Federal learning excitation method based on deposit conversion reputation

Abstract

The invention discloses a federated learning excitation method based on guarantee deposit conversion reputation, relates to the technical field of federated learning, and enables a new client to convert capital into reputation in a system, break zero reputation and quickly obtain task participation qualification through conversion of guarantee deposit into initial reputation. And secondly, a dynamic guarantee deposit conversion mechanism is designed, and the client is allowed to convert part of guarantee deposit into additional reputation according to the amount of guarantee deposit and real-time performance after each round of training, so that reasonable and rapid increase of the reputation is realized, and the Mortai effect is effectively relieved. And finally, a contribution integral system linked with the reputation is introduced, the reputation is taken as an influence factor of integral calculation while single task performance is quantified and taken as a direct reward basis, and a virtuous circle with high reputation, high integral and higher reputation is formed. Not only is the reputation as the core location of the trust voucher and task admission clear, but also a federal learning ecosystem which is compatible in excitation and can continuously operate is constructed.

Description

Technical Field

[0001] This application relates to the field of federated learning technology, specifically to a federated learning incentive method based on deposit-based reputation conversion. Background Technology

[0002] As artificial intelligence enters the era of big data, data privacy and security have increasingly become a focus of social concern. Traditional centralized machine learning models require submitting raw data to a central server, which not only faces a huge risk of data leakage but also contradicts increasingly stringent data privacy regulations.

[0003] Federated learning, as an emerging distributed machine learning paradigm, has emerged. Its core idea is "data doesn't move, model moves," meaning client devices train the model locally, only uploading updates to a central server for aggregation, thus theoretically avoiding direct exposure of the original data. However, while federated learning can solve data privacy issues and achieve data privacy protection, incentivizing clients to continuously provide high-quality data contributions remains difficult. Without sufficient compensation, clients may be unwilling to participate in federated learning due to factors such as local costs.

[0004] Current federated learning incentive mechanisms suffer from three significant flaws: First, new clients face a "cold start" dilemma, with initial reputation at zero or extremely low levels. Unable to accumulate reputation due to failing to meet task thresholds, they enter a vicious cycle of "no reputation → no tasks → no reputation." Second, reputation systems generally rely on automatic accumulation and growth after each training round, leading to early participants establishing insurmountable reputation barriers and a "the strong get stronger" Matthew effect, weakening the participation enthusiasm of new clients. Third, the role of reputation in contribution measurement is ambiguous; most schemes fail to effectively integrate it with incentive mechanisms, reducing the actual value and guiding role of reputation. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this application proposes the following technical solution: In a first aspect, embodiments of this application provide a federated learning incentive method based on deposit-to-reputation conversion, including: The smart contract on the blockchain side receives the deposit paid by the new client upon registration and allocates an initial reputation to the new client based on the deposit. After gaining an initial reputation, the new client makes participation decisions based on the training task information released by the server. The server distributes the global model to the training participating clients, the clients perform local training, and upload the local model parameters after completion; The server evaluates the client's contribution based on the model parameters uploaded by the participating client and records it in the blockchain; The blockchain converts client deposits based on a deposit-to-reputation function and the client's contribution percentage, and updates the client's reputation and contribution points accordingly. After the training task is completed, the client uses the contribution points to redeem rewards from the blockchain.

[0006] In one possible implementation, the smart contract on the blockchain side receives the deposit paid by a new client upon registration and allocates an initial reputation to the new client based on the deposit, including: Before joining the federated learning system, new clients pay a pre-set monetary deposit to the smart contract on the blockchain. The monetary deposit serves as a guarantee of good faith participation in the training; Smart contracts confirm deposits After the funds are received, the initial reputation function will be used as a preset method. An initial reputation value is assigned to the client, and the initial reputation function uses a linear transformation function: For the client's initial reputation, This is the reputation conversion factor; the converted reputation value must meet the following requirements. This is to ensure that new clients are able to participate in subsequent training and contribute initial points. When set to 0, the converted reputation value threshold remains unchanged. The deposit is calculated according to a predetermined ratio. After conversion into reputation points, the remaining deposit is: in: Excess deposit is used to actually convert reputation value. It can be used in subsequent training.

