Collaborative fair federated learning framework based on trust evaluation mechanism

Abstract

Distributed cooperation and privacy protection characteristics enable federal learning to have unique technical advantages in the aspect of data sharing. In order to improve the reliability of the global model, the behavior of the user expected to participate is trustworthy, the user hopes to obtain fair return in model training, and the influence of resource and data distribution and computing power difference is reduced. An existing federated learning trust evaluation scheme has the problems that consideration factors are single, a coarse-grained calculation method is adopted, and it is difficult to reasonably evaluate the user trust degree and give fair return. For this purpose, the invention provides a collaborative fair federated learning framework based on a trust evaluation mechanism. A trust evaluation mechanism is composed of direct and recommended trust information, a user behavior mode is designed according to multiple attributes, and behaviors are described with fine grit. And according to an evaluation result, through fair optimization algorithm aggregation, a fair training weight is given to a user, and the accuracy of a global model is improved. The result shows that the method can accurately evaluate the trust level of the users in different behavior modes, the accuracy of the model is improved, and the anti-attack performance is better.

Description

Technical Field

[0001] The present invention belongs to the fields of federated learning privacy security and federated learning fairness, and particularly relates to a collaborative fair federated learning framework based on a trust evaluation mechanism. Background Art

[0002] As a new framework for distributed machine learning, federated learning can jointly train machine learning models on multiple local devices while only sharing local model parameters, effectively avoiding data leakage problems caused by direct transmission of original local data from local devices to the aggregation server. Although federated learning has provided a method for decentralized data training and model aggregation, due to the fact that computing nodes are often in an untrusted environment, and the communication environments, computing resources, etc. among computing nodes are not exactly the same, the following problems are specifically faced:

[0003] The first problem is that there is a lack of division of malicious users in the federated learning scheme. In federated learning, participants may be subject to external attacks or affected by their limited resources, resulting in unreliable behaviors. Malicious users in federated learning may spread false data or low-quality models to the aggregation server, thus affecting the aggregation result of the global model. These are all behaviors that affect the aggregation result, but there are currently no excessive evaluation criteria for such behaviors.

[0004] The second problem is that in the federated learning scheme, there may be significant differences in the data quality and computing capabilities of different participants. For all participants, data imbalance and diversity may lead to poor performance of the global model on the data of some participants. Such users are not considered malicious users, but their computing results contribute less to the global aggregation. There is currently a lack of a reasonable evaluation mechanism for distinguishing such users.

[0005] The third problem is that the contributions of different participants to the model may be different. If some participants feel that their contributions are not fairly recognized, their enthusiasm for participation may be reduced. Therefore, existing federated learning schemes lack effective methods for measuring the reliability and stability of entity behaviors.

[0006] The fourth problem is that most of the existing federated learning trust evaluation methods only use the interaction results (positive or negative) of participants in each round of model training as trust evaluation factors. Trust evaluation is an effective method to measure the reliability of entity behavior. The aggregation server selects the local models submitted by users with high trust levels to update the global model according to the trust levels of users, thereby improving the reliability of the global model. However, the existing solutions lack fine-grained modeling of participants' behaviors. In a federated learning system, the behavior data of participants is multi-dimensional and heterogeneous. Most of the existing trust evaluation schemes use subjective logic models or simply use four arithmetic operations to calculate the trust values of participants, lacking correlation analysis of behavior data.

[0007] In summary, while the existing federated learning solutions provide efficient aggregation, they still have deficiencies in differentiating user quality, maintaining the fairness of federated learning, and improving the enthusiasm of user training. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to propose a collaborative fair federated learning framework based on a trust evaluation mechanism. Specifically, this paper designs a method to evaluate the trust level of users participating in federated learning, which takes into account multiple behavioral attributes of participants and the temporal correlation of behavior data. In this scheme, the local (direct) trust value of a user is combined with the recommended (indirect) trust value to obtain the final trust level of the user. And a user behavior model is designed from multiple dimensions to comprehensively evaluate the local trust value and the recommended trust value of the user. In addition, this paper proposes a temporal correlation analysis method to measure the behavioral stability of users and their familiarity with the aggregation server, so as to more accurately describe their trust level.

