Federal forest credible participant screening and auditing method and system

Abstract

The invention discloses a federal forest trusted participant screening and auditing method and system. The method comprises the steps that a model consumer puts forward a model service demand to a cloud platform; the cloud platform issues tasks; when a certain participant Pi participates in the task for the first time, registration is completed at the LGPA; the cloud platform performs identity authentication on the Pi according to the received task request message of the participant Pi; after the participant Pi passes the verification, the cloud platform determines whether the request of the participant is passed or not; the cloud platform and a plurality of participants authorized through the task requests cooperate together to complete construction of a random forest tree model; the model consumer submits rewards to the cloud platform, and the cloud platform gives a result and feeds back the result; the cloud platform performs reputation evaluation on the participant Pi through a subjective logic model to update the reputation level; the cloud platform allocates remuneration according to the latest reputation level of the participant Pi; compared with a traditional method, the reliability of participants in the federated learning system can be improved.

Description

Technical Field

[0001] The present invention relates to a trusted participant screening and auditing method and system, and in particular to a trusted participant screening and auditing method and system for a federal forest, belonging to the technical fields of artificial intelligence, federated learning and data privacy protection. Background Art

[0002] With the widespread application of artificial intelligence and big data analysis technologies in the fields of autonomous driving, smart security, and smart medical care, data sharing and data security issues have become hot issues in the field. On the one hand, federated learning solves the "data island" problem and achieves the dual purpose of local data retention and remote model sharing; on the other hand, it avoids the excessive loss of communication resources and privacy leakage caused by migrating a large amount of local terminal data to a remote cloud platform, thus further promoting the vigorous development of artificial intelligence.

[0003] Despite the huge advantages of federated learning, it still faces some serious challenges. First, local participants are unwilling to sacrifice a lot of resources, including data, computing, communication, etc., especially to participate in learning tasks without any rewards or benefits, including further continuous and efficient participation in learning tasks. Second, publicly sharing model parameters and gradient updates still faces the risk of local data and privacy leakage of participants. In order to avoid this risk, local participants refuse to cooperate to a certain extent.

[0004] In order to motivate more local participants to participate in federated learning, researchers have designed various incentive mechanisms, such as auction theory, contract theory, and game theory. These mechanisms focus more on the participants' contributions in terms of data volume, computing cost, communication cost, training time, and model accuracy, and pay little attention to the reliability of the participants. As the main entity of federated learning, the quality of data and the reliability of model parameters provided by local participants directly affect the performance of federated learning. Therefore, it is of great significance to study the reliability evaluation of participants. Summary of the invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a method and system for screening and reviewing trusted participants in a federated forest that can improve the reliability of participants in a federated learning system.

[0006] Technical solution: The method for screening and auditing federal forest credible participants described in the present invention includes:

[0007] (1) Model consumers submit model service requirements to the cloud platform;

[0008] (2) The cloud platform publishes tasks to the local participant group manager LGPA and candidate participants in the field in a broadcasting manner;

[0009] (3) When a participant Pi When participating in a task for the first time, complete the registration at LGPA, join the task group, and obtain the group member private key from LGPA through the interactive protocol;

[0010] (4) The cloud platform receives the information of the participant P i The task request message to P i Perform identity authentication and determine P i The legitimacy and authenticity of the credit rating;

[0011] (5) When the participant P i After verification, the cloud platform extracts the participant P from the task request i The credibility level of the participant is determined, and whether to approve the participant's request is determined based on the current model task volume;

[0012] (6) The cloud platform and several participants who have obtained authorization through task requests work together to complete the construction of the random forest tree model;

[0013] (7) Model consumers pay remuneration to the cloud platform, and the cloud platform provides services based on the results of model construction, gives results and feedback;

[0014] (8) The cloud platform uses the subjective logic model to identify the participant P i Conduct credit assessment, record the latest credit value, and update the credit rating;

[0015] (9) The cloud platform is based on the participant P i The most recent credit rating reward distribution.

[0016] Furthermore, the step (2) also includes: the cloud platform pre-sets a reputation threshold; when the reputation value of a participant is lower than the reputation threshold, the participant is regarded as a malicious user.

