receives the encrypted file, decrypts the encrypted file by using a private key of the data requester R, and acquires a shared model M;

(10) the data sharing transaction is complete.

Step 3, when a data sharing request occurs, in order to reduce the time consumption of the system in the data sharing process and improve the accuracy of the data sharing model, the system considers that the data request is not directly forwarded to all data holders in the system, but multi-party data retrieval is executed on a block chain, similarity matching is carried out on data request information and data description information uploaded by the data holders, and a specific data requester set P { P ═ P is screened by combining the matching similarity result₁,P₂,…,P_nAnd locating a related data provider, wherein the matching stage is mainly realized in a text clustering mode and specifically comprises the following steps:

(1) cutting words and removing stop words from the text;

(2) calculating text features and constructing a VSM (vector space model);

(3) clustering was performed using the K-means algorithm.

And 4, in the model updating stage, in order to effectively guarantee the data privacy of the user, carrying out model training of distributed machine learning by using a federal learning technology. The federated learning can ensure that all users participating in model training do not need to exchange data, but use local data sets to jointly learn the same machine learning model, and the realization effect of the final model is equivalent to the optimal model established by aggregating the data together.

Further, the VSM vector space model of step 3 is difficult to be processed by a computer because chinese text information is mostly unstructured and uses natural language. It is therefore desirable to pre-process the text to represent it in a mathematically analyzable form so that the machine can identify and convert the text similarity problem described above to a mathematical vector matrix problem. A Vector Space Model (VSM) is adopted to represent texts, a given text is converted into a vector, a feature item is used as a basic unit of text representation, each dimension of the vector corresponds to one feature item in the text, and each dimension represents the weight value of the corresponding feature item in the text. The weight value represents the importance degree of the feature item to the text, namely how much the feature item can reflect the category of the document in which the feature item is located. The VSM is defined in part as follows:

characteristic items: the feature item is the smallest inseparable language unit in the VSM, and may be a word, a phrase or a phrase, etc. The content of a text is regarded as a set of feature items it contains, expressed as: t (T)₁，t₂,…，t_n) Wherein t is_kIs a characteristic term, k is more than or equal to 1 and less than or equal to n.

Weight of term: for a text T (T) containing n feature items₁,t₂,…,t_n) Each feature term t_kAre given a weight w according to certain principles_kIndicating how important they are in the text. Such a text T can be represented by the feature items it contains and the weights corresponding to the feature items: text ═ T: (t₁,w₁；t₂,w₂；…；t_n,w_n) Wherein w is_kIs the feature item t_kThe weight of (c).

For a given Text T (T)₁,w₁；t₂,w₂；…；t_n,w_n) The following two conditions are satisfied:

(1) each feature item t_k(1. ltoreq. k. ltoreq. n) are different (no repetition is present);

(2) each feature item t_kNonsequential relationships (i.e., without regard to the internal structure of the document)

Satisfying the above two condition constraints, and combining the feature item t₁,t₂,…,t_nViewed as an n-dimensional coordinate system, and the weight w₁,w₂,…,w_nAre the corresponding coordinate values, and thus, a text is represented as a vector in an n-dimensional space.

Further, the federal learning in step 4 is the data provider P participating in the federal learning_iForming a federal community C, wherein each member in the community has own local data set D_iAnd need not be sent to other members. The goal of federal learning is to have all data sets D ═ D in the community₁∪D₂，…，∪D_nTraining the same machine learning model to realize the process, wherein a given loss function F (w) needs to find a parameter w which minimizes the model precision, and the target function F (w) is as follows:

wherein, F_i(w) is a data provider P_iAt its data set D_iThe loss function of (a) to (b),

wherein f is_j(h (w, x), y) is the data entry with model parameter w at jth(x_i,y_j) The value of the loss function of (c).

The objective of the above mentioned federated learning mechanism can be expressed as an optimization problem, for a given loss function F_i(w), finding the parameter w that minimizes the loss of accuracy of the model prediction, requires minimizing the ratio of f (w):

w (T) is the set of model parameters aggregated in the T-th iteration, T is the maximum number of iterations, and w (T) is calculated by the formula:

wherein, Δ w_iIs the model parameter that the user pi updated in the t-th iteration.

To efficiently achieve a minimization of the loss function, the objective function is minimized in a gradient-decreasing manner, i.e. the parameters are updated in the negative direction of the gradient of the objective function- Δ f (w), i.e. for each data provider P_iThe parameters are adjusted using the following formula in the local training process to find the optimal model parameters.

The quality of model training was demonstrated by using the Mean Absolute Error (MAE) of the model, with lower MAE indicating better quality of the trained model:

further, FL-SMPC-DPP, which is a data privacy protection algorithm based on federated learning and SMPC in step 4, is an improvement on the traditional federated learning algorithm, and the parameter summary of the third party organization is removed, and instead, secure multiparty computation (SMPC) is performed between data providers to achieve the aggregation of model parameters. At the heart of SMPC is that multiple holds own private data<x₁，x₂,…x_n>The participants of (1) perform a function of pre-agreed consensus and obtain the calculation result without depending on a trusted third party, as follows:

f(x1,x2,…,xn)＝(y1,y2,…,yn)

in the calculation process, each party involved does not leak the calculation of respective data, after the calculation is completed, the two parties can know the calculation result, but do not know the data of the other party and the intermediate data in the calculation process, and the specific process is as shown in fig. 4.

Claims (2)

1. The data pricing incentive mechanism based on the evolutionary game theory is characterized in that: the game theory is a mathematical method for researching and predicting social interaction evolution, and can be used for enabling an individual to become rational and then analyzing policy selection and game balance of the individual; establishing an incentive strategy, wherein the strategy depends on a user with limited information acquisition capacity, and gradually adjusts the strategy of the user through continuous learning and repeated attempts; under the development of the industrial internet, edge devices need to rely on data cooperation and exchange among the devices for realizing better intelligent application, and need to share data of each data holder for realizing industrial big data analysis, and as the data cannot be guaranteed, an excitation mechanism based on a block chain needs to be established to promote an intelligent contract of data sharing and help promote the interconnection and intercommunication characteristics of the industrial internet; the specific strategy is as follows: by combining the idea of the evolutionary game theory, the participation of users in data sharing is promoted by dynamically adjusting the incentive/participation cost, the users are encouraged to join in the alliance chain for model training, better model precision is obtained, meanwhile, the concept of credit value is introduced, the transmission of invalid and harmful data in the network is reduced, and the digital signature is used for encrypting the transaction to guarantee privacy.

2. The data sharing model based on the federal learning and the SMPC is characterized in that: centralized approaches to machine learning often suffer from the scalability of ever-expanding data sets and lack of privacy safeguards for the customers providing the data; the distributed machine learning method can realize the localization training of the model, overcome the problem of insufficient expansibility of the centralized machine learning model, and simultaneously can avoid the safety problem caused by data sharing; the invention relates to a method for realizing distributed machine learning, which aims to adopt the federal learning as an implementation scheme of a model and integrate the federal learning into a alliance chain; federal learning allows multiple data owners to collaboratively train a global model without sharing their raw data to respect privacy concerns when sharing data, and blockchain-based architectures allow training data to be kept secret, distributed, and shared within the system without sharing actual data; and provides audit trails, tracking changes during the learning process, in this way, all participating users can observe a unified model, while preserving privacy of sensitive input data.

Images