[0009] To achieve the above object, the present invention provides the following technical solutions for implementation:

[0010] A collaborative fair federated learning framework based on a trust evaluation mechanism, including the following parts: a trust behavior model, a trust evaluation mechanism, and a fair aggregation algorithm.

[0011] The trust behavior model is used to evaluate the trust level of users participating in federated learning. The aggregation server needs to record as comprehensively as possible the behaviors of users participating in federated learning. In the task running background In the task running background, the trust evidence of users recorded by the aggregation server can be formalized as the quadruple in the formula form.

[0012]

[0013] Task running background Refers to information related to user behavior, such as the time when the behavior occurs, the type of learning task being executed when the behavior occurs, etc. In this paper, we only consider using behavior and recommendation trust information in the same context to complete trust evaluation. Therefore, we do not give the specific content of the context here, and the specific content should be determined according to the specific application scenario. Among them, is an ordered set of behaviors recorded when the aggregation server interacts with the user . is a set of recommendation trust values of other users for the user .

[0014] The aggregation server uses the information recorded in and to obtain the local trust profile and non-local trust profile of the user respectively, and these two are used to participate in the calculation of the trust function in this paper.

[0015] The behavior record set can be expressed by the following formula . Due to storage resource limitations, the length of the behavior record set is denoted as , and the maximum value of is denoted as records the behavior of the user in the th interaction with the aggregation server.

[0016]

[0017] can be expressed by the following formula . Among them, represents the degree of abnormality of the user in the th interaction with the aggregation server. The more abnormal the detection result, the smaller it is, and the overall value range is. represents the delay of the user in uploading the local model during the th participation in the federated learning iteration.

[0018]

[0019] The recommendation trust record set can be expressed by the following formula . Each element All are a piece of news, and the main content of the news is the recommended trust evidence of other users for the user .

[0020]

[0021] It can be expressed in the following formula . Among them is the identity of the recommending user, is the recommended trust value of the recommending user for the user , is the number of interactions between the recommending user and the user , is 's sending time.

[0022]

[0023] Based on the information recorded in, the aggregation server has the ability to describe the reliability and stability of user behavior. Specifically, it can be expressed by the following formula . Among them is the reliability of the user , is the stability of the user .

[0024]

[0025] is an ordered set of reliability information , which records the reliability of the user , and is expressed by the following formula .

[0026]

[0027] And 's calculation method is expressed by the following formula. Among them , are parameters used to describe the abnormality and delay conditions of the user respectively.

[0028]

[0029] is the abnormality factor of the user . The calculation formula is as formula As shown. The probability of the user exhibiting abnormal behavior in the future is closely related to their recent behavior. Therefore, we use to represent this property. Among them, is the time forgetting factor within the range. is the user The interval between the occurrence time of the interaction with other users or the aggregation server at the th interaction and the current time. Therefore, the earlier the interaction occurs, the lower the proportion of the anomaly factor in the calculation.

[0030]

[0031] Similarly, we calculate the delay factor through the formula.

[0032]

[0033] is the stability of the user , and the specific calculation method is as shown in formula . The more similar the reliability of the user is during the interaction or the aggregation process of federated learning training, the higher the behavioral stability of the user . For users with highly stable behaviors, there may be two behavioral patterns. The first pattern is that the user has always performed poorly (all values are small), and the other pattern is that the user has always performed well (all values are large).

[0034]

[0035] The calculation formula of the non-local trust profile is as shown in the formula. It consists of two aspects. The first is the total number of interactions between the user and other users except the aggregation server. The second is the latest recommended trust value of the user .

