High-reliability semi-supervised federated learning method and system

Abstract

The application relates to a high-reliability semi-supervised federated learning method and system, solves the problems of poor processing effect of existing technologies on class imbalance data in semi-supervised learning and mutual influence of common supervised learning parameters and unsupervised learning parameters in semi-supervised learning, provides a good model aggregation method based on self-defined reliability, and enhances the ability to process class imbalance data; the model parameters are subjected to sparse additive decomposition, so as to reduce the interference between supervised and unsupervised tasks, thereby improving the performance of semi-supervised learning in federated learning; meanwhile, the communication cost between the server and the client is reduced through parameter decomposition, so that the communication efficiency of federated learning is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of federal semi-supervised learning, in particular to a high-reliability semi-supervised federal learning method and system. BACKGROUND

[0002] Semi-supervised learning is a machine learning method that uses a small amount of labeled data and a large amount of unlabeled data for model training. On the one hand, considering the definition of semi-supervised learning, it can solve the problem of a small amount of labeled data. On the other hand, compared with unsupervised learning, it can still use a part of the label information to reduce the difficulty of training the model, so the research on semi-supervised learning has very important practical significance. Semi-supervised learning actually uses a large amount of unlabeled data to make up for the overfitting phenomenon (improve the generalization ability) caused by a small amount of labeled data guiding model training.

[0003] Federated learning is a learning paradigm that ensures the privacy and security of local client data and jointly trains the model through a centralized server calling multiple local clients. One of the most important roles of federated learning is to provide a large amount of training data for machine learning related tasks under extremely high privacy protection requirements. This feature of federated learning plays a crucial role in machine learning methods that still need a large amount of data support today, as people pay more and more attention to privacy protection.

[0004] Federated semi-supervised learning (FSSL) is actually the application of semi-supervised learning methods to the application scenario of federated learning, combining the advantages of both technologies to better solve real-world problems. This technology can ensure sufficient training data through federated learning, and can also alleviate the problem of high data annotation overhead for each client through semi-supervised learning.

[0005] Currently, existing federated semi-supervised techniques mainly have the following problems: poor performance on class imbalance data, mutual influence between supervised learning and unsupervised learning processes, how to design a good model aggregation method, and communication bottlenecks. SUMMARY

[0006] In order to overcome at least one of the deficiencies in the prior art, the present application provides a high-reliability semi-supervised federal learning method and system.

[0007] In a first aspect, a high-reliability semi-supervised federal learning method is provided, comprising:

[0008] training the plurality of client semi-supervised learning models to obtain a supervised learning parameter and an unsupervised learning parameter of each client semi-supervised learning model;

[0009] selecting at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameter and the unsupervised learning parameter, wherein the supervised learning parameter and the unsupervised learning parameter of each high-reliability model are used to assist in training the plurality of client semi-supervised learning models in the next round of training;

[0010] aggregating the supervised learning parameters of all client semi-supervised learning models to obtain aggregated supervised learning parameters, aggregating the unsupervised learning parameters of all client semi-supervised learning models to obtain aggregated unsupervised learning parameters, and merging the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to obtain merged parameters, wherein the merged parameters are used to update the server model;

[0011] returning the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to the plurality of client semi-supervised learning models as initial supervised learning parameters and initial unsupervised learning parameters for the next round of model training.

[0012] In one embodiment, the plurality of client semi-supervised learning models are trained to obtain a supervised learning parameter and an unsupervised learning parameter of each client semi-supervised learning model, including:

[0013] The training includes supervised learning and unsupervised learning;

[0014] For each client semi-supervised learning model, the model overall parameter θ is decomposed into a supervised learning parameter α and an unsupervised learning parameter β, such that θ = α + β;

[0015] The supervised learning is trained based on a labeled data set, and the unsupervised learning parameter β is frozen during the training process to obtain an updated supervised learning parameter; the unsupervised learning is trained based on an unlabeled data set; and the supervised learning parameter α is frozen during the training process to obtain an updated unsupervised learning parameter;

[0016] After each round of unsupervised learning, the pseudo-labels generated by the model are sampled, and the data corresponding to the sampled pseudo-labels are added to the labeled data set for the next round of supervised learning.

