A method, system, device and medium for trust assessment of a machine learning participant

Abstract

A trust degree evaluation method, system, device and medium of a machine learning participant, based on a transfer learning technology, the trust degree evaluation model deployed on the server side when the system is initialized is migrated to different context scenarios, for the context c t If the distance between the source domain and the target domain is less than a predefined threshold, the trust degree evaluation model migration based on the TrAdaBoost algorithm is adopted, the participant trust degree evaluation model in the target domain is obtained, and the trust degree of the participant in the target domain is evaluated based on the model; the model obtained by migration is used by the aggregation server in the target domain, the trust evidence of the participant is input, and the trust degree of the participant in the target domain is obtained; the trust degree of the federated learning participant can be accurately evaluated in the case of insufficient trust evidence, and the precision of the trust evaluation result of the same participant in different context scenarios is improved.

Description

Technical Field

[0001] This invention belongs to the field of cyberspace security technology, and in particular relates to a method, system, device and medium for assessing the trust level of participants in machine learning. Background Technology

[0002] Federated learning achieves privacy-preserving multi-party collaborative modeling by sharing models instead of data, without requiring data to leave the local machine. However, in real-world applications, federated learning still faces significant security challenges. We cannot guarantee that participants will always be honest and trustworthy; the reliability of their behavior cannot be guaranteed during the learning process, thus compromising the reliability of the federated learning system and the integrity of the global model. Therefore, preventing unreliable participants from involvement in federated learning and improving its reliability and the credibility of its results have become crucial issues that federated learning urgently needs to address.

[0003] Trust assessment is a technique for evaluating entity reliability in uncertain network environments and has been widely applied in various scenarios. Trust assessment technology can be used in federated learning systems to determine the reliability of participants by evaluating their trustworthiness. Based on this, participants in the federated learning process are selected, avoiding unreliable participants and thus ensuring the reliability of the federated learning system and the credibility of the learning results. With the rapid development of mobile communication, online services, and other technologies, the data used for trust assessment is becoming increasingly massive. More and more researchers at home and abroad are beginning to explore the use of machine learning to achieve intelligent, accurate, and efficient trust assessment. The basic idea of ​​a machine learning-based trust assessment scheme is as follows: first, a model is trained using existing trust evidence; then, the trust evidence generated during subsequent learning is directly used as input to the machine learning model for trust assessment.

[0004] However, most existing trust assessment schemes for federated learning participants assume that all trust evidence comes from the same context, neglecting the impact of context on trust assessment. Context, as one of the most crucial factors in the trust assessment process, cannot be ignored in evaluating the trust levels of participants in a federated learning system. Federated learning participants may possess different resources and exhibit varying levels of behavioral reliability in different contexts; therefore, trust evidence generated from different contexts cannot be directly integrated. Furthermore, trust assessment models trained from trust evidence generated from different contexts are not directly equivalent.

[0005] Aggregator server S in context c s The federated learning task A released in [the document] generates a large amount of trust evidence for training the training context c due to the participation of numerous stakeholders in the model training. sThe trust assessment model Ma is used in the subsequent federated learning process. The server can then utilize model Ma to efficiently and accurately assess the trust levels of the participants; and when S is in context c... t The new federated learning task B released in [the document] presents a challenge: if the trust evidence generated by the participants in task B is insufficient, directly using this trust evidence to train the trust evaluation model will result in low accuracy. Conversely, directly using trust evaluation models obtained from other contexts (e.g., Ma) may lead to significant biases in the evaluation results. Therefore, how to tailor the trust evaluation model to a specific context (c) on the server side is crucial. t In cases where there is insufficient evidence of trust among the participants in (c), these participants can be utilized in other contexts (c) s The high-precision trust assessment model and trust evidence generated in (s≠t) are used to assess the trust of participants in c. t Accurately assessing trust levels is one of the challenges in building trust assessment models for participating parties in federated learning.

[0006] Existing methods for assessing trust among participants in federated learning, whether function-based or machine learning-based, assume that all trust evidence originates from a single context, ignoring the significant impact of context on trust assessment results. This can lead to biases in participant trust assessments across cross-context domains. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the present invention aims to provide a method, system, device, and medium for evaluating the trust level of machine learning participants. When trust evidence obtained by the server originates from different contexts, the TrAdaBoost algorithm is used to transfer the trust level to context c. s The trust assessment model is applied to context c. t In this study, the trust assessment model is made transferable across contexts, enabling the establishment of a trust assessment model even when trust evidence samples are insufficient. This allows for accurate assessment of the trust level of participants in federated learning when trust evidence samples are insufficient, thereby improving the accuracy of trust assessment results for the same participant in different context scenarios.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A machine learning participant trust assessment method is proposed. Based on transfer learning, the trust assessment model deployed on the server during system initialization is transferred to different context scenarios. For situations where there is insufficient trust evidence samples in a particular context, resulting in no suitable trust assessment model, the distance between the source and target domains is measured. If the distance between the source and target domains is less than a predefined threshold, a trust assessment model transfer based on the TrAdaBoost algorithm is used to obtain a participant trust assessment model in the target domain. Trust assessment of participants in the target domain is then performed based on this obtained participant trust assessment model. Finally, in the target domain, the aggregation server uses the transferred model, inputs participant trust evidence, and obtains the participant's trust level in the target domain.

[0010] A method for assessing trust among participants in machine learning, based on the following system assumptions:

[0011] 1) A federated learning system consists of an aggregation server and participants. The aggregation server can publish federated learning tasks and aggregate models. It can also evaluate the trustworthiness of participants and make corresponding decisions based on the evaluation results. The aggregation server is trustworthy. Participants are users who participate in federated learning tasks. There are multiple participants in a federated learning system.

[0012] 2) Trust is defined as the server's expectation of the reliability of the behavior of the participants in federated learning, that is, the probability that the behavior of the participants is reliable. The higher the probability that the behavior of the participants is reliable in the subsequent learning process, the higher the trust of the aggregation server for the participants, and vice versa. The trust space is a closed continuous interval [0,1], that is, the trust level is any real number in [0,1], where 0 means that the aggregation server does not trust the participants at all, 1 means that the aggregation server trusts the participants completely, and 0.5 means that the degree of trust of the aggregation server for the participants cannot be judged.

[0013] 3) Context is an important consideration when conducting trust assessments of participants in federated learning; context refers to the environmental information related to the participants' trust evidence; context is formally represented as C =<dy,T> Where dy is the learning task, including image classification tasks and text classification tasks; T = <T begin ,T end > is the study time, where T begin and T end These represent the start and end times of learning, respectively, for time T. begin and T end All are described using four dimensions, T begin =T end =<day,hour,minute,holiday> These are the day, hour, minute, and holiday signs, respectively.