[0017] Furthermore, the step (3) is specifically as follows:

[0018] (31) When a participant P i When you want to participate in a task, you need to send a private message to LGPA and obtain task authorization; the private information includes the sample size, feature quantity and local computing resource information of the data set that can be used for local model training. LGPA will assign a task authorization to the new participant P based on the private information. i Assigning an initial reputation value;

[0019] (32) LGPA calculation participant P i The hash value of the identity and its latest reputation value is used to identify the participant P based on the hash value. i Generate a task request message and send the group member's private key msk i After encryption, the data is fed back to the cloud platform.

[0020] Furthermore, the step (32) is specifically as follows:

[0021] LGPA first calculates the hash value h i =H(P i |R(P i )), the participant P i Its latest reputation value R(P i ) binding, where H is a one-way hash function;

[0022] Then LGPA is calculated by Trm i ={h i ,L(P i )} skGM P i Construct task request Trm i ; where L(P i ) is the participant P i credibility rating.

[0023] Furthermore, the step (4) is specifically as follows: the participant P i Send task request R anonymously to the cloud platform i , in By P i A randomly generated pseudonym for anonymously submitting task requests R i ; YesP i Use group member private key msk i The obtained group signature, Trm i It is a task request; through the task request information, the cloud platform completes the task request. i Anonymous authentication of identity to verify whether it has been registered with LGPA; specifically, the cloud platform uses LGPA’s public key pk GM and Trm i , verify the group signature Is it correct? If correct, judge P i Is a legitimate group member and further verifies the task request Trm i Whether it is signed by LGPA, and verify whether the reputation information is P reviewed and recorded by LGPA i The latest reputation value R(P i ) and the corresponding credit level L(P i ).

[0024] Furthermore, the step (5) is specifically as follows:

[0025] Participant P i After the identity verification is legal, the cloud platform requests Trm from the task iExtract its credit rating L(P i ), and decide whether to pass P based on the minimum reputation level and number of participants specified in the task requirements. i Task request;

[0026] When P i If the minimum credit rating is met and the number of users currently executing the task is less than the specified number, the cloud platform will authorize the task to the user and store Trm i , and give P i Send authorization information in, For cloud platform i Mission authorization information, It is an anonymous reputation certificate ARC, used by the participant P i Anonymously indicate the reputation level to the cloud platform during the trust assessment phase;

[0027] Otherwise, feedback to P i A rejection message ReM i :

[0028] Furthermore, the step (6) is specifically as follows: CS sends a randomly selected feature subset F′ to LGPA and all participants in the group, F is a set of all features, let κ represent the index of the feature in subset F′, κ∈{1, 2, Λ, |F′|}; local participant P i Recommend the best candidate splits based on the selected feature data κ∈{1, 2, Λ, |F′|}, the optimal candidate split is given by P i Based on the Fayyad boundary point determination theorem, the boundary points that minimize the GINI coefficient are selected, where the local data set D i Split into several sub-class sets based on the k-th feature GINI coefficient for 1≤j≤l(l≥2) For feature data P not in F′ i No analysis will be done when a feature data in F′ is missing locally. LPGA collects the best split candidates for the κth feature from all participants in the group to form the best candidate split set i∈{1,2,ΛN}, N is the number of participants; then LPGA randomly selects a member P from the group j The private key msk j For S κ Encrypt and send to CS; CS to received Decrypt to get S κ , for S κAll elements in are evaluated to obtain the best split with the highest quality and will Stored as a new random decision tree RDT node in a fixed specified format; then CS encrypts the optimal split And feedback to LPGA; for eligible group participants, storage Take it as the new node pair local sample set D i Based on feature data Divide; if ι∈{1,2,Λ,|D i |}, then the corresponding samples are divided into the left subtree set Otherwise, partition into the right subtree set

[0029] Finally, the cloud platform Delete the κth feature from F′, iterate all the above steps to assign the corresponding samples to the left subtree set or right subtree set Until the boundary conditions are met: trd c >trd max Or F′ is an empty set, trd max Indicates the maximum depth of the tree, trd c Indicates the current depth of the tree.

[0030] Furthermore, step (7) also includes, when the participant has no local resources or data, proposing a demand to purchase resources and services from the cloud platform.