[0017] In one embodiment, the loss function used in the unsupervised learning is:

[0018]

[0019] wherein Φ(·) is a loss function, CE is a cross-entropy loss function, is a label, and y is an output, Local client semi-supervised learning model For the output of input γ(x), γ(x) is the data enhanced data of input x, H is the number of high-reliability models, and KL is the relative entropy function, The jth high-reliability model For the output of input x, Local client semi-supervised learning model For the output of input x;

[0020] Wherein, the label The following formula is used to represent:

[0021]

[0022] Wherein, 1(·) is one-hot function.

[0023] In one embodiment, screening at least one high-reliability model from a plurality of client semi-supervised learning models according to supervised learning parameters and unsupervised learning parameters comprises:

[0024] According to the supervised learning parameters and the unsupervised learning parameters, the classification accuracy score of each client semi-supervised learning model on the validation set is calculated;

[0025] According to the classification accuracy score, the reliability of each client semi-supervised learning model is calculated, and the formula used is as follows:

[0026]

[0027] Wherein, K i The reliability of the i-th client semi-supervised learning model, Indicates the classification accuracy score of the i-th client semi-supervised learning model on the validation set, Indicates the sum of the classification accuracy scores of all client semi-supervised learning models from No. 1 to No. I;

[0028] According to the reliability, at least one high-reliability model is screened.

[0029] In one embodiment, the supervised learning parameters of all client semi-supervised learning models are aggregated to obtain aggregated supervised learning parameters, and the formula used is as follows:

[0030]

[0031] Wherein, The aggregated supervised learning parameters of all client semi-supervised learning models from No. 1 to No. I, K i The reliability of the i-th client semi-supervised learning model, The supervised learning parameter of the client semi-supervised learning model numbered i, I is the number of client semi-supervised learning models;

[0032] The unsupervised learning parameters of all client semi-supervised learning models are aggregated to obtain aggregated unsupervised learning parameters, and the formula used is as follows:

[0033]

[0034] Wherein, The aggregated unsupervised learning parameters of all client semi-supervised learning models numbered 1 to I, K i The reliability of the client semi-supervised learning model numbered i, The unsupervised learning parameter of the client semi-supervised learning model numbered i, I is the number of client semi-supervised learning models.

[0035] In a second aspect, a high-reliability semi-supervised federated learning system is provided, comprising: a plurality of client modules, a parameter uploading module and a server, the server comprising a high-reliability model selection module, a parameter aggregation and updating module and a parameter feedback module, each client module is provided with a client semi-supervised learning model, and the server is provided with a server model;

[0036] Each client module is configured to train the client semi-supervised learning model to obtain the supervised learning parameter and the unsupervised learning parameter of the client semi-supervised learning model;

[0037] The parameter uploading module is configured to upload the supervised learning parameter and the unsupervised learning parameter of the client semi-supervised learning model obtained by each client module to the server;

[0038] The high-reliability model selection module is configured to select at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameter and the unsupervised learning parameter, and to feed back the supervised learning parameter and the unsupervised learning parameter of each high-reliability model to the plurality of client modules for assisting in training the plurality of client semi-supervised learning models in the next round of training;

[0039] The parameter aggregation and updating module is configured to aggregate the supervised learning parameters of all client semi-supervised learning models to obtain aggregated supervised learning parameters, aggregate the unsupervised learning parameters of all client semi-supervised learning models to obtain aggregated unsupervised learning parameters, combine the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to obtain combined parameters, and update the server model according to the combined parameters;

[0040] a parameter returning module, configured to return the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to the plurality of client modules as initial supervised learning parameters and initial unsupervised learning parameters for a next round of model training.

[0041] In one embodiment, each client module is further configured to:

[0042] the training comprises supervised learning and unsupervised learning;

[0043] for each client semi-supervised learning model, the model overall parameters θ are decomposed into supervised learning parameters α and unsupervised learning parameters β, such that θ = α + β;

[0044] the supervised learning is trained based on a labeled data set, and the unsupervised learning parameters β are frozen during the training process to obtain updated supervised learning parameters; the unsupervised learning is trained based on an unlabeled data set; and the supervised learning parameters α are frozen during the training process to obtain updated unsupervised learning parameters;

[0045] After each round of unsupervised learning, the pseudo labels generated by the model are sampled, and the data corresponding to the sampled pseudo labels are added to the labeled data set for the next round of supervised learning.

