A blockchain-based artificial intelligence model training method and system

Abstract

The present application relates to the field of blockchain technology, and in particular to a blockchain-based artificial intelligence model training method and system, including: creating an initial artificial intelligence model in a common training node; based on the initial / current artificial intelligence model, the common training node sends a data read request to a common data node and a private connection node; the common data node organizes the corresponding common data, and the common mechanism node reaches a consensus on the common data; the common data after consensus is sent to the common training node; the private connection node organizes sensitive data from the private data node, and the private mechanism node reaches a consensus and desensitizes the sensitive data; the consensus and desensitized sensitive data are sent to the common training node; the common training node performs the next round of training on the initial / current artificial intelligence model based on the received common data and the desensitized sensitive data. The present application can avoid the leakage of sensitive data and ensure that the data involved in the training is relatively comprehensive.

Description

Technical Field

[0001] The present application relates to the field of blockchain technology, and in particular to a blockchain-based artificial intelligence model training method and system. Background Art

[0002] With the rapid development of artificial intelligence technology, the scale of artificial intelligence models is getting larger, the number of parameters is increasing, the number of participants is increasing, and the structure is becoming more and more complex, which makes it difficult for a single participant to train a corresponding artificial intelligence model according to its own needs.

[0003] Blockchain is an emerging information technology in recent years. It is a decentralized, trustless, collectively maintained, reliable database. The data or information stored in it has the characteristics of being unforgeable, traceable, traceable, open and transparent, and collectively maintained. Based on these characteristics, blockchain technology has laid a solid "trust" foundation, created a reliable "cooperation" mechanism, and has broad application prospects. With the continuous maturity of blockchain technology, using the blockchain's cooperation mechanism, multiple participants can work together to train artificial intelligence models, which can not only obtain more quantities and types of parameters for training artificial intelligence models, but also train artificial intelligence models with more complex structures and better performance.

[0004] However, some data of each participant is sensitive data. Once leaked, it will cause huge losses to the participants. Therefore, the participants usually do not share sensitive data with the outside world. This makes it difficult to fully utilize the data resources of each participant, which leads to the data involved in training the artificial intelligence model is not comprehensive enough, and it is difficult to train an artificial intelligence model with more complex structure and better performance.

[0005] Therefore, how to avoid the leakage of sensitive data of the participants to ensure that the data involved in the training of the artificial intelligence model is relatively comprehensive, so that the trained artificial intelligence model has a more complex structure and better performance, is a technical problem that technical personnel in this field urgently need to solve. Summary of the invention

[0006] The present application provides a blockchain-based artificial intelligence model training method and system to avoid the leakage of sensitive data of participants, to ensure that the data involved in training the artificial intelligence model is relatively comprehensive, so that the structure of the trained artificial intelligence model is more complex and the performance is better.

[0007] In order to solve the above technical problems, this application provides the following technical solutions:

[0008] A blockchain-based artificial intelligence model training method comprises the following steps: step S110, based on the artificial intelligence problem to be solved, creating an initial artificial intelligence model in a common training node of a public block community of the blockchain; step S120, based on the initial / current artificial intelligence model, the common training node sends a data read request to the common data node of the public block community of the blockchain and the private connection nodes of all private block communities of the blockchain; step S130, in response to receiving the data read request, the common data node of the public block community organizes the corresponding common data according to the data read request, and the common mechanism node of the public block community of the blockchain reads the organized common data consensus; step S140, the ordinary data node sends the common data that has been agreed upon to the common training node; step S150, in response to receiving the data read request, the private connection node of the private block community organizes the sensitive data of the participating parties from the private data node of the private block community according to the data read request, and the private mechanism node of the private block community of the blockchain reaches a consensus and desensitizes the sensitive data; step S160, the private connection node sends the consensus and desensitized sensitive data to the common training node; step S170, the common training node performs the next round of training on the initial / current artificial intelligence model based on the received common data and the desensitized sensitive data.

[0009] In the blockchain-based artificial intelligence model training method as described above, preferably, step S130 includes the following sub-steps: after receiving the data reading request from the common training node, the ordinary data node interprets the data reading request to obtain the training state parameters of the current artificial intelligence model; the ordinary data node obtains the data parameters of the ordinary data currently stored by the corresponding participant; the ordinary data node calculates the ordinary data provision parameters for the next round of training of the public block community based on the training state parameters of the current artificial intelligence model and the data parameters of the ordinary data currently stored by the participant; the ordinary data node selects the amount of ordinary data to be provided by all participants from the pre-created ordinary data provision standard library according to the ordinary data provision parameters for the next round of training of the public block community; the ordinary data node organizes the corresponding amount of ordinary data from the ordinary nodes based on the amount of ordinary data to be provided by all selected participants, and the ordinary mechanism node uses the public consensus mechanism therein to reach consensus on the organized ordinary data.