[0007] In one possible implementation, after acquiring an initial reputation, the new client makes a participation decision based on training task information published by the server, including: Before each training round begins, the server will release training task information, including: model structure, data requirements, and deadline; The client first checks its own reputation value. Does it meet the requirements? If the conditions are met, a decision is made after considering one's own factors: like If so, the client will join this round of training; like If so, the client will not participate in this round of training; The server then collects all participation decisions and generates a participant set. .

[0008] In one possible implementation, the server distributes a global model to training participating clients, and the clients perform local training, including: The server will be the first Round global model parameters Distributed to the set of participants All clients in; Before each training round begins, each client can choose to deposit additional funds into the smart contract. In order to enhance its potential for subsequent reputation conversion; The cumulative deposit balance of the smart contract update client: in: For the client deposit; After the deposit balance is updated, each client uses its local dataset. For the model Perform multiple rounds of local training to minimize the local loss function. .

[0009] In one possible implementation, the client uploads the local model parameters after training, including: After the client completes local training, it obtains the updated local model parameters. ; The client then calculates its local model update: The updated Uploaded to the server, the client's original private data remains locally throughout the upload process.

[0010] In one possible implementation, the server evaluates the client's contribution based on the model parameters uploaded by the participating client and records it in the blockchain, including: For any subset of clients Its efficacy Defined as the model update after aggregation of this subset With global update Cosine similarity: in: ; The server dynamically determines a cutoff point. This makes the scale larger than The average utility of a subset is below the threshold ; The server then sampled. The clients are arranged in order, among which: Based on Monte Carlo sampling numbers, To ensure the minimum number of samples that meet the requirements; For each sampling permutation Only the first part is calculated. Marginal contribution per client Subsequent client contributions are recorded as 0. Approximate Shapley value in this round The average marginal contribution across all sampling permutations; Normalize the approximate Shapley value to obtain the contribution percentage for each client. : server according to Calculate contribution ranking and will Recorded on the blockchain.

[0011] In one possible implementation, the blockchain converts a client's deposit based on a deposit-to-reputation function and the client's contribution percentage, and updates the client's reputation and contribution score accordingly, including: According to the client Reputation value from deposit conversion And the reputation gained from this round of training Gain reputation value According to the above Implement client-side reputation value updates; Based on the contribution percentage of each client after training. The client's contribution points are updated along with the updated reputation value.

[0012] In one possible implementation, the statement based on the client... Reputation value from deposit conversion And the reputation gained from this round of training Gain reputation value According to the above Implement client-side reputation value updates, including: Determine the client Reputation value from deposit conversion : in: , is the maximum deposit value that can be converted into reputation on the client side, of which This is the maximum deposit amount. This indicates that if the client's deposit value Then take ,like Then take , As a function, where This indicates the proportion of the client's contribution to the global model in this round of training, ranging from... between; It is a preset contribution percentage threshold, which is the inflection point of the efficiency function. At that time, the conversion efficiency was very low. hour, ;when At that time, efficiency will increase dramatically; It is a slope parameter, a control. Function at threshold The parameter indicating the steepness of the change in the vicinity, when The larger the value, the steeper the function curve, meaning that once the contribution percentage exceeds the threshold... Conversion efficiency will increase; It is the maximum efficiency coefficient, serving as the upper limit of the efficiency function. It is a binary variable involved in decision-making, where 1 indicates participation in this round of training and 0 indicates non-participation; Determine the reputation of training growth : in: It is the maximum value of reputation increment preset by the system for each round, which controls the expansion rate of the entire system's reputation; This indicates the client's ranking in terms of contribution among all participants in this round. It is an adjustable parameter. It is a decay factor used to control the rate at which reputation declines as ranking drops. It is a function that decays exponentially as the ranking decreases; In conjunction with the above and The total reputation increment is: Finally, the client's reputation value was updated as follows: in: This will be used to calculate the reputation value for the next round after the update.

[0013] In one possible implementation, the contribution percentage of each client after training is used as the basis for the decision. The client's contribution score is updated in accordance with the updated reputation value, including: after each training round, the smart contract updates the client's contribution score: in: As a basic integral coefficient, it is the integral base set by the system and determines the scale of the entire integral system; This refers to the client's current total reputation, which is the total reputation value the client possesses at the start of this round. This is the reputation gain coefficient, which controls the strength of the influence of reputation on the integral gain. The larger the value, the more points you gain from a high reputation. By taking the logarithm of the reputation value, a balance is achieved between the reputations of early reputation builders and those of later builders, preventing early clients from having excessively high reputations and disrupting the balance. After each round of conversion, the client's contribution points will be updated to: in: Contribute points to the next round after the update.

