Blockchain-assisted incentivization method for artificial intelligence learning

Abstract

The application relates to a blockchain-assisted incentive method for artificial intelligence learning, and belongs to the field of artificial intelligence.The method judges the level of a user terminal according to the contribution degree and model attitude positivity of the user terminal, and requires the user terminal to perform corresponding obligations according to the level.An edge server rewards and adjusts the level according to the behavior of the user terminal and the degree of obligation performance, and adjusts the weight of the first-type user according to the feedback of the second-type user.The reward is recorded in the form of a blockchain in a DAG structure.The method can improve the model precision of the edge server, guarantee the safety of the model, and reduce the training time and cost.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and relates to a blockchain-assisted incentive method for artificial intelligence learning. Background Technology

[0002] Technically, federated learning allows multiple participants to collaboratively train a model while protecting data privacy. It is currently used in situations where the number of devices and their computational speeds are unknown, aiming to protect user device privacy and efficiently utilize edge servers for model computation, thus optimizing both practical and user-end applications. However, federated learning currently faces several challenges: it requires significant network resources and powerful edge servers for frequent training and computation of local models. Furthermore, during this phase, there are instances where the rewards received by users do not align with their individual weight contributions, leading to an imbalance between user effort and reward. Additionally, using averaging techniques ensures that user effort and reward are aligned, hindering more proactive and effective model training.

[0003] The existing technical solution is as follows: The FedAvg algorithm uses the amount of data from the user end to allocate weights, and through the correlation between different user ends, it helps to improve the accuracy of the local model by uploading to the edge server to calculate the model. In this way, the local model is continuously iterated. This averaging method allows all the data to be trained and considered.

[0004] The drawbacks of the averaging algorithm are: Since it considers average data and the model, it doesn't account for users with high contributions; their rewards and weights should be increased. Furthermore, this algorithm is easily influenced by users with large amounts of duplicate data, making edge server training susceptible to interference from low-weight or low-contribution users, leading to difficulties in model convergence and vulnerability to malicious attacks. Because it considers training models for all users, frequent network bandwidth data exchange and numerous training iterations at the central model significantly increase network operation and model training costs and energy consumption. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a blockchain-assisted incentive method for artificial intelligence learning. This method uses the contribution and feedback obtained during model training when users upload models to classify and position users, primarily dividing them into two levels, each with different tasks and evaluation criteria. Rewards or transaction fees will be charged for the actions of users at different levels. Different rewards will be given to users who upload models and have a good user experience. During the model training period, if the number of users and their training contributions change, the user level and the weight of Class I users will be adjusted.

[0006] Federated learning requires users to train models locally and upload them. High rewards and support encourage willing and proactive users to share their models. This incentive helps users to communicate actively, making them more active and helping the server achieve higher model accuracy and more realistic model parameters.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A blockchain-assisted incentive method for artificial intelligence learning, comprising the following steps:

[0009] S1: Send the edge model to the user terminal and request training;

[0010] S2: The model obtained from the edge server aggregation is used to filter out user-end devices that improve the accuracy of the global model, and the degree of contribution and device number are recorded.

[0011] S3: Based on the feedback from the survey push information, determine the device's positive attitude towards the model; define the positive attitude towards the model as Att, and set the percentage of question answers and answer completeness to help calculate Att;

[0012] S4: By dividing the data into tables, inputting the contribution level and the positive attitude towards the model, the corresponding user level classification is obtained;

[0013] S5: Based on the corresponding user level, according to the user-end level obligation standard, the current device is required to share the model or upload the experience score. The weight of Class I users in federated learning is modified through the weight allocation method, and the model is trained using the edge server.

[0014] S6: Reward users based on the model uploaded by users of type I, and reward users of type II who provide feedback, and record the results using a DAG.

[0015] S7: After several weight adjustments, test the model accuracy on the edge server; stop training when the minimum threshold ACC1 is reached.