[0010] In the blockchain-based artificial intelligence model training method as described above, preferably, step S150 includes the following sub-steps: after receiving the data reading request from the common training node, the private connection node interprets the data reading request to obtain the training state parameters of the current artificial intelligence model; the private connection node obtains the sensitive parameters of the sensitive data currently stored by the affiliated participants; the private connection node calculates the sensitive data provision parameters for the next round of training of the private block community based on the training state parameters of the current artificial intelligence model and the sensitive parameters of the sensitive data currently stored by the affiliated participants; each private connection node selects the amount of sensitive data to be provided by the corresponding participant from the pre-created sensitive data provision standard library according to the sensitive data provision parameters for the next round of training of the private block community to which it belongs; the private connection node organizes the corresponding amount of sensitive data from the private data nodes of the corresponding private block community based on the amount of sensitive data to be provided by the selected affiliated participants, and the private mechanism node uses the private consensus mechanism therein to reach consensus on the organized sensitive data; and the private mechanism node uses the private processing mechanism therein to desensitize the organized sensitive data.

[0011] As described above, in the blockchain-based artificial intelligence model training method, preferably, the desensitizing processing includes the following sub-steps: the private mechanism node extracts the characteristic parameters of the organization's sensitive data; the private mechanism node calculates the desensitized data corresponding to the sensitive data based on the characteristic parameters of the organization's sensitive data and the private processing mechanism.

[0012] In the blockchain-based artificial intelligence model training method as described above, preferably, all participants jointly build a blockchain for training artificial intelligence models; the participants who initiate the joint training of artificial intelligence models create an initial artificial intelligence model in the common training node of the public block community of the blockchain based on the artificial intelligence problem to be solved.

[0013] A blockchain-based artificial intelligence model training system, comprising: a public block community and multiple private block communities; the public block community comprises: ordinary data nodes, ordinary mechanism nodes and common training nodes; the private block community comprises: private data nodes, private mechanism nodes and private connection nodes; according to the artificial intelligence problem to be solved, the common training node of the public block community of the blockchain creates an initial artificial intelligence model therein; according to the initial / current artificial intelligence model, the common training node sends a data read request to the ordinary data nodes of the public block community of the blockchain and the private connection nodes of all private block communities of the blockchain; in response to receiving the data read request, the ordinary data nodes of the public block community send a data read request to the private connection nodes of all private block communities of the blockchain according to the data read request. The request is received to organize the corresponding common data, and the common mechanism nodes of the public block community of the blockchain reach a consensus on the common data of the organization; the common data nodes send the common data that has been agreed upon to the common training nodes; in response to receiving the data reading request, the private connection nodes of the private block community organize the sensitive data of the participants from the private data nodes of the private block community according to the data reading request, and the private mechanism nodes of the private block community of the blockchain reach a consensus and desensitize the sensitive data; the private connection nodes send the consensus and desensitized sensitive data to the common training nodes; the common training nodes conduct the next round of training on the initial / current artificial intelligence model based on the received common data and the desensitized sensitive data.

[0014] In the blockchain-based artificial intelligence model training system as described above, preferably, after receiving the data reading request from the common training node, the ordinary data node interprets the data reading request to obtain the training state parameters of the current artificial intelligence model; the ordinary data node obtains the data parameters of the ordinary data currently stored by the corresponding participant; the ordinary data node calculates the ordinary data provision parameters for the next round of training of the public block community based on the training state parameters of the current artificial intelligence model and the data parameters of the ordinary data currently stored by the participant; the ordinary data node selects the amount of ordinary data to be provided by all participants from the pre-created ordinary data provision standard library according to the ordinary data provision parameters for the next round of training of the public block community; the ordinary data node organizes the corresponding amount of ordinary data from the ordinary nodes based on the amount of ordinary data to be provided by all selected participants, and the ordinary mechanism node uses the public consensus mechanism therein to reach consensus on the organized ordinary data.

[0015] In the blockchain-based artificial intelligence model training system as described above, preferably, after receiving the data reading request from the common training node, the private connection node interprets the data reading request to obtain the training state parameters of the current artificial intelligence model; the private connection node obtains the sensitive parameters of the sensitive data currently stored by the affiliated participants; the private connection node calculates the sensitive data provision parameters for the next round of training of the private block community based on the training state parameters of the current artificial intelligence model and the sensitive parameters of the sensitive data currently stored by the affiliated participants; each private connection node selects the amount of sensitive data to be provided by the corresponding participant from the pre-created sensitive data provision standard library according to the sensitive data provision parameters for the next round of training of the private block community to which it belongs; the private connection node organizes the corresponding amount of sensitive data from the private data nodes of the corresponding private block community based on the amount of sensitive data to be provided by the selected affiliated participants, and the private mechanism node uses the private consensus mechanism therein to reach consensus on the organized sensitive data; and the private mechanism node uses the private processing mechanism therein to desensitize the organized sensitive data.

[0016] In the blockchain-based artificial intelligence model training system as described above, preferably, the private mechanism node extracts characteristic parameters of the organization's sensitive data; the private mechanism node calculates the desensitized data corresponding to the sensitive data based on the characteristic parameters of the organization's sensitive data and the private processing mechanism.

[0017] In the blockchain-based artificial intelligence model training system as described above, preferably, all participants jointly build a blockchain for training artificial intelligence models; the participants who initiate the joint training of artificial intelligence models create an initial artificial intelligence model in the common training node of the public block community of the blockchain based on the artificial intelligence problem to be solved.