[0014] In one possible implementation, after the training task is completed, the client uses the contribution points to redeem rewards from the blockchain, including: in: Indicates the client The number of reward rounds to redeem. The client contributes a total score after training. This refers to the exchange rate, which is the number of points required to redeem one round of rewards.

[0015] In this embodiment, initial reputation is converted through a deposit, enabling new clients to transform capital into reputation within the system, breaking the zero-reputation barrier and quickly gaining eligibility to participate in tasks. Secondly, a dynamic deposit conversion mechanism is designed, allowing clients to convert a portion of their deposit into additional reputation after each training round based on their deposit amount and real-time performance, thereby achieving reasonable and rapid reputation growth and effectively mitigating the Matthew effect. Finally, a contribution points system linked to reputation is introduced. While quantifying individual task performance and using it as a direct reward basis, reputation is used as an influencing factor in points calculation, forming a virtuous cycle of high reputation → high points → even higher reputation. This not only clarifies the core role of reputation as a trust credential and task access indicator but also constructs an incentive-compatible and sustainably operating federated learning ecosystem. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a federated learning incentive method based on deposit-to-reputation conversion, provided for an embodiment of this application; Figure 2 This is a schematic diagram of the interactive communication of a federated learning incentive method based on deposit-to-reputation conversion, provided as an embodiment of this application. Detailed Implementation

[0017] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0018] To ensure the operation of each component in this embodiment, the system initialization phase first ensures that all components of the federated learning system are correctly configured and in a working state. During this phase, the administrator first completes the system parameter configuration, including setting the global model structure, training parameters, and reputation system parameters (such as the minimum reputation threshold). and reputation conversion threshold (etc.). Subsequently, a smart contract is deployed and initialized on the blockchain. This contract initializes the new client's initial reputation value and contribution points to 0 and formally writes the pre-configured minimum reputation threshold. With reputation conversion threshold Finally, the server initializes the global model required for the first round of training. The initial model is recorded on the blockchain to ensure the immutability and traceability of the initial state of the model.

[0019] Based on the initialized system, see [link / reference] Figure 1 and Figure 2 The federated learning incentive method based on deposit-to-reputation conversion provided in this embodiment includes: S101, the smart contract on the blockchain side receives the deposit paid by the new client upon registration and allocates an initial reputation to the new client based on the deposit.

[0020] Before joining the federated learning system, new clients need to deposit a certain amount of monetary deposit into the smart contract. The deposit serves as a guarantee of the participant's integrity in the training; the specific amount is set by the system and can be dynamically adjusted.

[0021] Smart contracts confirm deposits After the funds are received, the initial reputation function will be used as a preset method. Assign an initial reputation value to the client. This system uses a linear transformation function: For the client's initial reputation, This is the reputation conversion coefficient. The converted reputation value must meet the following requirements. This is to ensure that new clients are capable of participating in subsequent training. Initial contribution points It is set to 0. The formula indicates that a minimum monetary deposit is required to reach a certain reputation value. That is, there is a minimum deposit threshold, but no upper limit, while the reputation conversion rate has upper and lower limits; specifically, the converted reputation value is capped at a certain level. The deposit is calculated according to a predetermined ratio. After conversion into reputation points, the remaining deposit is: in Excess deposit is used to actually convert reputation value. It can be used in subsequent training.

[0022] S102, after the new client obtains an initial reputation, it makes a participation decision based on the training task information released by the server.

[0023] Before each training round begins, the server publishes training task information, such as model structure, data requirements, and deadline. The client then first checks its own reputation score. Does it meet the requirements? If the conditions are met, a decision is made after considering one's own factors. like If so, the client will join this round of training; like If the client does not participate in this round of training, then the client will not participate in this round of training.

[0024] The server then collects all participation decisions and generates a participant set. .

[0025] S103, the server distributes the global model to the training participating clients, the clients perform local training, and upload the local model parameters after completion.