[0031] Furthermore, in step (8), the cloud platform adopts a subjective logic model to determine the i To conduct reputation evaluation, specifically, the cloud platform uses a subjective logic model to evaluate the participant P i The reputation of the participant P i In the t round of training, the optimal or near-optimal split is selected to minimize the GINI coefficient, which is considered positive, and the non-optimal or non-near-optimal split is selected to fail to minimize the GINI coefficient, which is considered negative, which is neg. At this time, the trust relationship is Non-trust relationship is And T ti +T di =1,T ti ,T di ∈[0,1]; get the participant P i Reputation R(LP i ) is expressed as R(LP i )=λT di +βT ti , the historical credit t-1 round R t-1 (LP i) and current event reputation t (LP i ) acts on the current credit round R t (LP i ), update R(LP i ) is R t (LP i )=αR t-1 (LP i )+(1-α)r t (LP i );

[0032] The updating of the reputation level also includes, when P in step (5) i If the task request is not approved, this information will be fed back to the GM, and the GM will know that P i Not authorized, keep P i The reputation value remains unchanged.

[0033] Based on the same inventive concept, the present invention also provides a federal forest trusted participant screening and review system, including:

[0034] The task publishing module is used by the cloud platform to collect model service requirements from model consumers and publish tasks to the local participant group manager LGPA and candidate participants in the field in a broadcast form;

[0035] The registration module is used when a participant P i When participating in a task for the first time, complete the registration at LGPA, join the task group, and obtain the group member private key from LGPA through the interactive protocol;

[0036] The identity authentication module is used by the cloud platform to i The task request message to P i Perform identity authentication and determine P i The legitimacy and authenticity of the credit rating;

[0037] The reputation audit module is used when the participant P i After verification, the cloud platform extracts the participant P from the task request i The credibility level of the participant is determined, and whether to approve the participant's request is determined based on the current model task volume;

[0038] The model building module is used for the cloud platform and several participants who have been authorized through task requests to jointly complete the construction of the random forest tree model;

[0039] The result prediction module is used by model consumers to submit remuneration to the cloud platform. The cloud platform provides services based on the results of model construction, gives results and feedback;

[0040] The evaluation update module is used by the cloud platform to evaluate the participant P through the subjective logic model. i Conduct credit assessment, record the latest credit value, and update the credit rating;

[0041] The reward distribution module is used by the cloud platform to allocate rewards according to the participants’ P i The most recent credit rating reward distribution.

[0042] Beneficial effects: Compared with the prior art, the present invention can effectively improve the reliability of participants in the federated learning system through an effective method for evaluating the reputation and reputation level of participants; and the task publisher has the initiative to judge and screen local participants who are more capable of completing this round of tasks based on the reputation level, thereby improving the efficiency of completing learning tasks and the quality of the model; the present invention also constructs an audit mechanism based on the joining agreement, which dynamically manages and distributes remuneration to the participants of the federated learning task based on the reputation level, making the rewards more fair, generating positive incentives for credible participants, and also effectively suppressing unreliable or malicious participants. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the workflow for reviewing and screening trusted local participants by a cloud platform in an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0046] As attached Figure 1 As shown, the federal forest trusted participant screening and review method of this embodiment includes:

[0047] (1) Model consumers submit model service requirements to the cloud platform;

[0048] (2) The cloud platform publishes tasks to the local participant group manager LGPA and candidate participants in the field in a broadcasting manner;

[0049] (3) When a participant P i When participating in a task for the first time, complete the registration at LGPA, join the task group, and obtain the group member private key from LGPA through the interactive protocol;

[0050] (4) The cloud platform receives the information from the participant Pi The task request message to P i Perform identity authentication and determine P i The legitimacy and authenticity of the credit rating;

[0051] (5) When the participant P i After verification, the cloud platform extracts the participant P from the task request i The credibility level of the participant is determined, and whether to approve the participant's request is determined based on the current model task volume;

[0052] (6) The cloud platform and several participants who have obtained authorization through task requests work together to complete the construction of the random forest tree model;

[0053] (7) Model consumers pay the cloud platform a fee, and the cloud platform provides services based on the results of model construction, giving results and feedback;

[0054] (8) The cloud platform uses the subjective logic model to identify the participant P i Conduct credit assessment, record the latest credit value, and update the credit rating;

[0055] (9) The cloud platform is based on the participant P i The most recent credit rating reward distribution.

