A trust management method for transverse federated learning

Abstract

The application discloses a trust management method for transverse federal learning, which comprises the following steps: a main aggregation server issues a federal learning task; the main aggregation server selects candidates higher than a certain threshold to participate in the federal learning process based on the trust degree of the candidates; the candidates are users interested in the federal learning task and selected to participate in the federal learning task; a set of the candidates higher than the certain threshold is a participant set; the main aggregation server receives local model parameters, aggregates the received local model parameter information according to an aggregation rule, and then sends the aggregated model update back to the participants; in the interaction process of each round of iteration of the aggregation server and the participants, the aggregation server detects the local model of each participant for abnormality, records the behavior information of each participant, evaluates the trust degree of the participants according to the behavior information of the participants, and the evaluation result is used for user selection when the aggregation server issues a task next time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of trust management, and particularly relates to a trust management method for horizontal federated learning. BACKGROUND

[0002] With the continuous development of artificial intelligence technology, the traditional data processing mode of uploading a large amount of user data and personal information to the cloud for centralized model training faces huge privacy security challenges, so Google proposes the concept of federated learning. As a distributed machine learning framework, federated learning trains the model locally by each participant, and then uploads the model information (parameters or gradients) to the aggregation server. The aggregation server aggregates the local model information received by using certain aggregation rules to obtain a global model for all participants to share. This machine learning paradigm can make the local training data of each participant not shared with other participants and the aggregation server, thereby protecting the data privacy of the participants.

[0003] However, federated learning also faces security challenges due to its distributed nature. The most common attack is to launch a poisoning attack to destroy the integrity of the global model. Most existing federated learning algorithms usually assume that each participant is honest and trustworthy, that is, the participant will upload the real model data generated by effective data training. However, in reality, participants may produce incorrect global models due to various internal reasons such as malicious attack behavior or external uncontrollable factors such as hardware device problems. If the reliability of each participant cannot be guaranteed, the federated learning system may produce unexpected behavior outputs due to the participants, resulting in a decrease in the reliability of the entire system. This problem cannot be solved by traditional cryptography-based security mechanisms, and trust management, as an effective theory for solving uncertain problems, has been widely studied and applied in various network environments. SUMMARY

[0004] In view of the problem of untrusted participants in the federated learning system, the present application provides a trust management method for horizontal federated learning.

[0005] To achieve the above purpose, the technical scheme of the present application is as follows:

[0006] A trust management method for horizontal federated learning, comprising:

[0007] The main aggregation server publishes a federated learning task

[0008] The main aggregation server selects a candidate above a certain threshold to participate in the federated learning process based on the trust degree of the candidate; the candidate is a user interested in the federated learning task and selects to participate in the federated learning task; the set of candidates above the certain threshold constitutes a participant set;

[0009] The main aggregation server receives local model parameters, aggregates the received local model parameter information according to the aggregation rule, and then sends the aggregated model update back to the participants. This process is repeated until the model converges or the maximum number of iterations is reached. The local model parameters are the initial global model received by each participant in the participant set from the main aggregation server, and then each participant trains a local model using its own local data.

[0010] Further, in the interaction process of each round of iteration between the main aggregation server and the participants, the main aggregation server performs anomaly detection on the local models of the participants and records the behavior information of the participants.

[0011] Further, in the interaction process of each round of iteration between the main aggregation server and the participants, the main aggregation server performs anomaly detection on the local models of the participants and records the behavior information of the participants.

[0012] The abnormal local model detected by the anomaly detection will not be used for this round of model aggregation, and at the same time, the main aggregation server will record the behavior information of the participants in each round of iteration;

[0013] After a learning is completed, the main aggregation server calculates the direct trust degree of each participant according to a given trust degree evaluation method combined with the behavior information of the participants; at the same time, the aggregation server generates the recommended trust degree of each participant according to the recommended trust information of each participant recommended by the recommendation server, and continuously updates the recommendation information table; then, the main aggregation server integrates the direct trust degree and the recommended trust degree of the participants to obtain the final trust information, and updates the historical interaction trust information table of the participants, which is used for user selection when the main aggregation server publishes the next learning task.

[0014] Further, when the main aggregation server publishes a new federated learning task, the participants are reselected according to the trust degree of each candidate.

[0015] Further, during the entire interaction phase, the recommendation server sends the recommended trust information of the participants with which it has interaction history to the aggregation server according to its own experience.

[0016] Further, the aggregation server internally maintains a historical interaction trust information table HL and a recommended trust information table RL with all participants; wherein HL stores the historical interaction behavior information between the aggregation server and each participant, the set of historical interaction trust information of each participant constitutes a historical interaction trust information queue, which is sorted in chronological order, and at most stores P items of information; RL stores the recommended trust information of the recommendation server to each participant, the set of recommended trust information of each participant constitutes a recommended trust information queue, which is also sorted in chronological order, and at most stores Q items of information; exceeding the maximum storage limit of the queue, considering that the queue is full, when receiving new information, the information that enters the queue earliest in the current queue is deleted, and then the latest information is inserted for updating.

[0017] Further, all federated learning tasks are performed in the same context C.

[0018] Further, in a federated learning process, the aggregation server S j calculates the trust degree of the participant U i The trust degree calculation includes:

[0019] The participant behavior model construction step is to calculate the U i behavior reliability and behavior stability

[0020] The trust degree calculation step is to calculate the U i current trust degree.