[0018] Compared with the above-mentioned background technology, the blockchain-based artificial intelligence model training method and system provided by this application can avoid the leakage of sensitive data of the participants and ensure that the data involved in the training of the artificial intelligence model is relatively comprehensive, thereby making the structure of the trained artificial intelligence model more complex and the performance better. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0020] Figure 1 It is a flowchart of the blockchain-based artificial intelligence model training method provided by this application;

[0021] Figure 2 It is a schematic diagram of the blockchain-based artificial intelligence model training system provided by this application. DETAILED DESCRIPTION

[0022] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be interpreted as limiting the present invention.

[0023] Embodiment 1

[0024] See also Figure 1 , Figure 1 This is a flowchart of the blockchain-based artificial intelligence model training method provided by this application.

[0025] This application provides an artificial intelligence model training method based on blockchain, comprising the following steps:

[0026] Step S110: creating an initial artificial intelligence model in a common training node of a public block community of the blockchain according to the artificial intelligence problem to be solved;

[0027] In order to jointly train the AI ​​model, all participants jointly built a blockchain for training the AI ​​model. In addition, a predetermined number of nodes are divided for each participant in the blockchain as private data nodes, private mechanism nodes, and private connection nodes belonging to the participant. The private data nodes, private mechanism nodes, and private connection nodes of each participant constitute the private block community of the participant. The remaining nodes are used as common data nodes, common mechanism nodes, and common training nodes shared by all participants. These common data nodes, common mechanism nodes, and common training nodes constitute the public block community of all participants.

[0028] The private data nodes of each private block community store the sensitive data of the corresponding participants, and the private mechanism nodes of each private block community store the private processing mechanism and private consensus mechanism belonging to the private block community. The common data nodes of the public block community store the common data of all participants, the common mechanism nodes of the public block community store the public consensus mechanism belonging to the public block community, and the common training nodes of the public block community store the created initial artificial intelligence model, and all participants will subsequently collaborate in training the initial artificial intelligence model in the common training nodes.

[0029] The participants who initiate the joint training of the artificial intelligence model create an initial artificial intelligence model in the common training node of the public block community of the blockchain based on the artificial intelligence problem to be solved (for example, how to predict public opinion about an event based on the evaluation data, browsing data, user data, etc. generated by various entertainment application platforms for an event. At this time, the sensitive data of the participants are various user data, while the evaluation data and browsing data are ordinary insensitive data). After the initial artificial intelligence model is trained, the trained artificial intelligence model can be used to make corresponding predictions.

[0030] Step S120: Based on the initial / current artificial intelligence model, the common training node sends a data read request to the common data nodes of the public block community of the blockchain and the private connection nodes of all private block communities of the blockchain;

[0031] Different artificial intelligence models require different data for training. Therefore, after the initial artificial intelligence model is created, the common training node will send a data read request to the ordinary data node of the public block community of the blockchain based on the data required for training the initial artificial intelligence model, so as to read the ordinary data of all participants stored in the ordinary data nodes of the public block community. It will also send a data read request to the private connection node of the private block community of the blockchain, so as to read the sensitive data of the corresponding participants stored in the private data node of the private block community of the blockchain. The private connection node is the only node for the private block community to which it belongs to communicate with the public block community, that is, any node of the public block community can be linked to the private connection node of each private block community.

[0032] After multiple rounds of training of the initial artificial intelligence model, the common training node will send data reading requests to the ordinary data nodes of the public block community of the blockchain based on the data required to train the current artificial intelligence model, and also send data reading requests to the private connection nodes of the private block community of the blockchain.

[0033] Step S130: In response to receiving the data reading request, the common data node of the public block community organizes the corresponding common data according to the data reading request, and the common mechanism node of the public block community of the blockchain reaches a consensus on the organized common data;

[0034] After receiving the data read request from the common training node, the common data node of the public block community interprets the data read request and obtains the training state parameters (A t1 , A t2 , A t3 …), where A t1 is the first training state parameter at the current time t, A t2 is the second training state parameter at the current time t, At3 is the third training state parameter at the current time t. For example, the training state parameter may be the number of model training rounds, the degree of model training completion, the amount of common data participating in the previous round of training, etc.

[0035] The common data nodes of the public block community also obtain the data parameters of the common data currently stored by the corresponding participants (B ti1 , B ti2 , B ti3 …), where B ti1 is the first data parameter of the common data stored by the ith participant at the current time t, B ti2 is the second data parameter of the common data stored by the ith participant at the current time t, B ti3 The third data parameter of the common data stored by the i-th participant at the current time t. For example, the data parameter may be the amount of common data, the data type of common data, etc.

[0036] The common data nodes of the public block community are based on the training status parameters (A t1 , A t2 , A t3 ...) and the data parameters of the common data currently stored by the participants (B ti1 , B ti2 , B ti3 …), calculate the parameters provided by the next round of training common data of the public block community Among them, A tn is the nth training state parameter at the current time t; μ n A tn The corresponding normal weight value; N is the number of training state parameters at the current time t; ε is the normal adjustment value of the training state parameter, which is a constant and takes the value of 1.21; σ t is the weight value of the current training round; B tim is the mth data parameter of the common data stored by the i-th participant at the current time t; β m For B tim The corresponding weight value; M is the amount of common data stored by the ith participant at the current time t; θ is the adjustment value of the common data, which is a constant and takes the value of 0.325; α i is the weight value of the ith participant; I is the number of participants; To round x upwards.