[0026] The server will be the first Round global model parameters Distributed to the set of participants All clients can choose to pay an additional deposit to the smart contract before the start of each training round. This is to enhance its potential for subsequent reputation conversion. The smart contract will update the client's accumulated deposit balance: in: For the client deposit; After the deposit balance is updated, each client uses its local dataset. For the model Perform multiple rounds of local training to minimize the local loss function. .

[0027] After the client completes local training, it obtains the updated local model parameters. Subsequently, the client calculates its local model update: This update will then be updated. Uploaded to the server. Throughout the process, the client's original private data remains locally and is not leaked.

[0028] S104, the server evaluates the client's contribution based on the model parameters uploaded by the participating client and records it in the blockchain.

[0029] After the server collects all model updates from participating clients, it uses... The semi-Monte Carlo sampling method within the framework fairly calculates the contribution percentage of each client. The specific steps are as follows: Define a utility function: for any subset of clients Its efficacy Defined as the model update after aggregation of this subset With global update Cosine similarity: in .

[0030] SMC Sampling and Approximate Calculation: The server dynamically determines a cutoff point. This makes the scale larger than The average utility of a subset is below the threshold .

[0031] The server then sampled. The clients are arranged in order (of which) Based on Monte Carlo sampling numbers, To ensure the minimum number of samples that meet the requirements.

[0032] For each sampling permutation Only the first part is calculated. Marginal contribution per client Subsequent client contributions are recorded as 0.

[0033] Client Approximate Shapley value in this round This represents the average marginal contribution across all sampling permutations.

[0034] Calculate the contribution percentage: Normalize the approximate Shapley value to obtain the contribution percentage for each client. : server according to Calculate contribution ranking and will Recorded on the blockchain.

[0035] S105, the blockchain converts the client's deposit based on the deposit conversion reputation function and the client's contribution ratio, and updates the client's reputation and contribution points.

[0036] The blockchain converts client deposits based on a deposit conversion reputation function and the client's contribution percentage, and updates the client's reputation and points accordingly.

[0037] (1) Reputation update Client Reputational gain in this round It consists of two parts: 1. Only when its current total reputation It can only be obtained in time. This avoids making deposits the main source of reputation, slows down reputation growth, and further enhances the role of reputation, thus ensuring its value.

[0038] The core purpose of this formula is to convert the "capital" (deposit) invested by the client into a fixed amount. This translates into "reputation," but the efficiency of each round of conversion is not fixed; rather, it depends on the proportion of its contribution. It's a decision. It ensures that only clients who actively contribute can efficiently convert their deposits into reputation.

[0039] in: , is the maximum deposit value that can be converted into reputation on the client side, of which This is the upper limit for deposits, i.e., the upper limit for deposits to convert into reputation, to prevent client reputation values ​​from being dominated by deposits. This indicates that if the client's deposit value Then take ,like Then take . As a function, where This indicates the proportion of the client's contribution to the global model in this round of training, determined by... The method calculates that the range is within The values ​​in between represent the client's "level of effort" and "quality of contribution." It is a preset threshold for the proportion of contribution, and is the "turning point" of the efficiency function. When At that time, the conversion efficiency was very low; when hour, ;when Efficiency increases dramatically at this point. This formula sets a "contribution threshold" for achieving high-efficiency conversion. It is a slope parameter, a control. Function at threshold Parameters indicating the steepness of the change in the vicinity. When The larger the value, the steeper the function curve, meaning that once the contribution percentage exceeds the threshold... The conversion efficiency will increase "explosively", and the rewards for high contributions will be stronger, which determines the incentive level for contributions. This is the maximum efficiency coefficient, serving as the upper limit of the efficiency function. It restricts the highest efficiency of converting deposits into reputation within a single round. The greater the contribution, the closer the efficiency approaches this maximum efficiency. . As a binary variable involved in the decision-making process, 1 represents participation in this round of training, and 0 represents non-participation. Only clients that actually participated in the training, i.e. Only then can their deposit be converted in this round.

[0040] 2. Training enhances reputation This formula directly rewards reputation based on the client's contribution ranking, allowing reputation to grow through the client's training results.