[0021] Further, the participant behavior model construction step includes:

[0022] The participant U i participates in federated learning in the context C, assuming that m rounds of iterations are required to participate in this federated learning; in the rth round of iteration, the U i produced behavior information includes local model abnormality information upload local model delay information set Wherein, The smaller the value is, the more abnormal the U The larger the value is, the more abnormal the U i behavior in this round of iteration is; after the rth round of iteration, the U i produced behavior information is recorded as:

[0023]

[0024] The main aggregation server S j records the behavior information of the U i after each round of learning iteration, which constitutes the behavior model of the U i in this federated learning:

[0025]

[0026] U i The m rounds of iterative events participating in this learning are denoted as {e1, e2, ···, em}, where e k , k+1 , m} The time of e k occurs earlier than e k+1 ; The abnormality information set of U j recorded by S i in this federated learning process The time delay information set

[0027] is calculated by acc i , the abnormality factor α i of U i ∈ [0, 1] :

[0028]

[0029] is calculated by d i , the time delay factor β i of U i ∈ [0, 1] :

[0030]

[0031] Where, θ represents the forgetting factor, θ ∈ [0, 1]; Δt r represents the time interval between the current time and the occurrence time of e r , the earlier the time of a certain round of iteration occurs, the smaller the proportion of the behavior information of U i in this federated learning in the abnormality factor α i and the time delay factor β i of U i ;

[0032] The behavior reliability and the behavior stability of U i in context C are calculated

[0033] The behavior reliability of U i in context C

[0034]

[0035] Where, the factors affecting the behavior reliability of U i are two: ①, the abnormality factor α i , the smaller the value of α i , the smaller the value of Ui Behavioral reliability The smaller the value; ②, the time delay factor β i ,β i The smaller the value, the more reliable the behavior. The smaller;

[0036] Computational behavior stability First, the historical interaction trust information queue needs to be... All elements in Sort according to the corresponding time to form U i Reliability time series set in, Indicates S in context C j Storage about U i The historical interaction trust information queue; at this time, calculate U i Behavioral stability in context C

[0037]

[0038] Adjacent time intervals The more similar, The higher the level, the more stable the participant's behavior becomes; therefore, the stability of participant behavior is only one of the factors affecting participant trust.

[0039] U was obtained i The behavioral model is represented by the quadruple shown in equation (7):

[0040]

[0041] Furthermore, the specific steps for calculating the trust level include:

[0042] S j Maintain U i Historical interaction trust information queue and recommendation trust information queue; U i With S j The historical interaction trust information queue in context C is defined as follows:

[0043]

[0044] Where t represents the time when the historical interaction occurred. For U i In context C, the reliability of behavior is such that at most P items of information are stored; meanwhile, S j Records and U i The number of historical interactions is expressed as...

[0045] U i The recommendation trust information queue in context C is defined as follows:

[0046]

[0047] where rsis the recommended server, rtis the recommended trust degree, is the historical interaction times between U i and rs, t is the time when the recommended server releases this piece of recommended trust information, and the maximum storage quantity is Q;

[0048] S j is the historical interaction times between U i and rs in the context C, g is expressed as:

[0049] T g (S j ,U i ,C) = ω r ·T r (U i ,C) + ω d ·T d (S j ,U i ,C) (10)

[0050] That is, the recommended trust degree T r is the weighted sum of the direct trust degree T d and the recommended trust degree, and the weight calculation method of the two is shown in equations (11)-(16):

[0051]

[0052] ω d = 1-ω r (12)

[0053]

[0054]

[0055]

[0056]

[0057] where: is the historical interaction times between U i and S j in the context C, is the sum of the historical interaction times between U i and all recommended servers in the recommended trust information queue, and δ is the threshold in the function f(σ i );

[0058] U iThe recommended trust degree T r of the context C is calculated as shown in equation (17):

[0059]

[0060] U i The direct trust degree T d of the context C is calculated as shown in equation (18), which is seen to be a weighted sum of the historical trust degree T hist and the current trust degree T curr , and the weights of the two are obtained by equations (19) and (20) respectively:

[0061] T d (S j ,U i ,C)=ω hist ·T hist (S j ,U i ,C)+ω curr ·T curr (S j ,U i ,C) (18)

[0062]

[0063] ω curr =1-ω hist (20)

[0064] The parameter in equation (19) is the familiarity of S j to U i , and its calculation method is shown in equation (21), which is seen to be determined by the number of interactions between S j and U i , the more the number of interactions, the higher the familiarity between S j and U i , and the larger the value of ; in addition, the adjustment factor φ is an arbitrary constant and satisfies φ≥2, which is used to control the speed of tending to 1; if there is no historical interaction behavior between the participant and the aggregation server, i.e. , at this time

[0065]

[0066] Equation (22) gives the calculation method of the historical trust degree T i of U hist in the context C; when calculating the trust degree of the participant based on the historical behavior, the time decay of the trust behavior is considered, As a time decay factor, the further back in time a historical interaction occurred, the lower the trust level generated by that interaction in time T. hist The smaller the proportion it occupies;

[0067]

[0068] Put U i With S j Reliability of the behavior generated by this interaction in context C Represented as U i Current trust level T curr The calculation method is shown in formula (23), where the adjustment function g(h) is shown in formula (24), where h represents the actual historical number of interactions between the participants and the aggregation server; as shown in formula (24), U represents the number of interactions between the participants and the aggregation server. i With S j The number of historical interactions in context C, at this point The method for calculating the adjustment factor ε is shown in formula (25):

[0069]

[0070]

[0071]

[0072]

[0073] Where 0<λ≤1, is used to control the size of the interaction threshold H, and the calculation method of H is shown in formula (26); the smaller the value of λ, the larger the value of the threshold H; in the trust calculation based on the current behavior, with an uncertainty trust level of 0.5 as the benchmark, if U i Number of historical interactions At this point, due to the limited number of historical interactions, the uncertainty regarding the trustworthiness of the participants is high. The reliability of the participants' current behavior cannot fully determine their character; that is, the fewer the historical interactions, the higher the uncertainty regarding the participants' trustworthiness, and the closer the trust level will be to 0.5. That is, once the number of historical interactions reaches the interaction threshold, the current trust level is no longer processed; therefore, g(h) is a relation to U. i Based on dynamic analysis of current trust levels, as the number of historical interactions between the participant and the aggregation server increases, the aggregation server's trustworthiness of the participant's current learning increases; however, if the reliability of the current behavior is calculated... Directly consider U i Current trust level T curr =0.