[0037] The common data node of the public block community calculates the common data providing parameter δ for the next round of training in the public block community. t+1 After that, the ordinary data nodes provide the parameter δ according to the next round of training ordinary data of the public block community. t+1The amount of common data to be provided by all participants is selected from a pre-created standard library of common data provision.

[0038] The ordinary data nodes of the public block community organize the corresponding amount of ordinary data from the ordinary nodes based on the amount of ordinary data to be provided by all selected participants, and then the ordinary mechanism nodes of the public block community of the blockchain apply the public consensus mechanism to reach a consensus on the organized ordinary data to verify whether the ordinary data has been tampered with.

[0039] Step S140: The common data node sends the common data that has been agreed upon to the common training node;

[0040] After the ordinary mechanism nodes have verified that the ordinary data organized from the ordinary data nodes have not been tampered with, the ordinary data nodes will send the organized ordinary data to the common training nodes for the next round of training of the artificial intelligence model.

[0041] Step S150: In response to receiving the data read request, the private connection node of the private block community organizes the sensitive data of the participant from the private data node of the private block community according to the data read request, and the private mechanism node of the private block community of the blockchain performs consensus and desensitization processing on the sensitive data;

[0042] After receiving the data read request from the common training node, the private connection node of the private block community interprets the data read request and obtains the training state parameters (A t1 , A t2 , A t3 …), where A t1 is the first training state parameter at the current time t, A t2 is the second training state parameter at the current time t, A t3 is the third training state parameter at the current time t. For example, the training state parameter may be the number of model training rounds, the degree of model training completion, the amount of common data participating in the previous round of training, etc.

[0043] The private connection node of the private block community also obtains the sensitive parameters of the sensitive data currently stored by the participating party (C tj1 , C tj2 , C tj3 …), where C tj1 is the first sensitive parameter of the sensitive data stored by the jth participant at the current time t, C tj2 is the second sensitive parameter of the sensitive data stored by the jth participant at the current time t, C tj3The third sensitive parameter of the sensitive data stored by the j-th participant at the current time t. For example, the sensitive parameter may be a sensitive data type, a sensitivity level, etc.

[0044] Private connection nodes are based on the training status parameters (A t1 , A t2 , A t3 …) and the sensitive parameters of the sensitive data currently stored by the participant (C tj1 , C tj2 , C tj3 …), calculate the sensitive data parameters for the next round of training of the private block community Among them, ωjt+1 is the parameter provided by the jth private block community for the next round of training sensitive data, j∈1, J, J is the number of private block communities in the blockchain, which is also the number of participants; Atn is the nth training state parameter at the current time t; θ n A tn Corresponding sensitive weight value; N is the number of training state parameters at the current time t; γ is the sensitive adjustment value of the training state parameter, which is a constant and takes the value of 6.05; σ t is the weight value of the current training round; C tjs is the sth sensitive parameter of the sensitive data stored by the jth participant at the current time t; ρ s C tjs The corresponding weight value; S is the amount of sensitive data stored by the jth participant at the current time t; Q is the adjustment value of sensitive data, which is a constant and takes the value of 0.265; α j is the weight value of the jth participant; Z zb is the standard correction value; To round x upwards.

[0045] Calculate the sensitive data provided by the next round of training of all private block communities in the private connection node to provide parameters ω j(t+1) After that, each private connection node provides parameter ω according to the sensitive data of the next round of training of the private block community to which it belongs. j(t+1) The amount of sensitive data to be provided by the corresponding party is selected from the pre-created sensitive data provision standard library.

[0046] The private connection node organizes the corresponding amount of sensitive data from the private data nodes of the corresponding private block community based on the amount of sensitive data to be provided by the selected participants. Then the private mechanism nodes of the private block community of the blockchain apply the private consensus mechanism to reach a consensus on the organized sensitive data to verify whether the sensitive data has been tampered with.

[0047] After determining that these sensitive data have not been tampered with, the private mechanism nodes in the private block community of the blockchain apply the private processing mechanism therein to desensitize the sensitive data of the organization, so as to reduce the risk of sensitive data leakage.

[0048] Specifically, the private mechanism nodes extract the characteristic parameters of the sensitive data of the organization {(MG j11 、MG j12 …), (MG j21 、MG j22 …)…}; where MG j11 is the first characteristic parameter of the first sensitive data of the j-th participant of the organization, MG j12 is the second characteristic parameter of the first sensitive data of the j-th participant of the organization, MG j21 is the first characteristic parameter of the second sensitive data of the j-th participant of the organization, and MG j22 is the second characteristic parameter of the second sensitive data of the j-th participant of the organization.