[0046] In one embodiment, the loss function used in the unsupervised learning is:

[0047]

[0048] wherein Φ(·) is a loss function, CE is a cross-entropy loss function, is a label, is an output of the local client semi-supervised learning model for an input γ(x), γ(x) is data obtained by data augmentation on an input x, H is the number of high-reliability models, and KL is a relative entropy function, is an output of the jth high-reliability model for an input x, is an output of the local client semi-supervised learning model for an input x;

[0049] wherein the label is expressed by the following formula:

[0050]

[0051] wherein 1(·) is a one-hot function.

[0052] In one embodiment, the high-reliability model selection module is further configured to:

[0053] calculate a classification accuracy score of each client semi-supervised learning model on a validation set according to the supervised learning parameters and the unsupervised learning parameters.

[0054] The reliability of each client semi-supervised learning model is calculated according to the classification accuracy score, and the formula used is as follows:

[0055]

[0056] Wherein, K i is the reliability of the client semi-supervised learning model numbered i, represents the classification accuracy score of the client semi-supervised learning model numbered i on the validation set, represents the sum of the classification accuracy scores of all client semi-supervised learning models numbered from 1 to I;

[0057] At least one high-reliability model is screened according to the reliability.

[0058] In one embodiment, the parameter aggregation update module is further configured to:

[0059] The aggregated supervised learning parameters are calculated using the following formula:

[0060]

[0061] Wherein, is the aggregated supervised learning parameters of all client semi-supervised learning models numbered from 1 to I, K i is the reliability of the client semi-supervised learning model numbered i, is the supervised learning parameters of the client semi-supervised learning model numbered i, and I is the number of client semi-supervised learning models;

[0062] The aggregated unsupervised learning parameters are calculated using the following formula:

[0063]

[0064] Wherein, is the aggregated unsupervised learning parameters of all client semi-supervised learning models numbered from 1 to I, K i is the reliability of the client semi-supervised learning model numbered i, is the unsupervised learning parameters of the client semi-supervised learning model numbered i, and I is the number of client semi-supervised learning models.

[0065] Compared with the prior art, the present application has the following beneficial effects:

[0066] The high-reliability semi-supervised federated learning method provided in the application solves the problems of poor processing effect of existing technologies on class imbalance data in semi-supervised learning and mutual influence of common supervised learning parameters and unsupervised learning parameters in semi-supervised learning. The application provides a good model aggregation method based on self-defined reliability, and enhances the ability to process class imbalance data. The model parameters are sparsely decomposed to reduce the interference between supervised and unsupervised tasks, thereby improving the performance of semi-supervised learning in federated learning. At the same time, the communication cost between the server and the client is reduced through parameter decomposition, thereby improving the communication efficiency of federated learning. BRIEF DESCRIPTION OF DRAWINGS

[0067] The application can be better understood by referring to the following description in conjunction with the accompanying drawings, which are incorporated in and form a part of the specification, and together with the detailed description, serve to explain the principles of the application. In the drawings:

[0068] Figure 1 A working principle diagram of a high-reliability semi-supervised federated learning system according to an embodiment of the application is shown;

[0069] Figure 2 A flowchart of a high-reliability semi-supervised federated learning method according to an embodiment of the application is shown. DETAILED DESCRIPTION

[0070] In the following, exemplary embodiments of the application will be described with reference to the accompanying drawings. In the specification, not all the features of the actual embodiments are described for the sake of clarity and conciseness. However, it should be appreciated that many embodiment-specific decisions can be made in the process of developing any such actual embodiment in order to achieve the specific goals of the developer, and these decisions can vary from embodiment to embodiment.

[0071] It should also be noted here that, in order to avoid obscuring the application due to unnecessary details, only the device structure closely related to the scheme according to the application is shown in the drawings, and other details not closely related to the application are omitted.

[0072] It should be understood that the application is not limited to the described embodiments only due to the following description with reference to the drawings. In this text, the embodiments can be combined with each other, features between different embodiments can be replaced or borrowed, and one or more features can be omitted in one embodiment, if possible.

[0073] The application embodiment provides a high-reliability semi-supervised federated learning system, Figure 1 A working principle diagram of a high-reliability semi-supervised federated learning system according to an embodiment of the application is shown, referring to Figure 1The system comprises a plurality of client modules, a parameter uploading module and a server, the server comprises a high-reliability model selection module, a parameter aggregation updating module and a parameter returning module, each client module is provided with a client semi-supervised learning model, and the server is provided with a server model; the functions of the various models are described in detail as follows:

[0074] Each client module is configured to train the client semi-supervised learning model to obtain supervised learning parameters and unsupervised learning parameters of the client semi-supervised learning model.