[0074] Compared with the prior art, the present application has the beneficial effects that:

[0075] In view of the problem of untrusted participants in a federated learning system, the present application proposes a trust management method for horizontal federated learning, which evaluates the trust degree of participants by an aggregation server according to the behavior of the participants in the interaction process, timely and effectively discovers participants whose behavior does not meet the expectations (low trust degree), selects more trusted participants (high trust degree) to participate in federated learning, conducts model training, and improves system reliability. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 A federated learning trust management system framework is provided for the embodiments of the present application.

[0077] Figure 2 A trust degree change trend graph for six behaviors of participants is provided.

[0078] Figure 3 A trust degree change trend graph for sudden failure of a good behavior participant in the interaction process is provided.

[0079] Figure 4 A graph showing the influence of familiarity on trust degree evolution is provided.

[0080] Figure 5 A graph showing the influence of a time decay factor on trust degree evolution is provided.

[0081] Figure 6 A graph showing the influence of a current interaction trust adjustment factor epsilon on trust degree evolution is provided.

[0082] Figure 7 A comparison graph of a subjective logic model and the present trust degree evaluation model is provided. DETAILED DESCRIPTION

[0083] Embodiment:

[0084] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0085] In the existing federated learning participant trust degree evaluation method, the trust factors generally considered are relatively single, and the interaction results produced by the participants after each learning, i.e., positive interaction or negative interaction, are usually taken as their trust attributes, lacking fine-grained modeling of participant behavior data. Moreover, due to the multi-dimensionality and heterogeneity of participant behavior data in the federated learning system, the existing trust degree evaluation methods mostly use simple quantification and addition to analyze the behavior of the evaluation object, lacking correlation analysis of participant behavior data. Therefore, the existing trust degree evaluation method cannot comprehensively and accurately evaluate the credibility of participant behavior.

[0086] In order to more comprehensively and completely measure and evaluate the trust degree of the participants in the federated learning system, the application analyzes and models the behaviors that the participants may generate in the federated learning (including the time delay information of uploading the local model, the local model abnormality information, etc.). In the trust degree evaluation model, the familiarity of the aggregation server with each participant and the time of generating the interaction behaviors are considered, and the comprehensive trust degree of each participant is calculated based on the historical trust degree, the current trust degree and the recommended trust degree of each participant. For the phenomenon that the participants generate behaviors that do not meet the expectations due to accidental external uncontrollable factors (network environment, hardware equipment, etc.), the aggregation server will not mistakenly judge that the participant is not trustworthy. Therefore, the trust management method proposed by the application has higher accuracy and stronger robustness.

[0087] participate Figure 1 As shown in the figure, it is a federated learning trust management system framework proposed by the application. Based on this framework, each participant and the aggregation server generate a global model through the following processes:

[0088] ① The aggregation server publishes a federated learning task, describes the model training budget, the training task category, the computing resource and other information;

[0089] ② Users who are interested in the learning task select to participate in the federated learning task, and these users constitute a federated learning candidate set;

[0090] ③ The aggregation server selects the candidates whose trust degrees are higher than a certain threshold to participate in the federated learning process based on the trust degrees of the candidates, at this time, the set composed of these candidates that meet the requirements is called a participant set, wherein the participant set is a subset of the candidate set;

[0091] ④ In the model training stage, each participant in the participant set receives the initial global model sent by the server, then each participant trains a local model using the local data of each participant, and sends the local model parameter information to the aggregation server, the aggregation server aggregates the received local model parameter information according to the aggregation rule, and sends the updated model after aggregation back to the participants, and the process is repeated until the model converges or reaches the predetermined maximum iteration number.

[0092] During each iteration of interaction between the aggregation server and the participants, as described in step ④, the aggregation server performs anomaly detection on the local models of each participant and records their behavioral information. Specifically, abnormal local models detected by anomaly detection will not be used for model aggregation in this round. Simultaneously, the aggregation server records the behavioral information of each participant during each iteration, including anomaly degree information and latency information. After one learning cycle, the aggregation server calculates the direct trust degree of each participant based on a given trust evaluation method and the participants' behavioral information. Simultaneously, the aggregation server generates the recommended trust degree of each participant based on the recommended trust information obtained from the recommendation server and continuously updates the recommendation information table. Then, the aggregation server combines the participants' direct trust degree and recommended trust degree to obtain the final trust information, and updates the participants' historical interaction trust information table. This trust information is used for user selection when the aggregation server releases the next learning task. When the aggregation server releases a new federated learning task, it can reselect participants based on the trust degree of each candidate user. Throughout the interaction phase, the recommendation server (other aggregation servers in the system) sends recommendation trust information of participants with whom it has an interaction history to the aggregation server based on its own experience, as shown by the gray dashed line in the figure.