[0036]

[0037] The above describes the establishment criteria of the trust behavior model. Next, the proposed trust evaluation mechanism will be introduced. The goal of our method is to calculate the trust value of users participating in federated learning. Here, we believe that the higher the probability of timely delivering a high-precision local model to the aggregation server, the higher the trust value of the user, and vice versa. In this way, the aggregation server can select users with high trust values to update the global model, thereby improving the training efficiency and accuracy of the global model as much as possible.

[0038] The workflow of the trust evaluation mechanism is shown in the attached figure Figure 1 as follows. When the aggregation server directly interacts with the user , the aggregation server will record the user's behavior and update its behavior record . Then, based on the latest set of behavior records, the behavior reliability of the user is updated. Using the latest obtained reliability ( the last element in it), the aggregation server can evaluate the trust level of the user based on the current iteration.

[0039] Then, the local trust of the user is calculated. When calculating the local trust, the aggregation server continuously receives the recommended trust information sent by other users. Once the recommended trust information is received, the recommended trust of the corresponding user is immediately updated. Finally, by combining the latest recommended trust value of the user and its local trust value, the global trust value of the user is obtained by Equation . .

[0040]

[0041] is the local trust value of the user weight, is the recommended trust value of the user weight. Both values are greater than 0, and the sum is 1. The reason for considering both local interaction information and recommendation information in the trust evaluation function is that we cannot guarantee that there will always be trustworthy local interaction information or recommendation information. Therefore, we need coefficients and to determine whether and to what extent local trust information and recommendation trust information can be relied on during the trust calculation process. Depending on the relative amounts of local trust information and recommendation trust information, and will take different values in different situations.

[0042] Formula gives the calculation method. After obtaining the value, we can also easily calculate the value according to the above assumptions, as shown in Equation . We can see that the value depends on two parameters , and a function . And The larger the value, the greater the weight of the local trust value. reflects the ratio of the number of direct interactions between the user and the aggregation server to the average number of direct interactions between other users and the aggregation server. Similarly, represents the ratio of the number of direct interactions between the user and other users to the average number of direct interactions between the aggregation server and other users. represents the relative quality of the local trust information and the recommended trust information of the user . We can see that when there is no direct interaction, and have a value of 0. Similarly, if there is no indirect interaction between the user and other users, then is 0 and has a value of 1. Here, we use the relative number of direct interactions between the user and the aggregation server and the indirect interactions between the user and other users to measure the weights of local trust and recommended trust, so that in the case where the number of direct and indirect interactions is both large or small, the weights of both can be measured more accurately. The calculation methods of the parameters , , and the function form are shown in Equations , , , respectively. The parameter in the formula is a predefined threshold used to determine the upper bound.

[0043]

[0044]

[0045]

[0046]

[0047]

[0048]

[0049] Formula gives the calculation method of the local trust value. Each time an interaction occurs, the aggregation server evaluates the trust value based on the user 's behavior, denoted as . However, this result cannot fully reflect the trust level of the user . Therefore, we must make a comprehensive evaluation combining the trust levels. The user 's interaction behavior with other users during a historical period is denoted as , , The value of is the weighted sum of the current trust value and the historical trust value , as shown in Equation .

[0050]

[0051] Formula shows the calculation method of the historical trust value , which is the weighted sum of the behavior reliability of the user . The weight ensures that the impact of the user's behavior reliability on their credibility within a given time period decreases over time. Among them is the consumption time for calculating .

[0052]

[0053]

[0054] The current trust value can be calculated according to the following formula. This is the result of the recent evaluation of the user 's reliability( ). If the reliability of the user is less than 0.5 during the most recent interaction with the aggregation server, we will set the current trust value Assign it to 0. Otherwise, the current trust value should be between and 0.5. If the aggregation server has sufficient direct interaction experience with the user (the number of interactions exceeds a certain threshold), then the aggregation server and the user The more interactions there are between the aggregation server and the user , the more confident the aggregation server is in the reliability evaluation result of the user. Therefore, the current trust value is closer to . Otherwise, the current trust value tends to be more uncertain. Use the function in the formula to adjust the value of the current trust value , which can be obtained from the formula . It can be seen from the formula that the smaller the value of , the slower the speed at which approaches 1. It ensures that the value of the current trust value increases slowly and decreases rapidly, which is consistent with the property that trust values are difficult to increase but easy to decrease. The formula gives the calculation method of the parameter in the formula .