[0075] The parameter uploading module is configured to upload the supervised learning parameters and the unsupervised learning parameters of the client semi-supervised learning model obtained by each client module to the server.

[0076] The high-reliability model selection module is configured to select at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameters and the unsupervised learning parameters, and return the supervised learning parameters and the unsupervised learning parameters of each high-reliability model to the plurality of client modules to assist in training the plurality of client semi-supervised learning models in the next round of training.

[0077] The parameter aggregation updating module is configured to aggregate the supervised learning parameters of all the client semi-supervised learning models to obtain aggregated supervised learning parameters, aggregate the unsupervised learning parameters of all the client semi-supervised learning models to obtain aggregated unsupervised learning parameters, combine the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to obtain combined parameters, and update the server model according to the combined parameters.

[0078] The parameter returning module is configured to return the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to the plurality of client modules as initial supervised learning parameters and initial unsupervised learning parameters for the next round of model training.

[0079] The embodiment of the application provides a high-reliability semi-supervised federated learning method, Figure 2 A flowchart of a high-reliability semi-supervised federated learning method according to an embodiment of the application is shown, and reference is made to Figure 2 The method comprises the following steps.

[0080] Step S1, training a plurality of client semi-supervised learning models to obtain supervised learning parameters and unsupervised learning parameters of each client semi-supervised learning model.

[0081] Step S2, selecting at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameters and the unsupervised learning parameters; the supervised learning parameters and the unsupervised learning parameters of each high-reliability model are used to assist in training the plurality of client semi-supervised learning models in the next round of training.

[0082] Step S3, aggregating the supervised learning parameters of all client semi-supervised learning models to obtain aggregated supervised learning parameters; aggregating the unsupervised learning parameters of all client semi-supervised learning models to obtain aggregated unsupervised learning parameters; merging the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to obtain merged parameters; and using the merged parameters to update the server model;

[0083] Step S4, returning the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to the plurality of client semi-supervised learning models as initial supervised learning parameters and initial unsupervised learning parameters for the next round of model training.

[0084] The embodiment of the present application solves the problem of mutual influence of shared supervised learning parameters and unsupervised learning parameters in semi-supervised learning, provides a good model aggregation method based on self-defined reliability, performs sparse additive decomposition on model parameters to reduce interference between supervised and unsupervised tasks, thereby improving the performance of semi-supervised learning in federated learning; and at the same time, the parameter decomposition reduces the communication cost between the server and the client, thereby improving the communication efficiency of federated learning.

[0085] In one embodiment, the plurality of client semi-supervised learning models are trained in step S1 to obtain the supervised learning parameters and the unsupervised learning parameters of each client semi-supervised learning model, and the training process includes supervised learning and unsupervised learning.

[0086] The model decomposes the overall parameter θ into the supervised learning parameter α and the unsupervised learning parameter β, so that θ = α + β, thereby alleviating the model forgetting phenomenon through independent updating of the two partial parameters. After the supervised learning parameter is obtained, the supervised learning part performs supervised learning on the labeled data, and the training process freezes the unsupervised learning parameter β (i.e., β remains unchanged during the training process). After the model makes a prediction, the loss between the predicted value and the true value is calculated through forward propagation, and the loss term is minimized. Then the error is returned to adjust the supervised learning parameter of the model.

[0087] The unsupervised learning part performs unsupervised learning on the unlabeled data, and the training process freezes the supervised learning parameter α (i.e., α remains unchanged during the training process). After the model makes a prediction, progressive distribution alignment is first performed to align the distribution of the labeled data, then the consistency regularization loss term is minimized, and then the error is returned to adjust the unsupervised learning parameter of the model. Finally, the pseudo-labels generated during training are class rebalanced and sampled, and the results of the sampling are incorporated into the labeled data set for the next round of iterative training, while playing a role in balancing the classes of the labeled data set. Finally, the client model in federated learning is iteratively trained in this way.

[0088] Detailed steps of supervised learning can be divided into:

[0089] (1) Data collection and division: according to the selection or preparation of the data set according to the downstream task, and then divided into training set and test set according to the ratio of 8:2; here, according to the difference of the downstream task, the selected data set is different, for example, the downstream task is the target detection task, here, the data set can include multiple pictures, and the pictures include the target to be detected.