[0016] Optionally, in S4, the relevant information from the user end is obtained through S2, and the criteria for judging the degree of contribution include: upload and download latency Dtrans. u Local model computation latency DL u The average value L of the model is given by the following formula:

[0017]

[0018] C n This represents the contribution score of the nth user, where u is the total number of rounds in which contribution scores are calculated, and A...nu denoted by , where is the improvement in accuracy for the m-th user during the u-th training round; 'a' is the average network latency under the corresponding network conditions; and 'b' is the average latency of local computation. Dtrans u It refers to the real-time latency of uploading and downloading, DL u L is the real-time latency of the local model computation, and L is the average loss value of the model; A nu It is a high-weight decision factor, and its value is calculated using the average contribution parameter αM in the balance formula;

[0019] The following requirements and formulas are used to evaluate the enthusiasm for feedback from the survey:

[0020] The parameter for the nth user's enthusiasm for the model is:

[0021]

[0022] Where X n % represents the percentage of questions completed by the nth user, D L Indicates local latency, D T This represents the total latency of device training; m% indicates the general user survey's willingness to accept the data; the percentage of survey questions completed is used as one of the reference values.

[0023] Users with a positive attitude towards optimizing the model have the following requirements:

[0024] Att>Att3, where Att3 is a parameter that represents the average of positive parameters measured by a relatively positive device, indicating that the user's attitude towards the model is positive.

[0025] Optionally, the input contribution level and the positive attitude towards the model are used to derive the corresponding user-side classification as follows:

[0026] First, the initiator conducted a survey on the willingness of high-contribution users identified in S1 to share data / models;

[0027] Based on the user's contribution, users with high contribution and a high willingness to share models are considered as Category I users. That is, Category I users are users with strong computing power, high contribution and willingness to upload models in federated learning. If Category II users need the global model of the edge server and train it locally and then feed it back to the edge server, these users are also Category II users.

[0028] After establishing the written standards, users are categorized into levels using a linear method. A linear approach is used to quantitatively categorize users based on their contributions and data upload activity, as detailed below:

[0029] Based on the user's positive attitude towards the model (Att) and the user's contribution (C), the users are divided into two levels.

[0030] If Att falls within the range of Att1 and Att2, it is considered a type II user. This type of user has a high contribution and a relatively high contribution value to the data training of type I users, i.e., model weight allocation.

[0031] If Att is above Att2 and contribution C is above c2, where c2 is the average contribution of users with the highest historical contribution, then this is considered a Class I user.

[0032] Att1 is the statistical average of Att for negative devices in historical data; Att2 is the statistical average of Att for devices with a relatively positive attitude towards the model in historical data.

[0033] Optionally, after classifying the existing user terminals using the linear classification method, user terminals of different levels will have different weights in the edge server training.

[0034] Before each training round, except for the first round, the weights of Class I users will be changed in the following way:

[0035] For models uploaded by users of type I, the following method is used to assign weights and perform federated learning:

[0036] S41: Initial weights are set to...

[0037] Define a weight score αs n The weights are calculated using the magnitude of the weight scores.

[0038] αs n =I*C n +Att n +K*SC n +H*Ah n

[0039] Assume α n Let be the weight of the nth user terminal in K1. Then the weights of the nth user terminal in K1 are as follows:

[0040]

[0041]

[0042] C n The contribution parameter for the nth user in category I;

[0043] I represents the variance of the contribution of users of category I in the contribution score.

[0044] Att n This refers to the motivation parameter of the nth user in category I.

[0045] SC n It is the score value obtained by training the model based on the nth upload of the class I user by the class II user;

[0046] K is the variance of the scores given by most users for the same model;

[0047] Ah n The historical contribution of the nth user terminal in category I is obtained by using historical information from network training.

[0048] H is the variance of the contribution in the historical contribution.