[0049] The private mechanism nodes calculate the desensitized data corresponding to the sensitive data based on the characteristic parameters of the sensitive data of the organization and the private processing mechanism. Specifically, the private processing mechanism is reflected in the following formula, that is, the private mechanism nodes desensitize and calculate the desensitized data of the j-th participant of the organization according to the formula where MG is the z-th characteristic parameter of the y-th sensitive data of the j-th participant of the organization; R jyz is the random number corresponding to MG z ; k is an adjustable factor, 0 < k < 1, and k gradually changes from 1 to 0 to achieve a smooth splicing of the overlapping area of the characteristic parameter and the random number; Z is the number of characteristic parameters of the sensitive data; E jyz is the weight value of the y-th sensitive data of the j-th participant of the organization; Y is the number of sensitive data of the j-th participant of the organization; τ jy is the privacy parameter, which is the constant 1.523. *

[0050] Step S160: The private connection nodes send the consensus and desensitized sensitive data to the co-training nodes;

[0051] After the private mechanism nodes verify that the sensitive data organized from the private data nodes has not been tampered with and desensitize the sensitive data organized from the private data nodes to obtain desensitized data, each private connection node in the private block community sends the desensitized data to the co-training nodes in the public block community for the next round of training of the artificial intelligence model.

[0052] ​Step S170: The common training node performs the next round of training on the initial / current artificial intelligence model based on the received common data and the desensitized sensitive data;

[0053] After the common training nodes of the public block community receive the ordinary data and the desensitized data, they conduct the next round of training on the initial artificial intelligence model in the common training node / the current artificial intelligence model after several rounds of training based on these data until the artificial intelligence model is trained to the required accuracy, thereby relying on the participation and collaboration of multiple parties in the blockchain to complete the training of the artificial intelligence model.

[0054] Embodiment 2

[0055] See also Figure 2 , Figure 2 This is a schematic diagram of the blockchain-based artificial intelligence model training system provided by this application.

[0056] The present application provides a blockchain-based artificial intelligence model training system 200, including: a public block community 210 and multiple private block communities 220; the public block community 210 includes: ordinary data nodes 211, ordinary mechanism nodes 212 and common training nodes 213; the private block community 220 includes: private data nodes 221, private mechanism nodes 222 and private connection nodes 223.

[0057] Depending on the artificial intelligence problem to be solved, the common training node 213 of the public block community 210 of the blockchain creates an initial artificial intelligence model therein.

[0058] In order to jointly train the AI ​​model, all participants jointly built a blockchain for training the AI ​​model. In addition, a predetermined number of nodes are divided for each participant in the blockchain as private data nodes, private mechanism nodes, and private connection nodes belonging to the participant. The private data nodes, private mechanism nodes, and private connection nodes of each participant constitute the private block community of the participant. The remaining nodes are used as common data nodes, common mechanism nodes, and common training nodes shared by all participants. These common data nodes, common mechanism nodes, and common training nodes constitute the public block community of all participants.

[0059] The private data nodes of each private block community store the sensitive data of the corresponding participants, and the private mechanism nodes of each private block community store the private processing mechanism and private consensus mechanism belonging to the private block community. The common data nodes of the public block community store the common data of all participants, the common mechanism nodes of the public block community store the public consensus mechanism belonging to the public block community, and the common training nodes of the public block community store the created initial artificial intelligence model, and all participants will subsequently collaborate in training the initial artificial intelligence model in the common training nodes.

[0060] The participants who initiate the joint training of the artificial intelligence model create an initial artificial intelligence model in the common training node of the public block community of the blockchain based on the artificial intelligence problem to be solved (for example, how to predict public opinion about an event based on the evaluation data, browsing data, user data, etc. generated by various entertainment application platforms for an event. At this time, the sensitive data of the participants are various user data, while the evaluation data and browsing data are ordinary insensitive data). After the initial artificial intelligence model is trained, the trained artificial intelligence model can be used to make corresponding predictions.

[0061] Based on the initial / current artificial intelligence model, the common training node 213 sends a data read request to the ordinary data nodes 211 of the public block community 210 of the blockchain and the private connection nodes 223 of all private block communities 220 of the blockchain.

[0062] Different artificial intelligence models require different data for training. Therefore, after the initial artificial intelligence model is created, the common training node will send a data read request to the ordinary data node of the public block community of the blockchain based on the data required for training the initial artificial intelligence model, so as to read the ordinary data of all participants stored in the ordinary data nodes of the public block community. It will also send a data read request to the private connection node of the private block community of the blockchain, so as to read the sensitive data of the corresponding participants stored in the private data node of the private block community of the blockchain. The private connection node is the only node for the private block community to which it belongs to communicate with the public block community, that is, any node of the public block community can be linked to the private connection node of each private block community.

[0063] After multiple rounds of training of the initial artificial intelligence model, the common training node will send data reading requests to the ordinary data nodes of the public block community of the blockchain based on the data required to train the current artificial intelligence model, and also send data reading requests to the private connection nodes of the private block community of the blockchain.

[0064] In response to receiving the data reading request, the ordinary data node 211 of the public block community 210 organizes the corresponding ordinary data according to the data reading request, and the ordinary mechanism node 212 of the public block community 210 of the blockchain reaches a consensus on the organized ordinary data.

[0065] After receiving the data read request from the common training node, the common data node of the public block community interprets the data read request and obtains the training state parameters (A t1 , A t2 , A t3 …), where A t1 is the first training state parameter at the current time t, A t2 is the second training state parameter at the current time t, A t3 is the third training state parameter at the current time t. For example, the training state parameter may be the number of model training rounds, the degree of model training completion, the amount of common data participating in the previous round of training, etc.