[0090] (2) Training preparation: when a new round of federated learning starts, the supervised learning parameter a and the unsupervised learning parameter β are obtained from the server. The unsupervised learning parameter β is kept unchanged during the training process, that is, β is frozen. Here, in the first round of federated learning, the supervised learning parameter a and the unsupervised learning parameter β obtained from the server are the initial parameters. From the second round, the supervised learning parameter a and the unsupervised learning parameter β obtained from the server are the initial supervised learning parameter and the initial unsupervised learning parameter returned by the parameter return module.

[0091] (3) Training the client semi-supervised learning model: input the data in the data set and its label (true value) into the model, and initialize the learning rate to 10 -3 , iterative training;

[0092] (4) Parameter adjustment: after the model makes a prediction, the cross-entropy loss of the model prediction value and the true value is calculated by forward propagation, and then the error is returned to adjust the supervised learning parameter of the model; the corresponding training target expression is as follows:

[0093] minimizeL s (α)=CE(y||p a+β (y|γ(x))) (1)

[0094] Where, L s is the objective function, γ represents data augmentation, CE is the cross-entropy loss function, p α+β (y|γ(x)) represents the local client semi-supervised learning model p α+β The output for input γ(x) is the data after data augmentation of input x, and y is the output.

[0095] Detailed steps of unsupervised learning can be divided into:

[0096] (1) Data collection and division: the same process as the "data collection and division" step in supervised learning;

[0097] (2) Training preparation: when a new round of federated learning starts, the supervised learning parameter a and the unsupervised learning parameter β are obtained from the server. The unsupervised learning parameter β is kept unchanged during the training process, that is, β is frozen. Here, in the first round of federated learning, the supervised learning parameter a and the unsupervised learning parameter β obtained from the server are the initial parameters. From the second round, the supervised learning parameter a and the unsupervised learning parameter β obtained from the server are the initial supervised learning parameter and the initial unsupervised learning parameter returned by the parameter return module.

[0098] (3) Training the client model: input the data in the dataset into the model, initialize the learning rate to 10 -3 , and perform iterative training;

[0099] (4) Parameter tuning: after the model makes a prediction, perform progressive distribution alignment on the prediction result to match the distribution of the labeled data. Based on the aligned result, combine the decision of the high-reliability model to determine the generated pseudo label on a class with a prediction probability higher than a threshold τ. Here, the model training is guided by a loss function, and the loss function used is:

[0100]

[0101] where the first term CE function part in the above formula is the consistency regularization loss function at the data level, and the second term is the consistency regularization loss function between clients; Φ(·) is the loss function, CE is the cross-entropy loss function, is the label, y is the output, is the local client semi-supervised learning model For the output of input γ(x), γ(x) is the data enhanced data of input x, H is the number of high-reliability models, KL is the relative entropy function, which represents the KL divergence, which is a method of quantifying the difference between probability distributions P and Q. is the jth high-reliability model For the output of input x, p θl (y|x) is the local client semi-supervised learning model For the output of input x.

[0102] where the label is represented by the following formula:

[0103]

[0104] where 1(·) is the one-hot function, that is, the position corresponding to the maximum value in the model output (which the model considers the input to belong to this class) is set to 1, and the values of other positions are set to 0; the Max() function is used to take the position corresponding to the maximum value from the output vector, which is the label of the output. Here, the final output label value is determined by two parts, one part is the output of the local model for the input, and the other part is the output of the high-reliability model for the specified input returned from the server. For each 1(·) function, it is essentially that the model casts a vote in the position of the output that it considers the input to belong to. Finally, the Max() function is used to select the position with the highest vote, which can be regarded as the class / category that the input belongs to. This embodies a consensus mechanism.

[0105] The high-reliability model is a batch of consensus models selected from the selected client models according to the custom reliability of the server, and the number of consensus models is N. The output of the high-reliability model can be regarded as a label, and the above formula can be understood as the difference between the prediction result of the local model and the label provided by each high-reliability model should be as small as possible.

[0106] The corresponding training target expression is as follows:

[0107]

[0108] Wherein, L u is the target function, Φ α+β (·) is the loss function when the model parameters are α+β, λ is the weight of the loss function, is the weight of the L2 regularization term, is the weight of the L1 regularization term.