[0056] Specifically, the FL system consists of four entities: local participants (LP), local participant group administrator (LPGA), cloud server (CS) and model consumer (MC). The cloud platform and several authorized local participants work together to complete the construction of the random forest tree model. The LP in the group is numbered from 1 to N, and the number of authorized local participants is M, M≤N. Local LP i Provide dataset D i , 1≤i≤N participate in the global model construction.

[0057] Among them, Figure 2 As shown, in step (1), MC proposes model service requirements to CS.

[0058] In step (2), CS publishes the task to LGPA in the form of broadcast. In order to improve the accuracy of the collaborative learning of the system, it is necessary to ensure the honesty of the participants. CS pre-sets the system parameter: the reputation threshold. If the reputation value is lower than the threshold, the participant is regarded as a malicious user.

[0059] Step (3) Participants register, when local P i When participating in a task for the first time, you must first register to apply to be included in the LGPA task group, and obtain the group member private key msk from LGPA through the interactive protocol i .

[0060] 1) If the local Pi The participants who are willing to participate in this learning task can send a private message to LGPA and obtain task authorization. The private information includes the sample size, feature quantity, local computing resources and other information of the dataset that can be used for local model training. LGPA will give the newly joined P i Assign an initial reputation value R(P i 0 ).

[0061] 2) LGPA calculation P i The hash value of the identity bound to its latest reputation value is used to generate a task request message for the participant, and the group member's private key msk i After encryption, it is fed back to CS. The specific process is as follows: First, h is calculated using the single-term hash function LGPA i =H(P i |R(P i )), the identity P i Its latest reputation value R(P i ) binding. Then, LGPA constructs a task request Where L(P i ) is P i The latest credit rating of i To implement anonymous reputation evaluation, i With the help of L(P i ) proves its credibility level to CS without leaking its specific credibility value.

[0062] Step (4) CS based on Trm i P i Perform identity authentication to determine whether the identity is legitimate, L(P i ) is true and valid. The specific process is as follows: i Send task request R anonymously to CS i , in, By P i A randomly generated pseudonym for anonymously submitting task requests R i ; YesP i Use group member private key msk i The obtained group signature. CS uses Trm i Complete P i Anonymous authentication of identity to verify whether it is a legitimate participating user. CS receives R i After that, first anonymously authenticate P i The specific method is: CS uses LGPA's public key pk GM and Trm i , verify the group signature Is it correct? If it is correct, then P i Is a legitimate group member. At the same time, the task request Trm can be further verified i It is signed by LGPA and the credit information is recorded by LGPA i The latest reputation value R(P i ) and the corresponding credit level L(P i ).

[0063] Step (5) When the participant P i If the minimum credit rating requirement of the cloud platform is met and the number of participants currently executing the task is less than the specified number, the cloud platform will authorize the task to the participant P i , send authorization information to it as an anonymous credibility credential;

[0064] When the participant P i If the minimum credit rating requirement of the cloud platform is not met and the number of participants currently executing the task is greater than or equal to the specified number, the cloud platform will reject the participant P i 's joining and feedback the rejection message.

[0065] Specifically, P i After the identity verification is passed, CS extracts L(P i ), and decide whether to approve the participant's request based on the current model task volume. There are two specific cases:

[0066] Case 1: P i If the minimum credit level requirement of CS is met and the number of participants currently executing the task is less than the specified number, the cloud platform will authorize the task to the participant and send it authorization information as an anonymous credit certificate. It is also the basis for the third-party regulatory agency to update the credit value of the current participant based on this message.

[0067] Case 2: P i If the minimum credit rating requirement of CS is not met and the number of participants currently executing the task is greater than or equal to the specified number, CS will reject the participant's joining and feedback a rejection message. CS decides whether to update the current participant's credit value based on this message. The specific process is as follows: i After the identity verification is legal, CS will i Extract L(P i ) and decide whether to approve the request based on the task requirements. The task requirements specify the minimum credit level and number of participating users. If a user meets the minimum credit level and the number of users currently executing the task is less than the specified number, the task will be authorized to the user. i , store Trm i After giving P iSend authorization information For cloud platform i Mission authorization information, It is an anonymous credit certificate ARC, P i In the subsequent trust evaluation phase, ARC is used to anonymously prove the reputation level to the self; on the contrary, if CS rejects P i , will give P i Feedback rejection information ReM i : In the reputation update phase, P i This information needs to be fed back to LGPA. LGPA receives this information and knows P i If it is not authorized, its reputation value remains unchanged.