[0014] 4) In a cross-context federated learning participant trust evaluation scheme, assume the context c s The domain in which it exists is called the source domain, and the context c t The domain in which it is located is called the target domain;

[0015] The source domain is formally defined as D. s ={χ s ,P(X s )},in, Indicates context c s The source domain has n samples of trust evidence among the participating parties. s samples, P(X) s ) represents the marginal probability distribution of the samples; the task in the source domain is Where, γ s For trust level, For context c s A model for predicting trust levels among participating parties; Let x be the set of labeled samples from the source domain. si The corresponding label is γ si Each element This represents the p trust attributes that each sample possesses. The trust attribute information includes the anomaly information, contribution information, and latency information of the federated learning participants.

[0016] The target domain is formally defined as D. t ={χ t ,P(X t )},in, This indicates that the participating parties trust the evidence sample, and the target domain has n. t samples, P(X) t () represents the marginal probability distribution of the sample; the task in the target domain Where, γ t For trust level, For context c t Trust prediction model in China; For the target domain, the set of labeled samples, sample x ti The corresponding label is γ ti Each element Let p represent the p trust attributes possessed by each sample. These trust attributes include anomaly information, contribution information, and latency information of the participants in the federated learning process. Assume 0 ≤ n. t <<n s It is necessary to utilize the source domain D s and its learning task T s Knowledge enhancement goal prediction model The accuracy.

[0017] A method for assessing the trust level of participants in machine learning, specifically including the following steps:

[0018] Step 1. Measure the similarity between the source and target domains;

[0019] Step 2. Trust evaluation model transfer based on TrAdaBoost algorithm;

[0020] Step 3. After successful model migration, the aggregation server in the target domain uses the migrated model, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain.

[0021] Step 1 specifically involves:

[0022] The maximum mean discrepancy (MMD) is used to measure the distance between the source and target domains. A smaller MMD value indicates a smaller distance between the two domains, signifying higher similarity in their data distributions, and vice versa. When measuring the similarity between the source and target domains, a confidence threshold 'd' is given as a threshold for whether model transfer is possible. If the distance between the source and target domains is not greater than the predefined confidence threshold 'd' (i.e., MMD ≤ d), then the similarity between the source and target domains is considered high. In this case, the model transfer method is used based on the formally defined source domain value 'd'. s and its learning task T s Perform model transfer to obtain a participant trust assessment model for the target domain. If the distance between the source domain and the target domain is greater than the predefined confidence threshold d, i.e., MMD>d, the model transfer method cannot obtain a good confidence evaluation model, and transfer is not performed in this case; the confidence level d is a hyperparameter.

[0023] The solution for the squared distance of the MMD of the source and target domain datasets, as shown in Equation (1), can measure the distance between two distinct but related distributions in the Reproducing Kernel Hilbert Space (RKHS); X in Equation (1) s and X t These are the source domain dataset and the target domain dataset, with dataset sizes n respectively. s and n t ; It is a mapping function used to map the original variables to the regenerated Hilbert space; H represents the MMD squared distance between the source and target domain datasets, which is... The data is mapped to a regenerated Hilbert space for measurement;

[0024]

[0025] Expanding formula (1), the squared distance between the source and target domain datasets is obtained by using the Gaussian kernel function k(·) in Support Vector Machine (SVM), as shown in formula (2); where, the Gaussian kernel function... ||uv|| 2 Let U be the Euclidean distance between samples u and v.

[0026]

[0027] Step 2 specifically involves:

[0028] If the distance between the source domain and the target domain in step 1 is no greater than a predefined confidence threshold d (i.e., MMD ≤ d), the model transfer method is used to evaluate the trust level of the participants in the target domain, as follows:

[0029] Trust evaluation model transfer based on TrAdaBoost algorithm; let T be the set of labeled samples in the source domain. source (of size n), the set of labeled samples in the target domain is denoted as T. target (Of size m), merge the labeled sample set from the source domain with the labeled sample set from the target domain to obtain the dataset T = T. source ∪T target The size of T is n+m, represented as T={(x1,y1),…,(x n ,y n ),(x n+1 ,y n+1 ),…,(x n+m ,y n+m Let T be all the labeled datasets after merging; each element in T is a sample instance, and each sample instance contains a feature space and a label; use dataset T as training data to train the trust evaluation model:

[0030] 2.1 Initialize the weight distribution of the training data

[0031] For dataset T = {(x1,y1),…,(x n ,y n ),(x n+1 ,y n+1 ),…,(x n+m ,y n+m The training samples in )} are assigned initial weights D1, as shown in formula (3):

[0032]

[0033] in, Represents data sample (x) i ,y i The weight of ) and has

[0034] 2.2 The maximum number of iterations is predefined as N. Given a learning algorithm, Learner, this learning algorithm Learner is used for iteration, with t = 1, ..., N, and then the following steps are performed:

[0035] 2.2.1 Based on training set T and weight distribution D t (D t (For the weight distribution of the training samples in the t-th iteration) call the Learner algorithm to obtain the weak classifier h. t :X→R, where X={x1,…,x n ,x n+1 ,…,x n+m}, R is the confidence level value, where R∈[0,1]; Formula (5) gives the weak classifier h t In distribution D t The maximum error on each sample instance x i Weak classifier h t In distribution D t Adjustment error As shown in formula (6);

[0036]

[0037]

[0038] 2.2.2 Calculate h t In distribution D t Total adjustment error ε t As shown in formula (7); here ε t It must be less than 0.5 if ε t If the value is greater than 0.5, then for ε t Make a correction, let ε t =0.5, to prevent the algorithm from stopping;

[0039]

[0040] 2.2.3 Calculate the intermediate value β t As shown in formula (8):

[0041] β t =ε t / (1-ε t (8)

[0042] 2.2.4 Calculate the weak classifier h using formula (9) t The weight α in the final classifier t The smaller the error, the greater the proportion of weak classifiers in the final classifier;

[0043] α t =1 / 2ln(1 / β) t (9)

[0044] 2.2.5 Update the weight distribution of the training samples using formula (10):

[0045]

[0046] Among them, D t (k) represents the distribution D t The k-th element in β and β t The setting is for D t Adjust the middle element, Z t Let be the normalization constant, then we have

[0047] The trust evaluation model based on TrAdaBoost performs N transfer iterations. After N iterations, data in the source domain that conforms to the target domain task has higher weights, while data that does not conform to the target domain task has lower weights. Data with higher weights in the source domain results in smaller errors in the target domain.

[0048] Step 3 specifically involves:

[0049] According to the weight α of the weak classifier t Combining all weak classifiers yields a strong classifier, as shown in formula (11):

[0050]

[0051] Where x is the evidence of trust, h f (x) represents the trust assessment model; at this point, the model transfer was successful.

[0052] Then, the aggregation server in the target domain uses the trust assessment model obtained from the migration, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain.

[0053] A trust assessment system for machine learning participants includes:

[0054] The similarity measurement module is used in step 1 to measure the distance between the source domain and the target domain using the maximum mean difference (MMD).