[0093] right Figure 1 The Federated Learning Trust Management System is configured as follows:

[0094] (1) The entire system involves multiple participants and M aggregation servers for federated learning model training. The set of aggregation servers is denoted as S = {S1, S2, ..., S...}. j ,···,S M A federated learning trust management system consists of a main aggregation server responsible for issuing federated learning tasks and completing model aggregation, and N participating parties, who jointly complete specific federated learning tasks. The set of participating parties in this system is denoted as U = {U1, U2, ..., U...}. i ,···,U N The remaining M-1 aggregation servers are responsible for completing other federated learning tasks, and at the same time provide the main aggregation server with recommendation trust information of each participant based on their own experience. In this invention, these M-1 aggregation servers are called recommendation servers, and for ease of distinction, the recommendation servers are denoted as RS = {rs1, rs2, ..., rs...} M-1}

[0095] (2) The aggregation server internally maintains a historical interaction trust information table HL and a recommendation trust information table RL for all participants in the system. HL stores historical interaction behavior information between the aggregation server and each participant. The collection of historical interaction trust information for each participant forms a historical interaction trust information queue, which is sorted chronologically and can store a maximum of P items. RL stores the recommendation trust information of the recommendation server for each participant in the system. The collection of recommendation trust information for each participant forms a recommendation trust information queue, which is also sorted chronologically and can store a maximum of Q items. If the queue exceeds its maximum storage limit, the queue is considered full. When new information is received, the oldest information in the current queue is deleted, and the latest information is inserted to update the queue.

[0096] (3) All federated learning tasks are performed within the same context C. The context can be represented as C = < dy, < T. begin ,T end >>, where dy is the learning task type, T begin and T end These represent the start and end times of learning, respectively. Here, time is represented by four dimensions: day, hour, minute, and holiday, denoted as the quadruple T. begin =T end =<day, hour, minute, holiday>.

[0097] (4) The trust level of a participant indicates the degree to which the participant's actual behavior matches the expected behavior. The higher the degree of matching, the higher the trust level of the participant, and vice versa. In this invention, the trust level of a participant is a real number in the closed interval [0,1], where a trust level of 0 indicates that the participant is not trusted at all, 1 indicates that the participant is completely trusted, and 0.5 indicates uncertainty.

[0098] (5) The relevant parameters and their meanings in the trust management system proposed in this invention are as follows:

[0099]

[0100] Based on the above settings, aggregation server S j After a federated learning task is released to the public, participants are selected from among the system's users to collaborate on training the global model. The following details the process of S during a federated learning process. j For participant U i The methods for calculating trust levels include:

[0101] Participant behavior model construction:

[0102] Participant U iIn context C, we participate in federated learning, assuming that m iterations are required for this process. During the r-th iteration, U... i The generated behavioral information includes local model anomaly information. Upload local model latency information set in, The smaller the value, The larger the value of U, the better. i The more abnormal the behavior in this iteration, the more likely the r-th iteration will end, U i The generated behavioral information is recorded as follows:

[0103]

[0104] S j Record U i The behavioral information from each round of learning iterations forms U i The behavioral model of this federated learning:

[0105]

[0106] U i The m iterations involved in this learning process are denoted as {e1, e2, ..., e...} k ,e k+1 ,···,e m}, where e k It occurred earlier than e k+1 Therefore, S in this federated learning process j Recorded U i Anomaly information set Delay information set

[0107] By acc i Calculations show that U i Abnormal factor α i ∈[0,1]:

[0108]

[0109] By d i Calculations show that U i The time delay factor β i ∈[0,1]:

[0110]

[0111] Where θ represents the forgetting factor, θ∈[0,1]. Δt r Indicates the current time and e r The time interval between occurrences; the earlier a particular iteration occurs, the longer the iteration U... iBehavioral information in U i The anomaly factor α generated by this federated learning i and delay factor β i The smaller the proportion it occupies.

[0112] Therefore, calculate U i Reliability of behavior in context C and behavioral stability

[0113] (1)U i Behavioral reliability in context C

[0114]

[0115] Among them, the influence of U i There are two factors affecting the reliability of behavior: ①, the anomaly factor α i α i The smaller the value of U i Behavioral reliability The smaller the value; ②, the time delay factor β i ,β i The smaller the value, the more reliable the behavior. The smaller.

[0116] (2) Calculate behavioral stability First, the historical interaction trust information queue needs to be... All elements in Sort according to the corresponding time to form U i Reliability time series set in, Indicates S in context C j Storage about U i The historical interaction trust information queue, which will be discussed in detail below. At this point, calculate U... i Behavioral stability in context C

[0117]

[0118] Adjacent time intervals The more similar, The higher the level, the better. This reveals that for highly stable participants, there are typically two scenarios: ① the participant's behavioral reliability remains consistently high; ② the participant's behavioral reliability remains consistently low. Therefore, participant behavioral stability is only one factor influencing participant trust levels.

[0119] Thus, U was obtained. i The behavioral model is represented by the quadruple shown in equation (7).

[0120]

[0121] The specific calculation process for trust level:

[0122] S j Maintain U i Historical interaction trust information queue and recommendation trust information queue. i With S j The historical interaction trust information queue in context C is defined as follows:

[0123]

[0124] Where t represents the time when the historical interaction occurred. For U i In context C, the reliability of behavior is such that at most P items of information are stored. Meanwhile, S... j Records and U i The number of historical interactions is expressed as...

[0125] U i The recommendation trust information queue in context C is defined as follows:

[0126]

[0127] Where rs represents the recommendation server and rt represents the recommendation trust level. For context C in U i The historical number of interactions with rs, where t is the time when the recommendation server published this recommendation trust information, and a maximum of Q items of information can be stored.

[0128] Therefore, S j For U i Trust level T in context C g Represented as:

[0129] T g (S j U i ,C)=ω r ·T r (U i ,C)+ω d ·T d (S j U i (C) (10)

[0130] That is, the recommendation trust level T r and direct trust level T d The weighted sum of the two. The weights of the two are calculated as shown in formulas (11)-(16).