[0109] In the above formula, L1 and L2 regularization terms are added. The role of the L2 regularization term is to retain as much knowledge as possible learned from the supervised learning parameter α. This part uses the L2 norm to make the difference between the parameters α and β as small as possible. The role of the L1 regularization term is to make the unsupervised parameter set β contain as many 0 terms as possible to improve the communication efficiency of federated learning (parameters with 0 can not be propagated to the server side). Finally, the error is returned to adjust the unsupervised learning parameters of the model.

[0110] (5) After unsupervised learning in each round, the pseudo-labels generated by the model are sampled, and the data corresponding to the sampled pseudo-labels are added to the labeled data set for the next round of supervised learning, which plays a role in balancing the labeled data set.

[0111] The embodiment solves the problem that the prior art has poor processing effect on class imbalance data in semi-supervised learning, and enhances the ability to process class imbalance data.

[0112] In one embodiment, in step S2, selecting at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameters and the unsupervised learning parameters can include:

[0113] First, calculate the classification accuracy score of each client semi-supervised learning model on the validation set according to the supervised learning parameters and the unsupervised learning parameters;

[0114] Then, calculate the reliability of each client semi-supervised learning model according to the classification accuracy score, and the formula used is as follows:

[0115]

[0116] Wherein, K ireliability of the client semi-supervised learning model numbered i, classification accuracy score of the client semi-supervised learning model numbered i on the validation set, sum of classification accuracy scores of all client semi-supervised learning models numbered from 1 to I;

[0117] Finally, at least one high-reliability model is screened according to the reliability. Here, the top N model parameters with the highest reliability are screened, and the screened high-reliability model parameters are returned to each client at the beginning of the next round of federated learning, and are used to assist in training the plurality of client semi-supervised learning models in the next round of training, that is, to participate in the generation of pseudo-labels in unsupervised learning, that is, the second term in formula (3).

[0118] In one embodiment, in step S3, the supervised learning parameters of all client semi-supervised learning models are aggregated to obtain aggregated supervised learning parameters, and the formula used is as follows:

[0119]

[0120] wherein, aggregated supervised learning parameters of all client semi-supervised learning models numbered from 1 to I, K i reliability of the client semi-supervised learning model numbered i, supervised learning parameters of the client semi-supervised learning model numbered i, and I is the number of client semi-supervised learning models;

[0121] The unsupervised learning parameters of all client semi-supervised learning models are aggregated to obtain aggregated unsupervised learning parameters, and the formula used is as follows:

[0122]

[0123] wherein, aggregated unsupervised learning parameters of all client semi-supervised learning models numbered from 1 to I, K i reliability of the client semi-supervised learning model numbered i, unsupervised learning parameters of the client semi-supervised learning model numbered i, and I is the number of client semi-supervised learning models.

[0124] In summary, the present application has the following beneficial effects:

[0125] The application solves the problems of poor processing effect of existing technologies on class imbalance data in semi-supervised learning and mutual influence of common supervised learning parameters and unsupervised learning parameters in semi-supervised learning. The high-reliability semi-supervised federated learning method provided by the application provides a good model aggregation method based on self-defined reliability, and enhances the ability to process class imbalance data. The model parameters are sparsely and additively decomposed, the interference between supervised and unsupervised tasks is reduced, and the performance of semi-supervised learning in federated learning is improved. At the same time, the communication cost between the server and the client is reduced through parameter decomposition, so as to improve the communication efficiency of federated learning.

[0126] The above is only various embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the application, which should be covered within the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims.