[0055] The model transfer module is used in step 2 to perform trust evaluation model transfer based on the TrAdaBoost algorithm;

[0056] The trust assessment module is used in step 3. After the trust assessment model migration is successful, the aggregation server in the target domain uses the migrated model, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain.

[0057] A trust assessment device for machine learning participants includes:

[0058] Memory, used to store computer programs;

[0059] A processor is used to implement the trust assessment method for a machine learning participant as described in steps 1 to 3 when executing the computer program.

[0060] A computer-readable storage medium for storing a computer program, which, when executed by a processor, is capable of assessing the trust level of a participant based on the trust assessment method for machine learning participants described in steps 1 to 3.

[0061] Compared with the prior art, the present invention has the following advantages:

[0062] 1. To address the shortcomings of existing federated learning methods for participant trust evaluation, which neglect the significant impact of context on trust evaluation results and lead to biases in participant trust scores across different contexts, this invention utilizes transfer learning to transfer existing trust evaluation models to different contextual scenarios; for context c t In cases where insufficient trust evidence prevents the development of a suitable trust assessment model for the given context, the TrAdaBoost algorithm is used to transfer the application to context c. s The trust assessment model is applied to context c. t In this study, a trust assessment model is made transferable across contexts, enabling accurate assessment of the trust levels of participants in federated learning even when trust evidence is insufficient.

[0063] 2. When assessing the trust level of participants in federated learning, this invention fully considers the impact of context on the assessment results. By analyzing different contextual scenarios, it uses transfer learning technology to achieve cross-context transfer of the trust assessment model, overcoming the difficulty of establishing a trust assessment model when trust evidence is insufficient, and improving the accuracy of trust assessment results for the same participant in different contextual scenarios.

[0064] In summary, this invention establishes a trust assessment model in the case of insufficient trust evidence, enabling accurate assessment of the trust level of participants in federated learning under such circumstances, and improving the accuracy of trust assessment results for the same participant in different contexts. Attached Figure Description

[0065] Figure 1This is a diagram of the cross-context federated learning participant trust evaluation framework of the present invention.

[0066] Figure 2 It is a measure of the similarity between the source domain and the target domain.

[0067] Figure 3 This is an error analysis based on scenario one of the Amazon dataset.

[0068] Figure 4 This is an error analysis based on scenario two of the Amazon dataset.

[0069] Figure 5 This is an error analysis based on scenario three of the Amazon dataset.

[0070] Figure 6 This is based on the experimental results of Scenario 1 on the MNIST dataset, in which... Figure 6 (a) Trust level for honest users; Figure 6 (b) Trust level for malicious users; Figure 6 (c) represents the proportion of malicious users; Figure 6 (d) represents the iteration time.

[0071] Figure 7 This is based on the experimental results of scenario two on the MNIST dataset, in which... Figure 7 (a) Trust level for honest users; Figure 7 (b) Trust level for malicious users; Figure 7 (c) represents the proportion of malicious users; Figure 7 (d) represents the global model accuracy. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings.

[0073] Considering that the source and target domains are in different contexts, there are two specific cases:

[0074] ①Source domain context c s and target domain context c t The learning tasks dy are the same, but the learning time T is different.

[0075] When the learning times in the source and target domains differ, participants in federated learning may exhibit different behavioral patterns. For example, the behavioral patterns of participants on weekdays are likely to differ significantly from those on holidays, or their historical and current behavioral patterns may differ considerably. Therefore, the trust level gained by a participant in one time period may not follow the same distribution as in another. Consequently, it is unsuitable to use the same trust assessment model to calculate participant trust levels in different contexts. Transfer learning is needed to transfer the trust prediction model from the source domain to the target domain to improve prediction accuracy.

[0076] ②Source domain context c s and target domain context c t The learning tasks dy are different, but the learning time T is the same.

[0077] When the learning tasks in the source and target domains differ, participants in federated learning may exhibit different interaction patterns. For example, the behavior patterns of participants in image classification tasks differ significantly from those in speech processing tasks. Therefore, the trust levels of participants under the same task and under other tasks will follow different distributions, making it unsuitable to use the same trust evaluation model for participant trust calculation. Transfer learning is needed to transfer the trust prediction model from the source domain to the target domain to improve prediction accuracy.

[0078] When performing trust assessment model transfer in the two different contexts described above, the source and target domains possess the same feature space but different data distributions, i.e., χ² s =χ t , P(X s )≠P(X t Furthermore, since both the source and target domain tasks involve calculating participant trust levels, they share the same label space and prediction task, i.e., γ. s =γ t , To address these two different scenarios, this invention proposes the following cross-context trust evaluation scheme based on transfer learning.

[0079] A method for evaluating the trust level of machine learning participants is proposed. Based on transfer learning, a trust evaluation model deployed on the server during system initialization is transferred to different context scenarios. For situations where there is insufficient trust evidence in a particular context, making a suitable trust evaluation model unavailable, the distance between the source and target domains is measured. If the distance between the source and target domains is less than a predefined threshold, a trust evaluation model transfer based on the TrAdaBoost algorithm is used to obtain a participant trust evaluation model in the target domain. Trust evaluation of participants in the target domain is then performed based on this model. Finally, the server aggregates the transferred model in the target domain, inputting the participant's trust evidence to obtain the participant's trust level in the target domain.

[0080] A method for assessing trust among participants in machine learning, based on the following system assumptions:

[0081] 1) A federated learning system consists of an aggregation server and participants. The aggregation server can publish federated learning tasks and aggregate models. It can also evaluate the trustworthiness of participants and make corresponding decisions based on the evaluation results. The aggregation server is trustworthy. Participants are users who participate in federated learning tasks. There are multiple participants in a federated learning system.

[0082] 2) Trust is defined as the server's expectation of the reliability of the behavior of the participants in federated learning, that is, the probability that the behavior of the participants is reliable. The higher the probability that the behavior of the participants is reliable in the subsequent learning process, the higher the trust of the aggregation server for the participants, and vice versa. The trust space is a closed continuous interval [0,1], that is, the trust level is any real number in [0,1], where 0 means that the aggregation server does not trust the participants at all, 1 means that the aggregation server trusts the participants completely, and 0.5 means that the degree of trust of the aggregation server for the participants cannot be judged.

[0083] 3) Context is an important consideration when conducting trust assessments of participants in federated learning; context refers to the environmental information related to the participants' trust evidence; context is formally represented as C =<dy,T> Where dy is the learning task, including image classification tasks and text classification tasks; T = <T begin ,T end > is the study time, where T begin and T end These represent the start and end times of learning, respectively, for time T. begin and T end All are described using four dimensions, T begin =T end =<day,hour,minute,holiday> These are the day, hour, minute, and holiday signs, respectively.