[0068] Step (6) Model training phase: CS and several authorized LPs work together to complete the global model construction. The specific process is as follows: CS and several authorized local participants work together to complete the construction of the random forest tree model. First, CS sends a randomly selected feature subset F′ to LGPA and all participants in the group. F is a set containing all features, and κ represents the index of the feature in the subset F′, κ∈{1, 2, Λ, F′}. Then, P i Recommend the best candidate splits based on the selected feature data κ∈{1, 2, Λ, |F′|}, the best candidate split is given by P i Based on the Fayyad boundary point determination theorem, the boundary points that minimize the GINI coefficient are selected, where the local data set D i Split into several sub-class sets based on the k-th feature , GINI coefficient for 1≤j≤l(l≥2) For feature data P not in F′ i No analysis will be done, and if a feature data in F′ is missing locally LPGA collects the best split candidates for the κth feature from all participants in the group to form the best candidate split set i∈{1,2,ΛN}, N is the number of participants. Then, LPGA randomly selects a member’s private key msk from the group j For S κ Encrypt and send to CS. CS to the received Decrypt to get S κ , for S κ All elements in are evaluated to obtain the best split with the highest quality And store it in a fixed specified format as a new random decision tree RDT node. After that, CS encrypts the optimal split and feedback to the LPGA. ) of the group participants Take it as the new node pair local sample set D i Based on feature data If ι∈{1,2,Λ,|D i |}, then the corresponding samples are divided into the left subtree set On the contrary, partition into the right subtree set

[0069] Finally, CS Delete the κth feature from F′ and iterate the above process for the left and right subtree sets. and Continue dividing until the boundary conditions are met: trd c >trd max Or F′ is an empty set, here, trd max Indicates the maximum depth of the tree, trd c Indicates the current depth of the tree.

[0070] Based on the above process, LP and CS collaborated to complete the construction of the random forest tree model.

[0071] In step (7), during the model prediction phase, the MC or LP (some participants do not have local resources or data resources and need to purchase resources and services from the cloud platform) submits prediction requirements and remuneration to the CS. The CS provides services based on the results of the model training phase, gives prediction results and feedback.

[0072] Step (8) Reputation evaluation and update phase, use the subjective logic model to evaluate the reputation of LP, and record and update the most recent reputation value and reputation level update. The specific process of reputation evaluation is: CS, as the publisher of the task, uses the subjective logic model to evaluate the reputation of LP. During the t-round training, LP chooses to provide the optimal or near-optimal split that minimizes the GINI coefficient and is identified as positive (pos), and provides the non-optimal or non-near-optimal split that fails to minimize the GINI coefficient and is identified as negative (neg). Then the trust relationship is Non-trust relationship is and i T tid +T=1,T ti ,T di ∈[0,1]. Based on this, the reputation R(P i ) is expressed as R(P i )=λT di +βT ti , considering the historical reputation t-1 round R t-1 (P i) and current event reputation t (P i ) acts on the current credit round R t (P i ), update R(P i ) is R t (P i )=αR t-1 (P i )+(1-α)r t (P i ).

[0073] Step (9) Remuneration distribution stage, CS distributes remuneration based on LP’s latest credit rating.

[0074] Based on the same inventive concept, the present invention also provides a federal forest trusted participant screening and review system, including:

[0075] The task publishing module is used by the cloud platform to collect model service requirements from model consumers and publish tasks to the local participant group manager LGPA and candidate participants in the field in a broadcast form;

[0076] The registration module is used when a participant P i When participating in a task for the first time, complete the registration at LGPA, join the task group, and obtain the group member private key from LGPA through the interactive protocol;

[0077] The identity authentication module is used by the cloud platform to i The task request message to P i Perform identity authentication and determine P i The legitimacy and authenticity of the credit rating;

[0078] The reputation audit module is used when the participant P i After verification, the cloud platform extracts the participant P from the task request i The credibility level of the participant is determined, and whether to approve the participant's request is determined based on the current model task volume;

[0079] The model building module is used for the cloud platform and several participants who have been authorized through task requests to jointly complete the construction of the random forest tree model;

[0080] The result prediction module is used by model consumers to submit remuneration to the cloud platform. The cloud platform provides services based on the results of model construction, gives results and feedback;

[0081] The evaluation update module is used by the cloud platform to evaluate the participant P through the subjective logic model. i Conduct reputation assessment, record the latest reputation value, and update the reputation level;

[0082] The reward distribution module is used by the cloud platform to allocate rewards according to the participants’ P i The most recent credit rating reward distribution.