[0066] The common data nodes of the public block community also obtain the data parameters of the common data currently stored by the corresponding participants (B ti1 , B ti2 , B ti3 …), where B ti1 is the first data parameter of the common data stored by the ith participant at the current time t, B ti2 is the second data parameter of the common data stored by the ith participant at the current time t, B ti3 The third data parameter of the common data stored by the i-th participant at the current time t. For example, the data parameter may be the amount of common data, the data type of common data, etc.

[0067] The common data nodes of the public block community are based on the training status parameters (A t1 , A t2 , A t3 ...) and the data parameters of the common data currently stored by the participants (B ti1 , B ti2 , B ti3 …), calculate the parameters provided by the next round of training common data of the public block community Among them, A tn is the nth training state parameter at the current time t; μ n A tn The corresponding normal weight value; N is the number of training state parameters at the current time t; ε is the normal adjustment value of the training state parameter, which is a constant and takes the value of 1.21; σ t is the weight value of the current training round; B timis the mth data parameter of the common data stored by the i-th participant at the current time t; β m For B tim The corresponding weight value; M is the amount of common data stored by the ith participant at the current time t; θ is the adjustment value of the common data, which is a constant and takes the value of 0.325; α i is the weight value of the ith participant; I is the number of participants; To round x upwards.

[0068] The common data node of the public block community calculates the common data providing parameter δ for the next round of training in the public block community. t+1 After that, the ordinary data nodes provide the parameter δ according to the next round of training ordinary data of the public block community. t+1 The amount of common data to be provided by all participants is selected from a pre-created standard library of common data provision.

[0069] The ordinary data nodes of the public block community organize the corresponding amount of ordinary data from the ordinary nodes based on the amount of ordinary data to be provided by all selected participants, and then the ordinary mechanism nodes of the public block community of the blockchain apply the public consensus mechanism to reach a consensus on the organized ordinary data to verify whether the ordinary data has been tampered with.

[0070] The common data node 211 sends the common data that has been agreed upon to the common training node 213 .

[0071] After the ordinary mechanism nodes have verified that the ordinary data organized from the ordinary data nodes have not been tampered with, the ordinary data nodes will send the organized ordinary data to the common training nodes for the next round of training of the artificial intelligence model.

[0072] In response to receiving the data reading request, the private connection node 223 of the private block community 220 organizes the sensitive data of the participants from the private data node 221 of the private block community 220 according to the data reading request, and the private mechanism node 222 of the private block community 220 of the blockchain performs consensus and desensitization processing on the sensitive data.

[0073] After receiving the data read request from the common training node, the private connection node of the private block community interprets the data read request and obtains the training state parameters (A t1 , A t2 , A t3 …), where A t1 is the first training state parameter at the current time t, A t2 is the second training state parameter at the current time t, A t3is the third training state parameter at the current time t. For example, the training state parameter may be the number of model training rounds, the degree of model training completion, the amount of common data participating in the previous round of training, etc.

[0074] The private connection node of the private block community also obtains the sensitive parameters of the sensitive data currently stored by the participating party (C tj1 , C tj2 , C tj3 …), where C tj1 is the first sensitive parameter of the sensitive data stored by the jth participant at the current time t, C tj2 is the second sensitive parameter of the sensitive data stored by the jth participant at the current time t, C tj3 The third sensitive parameter of the sensitive data stored by the j-th participant at the current time t. For example, the sensitive parameter may be a sensitive data type, a sensitivity level, etc.

[0075] Private connection nodes are based on the training status parameters (A t1 , A t2 , A t3 …) and the sensitive parameters of the sensitive data currently stored by the participant (C tj1 , C tj2 , C tj3 …), calculate the sensitive data parameters for the next round of training of the private block community Among them, ω j(t+1) Provide parameters for the next round of training sensitive data for the jth private block community, j∈[1,J], where J is the number of private block communities in the blockchain and also the number of participants; A tn is the nth training state parameter at the current time t; θ n A tn Corresponding sensitive weight value; N is the number of training state parameters at the current time t; γ is the sensitive adjustment value of the training state parameter, which is a constant and takes the value of 6.05; σ t is the weight value of the current training round; C tjs is the sth sensitive parameter of the sensitive data stored by the jth participant at the current time t; ρ s C tjs The corresponding weight value; S is the amount of sensitive data stored by the jth participant at the current time t; Q is the adjustment value of sensitive data, which is a constant and takes the value of 0.265; α j is the weight value of the jth participant; Z zb is the standard correction value; To round x upwards.

[0076] Calculate the sensitive data provided by the next round of training of all private block communities in the private connection node to provide parameters ωj(t+1) After that, each private connection node provides the parameter ω for the sensitive data in the next round of training according to the private block community to which it belongs. j(t+1) Select the amount of sensitive data to be provided by the corresponding party from the pre-created standard library of sensitive data providers.

[0077] Based on the selected amount of sensitive data to be provided by the corresponding party, the private connection node organizes the corresponding amount of sensitive data from the private data nodes of the corresponding private block community. Then, the private mechanism node in the private block community of the blockchain applies the private consensus mechanism therein to conduct consensus on the organized sensitive data to verify whether these sensitive data have been tampered with.