[0055]

[0056]

[0057]

[0058] The current trust value and the historical trust value weights can be expressed as and . They are represented by the formula and the formula respectively. The more familiar the aggregation server is with the user , and the more stable the behavior of the user , the more confident the aggregation server is in the local trust value evaluation result of the user . So is equal to the product of and . Here, represents the degree of familiarity of the aggregation server with the user . The more interactions there are between the aggregation server and the user , the larger the value of , indicating that the aggregation server is more familiar with the user The higher the degree of familiarity. The formula gives the calculation method.

[0059]

[0060]

[0061]

[0062] Recommended trust value can be calculated by the formula It is the weighted sum of the recommended trust values of other users for user , where is the second element. We can see that the more direct interaction experience the recommender has with user ui, the greater the weight of the recommender's recommended trust when calculating the recommended trust value .

[0063]

[0064] In summary, the global trust value of user can be calculated. The calculation basis of the fair aggregation algorithm is the global trust value of the user. The specific calculation process is as follows. The global trust value is classified by Gap statistics. It evaluates the quality of the clustering result by comparing the clustering effect of the data under different numbers of clusters with the expected random distribution, so as to determine the optimal number of clusters . This means dividing users into different clusters, and at the same time means that the user quality of different clusters is different. If multiple different clusters can be divided, then perform K-means++ clustering classification on them. Divide them into two categories: suitable for participating in aggregation training and not suitable for participating in aggregation training. If the user is divided into the category suitable for participating in aggregation training, then increase the count. If it is assigned to the category not suitable for participating in aggregation training, then increase the count. If multiple clusters cannot be divided, then for all users increase the count. The specific model aggregation expression is as shown in the following formula .

[0065]

[0066] Based on the above formula, the global aggregation model , we represent the trust degree naturally through the form of beta distribution and the global trust value Combined, it can naturally represent the trust degree. Moreover, the combination of beta distribution and Bayesian inference is very close, and the trust degree of participants can be continuously adjusted through Bayesian update. This enables the trust evaluation to dynamically adapt to the behaviors of different participants. It is possible to adjust the weights of different users participating in the model aggregation training through a fine-grained trust evaluation mechanism. While providing efficient aggregation, it has effectively improved in aspects such as differentiating user quality, maintaining the fairness of federated learning, and enhancing the enthusiasm of user training. Brief Description of the Drawings

[0067] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be described in detail and preferably below in conjunction with the drawings, where:

[0068] Figure 1 It is a flowchart of the trust evaluation mechanism.

[0069] Figure 2 It is a graph of experimental results. Detailed Embodiments

[0070] The following specific examples illustrate the embodiments 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. Various details in this specification can also 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 diagrams provided in the following embodiments only illustrate the basic concept of the present invention schematically. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0071] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to the present invention: For better illustrating the embodiments of the present invention, some components in the drawings will be omitted, enlarged, or reduced, which do not represent the dimensions of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted. The same or similar reference numerals in the drawings of the embodiments of the present invention correspond to the same or similar components.

[0072] The present invention proposes a collaborative fair federated learning framework based on a trust evaluation mechanism. Specifically, this paper designs a method to evaluate the trust level of users participating in federated learning, which takes into account multiple behavioral attributes of the participants and the temporal correlation of behavioral data. In this scheme, the local (direct) trust value of the user is combined with the recommended (indirect) trust value to obtain the final trust level of the user. And a user behavior model is designed from multiple dimensions to comprehensively evaluate the local trust value and the recommended trust value of the user. In addition, this paper proposes a time-related analysis method to measure the behavioral stability of the user and their familiarity with the aggregation server to more accurately describe their trust level.