[0049] 42: Upload the data and models uploaded by users of type I to the edge server. The edge server aggregates the models based on the initial weights. The aggregation process is as follows:

[0050] For epochs∈{1,2,3…}

[0051]

[0052]

[0053]

[0054] S43: Distribute the trained global model to Type II users, collect suggestions and feedback, and analyze model accuracy through question design and performance evaluation. The specific score given by the nth Type II user for the same model is calculated using SC. n This means that the score will be returned to the edge server during a training round;

[0055] The model is sent to all Category II users, but these users have to pay a certain price. When they evaluate and score the global model, they receive a reward by sending feedback to the edge server.

[0056] By changing the scores, new weights are derived according to the weighting formula described above.

[0057] Optionally, in S6:

[0058] (1) Explanation of the rewards for models uploaded by Category I users:

[0059] When a Type I user uploads the trained model to the edge server, the user's contribution to the edge server will be analyzed, and a reward will be given based on the size of the contribution. The reward system will be used to recognize and trade the contributions of Type I users. During this process, the transactions of Type I users will be recorded, and the rewards will be recorded through a DAG-structured blockchain.

[0060] (2) For Category II users, the model usage experience feedback score submitted to the edge server after downloading the global model will be awarded based on the number of downloads and feedback on the corresponding global model by the Category II user. When a Category II user downloads the global model, the version number of the global model downloaded will be recorded. When the user uploads feedback again, the user will be required to upload the model and the corresponding model's score reference value, denoted as SC. n To prevent multiple identical feedbacks from the same global model without the user actually experiencing it, the system uses the average score of feedback from multiple users of the same global model to determine the reference score for feedback uploaded by a user of the same type. If the reference score given by a user of the same type is significantly lower than the average and exceeds an unreasonable range, it is considered that the user is at risk of maliciously uploading data, and the user will be warned or their suggestions will not be adopted. After completing the above process, the user of the same type will be rewarded and the results will be recorded using a Directed Acyclic Graph (DAG).

[0061] When a Type II user completes multiple downloads of the global model and provides corresponding feedback in accordance with the normal process, and the number of times this Type II user completes the download reaches a threshold Y, this user is considered to be valuable and willing to share their user experience with the model. This user will then be marked and followed. For models uploaded to the edge server by some Type I users with low weight, we will consider distributing the global model to this interested and compliant Type II user free of charge.

[0062] (3) Explanation of reward values: Different rewards are given to different users; the reward value for the model uploaded by user type I is reasonably allocated based on the weight of the nth user at the time of training, that is, the total global model value is Pay. total The reward for the nth user who uploads the model is:

[0063] Pay1 n =Pay total *α n

[0064] For Category II users, the reward is determined based on the similarity between their score and the average score of all Category II users, within the same global model's experience feedback framework. The reward value for the nth Category II user is:

[0065] Pay2 n =App*Pay base

[0066]

[0067] The app score is the degree of similarity between the score of the nth type II user and the average score of all type II users. avg is the average score of type II user feedback. SC n It is the specific score given by the nth user, Pay base It is a standard value;

[0068] Each time a user needs to retrieve the global model from the edge server, a certain transaction fee is required, which is a fixed value CH.

[0069] (4) For other users, if they are interested in the global model, they will be required to pay a certain transaction fee when they send a request to download the global model. This will be recorded in the DAG-structured blockchain system. If they upload feedback, they will be given a certain reward. Unlike Class II users, the reward will be less than that of Class II users. When their feedback is uploaded to another threshold Y1, they will be transformed from other users into Class II users and will receive more rewards when uploading feedback.

[0070] (5) The DAG-structured blockchain is recorded in federated learning as follows:

[0071] Directed Acyclic Graphs (DAGs) are a directional but closed-loop method of transaction recording. In incentive mechanisms, a blockchain with a DAG structure is used to help record the transaction fees required for a model to acquire data and the rewards received when uploading the model and receiving feedback.