[0078] After determining that these sensitive data have not been tampered with, the private mechanism node in the private block community of the blockchain applies the private processing mechanism therein to desensitize the organized sensitive data to reduce the risk of sensitive data leakage.

[0079] Specifically, the private mechanism node extracts the feature parameters of the organized sensitive data {(MG j11 、MG j12 …)、(MG j21 、MG j22 …)…}; where MG j11 is the first feature parameter of the first sensitive data of the j-th party in the organization, MG j12 is the second feature parameter of the first sensitive data of the j-th party in the organization, MG j21 is the first feature parameter of the second sensitive data of the j-th party in the organization, and MG j22 is the second feature parameter of the second sensitive data of the j-th party in the organization.

[0080] The private mechanism node calculates the desensitized data corresponding to the sensitive data based on the feature parameters of the organized sensitive data and the private processing mechanism. Specifically, the private processing mechanism is reflected in the following formula, that is, the private mechanism node calculates the desensitized data of the j-th party in the organization according to the formula by desensitization calculation. where MG jyz is the z-th feature parameter of the y-th sensitive data of the j-th party in the organization; R z is the random number corresponding to MG jyz ; k is an adjustable factor, 0 < k < 1, and k gradually changes from 1 to 0 to achieve smooth splicing of the overlapping area between the feature parameters and the random number; Z is the number of feature parameters of the sensitive data; E jy is the weight value of the y-th sensitive data of the j-th party in the organization; Y is the number of sensitive data of the j-th party in the organization; τ * is the privacy parameter, which is the constant 1.523.

[0081] The private connection node 223 sends the consensus-based and desensitized sensitive data to the common training node 213 .

[0082] After the private mechanism nodes have verified that the sensitive data organized from the private data nodes have not been tampered with, and the sensitive data organized from the private data nodes have been desensitized to obtain desensitized data, the private connection nodes of each private block community will send the desensitized data to the common training nodes of the public block community for the next round of training of the artificial intelligence model.

[0083] The common training node 213 performs the next round of training on the initial / current artificial intelligence model based on the received normal data and the desensitized sensitive data.

[0084] After the common training nodes of the public block community receive the ordinary data and the desensitized data, they conduct the next round of training on the initial artificial intelligence model in the common training node / the current artificial intelligence model after several rounds of training based on these data until the artificial intelligence model is trained to the required accuracy, thereby relying on the participation and collaboration of multiple parties in the blockchain to complete the training of the artificial intelligence model.

[0085] Since each participant in the present application has his own private block community, and each participant's private block community is linked to the public block community shared by all participants only through a private connection node, the security of the data stored in the private data nodes of the private block community is guaranteed, and the sensitive data in the private data nodes of the private blocks must be desensitized before participating in the training of the artificial intelligence model in the common training node of the public block community, further ensuring the security of the sensitive data in the private data nodes. Therefore, the present application can avoid the leakage of the participants' sensitive data, and can also ensure that the data involved in the training of the artificial intelligence model is relatively comprehensive, so that the structure of the trained artificial intelligence model is more complex and the performance is better.

[0086] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0087] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A blockchain-based artificial intelligence model training method, characterized in that: The steps include: Step S110: creating an initial artificial intelligence model in a common training node of a public block community of the blockchain according to the artificial intelligence problem to be solved; Step S120: Based on the initial / current artificial intelligence model, the common training node sends a data read request to the common data nodes of the public block community of the blockchain and the private connection nodes of all private block communities of the blockchain; Step S130: In response to receiving the data reading request, the common data node of the public block community organizes the corresponding common data according to the data reading request, and the common mechanism node of the public block community of the blockchain reaches a consensus on the organized common data; Step S130 includes the following sub-steps: After receiving the data reading request from the common training node, the common data node interprets the data reading request and obtains the training state parameters of the current artificial intelligence model; The common data node obtains the data parameters of the common data currently stored by the corresponding participant; The general data node calculates the general data provision parameters for the next round of training of the public block community based on the training status parameters of the current artificial intelligence model and the data parameters of the general data currently stored by the participants; The ordinary data node selects the amount of ordinary data to be provided by all participants from the pre-created ordinary data provision standard library according to the ordinary data provision parameters for the next round of training of the public block community; The ordinary data nodes organize the corresponding amount of ordinary data from the ordinary nodes according to the amount of ordinary data to be provided by all the selected participants, and the ordinary mechanism nodes use the public consensus mechanism to reach consensus on the organized ordinary data; Step S140: The common data node sends the common data that has been agreed upon to the common training node; Step S150: In response to receiving the data read request, the private connection node of the private block community organizes the sensitive data of the participant from the private data node of the private block community according to the data read request, and the private mechanism node of the private block community of the blockchain performs consensus and desensitization processing on the sensitive data; Step S160: The private connection node sends the consensus-based and desensitized sensitive data to the common training node; Step S170: The common training node performs the next round of training on the initial / current artificial intelligence model based on the received normal data and the desensitized sensitive data.