[0073] The specific implementation is as follows:

[0074] The aggregation server records the trust evidence of the user, formalized as the quadruple in Equation (1) form. And according to the above formula perform the calculation of the quadruple data.

[0075] With the above quadruple , according to the formula calculate the global trust value of the user.

[0076] With the above global trust value , perform Gap statistical classification on the global trust value. If it can be divided into multiple clusters, continue with K-means++ classification to divide the users. Finally, aggregate the global model according to the aggregation expression , which completes one aggregation process of this federated learning framework.

[0077] The above is a complete round of training process. The main symbols used in this paper are shown in Table 1.

[0078] Table 1 Symbol Explanation

[0079] Symbol Meaning #timg# Number of users (devices) in the system #timg# Identity of the #timg#-th user #timg# Number of users participating in modeling #timg# Behavior record of user #timg# when participating in the iterative learning of the aggregation server for the #timg#-th time under background #timg# #timg# Behavior record set of user #timg# under background #timg# #timg# Length of #timg# #timg# Maximum value of #timg# #timg# Global aggregation model #timg# Local model of user #timg# #timg# Recommendation trust of other users for user #timg# under background #timg# #timg# Recommendation trust record set #timg# of user #timg# under background #timg# #timg# Maximum length of #timg# #timg# Global trust value of user #timg# under background #timg# #timg# Local trust value of the aggregation server for user #timg# under background #timg# #timg# Historical trust value of the aggregation server for user #timg# under background #timg# #timg# Direct trust value of the aggregation server for user #timg# under background #timg# #timg# Recommendation trust value of the aggregation server for user #timg# under background #timg# #timg# Historical trust value weight #timg# Weight of the current trust value #timg# Weight of the local trust value #timg# Weight of the recommendation trust value #timg# Abnormality factor of user #timg# #timg# Latency factor of user #timg# #timg# Degree of abnormality of user #timg# when participating in the #timg#-th aggregation training #timg# Degree of latency of user #timg# when participating in the #timg#-th aggregation training #timg# Time forgetting factor when calculating the abnormality and latency factors #timg# Interval between the occurrence time of #timg# and the current time #timg# Time forgetting factor when calculating the historical trust value #timg# Reliable record set of user #timg# under background #timg# #timg# The #timg#-th reliable record of user #timg# under background #timg# #timg# Behavior stability of user #timg# under background #timg# #timg# Degree of familiarity between the aggregation server and user #timg# #timg# Familiarity adjustment factor #timg# Number of interactions between the aggregation server and user #timg# under background #timg# #timg# Number of interactions between user #timg# and other users under background #timg# #timg# Threshold for judging the reliability of the user in the current interaction #timg# Adjustment function for calculating the current trust value #timg#, #timg#, #timg# Coefficient for calculating #timg# #timg# Adjustment function for calculating #timg#