[0084] 4) In a cross-context federated learning participant trust evaluation scheme, assume the context c s The domain in which it exists is called the source domain, and the context c t The domain in which it is located is called the target domain;

[0085] The source domain is formally defined as D. s ={χ s ,P(X s )},in, Indicates context c s The source domain has n samples of trust evidence among the participating parties. s samples, P(X) s ) represents the marginal probability distribution of the samples; the task in the source domain is Where, γ s For trust level, For context c s A model for predicting trust levels among participating parties; Let x be the set of labeled samples from the source domain. si The corresponding label is γ si Each element This represents the p trust attributes that each sample possesses. The trust attribute information includes the anomaly information, contribution information, and latency information of the federated learning participants.

[0086] The target domain is formally defined as D. t ={χ t ,P(X t )},in, This indicates that the participating parties trust the evidence sample, and the target domain has n. t samples, P(X) t () represents the marginal probability distribution of the sample; the task in the target domain Where, γ t For trust level, For context c t Trust prediction model in China; For the target domain, the set of labeled samples, sample x ti The corresponding label is γ ti Each element This represents the p trust attributes that each sample possesses. The trust attribute information includes the anomaly information, contribution information, and latency information of the participants in the federated learning process.

[0087] A method for assessing the trust level of participants in machine learning, specifically including the following steps:

[0088] First, the distance between the source domain and the target domain is measured. If the distance between the source domain and the target domain is less than a predefined threshold, then model transfer is performed based on the knowledge of the source domain to obtain the participant trust assessment model in the target domain. Based on this model, the trust level of the participants in the target domain is assessed.

[0089] Step 1. Similarity measurement between source and target domains

[0090] To improve the transfer of knowledge from the source domain to the target domain, this invention uses Maximum Mean Discrepancy (MMD) to measure the distance between the source and target domains. A smaller MMD value indicates a smaller distance between the source and target domains, and a higher similarity in their data distributions, and vice versa. When measuring the similarity between the source and target domains, a confidence threshold 'd' is given as a threshold for whether model transfer is possible. If the distance between the source and target domains is not greater than the predefined confidence threshold 'd' (i.e., MMD ≤ d), the similarity between the source and target domains is considered high. In this case, the model transfer method is used based on the formally defined source domain value 'd'. s and its learning task T s Perform model transfer to obtain a participant trust assessment model for the target domain. based on Trust assessment of stakeholders in the target domain: Inputting information about stakeholders in the target domain into... The output is the trust level of the participants; if the distance between the source domain and the target domain is greater than the predefined confidence threshold d, i.e., MMD>d, the model transfer method cannot obtain a good trust evaluation model, and no transfer is performed in this case; the confidence level d is a hyperparameter; its value is determined by experiments.

[0091] The squared distance between the source and target domain datasets is shown in Equation (1). The method for solving the squared distance between the source and target domain datasets can measure the distance between two different but related distributions in the Reproducing Kernel Hilbert Space (RKHS); X in Equation (1) s and X t These are the source domain dataset and the target domain dataset, with dataset sizes n respectively. s and n t ; It is a mapping function used to map the original variables to the regenerated Hilbert space; H represents the MMD squared distance between the source and target domain datasets, which is... The data is mapped to a regenerated Hilbert space for measurement;

[0092]

[0093] Because it is difficult to find a suitable mapping function Furthermore, the mapping function is not fixed in different tasks. Therefore, by expanding formula (1) and selecting the Gaussian kernel function k(·) in Support Vector Machine (SVM), the MMD squared distance between the source and target domain datasets can be obtained, as shown in formula (2). ||uv|| 2 Let U be the Euclidean distance between samples u and v.

[0094]

[0095] Step 2. Trust evaluation model transfer based on TrAdaBoost algorithm;

[0096] If the distance between the source domain and the target domain in step 1 is no greater than a predefined confidence threshold d (i.e., MMD ≤ d), the model transfer method is used to evaluate the trust level of the participants in the target domain, as follows:

[0097] Trust evaluation model transfer based on TrAdaBoost algorithm; let T be the set of labeled samples in the source domain. source (of size n), the set of labeled samples in the target domain is denoted as T. target (Of size m), merge the labeled sample set from the source domain with the labeled sample set from the target domain to obtain the dataset T = T. source ∪T target The size of T is n+m, represented as T={(x1,y1),…,(x n ,y n ),(x n+1 ,y n+1 ),…,(x n+m ,y n+m Let T be all the labeled datasets after merging; each element in T is a sample instance, and each sample instance contains a feature space and a label; use dataset T as training data to train the trust evaluation model:

[0098] 2.1 Initialize the weight distribution of the training data

[0099] For dataset T = {(x1,y1),…,(x n ,y n ),(x n+1 ,y n+1 ),…,(x n+m ,y n+m The training samples in )} are assigned initial weights D1, as shown in formula (3):

[0100]

[0101] in,

[0102] 2.2 The maximum number of iterations is predefined as N. Given a basic learning algorithm, Learner, this learning algorithm Learner is used for iteration. Let t = 1, ..., N, and then the following steps are performed:

[0103] 2.2.1 Based on training set T and weight distribution D t (D t (For the weight distribution of the training samples in the t-th iteration) call the Learner algorithm to obtain the weak classifier h. t :X→R, where X={x1,…,x n ,x n+1 ,…,x n+m}, R is the confidence level value, where R∈[0,1]; Formula (5) gives the weak classifier h t In distribution D t The maximum error on each sample instance x i Weak classifier h t In distribution D t Adjustment error As shown in formula (6);

[0104]

[0105]

[0106] 2.2.2 Calculate h t In distribution D t Total adjustment error ε t As shown in formula (7); here ε t It must be less than 0.5 if ε t If the value is greater than 0.5, then for ε t Make a correction, let ε t =0.5, this operation is to prevent the algorithm from stopping;

[0107]

[0108] 2.2.3 Calculate the intermediate value β t As shown in formula (8):

[0109] β t =ε t / (1-ε t (8)

[0110] 2.2.4 Calculate the weak classifier h using formula (9) t The weight α in the final classifiert The smaller the error, the greater the proportion of weak classifiers in the final classifier;

[0111] α t =1 / 2ln(1 / β) t (9)

[0112] 2.2.5 Update the weight distribution of the training samples using formula (10):

[0113]

[0114] Among them, D t (k) represents the distribution D t The k-th element in β and β t The setting is for D t Adjust the middle element, Z t Let be the normalization constant, then we have

[0115] The trust evaluation model based on the TrAdaBoost algorithm performs N transfer iterations. After N iterations, data in the source domain that conforms to the target domain task has higher weights, while data that does not conform to the target domain task has lower weights. Data with higher weights in the source domain results in smaller errors in the target domain.

[0116] Step 3: According to the weight α of the weak classifier t Combining all weak classifiers yields a strong classifier, as shown in formula (11):

[0117]

[0118] Where x is the evidence of trust, h f (x) represents the trust assessment model; at this point, the model transfer was successful.

[0119] Then, the aggregation server in the target domain uses the final model obtained from the migration, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain.