[0131]

[0132] ω d = 1 - ω r (12)

[0133]

[0134]

[0135]

[0136]

[0137] wherein: is the U i and S j the number of historical interactions in the context C, is the U i and the sum of the number of interactions with all the recommended servers in the recommended trust information queue, δ is the threshold value in the function f(σ i ) to prevent unreasonable values due to the too large value of σ i .

[0138] U i The calculation method of the recommended trust degree T r of the context C is shown in formula (17).

[0139]

[0140] U i The calculation method of the direct trust degree T d of the context C is shown in formula (18), which can be seen as the weighted sum of the historical trust degree T hist and the current trust degree T curr , and the weights of the two can be obtained by formula (19) and (20).

[0141] T d (S j ,U i ,C) = ω hist ·T hist (S j ,U i ,C) + ω curr ·T curr (S j ,U i ,C) (18)

[0142]

[0143] ω curr = 1 - ω hist(20)

[0144] Parameters in formula (19) For S j For U i The degree of familiarity with it is calculated as shown in formula (21). It can be seen that its value depends on S. j with U i The more interactions, the higher the S j with U i The higher the level of familiarity between them, The larger the value, the better. Furthermore, the adjustment factor φ is an arbitrary constant that satisfies φ≥2, and is used for control. The velocity approaches 1. The larger the value of φ, the greater the velocity. The slower the speed approaches 1. If there is no historical interaction between the participants and the aggregation server, that is... at this time

[0145]

[0146] Formula (22) gives U i Historical trust level T in context C hist The calculation method proposed in this invention considers the time decay of trust behavior when calculating the trust level of participants based on historical behavior. This is the time decay factor. The further back in time a historical interaction occurred, the lower the trust level generated by that interaction in time (T). hist The smaller the proportion of [something], the less reliable the trust level. For example, if the last interaction between a participant and the aggregation server occurred ten years ago, then this historical interaction has almost no reference value for assessing the current trust level of the participant.

[0147]

[0148] To make it easier to distinguish, U i With S j Reliability of the behavior generated by this interaction in context C Represented as Therefore, U i Current trust level T curr The calculation method is shown in formula (23), where the adjustment function g(h) is shown in formula (24), where h represents the actual historical number of interactions between the participants and the aggregation server. Specifically, as shown in formula (24), U represents the number of interactions between the participants and the aggregation server. i With S j The number of historical interactions in context C, at this point The method for calculating the adjustment factor ε is shown in formula (25).

[0149]

[0150]

[0151]

[0152]

[0153] wherein 0 < λ ≤ 1, used to control the size of the interaction threshold H of the system, the calculation method of H is shown in formula (26). The smaller the value of λ, the larger the value of threshold H. In the trust degree calculation based on the current behavior, taking the uncertainty trust degree 0.5 as the benchmark, if U i the number of historical interactions At this time, due to too few historical interactions, the uncertainty of the trustworthiness of the participants is high, and the reliability of the current behavior of the participants cannot fully judge the good or bad of the participants, that is, the fewer the number of historical interactions, the higher the uncertainty of the trustworthiness of the participants, and the trust degree will tend to 0.5; if that is, the number of historical interactions reaches the interaction threshold, at this time, the current trust degree is no longer processed. Therefore, g(h) is the processing of U i Based on the dynamic analysis of the current trust degree, with the increase of the number of historical interactions between the participants and the aggregation server, the trustworthiness of the participants in this learning of the aggregation server is increased. However, if the current behavior reliability directly consider U i the current trust degree T curr = 0. In this way, it can ensure that the trust degree realizes the characteristics of slow growth and rapid decline, so as to resist the attack behavior of the attacker.

[0154] The rationality of the trust management system and the trust degree calculation method of the participants is verified by the experiment, and the performance of the trust management system in accurately identifying malicious participants and improving the reliability of federated learning is verified.

[0155] The experimental environment is Intel Core i5-10400 CPU, 160G memory, Windows 10 operating system. The programming environment is Python 3.8.8, anaconda 4.10.1.

[0156] Table 1 experimental parameter setting

[0157]

[0158]

[0159] During the experiment, 100 aggregation servers and 100 participants are deployed in the whole system. The aggregation server S1 is the server for publishing tasks, and the remaining 99 aggregation servers are recommendation servers that provide recommendation trust information to S1. The initial trust degree of all participants is 0.5. It is assumed that the aggregation server is trustworthy, and the recommendation server's recommended trust degree for the participant is accurate. There is no recommendation server whose recommended trust degree for a participant is higher or lower than the participant's real trust degree, and there is no collusion behavior. At the same time, the participants are online throughout the process, and the problem of dropping out in the middle is not considered. In this experiment, participants with a trust degree of 0.8-1 are considered to be good-behavior participants, participants with a trust degree of 0-0.4 are considered to be malicious-behavior participants, and the rest are considered to be uncertain-behavior participants.

[0160] During the experiment of the present application, it is assumed that the participants may exhibit six different behaviors, namely, behavior one, good-behavior participant, behavior good, high reliability, high stability; behavior two, malicious behavior, poor reliability, high stability; behavior three, malicious behavior, 10 good behaviors, 5 malicious behaviors alternately; behavior four, malicious behavior, 10 good behaviors, 10 malicious behaviors alternately; behavior five, malicious behavior, 1 good behavior, 1 malicious behavior alternately; behavior six, malicious behavior, 5 good behaviors, 10 malicious behaviors alternately. At the same time, in order to track the trust degree change trend of different behavior participants in the process of multiple federal learning interactions with the aggregation server, it is assumed that all participants participate in each learning process, whether they are malicious or not. In the following experiment, one federal learning iteration is called an interactive learning cycle.