[0080] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A collaborative fair federated learning framework based on trust evaluation mechanism, including trust behavior model, trust evaluation mechanism, and fair aggregation algorithm; The trust behavior model is characterized by: The trust behavior model is used to evaluate the degree of trust of users in participating in federated learning. The aggregation server needs to record the user's behavior in participating in federated learning as comprehensively as possible. In the background of the task running In the aggregation server, users recorded The trust evidence can be formalized as The four-tuple in form; Task running background Refers to information related to user behavior, such as the time when the behavior occurs, the type of learning task being performed when the behavior occurs, etc. Is to record the aggregation server and the user An ordered set of behaviors during interaction; Is other users to users The set of recommended trust values; Aggregation server exploits are recorded in and In the information, we get the user Local trust configuration file and non-local trust profiles , these two are used to participate in the calculation of the trust function in this paper; Behavior Record Set It can be expressed as the following formula ; Due to storage resource limitations, the behavior record set The length of , The maximum value of ; Recorded user Follow the aggregation server Behavior during the first interaction; It can be expressed as the following formula ;in Indicates the user With the aggregation server in The more abnormal the detection result is, the The smaller it is, the overall value range is [0,1]; Indicates the user In the The delay in uploading the local model when participating in the federated learning iteration; Recommended Trust Record Set It can be expressed as the following formula ; each element It is a message, and the main message content is other users' Evidence of trust in recommendations; It can be expressed as the following formula of the form; For the identity of the recommending user, Recommend user to user The recommended trust value, Recommend users and users The number of interactions between for The time of sending; based on The information recorded in the aggregation server has the ability to describe the reliability and stability of user behavior; specifically, the following formula can be used expression; For users Reliability, For users stability; Reliability information An ordered collection that records the user The reliability of To express; and The calculation method of is expressed by the following formula; , are used to describe users Parameters for exception and delay situations; Is a user The abnormal factor of As shown in ; the probability of a user's future abnormal behavior is closely related to his recent behavior; so we use To express this property; is the temporal forgetting factor in the range [0,1]; For users Communicate with other users or aggregate servers in The interval between the occurrence time of the first interaction and the current time; therefore, the earlier the interaction occurs, the lower the proportion of the anomaly factor in the calculation; Similarly, we calculate the delay factor by the formula ; For users The stability of As shown; User Reliability of behavior during interactive or federated learning training aggregation The more similar they are, the more likely users are to The higher the stability of the behavior, the higher the stability of the behavior. For users with high stability behavior, there may be two behavior patterns. The first pattern is that the user has always performed poorly (all The other mode is that the user has been behaving well (all are all very large); Non-local trust profiles The calculation formula is as shown in the formula; it consists of two aspects; the first is the user Total number of interactions with other users besides the aggregator server ; The second is the user The latest recommended trust value ; 。 2. The collaborative fair federated learning framework based on the trust evaluation mechanism according to claim 1 comprises a trust behavior model, a trust evaluation mechanism, and a fair aggregation algorithm; the trust evaluation mechanism is characterized by: When the aggregation server and the user When interacting directly, the aggregation server will record the user and update its behavior record ; Then, based on the latest set of behavior records, update the user Behavioral reliability; Using the most recently obtained reliability ( The last element in ), the aggregation server can evaluate the user based on the current iteration The level of trust; Then, calculate the user When calculating local trust, the aggregation server will continuously receive the recommended trust information sent by other users; once the recommended trust information is received, the recommended trust of the corresponding user will be updated immediately; finally, combined with the user The latest recommendation trust value and its local trust value are given by Get User The global trust value ; For users The local trust value Weight, For users Recommended trust value Weight; both values ​​are greater than 0, and the sum is 1; the reason why both local interaction information and recommendation information are considered in the trust evaluation function is that we cannot guarantee that there will always be trustworthy local interaction information or recommendation information; Therefore, we need the coefficient and To decide whether and to what extent local trust information and recommended trust information can be relied upon in the trust calculation process; Depending on the relative amount of local trust information and recommended trust information, and The value of will be different in different situations; formula Given The calculation method of We can also easily calculate the value of The value of As shown; we can see The value depends on two parameters , and a function ; and The larger the value, the greater the weight of the local trust value; Reflects the user The ratio of the number of direct interactions with the aggregator to the average number of direct