[0120] A trust assessment system for machine learning participants includes:

[0121] The similarity measurement module is used in step 1 to measure the distance between the source domain and the target domain using the maximum mean difference (MMD).

[0122] The model transfer module is used in step 2 to perform trust evaluation model transfer based on the TrAdaBoost algorithm;

[0123] The trust assessment module is used in step 3. After the trust assessment model migration is successful, the aggregation server in the target domain uses the migrated model, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain.

[0124] A trust assessment device for machine learning participants includes:

[0125] Memory, used to store computer programs;

[0126] A processor is used to implement the trust assessment method for a machine learning participant as described in steps 1 to 3 when executing the computer program.

[0127] A computer-readable storage medium for storing a computer program, which, when executed by a processor, is capable of assessing the trust level of a participant based on the trust assessment method for machine learning participants described in steps 1 to 3.

[0128] Simulation Experiment

[0129] This invention uses the Python programming language and employs the PyCharm platform to build a simulation experimental environment. The experimental environment consists of an Intel(R) Core(TM) i7-10700 CPU, 16GB of RAM, and a Windows 10 operating system. The programming environment includes Python 3.8.8, Anaconda 2021.05, Torch 1.8.1, and NumPy 1.20.1.

[0130] The experiments of this invention are conducted in two parts. The first part uses the Amazon dataset for simulation experiments to verify the effect and necessity of model transfer in a centralized environment. The second part uses the MNIST dataset to verify that the trust evaluation scheme based on transfer learning proposed in this invention can effectively improve the accuracy of trust evaluation of federated learning participants. In this part of the experiment, the local models in both the source and target domains adopt Convolutional Neural Network (CNN). The CNN model structure includes two convolutional layers and two pooling layers.

[0131] 3.1 Simulation Experiment Based on Amazon Dataset

[0132] The purpose of the experiment using the Amazon dataset is to verify that evidence generated in different contexts is not suitable for direct fusion, and that trust evaluation models generated in different contexts are not suitable for mutual applicability. This demonstrates that using transfer learning for context-aware model transfer is a feasible approach to solve the problem described in this invention (i.e., in federated learning systems where there is insufficient trust evidence among participants, using transfer learning to transfer models from other contexts to the current context for trust evaluation).

[0133] This part of the experiment uses the Amazon Review Dataset 2013. This dataset contains product review information from Amazon, including product information, user information, and user ratings of products (effective review rate, rating, and review time). The data spans 18 years, including approximately 35 million reviews up to March 2013. This dataset is divided into sub-datasets based on product categories, such as Books, Electronics, Movies and TV, and CDs and Vinyl. The format of this dataset was modified for this experiment. The feature space of the samples in the dataset includes: ① product ID, ② user rating information for this product (effective review rate, user rating, and review time), containing a total of 10 such user rating information; the label space is a closed interval [0,5], representing the user's overall rating for this product.

[0134] Table 1. Three strategies based on simulation experiments using the Amazon dataset.

[0135]

[0136]

[0137] To simulate the problem addressed by this invention (i.e., in a federated learning system where there is insufficient trust evidence among participants, using transfer learning to transfer models from other contexts to the current context for trust evaluation), it is assumed that the source domain possesses a relatively good product rating model f suitable for the source domain due to the existence of a large number of labeled samples. s However, due to missing samples in the target domain, there is no suitable product rating model. Only 5% of the samples in the target domain are labeled, while the rest are unlabeled. Therefore, to obtain the comprehensive rating for all products in the target domain, the following three strategies are implemented, as shown in Table 1. Furthermore, some overlapping samples are included in the source and target domains. These overlapping samples have the same data distribution in both domains, and while they may be labeled in the source domain, they may not be labeled in the target domain.

[0138] Corresponding to the two model transfer scenarios given above (source domain context c) s and target domain context c t The learning task dy is the same, but the learning time T and the source domain context c are different. s and target domain context c t(The learning tasks dy are different, but the learning time T is the same). Experimental simulation scenarios are set up. Due to the different times T, two different contexts will be generated: ① two different contexts generated by historical time and current time; ② two different contexts generated by weekday time and weekend time. User reviews of products will also vary depending on the context. Therefore, we set up three experimental simulation scenarios:

[0139] Scenario 1: The context c of the source domain s In the middle, T represents a certain historical time period. hist (January 2008 to January 2010), the context c of the target domain t In this context, T represents a specific time period. curr (January 2010 to March 2013). To ensure that data in the source and target domains are under the same task dy, in this scenario, the context task dy for both the source and target domains is the product rating task dy from the subset dataset Electronics. ele The source and target domain contexts are represented as c, respectively. s = <dy ele ,T hist >,c t = <dy ele ,T curr In this scenario, the source domain considers time decay of scores as of January 2010 when evaluating the overall scores of products, while the target domain does not. Therefore, samples in the source domain from January 2010, which do not yet meet the conditions for time decay, will not be considered in the overall score evaluation. These samples will have the same data distribution in both the source and target domains, constituting overlapping samples in this scenario.

[0140] Scenario 2: The context c of the source domain s In this context, T represents the working day time period. work (Monday to Friday, between January 2008 and March 2013), the context c of the target domain. t T represents the weekend time period. weekend (Saturday to Sunday period between January 2008 and March 2013). Similarly, to ensure that data in the source and target domains are under the same task dy, in this scenario, the context task dy for both the source and target domains is the product rating task dy from the subset dataset Electronics. ele The source and target domain contexts are represented as c, respectively. s = <dy ele ,T week >,c t = <dy ele,T weekend In this scenario, it is assumed that data samples generated during holidays (such as New Year's Day and Spring Festival) have the same data distribution as data samples generated on weekends. Therefore, due to the existence of holidays, some overlapping samples will occur in this scenario. For example, on January 1, 2008 (Tuesday), New Year's Day, the data samples generated in the source domain on this day have the same distribution as the samples in the target domain; these samples are overlapping samples.

[0141] Scenario 3: The context c of the source domain s In the middle, dy represents the product rating task in the Electronics dataset. ele The context c of the target domain t In the dataset dy, dy represents the task dy under the Cell_Phones_&_Accessories dataset. cell In order to ensure that the data in the source domain and the target domain are at the same time T, the context time T of both the source domain and the target domain is T in this scenario. v (January 2008 to January 2010). The source and target domain contexts are represented as c. s = <dy ele ,T v >,c t = <dy cell ,T v In this scenario, although the tasks dy in the source and target domains differ due to the use of datasets with different product categories, some samples overlap in product names due to the similarity between the source and target domain datasets. These samples are considered to have the same data distribution in both the source and target domains, and are therefore considered overlapping samples in this scenario. For example, there might be a sample with the product name "Car Charger" in both the target and source domain datasets. Although they are not the same model, they belong to the same category. This sample has the same data distribution in both the source and target domains, and is therefore an overlapping sample.

[0142] The relevant settings for the above three simulation scenarios are shown in Table 2.