[0161] Based on the above system deployment and assumptions, the model and method proposed in the present application are implemented, and the effectiveness of the proposed method is verified. The trust degree change trend of different behavior participants, the influence of accidental failure of good-behavior participants on their trust values, and the influence of the values of parameters in the proposed scheme on the trust evaluation results are analyzed. At the same time, the effectiveness of the existing trust evaluation scheme is compared to prove the advantages of the proposed scheme.

[0162] Trust degree change of different behavior participants:

[0163] In order to study and analyze whether the trust management system can effectively identify malicious behavior participants, the trust degree change trend of six different behavior participants in the process of 100 federal learning interactions with the aggregation server is tracked, and the results are shown in Figure 2 .

[0164] Figure 2, except for behavior one, the trust degree of the rest of the behavior participants will decrease to below 0.4 in a short time, thus being defined as malicious participants by the aggregation server, achieving the expected effect. The participant with behavior one (good behavior) will have its trust degree increase in a short time because the recommended trust degree of the good behavior participant is high, and the recommended trust degree will increase the trust degree of the participant to a certain extent in a short time, and then the trust degree of the participant will slowly increase for a long time; the trust degree of the participant with behavior two (continuous malicious behavior) will decrease linearly after one interaction with the aggregation server, and for such a malicious behavior participant, the aggregation server can quickly identify after one interaction; compared with the trust degree change trend of the participants with behaviors three, four and six, they are all malicious participants who alternate between good behavior and malicious behavior, and as the number of good behavior interactions decreases and the number of malicious behavior interactions increases, the trust degree of the participant decreases; compared with the trust degree change trend of the participants with behaviors four and five, although behaviors four and five are the same number of good behavior and malicious behavior alternately, the trust degree change trends of behaviors four and five are quite different, that is, as the interval between the alternation of good behavior and malicious behavior decreases, the trust degree of the participant will have a small increase trend, but it can also be identified by the aggregation server in the first few interactions.

[0165] The impact of accidental failure on the trust degree of the participant:

[0166] To illustrate the impact of accidental failure on the trust degree of the participant, a scenario of a good behavior participant suddenly failing during interaction with the aggregation server is simulated. It is assumed that a sudden failure occurs at the 43rd federated learning, and the trust degree change trend of the participant during the 100 learning interactions with the aggregation server is tracked, and the result is shown in Figure 3 .

[0167] From Figure 3 , during the interaction of the participant with the aggregation server, because of the sudden failure of the participant, the trust degree decreases sharply after the 43rd interaction learning period, but in the subsequent continuous good behavior interaction learning, the trust degree first has a slow rising process, and after accumulating multiple rounds of good behavior interaction, the trust degree quickly reaches the previous height. Therefore, if a good behavior participant has a bad behavior due to occasional sudden failure during interaction with the aggregation server, although the trust degree will decrease sharply at this time, the participant can still return to the trust degree level before the failure after a few good behavior interactions, and will not be judged as a malicious participant by the aggregation server, thus the trust degree evaluation model has good robustness.

[0168] The impact of parameters in the proposed trust management system on the trust degree evaluation:

[0169] To study the influence of familiarity Ω on the change of trust degree in the trust management system, the following experiment is conducted on the good behavior participants. Given 100 good behavior participants, set S A contains 50 participants, numbered 1-50, set S B contains 50 participants, numbered 51-100. In the first 50 interaction learning periods, set S A participants in set S B do not participate in learning, at this time set S B participants always maintain the initial state of 0.5; taking the 51st period as a benchmark, at this time set S A participants have a history of 50 interactions with the aggregation server, set S B participants have a history of 0 interactions with the aggregation server, different history interaction times result in different familiarity Ω. Under the premise of different familiarity Ω, all 100 participants in set S A and set S B participate in 100 federated learning, track the change of participant trust degree, the results are shown in Figure 4 , wherein the change of trust degree of participants in set S A is shown in the Figure 4 curve , and the change of trust degree of participants in set S B is shown in the Figure 4 curve .

[0170] From Figure 4 , in the comparison chart, the curve and the curve, as the number of historical interactions between participants and the aggregation server increases, the familiarity Ω between participants and the aggregation server increases, and the trust degree of participants is closer to the true value of the participant.

[0171] To study the influence of time decay factor on the change of trust degree in the trust management system, a behavior malicious participant is designed. In order to accumulate its own trust degree at a high level, the participant always maintains good behavior in the first 88 interaction learning periods with the aggregation server, and starts to perform malicious behavior from the 89th learning. At this time, based on the trust management system, the change trend of the trust degree of the malicious behavior participant is compared between the case where the time decay factor exists and the case where the time decay factor does not exist. The results are shown in Figure 5 . Among them, for the trust management system with time decay factor In cases where the participants' trust level is calculated, the trust level calculation method proposed in this invention is used; for cases where there is no time decay factor... In the case where all historical behaviors have the same weight regardless of how far back in time the historical interaction occurred, the historical trust level of a participant is the average of the reliability of all the participant's historical behaviors.

[0172] Depend on Figure 5 It is understood that the participant began to perform malicious behavior in the 89th cycle, at which point a time decay factor exists. Participants in the trust management system (as shown in the diagram) (The curve) indicates that trust can drop rapidly, and the aggregation server can detect the malicious participant within a short period of time; however, for cases without a time decay factor... Participants in the trust management system (as shown in the diagram) (The curve) Although the trust level will also decrease, the decrease is too small. Even after the participant has performed 10 malicious acts, the trust level still does not drop below 0.5. Therefore, the aggregation server will find it difficult to detect the malicious participant in a short period of time.

[0173] Furthermore, to investigate the impact of the current interactive trust moderating factor ε on the evolution of trust levels among different behavioral participants, based on this trust management system, the trust level change trends of six different behavioral participants were compared under two scenarios: the presence and absence of the current interactive trust moderating factor ε. The results are as follows: Figure 6 As shown. Specifically, when the trust management system has a current interaction trust adjustment factor ε, the trust level calculation of the participants is the trust level calculation method proposed in this invention; when there is no current interaction trust adjustment factor ε, ε is always set to 1 in the trust level calculation of the participants.