interactions of other users with the aggregator; same, Indicates user The ratio of the number of direct interactions with other users to the average number of direct interactions between the aggregating server and other users; Indicates user The relative quality of local trust information and recommended trust information; we can see that when there is no direct interaction, and The value of is 0; Likewise, if the user Without indirect interaction with other users, is 0, The value is 1; here, we use the user Direct interaction with the aggregation server and users The relative number of indirect interactions with other users is used to measure the weights of local trust and recommendation trust, so that the weights of both direct and indirect interactions can be measured more accurately when the number of both is large or small; parameter , , And the functional form The calculation methods are shown in Formula , , , ; Parameters in the formula is a predefined threshold used to determine The upper bound of formula Given a local trust value The calculation method is that each time an interaction occurs, the aggregation server calculates the The behavior evaluation trust value is denoted as ; But this result cannot fully reflect the user So we have to make a comprehensive evaluation combining the trust level of users The interaction behavior with other users in a historical period is recorded as , The value of is the current trust value Historical trust value The weighted sum of As shown; formula Shows historical trust value The calculation method is user The weighted sum of the behavioral reliabilities; weight Ensure that the impact of the reliability of user behavior in a given period of time on its credibility decreases over time; It is calculated The time consumed; Current Trust Value The value of can be calculated according to the following formula; this is the most recent reliability( ) evaluation result; if the user If the reliability of the most recent interaction with the aggregation server is less than 0.5, we set the current trust value Assign a value of 0; otherwise, the current trust value The value should be between and 0.5; if the aggregation server and the user If there is enough direct interaction experience (the number of interactions exceeds a certain threshold), the aggregation server and the user The more interactions there are, the more confident the aggregation server is in its reliability assessment results, so the current trust value The closer the value is to ; Otherwise, the current trust value The value of is more uncertain; using the formula Functions in Adjust the current trust value The value of can be obtained from It can be seen that The smaller the value of The slower it approaches 1, the more likely it is to reach the current trust value. The value of increases slowly but decreases rapidly, which is consistent with the property that trust value is difficult to increase but easy to decrease; Formula Given the formula parameter Calculation method of Current Trust Value and historical trust value The weight can be expressed as and ; Use the formula and formula Aggregation server for users The more familiar, the user The more stable the behavior of the aggregation server, the more The more confident the local trust value evaluation result is; so The value is equal to and Here, Indicates the aggregation server for users Familiarity with the server and the user The more interactions you have, The larger the value, the more the aggregation server supports users. The more familiar you are with the formula Given Calculation method of Recommended Trust Value It can be expressed by the formula to perform calculations; it is other users' The weighted sum of the recommendation trust values ​​of yes The second element of; we can see that the recommender and the user The more direct interaction experience with the UI, the more trust the recommender has in his recommendation. Calculating the recommendation trust value The greater the weight; 。 3. The collaborative fair federated learning framework based on the trust evaluation mechanism according to claim 1 comprises a trust behavior model, a trust evaluation mechanism, and a fair aggregation algorithm; the fair aggregation algorithm is characterized by: The specific calculation process is as follows: Gap statistical classification is performed on the global trust value, which evaluates the pros and cons of the clustering results by comparing the clustering effects of data under different cluster numbers with the clustering effects under the expected random distribution, thereby determining the optimal number of clusters. ; This means that users are divided into different clusters, which also means that the quality of users in different clusters is different; if multiple clusters can be divided, K-means++ clustering classification is performed on them; they are divided into two categories: suitable for participating in aggregation training and unsuitable for participating in aggregation training; if users are divided into the category of suitable for participating in aggregation training, then Increase the count; if it is classified as unsuitable for participating in aggregation training, then Increase the count; if it cannot be divided into multiple clusters, all users Increase the count; the specific model aggregation expression is as follows As shown; According to the above formula, the global aggregation model is obtained , we use the form of Beta distribution and the global trust value Combined with the Bayesian inference, the trust can be naturally expressed; the Beta distribution is closely combined with the Bayesian inference, and the trust of the participants can be continuously adjusted through Bayesian updating; this enables the trust evaluation to dynamically adapt to the behavior of different participants; the weights of different users participating in the model aggregation training can be adjusted through a fine-grained trust evaluation mechanism; while providing efficient aggregation, effective improvements have been made in distinguishing user quality, maintaining the fairness of federated learning, and increasing the enthusiasm of user training.

Images