[0072] The beneficial effects of this invention are as follows:

[0073] This invention employs a weighted analysis and incentive method to classify and assess the engagement of users who exchange and upload models. All users are divided into two levels, each with different responsibilities and assessment methods. This classification allows for better management and control of users within a limited timeframe, resulting in a more practical and accurate global model. The main objective is to reward high-contributing users, encouraging them to perform better data processing and model sharing. Users with average computing power or low contribution are screened and required to pay a certain cost when requesting global model usage. If they share their experience with the edge server after using the global model, they are incentivized to continue sharing their global model usage experience, thus stimulating them to make higher contributions. This classification algorithm addresses the issues of average considerations and unreasonable model calculation weight allocation in the FedAvg algorithm, enhancing user engagement and encouraging users to share models or their experiences.

[0074] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0075] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0076] Figure 1 A block diagram of a blockchain system with a DAG structure;

[0077] Figure 2 A system flowchart for a user-feedback-based federated learning incentive method;

[0078] Figure 3 This is a system model. Detailed Implementation

[0079] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0080] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0081] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0082] This invention is a federated learning method that utilizes user-end contribution and user engagement with the model to classify users using a linear classification method, and then assigns different tasks based on their level. This method, through classification, results in: models being uploaded to Level I devices, shared experiences on Level II devices, and sufficient training of the model on the edge server. User ends are selected, and rewards are given according to their contribution levels, recorded through a DAG-structured blockchain. This ensures the security of the edge server model and reduces training time and costs on the edge server. The method includes the following steps:

[0083] S1: Send the edge model to the user and request them to train it.

[0084] S2: The model obtained from the edge server aggregation is used to filter out user-end devices that significantly improve the accuracy of the global model, and the degree of contribution and device number are recorded.

[0085] S3: Based on the feedback from the survey push, determine the device's positive attitude towards the model. Define the positive attitude towards the model as Att, and use the percentage of complete answers to questions to help calculate Att.

[0086] S4: By dividing the data into tables, inputting contribution level and positive attitude towards the model, we can obtain the corresponding user level classification.

[0087] S5: Based on the corresponding user level, according to the user-end level obligation standard, the current device is required to share the model or upload the experience score. The weight of Class I users in federated learning is modified through the weight allocation method, and the model is trained using the edge server.

[0088] S6: Reward users of type I who upload models, and also reward users of type II who provide feedback. The reward amounts are calculated using the methods described in the manual and recorded using a Directed Acyclic Graph (DAG).

[0089] S7: After multiple weight adjustments, test the model accuracy on the edge server. Stop training when the minimum threshold ACC1 is reached.

[0090] The grading system will undergo rigorous analysis, classifying users based on two main parameters: contribution level and survey feedback engagement. Below is a detailed grading breakdown based on different scenarios:

[0091] Contribution and user engagement with the model

[0092] The relevant information obtained from the user's end via S2, and the criteria for judging the degree of contribution include: upload and download latency (Dtrans). u ), local model computation latency (DL) u The average Loss(L) of the model, and its contribution, are jointly influenced by these parameters, and the specific formula is as follows:

[0093]

[0094] C n This represents the contribution score of the nth user, where u is the total number of rounds in which contribution scores are calculated, and A... nu denoted by , where 'a' represents the improvement in accuracy for the m-th user during the u-th training round, 'b' is the average network latency under the corresponding network conditions, and 'a' is the average latency of local computation. u It refers to the real-time latency of uploading and downloading, DL u L is the real-time latency of the local model calculation, and L is the average loss value of the model. nu The contribution average parameter αM is a high-weight decision factor, so its calculated value in the balancing formula is used. Other factors have varying degrees of influence on the timeliness of FL training and also affect the contribution of Class I users. To improve the accuracy of the final model, these factors are treated as lower-weight decision factors.