[0041] in: It is the maximum value of reputation increment preset by the system for each round. It controls the rate at which the reputation of the entire system expands. It is the client's ranking of contributions among all participants in this round. This is an adjustable parameter, usually set to 1. Ensure that when... At that time, the exponential part is Thus obtaining the full amount award. It is a decay factor used to control the rate at which reputation decays as ranking declines. The smaller the value, the steeper the curve, and the reputation reward will drop sharply if the ranking is slightly lower. This is a function that decays exponentially as the ranking decreases. This function converts the ranking into a reward coefficient. The higher the ranking, the closer the coefficient is to 1, and the closer the reward is to... The lower the ranking, the closer the coefficient is to 0, and the reward is also close to 0. Only clients that actually participated in the training, i.e. Only then can their deposit be converted in this round.

[0042] The total reputation increment is determined as follows: Final round Client The reputation value that has increased is: Therefore, the client's reputation value is updated as follows: in This will be used to calculate the reputation value for the next round after the update.

[0043] (2) Update of contribution points After each round of training, the smart contract updates the client's contribution points: This formula will calculate the contribution percentage of each client. With the updated All of these factors influence contribution points, making the final reward distribution more reasonable.

[0044] in: As a basic integral coefficient, it is the integral base set by the system. It determines the scale of the entire integral system. This represents the client's current total reputation, which is the total reputation value the client possesses at the start of this round. This is the reputation gain coefficient, which controls the strength of the influence of reputation on the integral gain. The higher the level, the more points you gain from a high reputation. Taking the logarithm of the reputation value balances the reputation of early reputation builders and later builders, preventing early clients from having an excessively high reputation and disrupting the balance.

[0045] After each round of conversion, the client's contribution points will be updated to: in: Contribute points to the next round after the update.

[0046] S106, After the training task is completed, the client uses the contribution points to exchange for rewards on the blockchain.

[0047] After completing the training mission, you can use contribution points to redeem the final reward. The reward is the same for each round, and contribution points are used to increase the number of reward rounds. in: Indicates the client The number of reward rounds to redeem. The client contributes a total score after training. This is the redemption ratio, which is the number of points required to redeem one round of rewards. Upon successful redemption, the corresponding points will be deducted.

[0048] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, and c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0049] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for federated learning incentive based on deposit conversion reputation, characterized in that, The blockchain-side smart contract receives new client registration and the deposit paid, and allocates an initial reputation to the new client according to the deposit; The new client makes participation decisions according to the training task information published by the server after obtaining the initial reputation; The server distributes a global model to the training participant client, the client performs local training, and uploads the local model parameters after completion; The server evaluates the contribution of the participating client according to the uploaded model parameters and records it to the blockchain; The blockchain converts the deposit of the client according to the deposit-to-reputation function and the contribution proportion of the client, and updates the reputation and contribution points of the client; After the training task is completed, the client exchanges rewards with the blockchain using the contribution points. The blockchain-side smart contract receives new client registration and the deposit paid, and allocates an initial reputation to the new client according to the deposit, including:

2. The method of claim 1, wherein, The new client makes participation decisions according to the training task information published by the server after obtaining the initial reputation, including: A new client pays a preset amount of currency deposit to a smart contract on the side of the blockchain before joining the federated learning system , which serves as a guarantee for honest participation in training; The smart contract confirms the deposit after the deposit is credited, according to a preset initial reputation function assigning an initial reputation value to the client, the initial reputation function employing a linear transformation function: Initial reputation for a client, Reputation conversion coefficient, converted reputation must satisfy , to ensure that the new client has the ability to participate in subsequent training, initial contribution points is set to 0, the limit of the converted reputation value is kept at , the deposit is converted into reputation value according to the established proportion , the remaining deposit after conversion: wherein: is the actual conversion reputation value, excess deposit can be used in subsequent training.

3. The method of claim 2, wherein, Before the start of each round of training, the server will publish training task information, including: model structure, data requirements and deadline; The server distributes a global model to the training participant client, the client performs local training, including: The client first checks its own reputation value whether it satisfies If it does, it makes a decision taking into account its own factors: If then the client joins the current round of training; If then the client does not participate in this round of training; Subsequently the server collects all the participations in the decision and generates a set of participants .