[0143] Table 2 Simulation Experiment Scenarios Based on Amazon Dataset

[0144]

[0145]

[0146] First, the performance of the models obtained under the three strategies in Scenario 1, applicable to the target domain, is verified. Tests are conducted in the target domain, and the experimental results are as follows: Figure 3As shown in the figure, the number of test samples is 30% (96) of the number of labeled samples in the target domain. The vertical axis represents the error between the original label of the sample and the predicted label. In the figure, the blue scatter points represent the error in evaluating the overall product score in the target domain under Strategy 1 in Scenario 1, the green scatter points represent the error in evaluating the overall product score in the target domain under Strategy 2 in Scenario 1, and the orange scatter points represent the error in evaluating the overall product score in the target domain under Strategy 3 in Scenario 1. Furthermore, to make the error comparison more intuitive, a straight line with y=0 is drawn in the figure. It can be seen that compared to the blue scatter points representing Strategy 1 and the green scatter points representing Strategy 2, the orange scatter points representing Strategy 3 basically fluctuate around the straight line y=0, and the fluctuation range is smaller. Therefore, Strategy 3 has a smaller error. That is, the overall product score evaluation result in the target domain using the product rating model transferred from the source domain through transfer learning is more accurate and has a smaller error.

[0147] To verify the performance of the models obtained under the three strategies in Scenario 2, applicable to the target domain, tests were conducted in the target domain, and the experimental results are as follows. Figure 4 As shown, the number of test samples is 30% (53) of the number of labeled samples in the target domain. It can be seen that the orange scatter points representing strategy three have a smaller fluctuation range and smaller error.

[0148] To verify the performance of the models obtained under the three strategies in Scenario 3, applicable to the target domain, tests were conducted in the target domain, and the experimental results are as follows. Figure 5 As shown, the number of test samples is 30% (20) of the number of labeled samples in the target domain. Similarly, compared to strategies one and two, strategy three has a smaller error in the test samples.

[0149] On Figure 3 , Figure 4 , Figure 5 Intuitively, it's clear that the scoring model obtained by transferring the source domain's product scoring model performs better. However, to compare the performance of the models used for comprehensive product score evaluation in the target domain under the three strategies in the above three scenarios, we present the Mean Absolute Error (MAE), Mean Square Error (MSE), and Coefficient of Determination (R²) for each scoring model in the target domain. 2 As shown in Table 3, MAE represents the average magnitude of the difference between the actual and predicted values. A larger MAE indicates a larger prediction error, meaning a worse model performance. MSE represents the average magnitude of the squared difference between the actual and predicted values. A larger MSE also indicates a larger prediction error, meaning a worse model performance. R 2 It measures the degree of agreement between the predicted value and the actual value, and takes a value of [0,1]. When R...2 The closer the value is to 1, the better the fit and the better the model performance. As can be seen from Table 3, for all three simulation scenarios, the model performance is best under the third strategy, that is, the model obtained by transferring the source domain product rating model has the best performance.

[0150] Table 3 Comparison of model performance in three simulation scenarios based on the Amazon dataset.

[0151]

[0152] This verifies that trust evidence generated in different contexts is not suitable for direct fusion, and trust evaluation models generated in different contexts are not suitable for mutual applicability. However, when trust evidence is insufficient, using transfer learning to transfer models from other contexts to the current context for trust evaluation has good performance.

[0153] 3.2 Simulation Experiment Based on MNIST Dataset

[0154] The purpose of conducting experiments using the MNIST dataset is to verify, through simulation of a federated learning system, that the cross-context trust assessment scheme proposed in this invention can improve the accuracy of participant trust assessment in a federated learning system by transferring the trust assessment model when trust evidence is insufficient.

[0155] This part of the experiment primarily uses the MNIST dataset. The MNIST dataset is a classic machine learning dataset containing 60,000 training samples and 10,000 test samples. Each sample is a 28*28 grayscale image of a handwritten digit, labeled with any digit from 0 to 9. This experiment selects two factors as the trust attributes of the participants: ① the latency information of uploading the local model, and ② the anomaly information of the local model. The difference in the time dimension T between the source and target domains leads to differences in the latency information of the participants' uploaded local models, while the difference in the task dimension dy between the two contexts leads to differences in the anomaly information of the participants' local models. The difference in the latency information of the participants' uploaded local models is mainly due to differences in network environment and hardware devices, while the factors affecting the anomaly information of the participants' local models are mainly the number of training samples or whether they have been attacked.

[0156] Table 4 Two strategies for simulation experiments based on the MNIST dataset

[0157]

[0158] Similarly, to simulate the problem described in this invention (i.e., in a federated learning system where there is insufficient trust evidence among participants, transfer learning is used to transfer models from other contexts to the current context for trust evaluation), it is assumed that the source domain possesses a good trust evaluation model f suitable for the source domain due to the existence of abundant trust evidence. HL_s The aggregation server in the source domain can use f HL_s The trust level of the participants is assessed, and corresponding decisions are made based on the assessment results. This trust assessment model considers both latency and anomaly information from the participants' multiple historical interactions with the aggregation server. However, the target domain cannot be trained with a trust assessment model due to insufficient trust evidence; similarly, it is assumed that only 5% of the samples in the target domain are labeled. Therefore, to assess the trust level of the participants in the target domain, two strategies are set up, as shown in Table 4. Similar to the experiment based on the Amazon dataset, some overlapping samples are included in the source and target domains.

[0159] Similar to the experiments based on the Amazon dataset, this invention presents two model transfer scenarios (source domain context c). s and target domain context c t The learning task dy is the same, but the learning time T and the source domain context c are different. s and target domain context c t Two experimental simulation scenarios are set up, with different learning tasks (dy) but the same learning time (T).

[0160] Scenario 1: The context c of the source domain s In this context, T represents time period 1, denoted as T1, and c represents the context of the target domain. t In this context, T represents time period 2, denoted as T2. The difference in the context time dimension T between the source and target domains affects the behavior of the participants, specifically manifested in the difference in the latency information of uploading the local model. The factors influencing the change in latency information are mainly network performance and differences in hardware devices. For example, the time dimension T of the source domain context may be in the 2G network era, while the time dimension T of the target domain context may be in the 5G network era; or the time dimension T of the source domain context may be in a period of smooth network access, while the time dimension T of the target domain context may be in a period of network congestion. This leads to differences in the trust evaluation models between the source and target domains. Furthermore, to ensure that the data in the source and target domains are in the same task dy, in this scenario, the context task dy of both the source and target domains is the image classification task dy under the MNIST dataset. mnist The source and target domain contexts are represented as c, respectively. s = <dy mnist ,T1>,c t = <dy mnist ,T2>.