2. The blockchain-based artificial intelligence model training method according to claim 1 is characterized in that: Step S150 includes the following sub-steps: After receiving the data reading request from the common training node, the private connection node interprets the data reading request and obtains the training state parameters of the current artificial intelligence model; The private connection node obtains the sensitive parameters of the sensitive data currently stored by the participant to which it belongs; The private connection node calculates the sensitive data provision parameters for the next round of training of the private block community based on the training status parameters of the current artificial intelligence model and the sensitive parameters of the sensitive data currently stored by the participant. Each private connection node selects the amount of sensitive data to be provided by the corresponding participant from the pre-created sensitive data provision standard library according to the sensitive data provision parameters for the next round of training of the private block community to which it belongs; The private connection node organizes the corresponding amount of sensitive data from the private data nodes of the corresponding private block community according to the amount of sensitive data to be provided by the selected participants, and the private mechanism node uses the private consensus mechanism to reach consensus on the organized sensitive data; And the private mechanism node applies the private processing mechanism therein to desensitize the organization's sensitive data.

3. The blockchain-based artificial intelligence model training method according to claim 2 is characterized in that: Desensitization treatment includes the following sub-steps: The private mechanism node extracts characteristic parameters of the organization’s sensitive data; The private mechanism node calculates the desensitized data corresponding to the sensitive data based on the characteristic parameters of the organization's sensitive data and the private processing mechanism.

4. The blockchain-based artificial intelligence model training method according to any one of claims 1 to 3, characterized in that: All parties jointly build a blockchain for training AI models; Participants who initiate joint training of AI models create initial AI models in the joint training nodes of the public block community of the blockchain based on the AI ​​problems to be solved.

5. An artificial intelligence model training system based on blockchain, characterized in that: include: A public block community and multiple private block communities; The public block community includes: ordinary data nodes, ordinary mechanism nodes and common training nodes; the private block community includes: private data nodes, private mechanism nodes and private connection nodes; Depending on the AI ​​problem to be solved, the common training nodes of the blockchain’s public block community create an initial AI model within it; Based on the initial / current AI model, the co-training node sends data read requests to the common data nodes of the public block community of the blockchain and the private connection nodes of all private block communities of the blockchain; In response to receiving the data read request, the common data node of the public block community organizes the corresponding common data according to the data read request, and the common mechanism node of the public block community of the blockchain reaches a consensus on the organized common data; After receiving the data reading request from the common training node, the common data node interprets the data reading request and obtains the training state parameters of the current artificial intelligence model; The common data node obtains the data parameters of the common data currently stored by the corresponding participant; The general data node calculates the general data provision parameters for the next round of training of the public block community based on the training status parameters of the current artificial intelligence model and the data parameters of the general data currently stored by the participants; The ordinary data node selects the amount of ordinary data to be provided by all participants from the pre-created ordinary data provision standard library according to the ordinary data provision parameters for the next round of training of the public block community; The ordinary data nodes organize the corresponding amount of ordinary data from the ordinary nodes according to the amount of ordinary data to be provided by all the selected participants, and the ordinary mechanism nodes use the public consensus mechanism to reach consensus on the organized ordinary data; Ordinary data nodes send the common data that has been agreed upon to the common training nodes; In response to receiving the data read request, the private connection node of the private block community organizes the sensitive data of the party to which it belongs from the private data node of the private block community according to the data read request, and the private mechanism node of the private block community of the blockchain performs consensus and desensitization processing on the sensitive data; The private connection node sends the consensus-based and desensitized sensitive data to the common training node; The joint training node conducts the next round of training for the initial / current artificial intelligence model based on the received normal data and desensitized sensitive data.

6. The blockchain-based artificial intelligence model training system according to claim 5 is characterized in that: After receiving the data reading request from the common training node, the private connection node interprets the data reading request and obtains the training state parameters of the current artificial intelligence model; The private connection node obtains the sensitive parameters of the sensitive data currently stored by the participant to which it belongs; The private connection node calculates the sensitive data provision parameters for the next round of training of the private block community based on the training status parameters of the current artificial intelligence model and the sensitive parameters of the sensitive data currently stored by the participant. Each private connection node selects the amount of sensitive data to be provided by the corresponding participant from the pre-created sensitive data provision standard library according to the sensitive data provision parameters for the next round of training of the private block community to which it belongs; The private connection node organizes the corresponding amount of sensitive data from the private data nodes of the corresponding private block community according to the amount of sensitive data to be provided by the selected participants, and the private mechanism node uses the private consensus mechanism to reach consensus on the organized sensitive data; And the private mechanism node applies the private processing mechanism therein to desensitize the organization's sensitive data.

7. The blockchain-based artificial intelligence model training system according to claim 6 is characterized in that: The private mechanism node extracts characteristic parameters of the organization's sensitive data; the private mechanism node calculates the desensitized data corresponding to the sensitive data based on the characteristic parameters of the organization's sensitive data and the private processing mechanism.

8. The blockchain-based artificial intelligence model training system according to any one of claims 5 to 7, characterized in that: All participants jointly build a blockchain for training artificial intelligence models; participants who initiate joint training of artificial intelligence models create an initial artificial intelligence model in the joint training node of the blockchain's public block community based on the artificial intelligence problem to be solved.

Images