[0174] Depend on Figure 6 For participants with good behavior ( Figure 6 (a) The current interaction trust moderating factor ε has no effect on changes in trust level because the good behavior participant maintains the reliability R of the current behavior. curr >0.5; while for malicious actors ( Figure 6 (b),(c),(d),(e),(f)), regardless of whether the malicious actor is a continuous malicious actor or an alternating malicious actor, the existence of the current interactive trust moderating factor ε can inhibit the growth of the malicious actor's trust level.

[0175] Comparative experiment:

[0176] To compare the rationality of the subjective logic model and the trust calculation method proposed in this invention in assessing participant trust, a comparative experiment was conducted based on six different participant behaviors. The results are as follows...Figure 7 The curves are the trust degree change trends of the six behavior participants tracked using the subjective logic model. The curves are the trust degree change trends of the six behavior participants tracked using the subjective logic model. The curves are the trust degree change trends of the six behavior participants tracked using the subjective logic model.

[0177] The subjective logic model participant's trust degree can reach the maximum after one good interaction and then keep the maximum trust degree for the good behavior participant (a), while the subjective logic model participant's trust degree can reach the minimum after one interaction and then keep the minimum trust degree for the continuously malicious behavior participant (b). Figure 7 The subjective logic model participant's trust degree can reach the maximum after one good interaction and then keep the maximum trust degree for the good behavior participant (a), while the subjective logic model participant's trust degree can reach the minimum after one interaction and then keep the minimum trust degree for the continuously malicious behavior participant (b). Figure 7 The subjective logic model participant's trust degree can reach the maximum after one good interaction and then keep the maximum trust degree for the good behavior participant (a), while the subjective logic model participant's trust degree can reach the minimum after one interaction and then keep the minimum trust degree for the continuously malicious behavior participant (b). Figure 7 The subjective logic model participant's trust degree can reach the maximum after one good interaction and then keep the maximum trust degree for the good behavior participant (a), while the subjective logic model participant's trust degree can reach the minimum after one interaction and then keep the minimum trust degree for the continuously malicious behavior participant (b). Figure 7 The subjective logic model participant's trust degree can reach the maximum after one good interaction and then keep the maximum trust degree for the good behavior participant (a), while the subjective logic model participant's trust degree can reach the minimum after one interaction and then keep the minimum trust degree for the continuously malicious behavior participant (b).

[0178] The above examples are only for illustrating the technical concept and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made according to the essence of the present application should be covered within the protection scope of the present application.

Claims

1. A trust management method for horizontal federated learning, characterized in that, include: The main aggregation server publishes federated learning tasks; The main aggregation server selects candidates whose trust level is higher than a certain threshold to participate in the federated learning process. The candidates are users who are interested in the federated learning task and choose to participate in it; the set of candidates above a certain threshold is the participant set. The main aggregation server receives local model parameters, aggregates the received local model parameter information according to the aggregation rules, and then sends the aggregated model update back to the participants. This process is repeated until the model converges or reaches the predetermined maximum number of iterations. The local model parameters are the initial global model sent by the main aggregation server to each participant in the participant set. Then, each participant uses its own local data to train its local model. During each round of iterative interaction between the main aggregation server and the participants, the main aggregation server performs anomaly detection on the local models of each participant and records the behavioral information of each participant, including: Abnormal local models detected by anomaly detection will not be used in this round of model aggregation. At the same time, the main aggregation server will record the behavior information of the participants in each round of iteration. After a learning session, the main aggregation server calculates the direct trust score of each participant based on a given trust assessment method and the participants' behavioral information. Simultaneously, the main aggregation server generates the recommended trust score of each participant based on the recommended trust information obtained from the recommendation server, and continuously updates the recommendation information table. Then, the main aggregation server combines the participants' direct trust scores and recommended trust scores to obtain the final trust information, and updates the participants' historical interaction trust information table. This trust information is used by the main aggregation server for user selection when it releases the next learning task. The specific steps for calculating the trust level include: Main aggregation server S j Maintenance Participant U i Historical interaction trust information queue and recommendation trust information queue; U i With S j The historical interaction trust information queue in context C is defined as follows: Where t' represents the time when the historical interaction occurred. For U i In context C, the reliability of behavior is such that at most P items of information are stored; meanwhile, S j Records and U i The number of historical interactions is expressed as... U i The recommendation trust information queue in context C is defined as follows: Where rs represents the recommendation server and rt represents the recommendation trust level. For context C in U i The number of historical interactions with rs, where t is the time when the recommendation server published this recommendation trust information, and a maximum of Q items of information can be stored; S j For U i Trust level T in context C g Represented as: T g (S j ,U i ,C)=ω r ·T r (U i ,C)+ω d ·T d (S j ,U i ,C) (10) That is, the recommendation trust level T r and direct trust level T d The weighted sum of the two is calculated as shown in formulas (11)-(16): oh d =1-h r (12) in: For U i With S j The number of historical interactions in context C. For U i The sum of the number of interactions with all recommendation servers in the recommendation trust information queue, δ is the function f(σ) i The threshold in ); U i In context C, the recommendation trust level T r The calculation method is shown in formula (17): U i In context C, the degree of direct trust T d The calculation method is shown in formula (18), and it can be seen that its value is the historical trust level T. hist And current trust level T curr The weighted sum of the two can be obtained from formulas (19) and (20) respectively: T d (S j ,U i ,C)=ω hist ·T hist (S j ,U i ,C)+ω curr ·T curr (S j ,U i ,C) (18) oh curr =1-h hist (20) Parameters in formula (19) For S j For U i The degree of familiarity with it is shown in formula (21), and its value depends on S. j with U i The more interactions, the higher the S j with U i The higher the level of familiarity between them, The larger the value, the better; furthermore, the adjustment factor φ is an arbitrary constant and satisfies φ≥2, used for control. The speed tends to 1; if there is no historical interaction between the participants and the main aggregation server, that is, at this time Formula (22) gives U i Historical trust level T in context C hist The calculation method considers the time decay of trust behavior when calculating the trust level of participants based on historical behavior. As a time decay factor, the further back in time a historical interaction occurred, the lower the trust level generated by that interaction in time T. hist The smaller the proportion it occupies; Put U i With S j Reliability of the behavior generated by this interaction in context C Represented as U i Current trust level T curr The calculation method is shown in formula (23), where the adjustment function g(h) is shown in formula (24), where h represents the actual historical number of interactions between the participants and the main aggregation server; if it represents U i With S j The number of historical interactions in context C, at this point The calculation method for the adjustment parameter ε is shown in formula (25): Where 0<λ≤1, is used to control the size of the interaction threshold H, and the calculation method of H is shown in formula (26); the smaller the value of λ, the larger the value of the threshold H; in the trust calculation based on the current behavior, with an uncertainty trust level of 0.5 as the benchmark, if U i Number of historical interactions At this point, due to the limited number of historical interactions, the uncertainty regarding the trustworthiness of the participants is high. The reliability of the participants' current behavior cannot fully determine their character; that is, the fewer the historical interactions, the higher the uncertainty regarding the participants' trustworthiness, and the closer the trust level will be to 0.