4. The method of claim 3, wherein, The smart contract updates the cumulative deposit balance of the client: The server will send the first wheel global model parameters to all clients in the set of participants ; Each client can choose to top up the deposit to the smart contract at the beginning of each round of training to enhance its subsequent reputation conversion potential; The client uploads the local model parameters after completing the training, including: wherein: for the client deposit; The balance of the deposit is updated after each client uses its local dataset To the model Performing multiple rounds of local training to minimize the local loss function .

5. The method of claim 4, wherein, Then the client calculates its local model update: The client obtains updated local model parameters after completing local training ; The server evaluates the contribution of the participating client according to the uploaded model parameters and records it to the blockchain, including: The updated is uploaded to the server, and the original private data of the client is always kept locally during the uploading process.

6. The method of claim 5, wherein, The blockchain converts the deposit of the client according to the deposit-to-reputation function and the contribution proportion of the client, and updates the reputation and contribution points of the client, including: For any client subset whose utility is defined as the aggregated model update of the subset cosine similarity to the global update wherein: ; The server dynamically determines a truncation position such that the average utility of a subset of size is below a threshold ; The server then samples The number of clients is arranged, wherein: The number of base Monte Carlo samples, To ensure the minimum number of samples that meet the conditions; For each sample permutation , only the marginal contribution of the first clients is calculated , and the contribution of the subsequent clients is recorded as 0, and the client's approximate Shapley value for this round is the average marginal contribution across all sample permutations ; Normalizing the approximate Shapley values to obtain the contribution proportion of each client : The server computes a contribution rank based on the contribution score and records it on the blockchain. ​ 7. The method of claim 6, wherein, Finally, the reputation value of the client is updated to: According to the client Reputation value converted from the deposit And the reputation growth of this round of training Obtain reputation value-added , according to the Realize the reputation value update of the client; According to the proportion of each client's participation contribution after the training ends The total reputation value of the client The contribution points of the client are updated.

8. The method of claim 7, wherein, The method according to the client Reputation value converted from the deposit And the reputation of the current round of training growth Obtain reputation value-added According to the method Implementing reputation value update of the client includes: Determining a client Reputation value for conversion of deposit : wherein: is the maximum stake value of the client's stake conversion reputation, wherein is the upper limit value of the stake, represents that if the client's stake value , then is taken, if , is a function, wherein represents the proportion of the client's contribution to the global model in this round of training, ranging between ; is a preset contribution proportion threshold, which is the turning point of the efficiency function, when , the conversion efficiency is very low, when , ; when , the efficiency will increase sharply; is a slope parameter, a parameter that controls the steepness of the function near the threshold , the greater , the steeper the function curve, meaning that once the contribution proportion exceeds the threshold , the conversion efficiency will increase; is the maximum efficiency coefficient, which is the upper limit of the efficiency function, is a binary variable for decision-making, wherein 1 represents participating in this round of training, and 0 represents not participating; Determining reputation of training growth : wherein: is the maximum value of the reputation increment for each round, which controls the inflation rate of the overall system reputation; represents the contribution rank of the client among all participants in the current round, is a tuning parameter, is a decay coefficient, which controls the speed of reputation decay as the rank decreases, is a function that exponentially decays as the rank decreases; In conjunction with the and The total reputation increment is obtained as: After each conversion round, the contribution points of the client are updated to: wherein: is the updated next round reputation value.

9. The method of claim 8, wherein, The proportion of each client's participation contribution after the end of training The contribution points of the client are updated according to the updated reputation value, including that the smart contract updates the contribution points of the client after each round of training ends: Wherein: As the basic integral coefficient, it is a system set integral base, which determines the size of the entire integral system; Total reputation of the current client, that is, the total reputation value that the client has at the beginning of this round, Reputation gain coefficient, which controls the strength of the influence of reputation on integral gain, The larger the reputation gain coefficient, the more integral addition brought by high reputation; Taking the logarithm of the reputation value, it balances the reputation of early accumulators and later ones, and avoids the reputation of early clients being too high to break the balance; After the training task is completed, the client exchanges rewards with the blockchain using the contribution points, including: wherein: is the updated next round contribution score.

10. The method of claim 9, wherein, ​ Wherein: represents the client the number of rounds of rewards exchanged, is the total value of points contributed by the client after the end of training, is the exchange ratio, i.e. the value of points required to exchange one round of rewards.

Images