[0161] In Scenario 1, there are 200 users in both the source and target domains. In each iteration, the aggregation server selects 130 users as participants in the learning process based on their trust levels. It is assumed that all users are honest and trustworthy during model training in both domains, the federated learning system is immune to malicious attacks, and each user has the same number of data samples; therefore, the local model accuracy and anomaly rate remain unchanged. However, due to the different time dimensions T of the two domain contexts, the latency of each user will vary. In this scenario, honest users perceive lower anomalies and lower latency, while malicious users perceive lower anomalies and higher latency. Furthermore, the overlapping samples in this scenario arise because users in both the source and target domains have the same hardware and network performance; the samples generated by these users have the same data distribution in both domains.

[0162] Scenario 2: The context c of the source domain s In the middle, dy represents the image classification task under the CIFAR-10 dataset. cifar The context c of the target domain t In the image classification task dy on the MNIST dataset, dy represents the image classification task dy. mnist The difference in contextual task dimension dy between the source and target domains affects the behavior of participants, specifically manifested in the difference in the local model anomaly score. The factors influencing the change in local model anomaly score are mainly the number of training samples or whether an attack has occurred, leading to differences in the trust evaluation models between the source and target domains. In this scenario, there are no differences in the network environment, so the impact on the latency information of participants uploading their local models is negligible. Furthermore, to ensure that the data in the source and target domains are at the same time T, the contextual time T of both the source and target domains in this experiment is within the time interval T. v The source and target domain contexts are represented as c, respectively. s = <dy cifar ,T v >,c t = <dy mnist ,T v >

[0163] In Scenario 2, the source and target domain aggregation servers also select 130 users from all 200 users as participants in each iteration based on their trust level. These 200 users participate in the learning tasks for both the source domain CIFAR-10 dataset and the MNIST dataset. Each user has a different number of samples in the CIFAR-10 and MNIST datasets. The 200 users are divided into three groups, as shown in Table 5. Since both domains are within the same time period, it is assumed that the network performance is the same, and the latency for each user to upload their local model in the source and target domains is basically the same. However, due to the difference in task dimension dy, the context of the two domains differs, resulting in variations in the anomalousness of each user. Specifically, this difference is reflected in the anomalousness caused by the different number of samples each user possesses, thus causing variations in the accuracy of the local model. In this scenario, honest users perceive a low anomalousness and low latency, while malicious users perceive a high anomalousness and low latency, and malicious users upload malicious model parameters. Meanwhile, the overlapping samples in this scenario are generated because some users have the same proportion of data samples in both the source and target domains. The samples generated by these users have the same data distribution in both the source and target domains, and thus constitute the overlapping samples in this scenario.

[0164] The relevant settings for the two simulation scenarios are shown in Table 6.

[0165] Table 5 User settings for simulation experiments based on the MNIST dataset

[0166]

[0167] Table 6 Simulation Experiment Scenarios Based on MNIST Dataset

[0168]

[0169]

[0170] The simulation results in Scenario 1 are as follows Figure 6 As shown in the figure. The trust levels of all 200 users in the target domain under the two strategies described above are statistically analyzed. The trend of trust level changes among honest users in the target domain under the two strategies is shown in the figure. Figure 6 As shown in (a), the trend of trust changes of malicious users under the above two strategies is as follows: Figure 6 As shown in (b), this trust level is the average of users in the same category during each iteration. It can be seen that, regardless of whether the user is honest or malicious, the trust level varies significantly under both strategies due to the different trust assessment models used in the target domain.

[0171] Figure 6(c) shows the change in the proportion of malicious users in the federated learning system as the number of iterations increases in Scenario 1. It can be seen that Strategy 2 can effectively identify the presence of malicious users in the system through trust evaluation in the first few iterations, thus avoiding selecting these malicious users in subsequent iterations, and the proportion of malicious users in the system can quickly drop to 0. Strategy 1, because it uses a trust evaluation model from the source domain, will have a large bias in the trust evaluation results for users in the target domain, and therefore cannot accurately identify malicious users in the system. In all 30 iterations, malicious users participated in the training of the federated learning model.

[0172] Figure 6 (d) shows the time variation of each iteration in the federated learning system as the number of iterations increases in Scenario 1. Similarly, due to the adaptability of the trust evaluation model to this domain, Strategy 2 can accurately assess the trust level of each user in the federated learning system, so as to promptly detect the presence of malicious users and disallow them from participating in subsequent iterations. Therefore, the iteration time of Strategy 2 can decrease rapidly and remain at the same level thereafter. In contrast, Strategy 1 cannot accurately identify malicious users, so the addition of malicious users causes excessively long iteration times, resulting in poor federated learning performance.

[0173] The simulation results in scenario two are as follows: Figure 7 As shown in the figure. The trust levels of all 200 users in the target domain under the two strategies described above are statistically analyzed. The trend of trust level changes among honest users in the target domain under the two strategies is shown in the figure. Figure 7 As shown in (a), the trend of trust changes of malicious users under the above two strategies is as follows: Figure 7 As shown in (b), this trust level is the average of users of the same category in each iteration. It can be seen that, similar to the experimental results in scenario one, the trust level differs significantly between honest and malicious users under both strategies.

[0174] Figure 7 (c) shows how the proportion of malicious users in the federated learning system changes as the number of iterations increases in Scenario 2. Figure 7 (d) shows the time variation of each iteration of the federated learning system in Scenario 2 as the number of iterations increases. Similar to Scenario 1, the inapplicability of the trust evaluation model in the target domain causes a deviation in the user trust evaluation results in the target domain. As a result, malicious users exist in each iteration of Strategy 1, causing a loss of global model accuracy. However, Strategy 2, by using transfer learning, makes the trust evaluation model applicable to the context of the target domain, thus enabling the timely detection of malicious users. This allows the federated learning system to ultimately obtain a global model with higher accuracy, improving the reliability of the federated learning system.