5. That is, once the number of historical interactions reaches the interaction threshold, the current trust level is no longer processed; therefore, g(h) is a relation to U. i Based on dynamic analysis of current trust levels, as the number of historical interactions between the participants and the main aggregation server increases, the main aggregation server's trustworthiness of the participants' current learning increases; however, if the reliability of the current behavior is calculated... Directly consider U i Current trust level T curr =0.

2. The trust management method for horizontal federated learning as described in claim 1, characterized in that, When the main aggregation server publishes a new federated learning task, the participants are reselected based on the trust level of each candidate.

3. The trust management method for horizontal federated learning as described in claim 1, characterized in that, Throughout the interaction phase, the recommendation server sends recommendation trust information about participants with whom it has an interaction history to the main aggregation server based on its own experience.

4. The trust management method for horizontal federated learning as described in claim 1, characterized in that, The main aggregation server internally maintains a historical interaction trust information table HL and a recommended trust information table RL with all participating parties. HL stores historical interaction information between the main aggregation server and each participant. The collection of historical interaction trust information for each participant forms a historical interaction trust information queue, which is sorted in chronological order and can store a maximum of P items. RL stores recommendation trust information from the recommendation server to each participant. The collection of recommendation trust information for each participant forms a recommendation trust information queue, which is also sorted in chronological order and can store a maximum of Q items. If the maximum storage limit of the queue is exceeded, the queue is considered full. When new information is received, the oldest information in the current queue is deleted, and then the latest information is inserted to update it.

5. The trust management method for horizontal federated learning as described in claim 4, characterized in that, All federated learning tasks are performed within the same context C.

6. The trust management method for horizontal federated learning as described in claim 5, characterized in that, During a federated learning process, the main aggregation server S j For participant U i Trust calculation includes: The steps to build a participant behavior model to calculate U i Reliability of behavior in context C and behavioral stability The specific steps for calculating trust level, to calculate U i Current level of trust.

7. The trust management method for horizontal federated learning as described in claim 6, characterized in that, The steps for constructing the participant behavior model include: Participant U i In context C, participate in federated learning, assuming that participation in this federated learning requires m iterations; during the r-th iteration, U... i The generated behavioral information includes local model anomaly information. Upload local model latency information set in, The smaller the value, The larger the value of U, the better. i The more abnormal the behavior in this iteration, the more abnormal the r-th iteration ends, U i The generated behavioral information is recorded as follows: Main aggregation server S j Record U i The behavioral information from each round of learning iterations forms U i The behavioral model of this federated learning: U i The m iterations involved in this learning process are denoted as {e1, e2, ..., e...} k ,e k+1 ,···,e m }, where e k It occurred earlier than e k+1 During this federal learning process, S j Recorded U i Anomaly information set Delay information set By acc i Calculations show that U i Abnormal factor α i ∈[0,1]: By d i Calculations show that U i The time delay factor β i ∈[0,1]: Where θ represents the forgetting factor, θ∈[0,1]; Δt r Indicates the current time and e r The time interval between occurrences; the earlier a particular iteration occurs, the longer the iteration U... i Behavioral information in U i The anomaly factor α generated by this federated learning i and delay factor β i The smaller the proportion it occupies; Calculate U i Reliability of behavior in context C and behavioral stability U i Behavioral reliability in context C Among them, the influence of U i There are two factors affecting the reliability of behavior: ①, the anomaly factor α i α i The smaller the value of U i Behavioral reliability The smaller the value; ②, the time delay factor β i ,β i The smaller the value, the more reliable the behavior. The smaller; Computational behavior stability First, the historical interaction trust information queue needs to be... All elements in Sort according to the corresponding time to form U i Reliability time series set in, Indicates S in context C j Storage about U i The historical interaction trust information queue; at this time, calculate U i Behavioral stability in context C Adjacent time intervals The more similar, The higher the level, the more stable the participant's behavior becomes; therefore, the stability of participant behavior is only one of the factors affecting participant trust. U was obtained i The behavioral model is represented by the quadruple shown in equation (7):

Images