[0095] The following requirements and formulas are used to evaluate the enthusiasm for feedback from the survey:

[0096] The parameter for the nth user's enthusiasm for the model is:

[0097]

[0098] Where X n % represents the percentage of questions completed by the nth user, D LIndicates local latency, D T This represents the total latency of device training, and m% indicates the general user survey's willingness to accept the data. The percentage of survey questions completed is used as one of the reference values.

[0099] Users with a positive attitude towards optimizing the model have the following requirements:

[0100] ATT>Att3, where Att3 is a parameter that represents the average value of positive parameters measured by a relatively positive device. This can be interpreted as indicating a positive attitude of the user towards the model.

[0101] Based on the above two parameter standards, a written description is provided for classifying and grading user terminals:

[0102] First, the initiator conducts a data / model sharing willingness survey among the high-contribution users identified in step 1.

[0103] Based on their contribution to the user end, users with high contribution and a strong willingness to share models are considered Type I users. Type I users are those with strong computational capabilities, high contribution rates in federated learning, and a willingness to upload their models. Type II users may not necessarily be willing to share their models to the edge server, but if they need the global model from the edge server and are willing to train locally and then provide feedback to the edge server, then they are also considered Type II users.

[0104] After establishing the written standards, user levels are categorized using the following mathematical linear method, a mathematical descriptive approach. This method uses a linear approach to quantitatively statistically categorize users based on their contribution and data upload activity, as shown in the diagram.

[0105] Based on the user's positive attitude towards the model (Att) and the user's contribution (C), the users are divided into two levels.

[0106] If Att falls within the range of Att1 and Att2, it is considered a type II user. This type of user has a high contribution rate and a relatively high contribution value to the data training of type I users, i.e., model weight allocation.

[0107] If Att is greater than Att2 and contribution C is greater than C2, where C2 is the average contribution of users with the highest historical contribution ranking, this user is considered to be of type I. These users are the main force behind model uploading in federated learning, helping the central model achieve better aggregation and fitting.

[0108] Att1 is the statistical average of Att values ​​for negative devices in historical data. Similarly, Att2 is the statistical average of Att values ​​for devices with a more positive attitude towards the model in historical data.

[0109] Third, after classifying the existing client terminals using a linear classification method, client terminals of different levels will have different weights in the edge server training.

[0110] Before each training round, except for the first round, the weights of Class I users will be changed in the following way.

[0111] For models uploaded by users of type I, the following method is used to assign weights and perform federated learning:

[0112] 1. Initial weights are set to Because everyone's initial weight score is the same.

[0113] Define a weight score αs n The weights are calculated using the magnitude of the weight scores.

[0114] αs n =I*C n +Att n +K*SC n +H*Ah n

[0115] Assume α n Let be the weight of the nth user terminal in K1. Then the weights of the nth user terminal in K1 are as follows:

[0116]

[0117]

[0118] C n This is the contribution parameter for the nth user in category I.

[0119] I represents the variance of the contribution of users of type I in the contribution score.

[0120] Att n This represents the motivation parameter for the nth user in category I.

[0121] SC n It is the score value obtained by training the model based on the nth upload of the class I user and the data from the class II user.

[0122] K is the variance of the scores given by most users for the same model.

[0123] Ah n The historical contribution of the nth user terminal in type I can be obtained through historical information from network training.

[0124] H is the variance of the contribution in the historical contribution.

[0125] 2. Next, the data and models uploaded by users in category I will be uploaded to the edge server. The edge server will aggregate the models based on the initial weights. The aggregation process is as follows:

[0126] For epochs∈{1,2,3…}

[0127]

[0128]

[0129]

[0130] 3. After training, the global models are distributed to Type II users to collect their suggestions and experiences. Through question design and model accuracy analysis, the specific score given by the nth Type II user for the same model is calculated using Scorecard (SC). n This indicates that the score will be returned to the edge server during a training round.

[0131] The model is sent to all Category II users, who will receive a reward after they rate and evaluate the global model.

[0132] By changing the scores, new weights can be derived using the weighting formula described above.