Claims

1. A method for assessing the trust level of participants in machine learning, characterized in that, The trust assessment model deployed on the server during system initialization is transferred to different context scenarios using transfer learning. For situations where there is insufficient trust evidence in a particular context, making a suitable trust assessment model unavailable, the distance between the source and target domains is measured. If the distance is less than a predefined threshold, a trust assessment model transfer based on the TrAdaBoost algorithm is used to obtain a participant trust assessment model in the target domain. Trust assessment of participants in the target domain is then performed based on this model. Finally, the aggregation server in the target domain uses the transferred model, inputs the participant's trust evidence, and obtains the participant's trust level in the target domain. The specific steps include: Step 1. Measure the similarity between the source and target domains; Step 2. Based on TrAdaBoost Trust evaluation model transfer of the algorithm; Step 3. After successful model migration, the aggregation server in the target domain uses the migrated model, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain. Step 2 specifically involves: For step 1, the distance between the source domain and the target domain is no greater than a predefined confidence threshold. ,Right now The trust evaluation model from the source domain is transferred using a model transfer method to make it suitable for the context of the target domain, as follows: Trust evaluation model transfer based on TrAdaBoost algorithm; the set of labeled samples in the source domain is denoted as... Size is The target domain has a set of labeled samples denoted as Size is The dataset is obtained by merging the labeled sample set from the source domain with the labeled sample set from the target domain. , The size is , represented as , For all labeled datasets after merging; Each element in the dataset represents a sample instance, and each sample instance contains a feature space and a label; Used as training data for training the trust assessment model: 2.1 Initialize the weight distribution of the training data For dataset The training samples are assigned initial weights. As shown in formula (3): in, Represents data samples The weights, and have ; 2.2 The predefined maximum number of iterations is Given a learning algorithm, Learner, iterate using this learning algorithm, Let... Then perform the following steps: 2.2.1 Based on the training set and weight distribution , for the first The weight distribution of the training samples is iterated through multiple iterations, and the Learner learning algorithm is called to obtain a weak classifier. ,in, , Here, a value is assigned to the trust level. Formula (5) gives the weak classifier. In distribution The maximum error over each sample instance Weak classifier In distribution Adjustment error As shown in formula (6); 2.2.2 Calculation In distribution Total adjustment error As shown in formula (7); here It must be less than 0.5, if Then for Make corrections, so that To prevent the algorithm from stopping; 2.2.3 Calculate the intermediate value As shown in formula (8): 2.2.4 Calculate the weak classifier using formula (9) Weight in the final classifier The smaller the error, the greater the proportion of weak classifiers in the final classifier; 2.2.5 Update the weight distribution of the training samples using formula (10): in, Represents distribution The Middle One element, , and The setting is for... Adjust the elements in the middle. Let be the normalization constant, then we have ; Trust evaluation model transfer loop based on TrAdaBoost algorithm Once; after After rounds of iteration, data in the source domain that conforms to the target domain task will have higher weights, while data that does not conform to the target domain task will have lower weights; data with higher weights in the source domain will result in smaller errors in the target domain. Step 3 specifically involves: By weak classifier weights Combining all weak classifiers yields a strong classifier, as shown in formula (11): in, To trust the evidence, This is a trust assessment model; at this point, the model transfer is successful. Then, the aggregation server in the target domain uses the trust assessment model obtained from the migration, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain.

2. The trust assessment method for machine learning participants according to claim 1, characterized in that, The system assumes the following: 1) A federated learning system consists of an aggregation server and participants. The aggregation server can publish federated learning tasks and aggregate models. It can also assess the trustworthiness of participants and make corresponding decisions based on the assessment results. Aggregation servers are all trustworthy. Participants are users who participate in federated learning tasks. There are multiple participants in a federated learning system. 2) Trust is defined as the server's expectation of the reliability of the behavior of the participants in federated learning, that is: the probability that the behavior of the participants is reliable; the greater the probability that the behavior of the participants is reliable in the subsequent learning process, the higher the trust of the aggregation server in the participants, and vice versa. Trust space is a closed continuous interval. That is, the level of trust is Any real number in the set, where 0 represents that the aggregation server has no trust in the participants, 1 represents that the aggregation server has complete trust in the participants, and 0.5 represents that the degree of trust of the aggregation server in the participants cannot be judged. 3) Context is an important consideration when conducting trust assessments of participants in federated learning; context refers to the environmental information related to the participants' trust evidence; context is formally represented as... ,in, The learning tasks include image classification and text classification. For study time, of which, and These represent the start and end times of learning, respectively, in terms of time. and All are described using four dimensions. These are the day, hour, minute, and holiday signs, respectively. 4) In a cross-context federated learning participant trust evaluation scheme, the following contextual environment is assumed. The domain in which it exists is called the source domain, and the context is called the context domain. The domain in which it is located is called the target domain; The source domain is formally defined as ,in, Representing context The source domain has the following characteristics: (This refers to a sample of evidence related to trust among participating parties.) One sample, The sample marginal probability distribution; the task in the source domain is ,in, For trust level, For context A model for predicting trust levels among participating parties; The set of labeled samples in the source domain, samples The corresponding tags are Each element , indicating that each sample possesses Each trust attribute contains information on the anomaly level, contribution level, and latency of the participants in the federated learning process. The target domain is formally defined as ,in, This indicates that the participating parties trust the evidence sample, and the target domain has... One sample, The marginal probability distribution of the sample; the task in the target domain. ,in, For trust level, For context Trust prediction model in China; For the target domain, a set of labeled samples, samples The corresponding tags are Each element , indicating that each sample possesses The trust attribute information includes the anomaly level, contribution level, and latency information of the participants in the federated learning process.

3. The method for assessing the trust level of machine learning participants according to claim 1, characterized in that, Step 1 specifically involves: The maximum mean discrepancy (MMD) is used to measure the distance between the source and target domains. A smaller MMD value indicates a smaller distance between the two domains, and a higher similarity in their data distributions, and vice versa. A confidence threshold is given when measuring the similarity between the source and target domains. As a threshold for whether model transfer is possible, if the distance between the source and target domains is not greater than a predefined confidence threshold... ,Right now If the similarity between the source and target domains is high, then the model transfer method is used based on the formal definition of the source domain. Perform model transfer to obtain a participant trust assessment model for the target domain. If the distance between the source domain and the target domain is greater than a predefined confidence threshold. ,Right now If the model transfer method cannot yield a good confidence evaluation model, then transfer is not performed; confidence level. It is a hyperparameter; The solution for the squared distance of the MMD of the source and target domain datasets, as shown in Equation (1), can measure the distance between two distinct but related distributions in the Reproducing Kernel Hilbert Space (RKHS); in Equation (1) and These are the source domain dataset and the target domain dataset, with dataset sizes of respectively. and ; It is a mapping function used to map the original variables to the regenerated Hilbert space; The MMD squared distance between the source and target domain datasets is represented by... The data is mapped to a regenerated Hilbert space for measurement; Expanding formula (1), the Gaussian kernel function in Support Vector Machine (SVM) is selected. The MMD squared distance between the source and target domain datasets is obtained as shown in formula (2); where the Gaussian kernel function is... , For the sample and The Euclidean distance; （2）。 4. A trust assessment system for machine learning participants, used to perform the method according to any one of claims 1 to 3; characterized in that, include: The similarity measurement module is used in step 1 to measure the distance between the source domain and the target domain using the maximum mean difference (MMD). The model transfer module is used in step 2 to perform trust evaluation model transfer based on the TrAdaBoost algorithm; The trust assessment module is used in step 3. After the trust assessment model migration is successful, the aggregation server in the target domain uses the migrated model, inputs the trust evidence of the participants, and obtains the trust level of the participants in the target domain.

5. A trust assessment device for machine learning participants, used to perform the method according to any one of claims 1 to 3; characterized in that, include: Memory, used to store computer programs; A processor, used to execute the computer program to implement steps 1 to 3 for a trust assessment method for a machine learning participant.

6. A computer-readable storage medium for storing a computer program, characterized in that, When the computer program is executed by a processor, it can perform a trust assessment of the participants based on the method described in any one of claims 1 to 3.

Images