Claims

1. A method for screening and auditing federal forest trusted participants, characterized in that: include: (1) Model consumers submit model service requirements to the cloud platform; (2) The cloud platform publishes tasks to the local participant group manager LGPA and candidate participants in the field in a broadcasting manner; (3) When a participant P i When participating in a task for the first time, complete the registration at LGPA, join the task group, and obtain the group member private key from LGPA through the interactive protocol; (4) The cloud platform receives the information of the participant P i The task request message to P i Perform identity authentication and determine P i The legitimacy and authenticity of the credit rating; (5) When the participant P i After verification, the cloud platform extracts the participant P from the task request i The credibility level of the participant is determined, and whether to approve the participant's request is determined based on the current model task volume; (6) The cloud platform and several participants who have obtained authorization through task requests work together to complete the construction of the random forest tree model; (7) Model consumers pay remuneration to the cloud platform, and the cloud platform provides services based on the results of model construction, gives results and feedback; (8) The cloud platform uses the subjective logic model to identify the participant P i Conduct reputation assessment, record the latest reputation value, and update the reputation level; (9) The cloud platform is based on the participant P i The most recent credit rating reward distribution.

2. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that: The step (2) also includes: the cloud platform pre-sets a reputation threshold; when the reputation value of a participant is lower than the reputation threshold, the participant is regarded as a malicious user.

3. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that , the step (3) is specifically as follows: (31) When a participant P i When you want to participate in a task, you need to send a private message to LGPA and obtain task authorization; the private information includes the sample size, feature quantity and local computing resource information of the data set that can be used for local model training. LGPA will assign a task authorization to the new participant P based on the private information. i Assigning an initial reputation value; (32) LGPA calculation participant P i The hash value of the identity and its latest reputation value is used to identify the participant P based on the hash value. i Generate a task request message and send the group member's private key msk i After encryption, the data is fed back to the cloud platform.

4. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that , the step (32) is specifically as follows: LGPA first calculates the hash value hi = H(Pi|R(Pi)) and binds the participant Pi to its latest reputation value R(Pi), where H is a one-way hash function; Then LGPA calculated P i Construct task request Trm i ; where L(P i ) is the participant P i credibility rating.

5. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that , the step (4) is specifically as follows: Participant P i Send task request R anonymously to the cloud platform i , in By P i A randomly generated pseudonym for anonymously submitting task requests R i ; YesP i Use group member private key msk i The obtained group signature, Trm i Request for tasks; Through task request information, the cloud platform completes the i Anonymous authentication of identity to verify whether it has been registered with LGPA; specifically, the cloud platform uses LGPA’s public key pk GM and Trm i , verify the group signature Is it correct? If correct, judge P i Is a legitimate group member and further verifies the task request Trm i Whether it is signed by LGPA, and verify whether the reputation information is P reviewed and recorded by LGPA i The latest reputation value R(P i ) and the corresponding credit level L(P i ).

6. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that , the step (5) is specifically as follows: When the participant P i If the minimum credit rating requirement of the cloud platform is met and the number of participants currently executing the task is less than the specified number, the cloud platform will authorize the task to the participant P i , send authorization information to it as an anonymous credibility credential; When the participant P i If the minimum credit rating requirement of the cloud platform is not met and the number of participants currently executing the task is greater than or equal to the specified number, the cloud platform will reject the participant P i Join and feedback the rejection message; Participant P i After the identity verification is legal, the cloud platform requests Trm from the task i Extract its credit rating L(P i ), and decide whether to pass P based on the minimum reputation level and number of participants specified in the task requirements. i Task request; When P i If the minimum credit rating is met and the number of users currently executing the task is less than the specified number, the cloud platform will authorize the task to the user and store Trm i , and give P i Send authorization information in, For cloud platform i Mission authorization information, It is an anonymous reputation certificate ARC, used by the participant P i Anonymously indicate the reputation level to the cloud platform during the trust assessment phase; Otherwise, feedback to P i A rejection message ReM i :

7. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that ,The step (6) is specifically as follows: CS sends the randomly selected feature subset F′ to LGPA and all participants in the group. F is a set of all features, let κ represent the index of the feature in subset F′, κ∈{1, 2, Λ, F′}; local participant P i Recommend the best candidate splits based on the selected feature data The optimal candidate split is given by P i Based on the Fayyad boundary point determination theorem, the boundary points that minimize the GINI coefficient are selected, where the local data set D i Based on the k-th feature, it is split into several sub-class sets D i j , GINI coefficient for 1≤j≤l(l≥2) For feature data P not in F′ i No analysis will be done when a feature data in F′ is missing locally. LPGA collects the best split candidates for the κth feature from all participants in the group to form the best candidate split set N is the number of participants; LPGA then randomly selects a member P from the group j The private key msk j For S κ Encrypt, send to CS; CS to received Decrypt to get S κ , for S κ All elements in are evaluated to obtain the best split with the highest quality and will Stored as a new random decision tree RDT node in a fixed specified format; then CS encrypts the optimal split And feedback to LPGA; for eligible group participants, storage Take it as the new node pair local sample set D i Based on feature data Divide; if Then divide the corresponding samples into the left subtree set Otherwise, partition into the right subtree set Finally, the cloud platform Delete the κth feature from F′, iterate all the above steps to assign the corresponding samples to the left subtree set or right subtree set Until the boundary conditions are met: trd c >trd max Or F′ is an empty set, trd max Indicates the maximum depth of the tree, trd c Indicates the current depth of the tree.

8. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that: Step (7) also includes, when the participant has no local resources or data, proposing a demand to purchase resources and services from the cloud platform.

9. The method for screening and auditing federal forest trusted participants according to claim 1 is characterized in that: Step (8) The cloud platform uses the subjective logic model to i To conduct reputation evaluation, specifically, the cloud platform uses a subjective logic model to evaluate the participant P i The reputation of the participant P i In the t round of training, the optimal or near-optimal split is selected to minimize the GINI coefficient, which is considered positive, and the non-optimal or non-near-optimal split is selected to fail to minimize the GINI coefficient, which is considered negative, which is neg. At this time, the trust relationship is Non-trust relationship is And T ti +T di =1,T ti ,T di ∈[0,1]; get the participant P i Reputation R(LP i ) is expressed as R(LP i )=λT di +βT ti , the historical credit t-1 round R t-1 (LP i ) and current event reputation t (LP i ) acts on the current credit round R t (LP i ), update R(LP i ) is R t (LP i )=αR t-1 (LP i )+(1-α)r t (LP i ); The updating of the reputation level also includes, when P in step (5) i If the task request is not approved, this information will be fed back to the GM, and the GM will know that P i Not authorized, keep P i The reputation value remains unchanged.

10. A federal forest trusted participant screening and auditing system, characterized in that: include: The task publishing module is used by the cloud platform to collect model service requirements from model consumers and publish tasks to the local participant group manager LGPA and candidate participants in the field in a broadcast form; The registration module is used when a participant P i When participating in a task for the first time, complete the registration at LGPA, join the task group, and obtain the group member private key from LGPA through the interactive protocol; The identity authentication module is used by the cloud platform to receive the identity of the participant P i The task request message to P i Perform identity authentication and determine P i The legitimacy and authenticity of the credit rating; The reputation audit module is used when the participant P i After verification, the cloud platform extracts the participant P from the task request i The credibility level of the participant is determined, and whether to approve the participant's request is determined based on the current model task volume; The model building module is used for the cloud platform and several participants who have been authorized through task requests to jointly complete the construction of the random forest tree model; The result prediction module is used by model consumers to submit remuneration to the cloud platform. The cloud platform provides services based on the results of model construction, gives results and feedback; The evaluation update module is used by the cloud platform to evaluate the participant P through the subjective logic model. i Conduct reputation assessment, record the latest reputation value, and update the reputation level; The reward distribution module is used by the cloud platform to allocate rewards according to the participants’ P i The most recent credit rating reward distribution.

Images