[0133] Thirdly, the reward methods for Category I users, who contribute to the model, will be explained below, as will the rewards for Category II users, who provide feedback on the global model. In addition, transaction fees will be charged and recorded for users who obtain the global model from the edge server.

[0134] 1. Explanation of the rewards for model uploads by Category I users:

[0135] When a Type I user uploads their trained model to the edge server, their contribution to the edge server will be analyzed, and rewards will be given based on the magnitude of that contribution. Rewards will be used to recognize and trade the contributions of Type I users. During this process, the transactions of these Type I users will be recorded, and the rewards will be tracked through a DAG-structured blockchain.

[0136] 2. For Category II users, a score will be awarded based on their feedback on the global model downloaded and submitted to the edge server. This score will be determined by the number of downloads and corresponding global model feedback submitted by the Category II user. When a Category II user downloads a global model, the version number of the downloaded model will be recorded. When they upload feedback again, they will be required to upload their model along with its corresponding score reference value, denoted as SC. nThis method prevents multiple identical feedbacks from the same global model from occurring without the user actually experiencing the product. Of course, for the same global model, the average score of feedback from multiple users of the same type II user's experience will be used to judge the reference score of feedback uploaded by a single user of that type II user. If the reference score given by a user of that type II user is significantly lower than the average, exceeding an unreasonable range, then this user is considered to have a risk of maliciously uploading data, and will be warned or have their suggestions rejected. After completing this process, the user of that type II user will be rewarded, and this will be recorded using a Directed Acyclic Graph (DAG).

[0137] When a Type II user repeatedly downloads the global model and provides corresponding feedback in accordance with the normal process, and the number of times this Type II user completes the download reaches a threshold Y, this user is considered valuable and willing to share their user experience with the model. This user is then marked and followed. For models uploaded to the edge server by some Type I users with lower weight, we will consider distributing the global model free of charge to interested and compliant Type II users.

[0138] 3. Explanation of reward values: Different rewards are given to different users. The reward value for the model uploaded by user type I is reasonably allocated based on the weight of the nth user at the time of training, that is, the total global model value is Pay. total The reward received by the nth user for uploading the model is

[0139] Pay1 n =Pay total *α n

[0140] For Category II users, rewards are calculated based on how closely their scores resemble the average scores of all Category II users, within the same global model's feedback framework. The reward value for the nth Category II user is:

[0141] Pay2 n =App*Pay base

[0142]

[0143] The app score is the degree of similarity between the score of the nth type II user and the average score of all type II users. avg is the average score of type II user feedback. SC n It is the specific score given by the nth user, Pay base It is a standard value.

[0144] Each time a user needs to retrieve the global model from the edge server, a certain transaction fee is charged, which is a fixed value CH.

[0145] 4. For other users, if they are interested in the global model, they will be required to pay a transaction fee when they send a request to download it. This fee will be recorded in the DAG-structured blockchain system. Of course, they will also receive a reward if they upload feedback, although the reward will be less than that of Category II users. When their feedback reaches another threshold Y1, they will be upgraded from other users to Category II users, and will receive more rewards for uploading feedback.

[0146] 5. The DAG-structured blockchain is recorded in federated learning as follows:

[0147] A Directed Acyclic Graph (DAG) is a directional but closed-loop transaction recording method. In this incentive mechanism, a blockchain-like approach using a DAG structure is employed to record the transaction fees required for the model to acquire data and the rewards received when uploading the model and receiving feedback. This process does not utilize the concept of block packaging to save time; however, each transaction requires relatively complex rules for verification. Finally, after all transactions are completed, every reward and transaction in the federated learning process is traceable. Through this method, the incentive mechanism enhances the user's enthusiasm for model training and ensures the security of the edge server.

[0148] A specific block diagram of a DAG-structured blockchain is shown below. Figure 1 As shown.

[0149] The system flowchart for the user feedback-based federated learning incentive method is as follows: Figure 2 As shown.

[0150] In smart homes and cities, federated learning can effectively alleviate network bandwidth pressure and edge server computation. The algorithm categorizes user terminals in smart homes and cities, using Category I users and models to assist edge servers in computational aggregation. Furthermore, Category II users help Category I users obtain user experience and operator surveys, allowing for more appropriate weight allocation through central aggregation of Category I users. A system flowchart is shown below. Figure 3 As shown.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A blockchain-assisted incentive method for artificial intelligence learning, characterized in that, Includes the following steps: S1: Send the edge model to the user terminal and request the user terminal to train it; S2: The edge server aggregates the models trained on the user terminals, filters out user terminal devices that improve the accuracy of the global model, and records the contribution and device number of the user terminal devices. S3: Based on the survey information pushed to the user terminal and the user terminal's feedback, determine the user terminal's positive attitude towards the model Att, wherein the positive attitude Att is determined by setting questions and based on the completeness of the user terminal's answers to the questions; S4: Classify users according to their contribution level C and attitude enthusiasm Att. Conduct a survey on the willingness of high-contribution users to share data or models. Users with high contribution level and willingness to share models are identified as Class I users. Users who need to obtain the global model from the edge server and train it locally before feeding it back to the edge server are identified as Class II users. Specifically, when the user's attitude enthusiasm Att is between the statistical average Att of negative devices and the statistical average Att of devices with a relatively positive attitude towards the model in historical information statistics, the corresponding user is identified as a Class II user. When the user's attitude enthusiasm Att is higher than the statistical average Att of the devices with a relatively positive attitude towards the model and the contribution level C is higher than the average contribution level of the top-ranked users in historical contribution, the corresponding user is identified as a Class I user. S5: Based on the corresponding obligation standards for user levels, Type I users are required to share models, and Type II users are required to upload model usage experience ratings. During the first round of training, all Type I users are assigned the same initial weights. Before each subsequent round of training, the weight score for Type I users is determined based on their current contribution, attitude, feedback ratings from Type II users, and historical contributions. The aggregate weight for each Type I user is then determined according to the proportion of their weight score to the total weight score of all Type I users. The edge server aggregates the models uploaded by Type I users according to these aggregate weights and distributes the resulting global model to Type II users. Type II users then generate their experience ratings based on the problem settings and model accuracy analysis and return them to the edge server. These experience ratings are used to update the aggregate weights of Type I users in subsequent rounds. S6: Rewards are given to Category I users based on their contributions to the uploaded model, and to Category II users based on their download of the global model and feedback on their experience with the global model. When a Category II user downloads the global model, the version number of the downloaded model is recorded. When a Category II user uploads feedback again, they are required to upload the model corresponding to the aforementioned version number and a reference score. For the same global model, the reference score uploaded by a single Category II user is judged based on the average of the experience scores from multiple Category II users. Rewards are given to Category II users after the feedback judgment is completed. Rewards for both Category I and Category II users are recorded using a DAG-structured blockchain. S7: After multiple weight adjustments, the accuracy of the global model on the edge server is tested. Training stops when the accuracy of the global model reaches the minimum threshold ACC1.

2. The blockchain-assisted incentive method for artificial intelligence learning according to claim 1, characterized in that, In S6, the reward for type I users is based on the training time... n The aggregate weight of each type I user The allocation is performed, and the total reward value corresponding to the global model is set to... Then the first n Rewards received by each Class I user for uploading a model satisfy: .

3. The blockchain-assisted incentive method for artificial intelligence learning according to claim 1, characterized in that, In S6, the reward for Category II users is based on the... n The degree of similarity between the experience rating of a single Type II user and the average experience rating of all Type II users is determined. Let the first type be... n The experience rating of the individual Category II users is SCn, and the average experience rating of all Category II users is [value missing]. avg The standard reward value is If the degree of proximity is defined as App, then... , No. n Rewards for each Category II user .

Images