Claims

1. A high-reliability semi-supervised federated learning method, characterized in that, The method comprises the following steps: training a plurality of client semi-supervised learning models to obtain supervised learning parameters and unsupervised learning parameters of each of the client semi-supervised learning models; selecting at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameters and the unsupervised learning parameters; the supervised learning parameters and the unsupervised learning parameters of each of the high-reliability models are used to assist in training the plurality of client semi-supervised learning models in the next round of training; aggregating the supervised learning parameters of all the client semi-supervised learning models to obtain aggregated supervised learning parameters; aggregating the unsupervised learning parameters of all the client semi-supervised learning models to obtain aggregated unsupervised learning parameters; and merging the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to obtain merged parameters; the merged parameters are used to update a server model; returning the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to the plurality of client semi-supervised learning models as initial supervised learning parameters and initial unsupervised learning parameters for the next round of model training; the training of the plurality of client semi-supervised learning models to obtain the supervised learning parameters and the unsupervised learning parameters of each of the client semi-supervised learning models comprises the following steps: the training comprises supervised learning and unsupervised learning; for each of the client semi-supervised learning models, model overall parameters θ are decomposed into supervised learning parameters α and unsupervised learning parameters β, so that θ = α + β; the supervised learning is trained based on a labeled data set, and the unsupervised learning parameters β are frozen during the training process to obtain updated supervised learning parameters; the unsupervised learning is trained based on an unlabeled data set; and the supervised learning parameters α are frozen during the training process to obtain updated unsupervised learning parameters; after each round of unsupervised learning, the pseudo-labels generated by the model are sampled, and the data corresponding to the sampled pseudo-labels are added to the labeled data set for the next round of supervised learning; the loss function used in the unsupervised learning is as follows: wherein, is a loss function, CE is a cross-entropy loss function, is a label, y is an output, is a local client semi-supervised learning model for an input an output, is an input data after data augmentation, H is a number of high-reliability models, KL is a relative entropy function, is a high-reliability model for an input an output, is a local client semi-supervised learning model for an input an output; wherein the label is expressed using the following equation: wherein is a one-hot function.

2. The method of claim 1, wherein, wherein, the selecting of at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameters and the unsupervised learning parameters comprises the following steps: calculating a classification accuracy score of each of the client semi-supervised learning models on a validation set according to the supervised learning parameters and the unsupervised learning parameters; calculating the reliability of each of the client semi-supervised learning models according to the classification accuracy score, and using the following formula: = wherein, reliability of a client semi-supervised learning model numbered i, denotes the classification accuracy score of the client semi-supervised learning model numbered i on the validation set, denotes the sum of the classification accuracy scores from all client semi-supervised learning models numbered 1 through i; selecting at least one high-reliability model according to the reliability.

3. The method of claim 1, wherein, wherein, the aggregating of the supervised learning parameters of all the client semi-supervised learning models to obtain the aggregated supervised learning parameters uses the following formula: wherein, is aggregated supervised learning parameters of all client semi-supervised learning models from No. 1 to No. I, is reliability of the client semi-supervised learning model No. i, is supervised learning parameters of the client semi-supervised learning model No. i, and I is the number of client semi-supervised learning models. the aggregating of the unsupervised learning parameters of all the client semi-supervised learning models to obtain the aggregated unsupervised learning parameters uses the following formula: wherein, is the aggregated unsupervised learning parameter of all client semi-supervised learning models from No. 1 to No. I, is the reliability of the client semi-supervised learning model No. i, is the unsupervised learning parameter of the client semi-supervised learning model No. i, and I is the number of client semi-supervised learning models.

4. A high-reliability semi-supervised federated learning system, characterized in that, the method comprises the following steps: A plurality of client modules, a parameter uploading module and a server, the server comprising a high-reliability model selection module, a parameter aggregation updating module and a parameter returning module, each of the client modules being provided with a client semi-supervised learning model, and the server being provided with a server model; Each of the client modules is configured to train the client semi-supervised learning model to obtain supervised learning parameters and unsupervised learning parameters of the client semi-supervised learning model; The parameter uploading module is configured to upload the supervised learning parameters and the unsupervised learning parameters of the client semi-supervised learning model obtained by each of the client modules to the server; The high-reliability model selection module is configured to select at least one high-reliability model from the plurality of client semi-supervised learning models according to the supervised learning parameters and the unsupervised learning parameters, and further configured to return the supervised learning parameters and the unsupervised learning parameters of each of the high-reliability models to the plurality of client modules, so as to assist in training the plurality of client semi-supervised learning models in the next round of training; The parameter aggregation updating module is configured to aggregate the supervised learning parameters of all the client semi-supervised learning models to obtain aggregated supervised learning parameters, aggregate the unsupervised learning parameters of all the client semi-supervised learning models to obtain aggregated unsupervised learning parameters, combine the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to obtain combined parameters, and update the server model according to the combined parameters; The parameter returning module is configured to return the aggregated supervised learning parameters and the aggregated unsupervised learning parameters to the plurality of client modules as initial supervised learning parameters and initial unsupervised learning parameters for the next round of model training; Each of the client modules is further configured to: The training comprises supervised learning and unsupervised learning; For each client semi-supervised learning model, model overall parameters θ are decomposed into supervised learning parameters α and unsupervised learning parameters β, so that θ = α + β; The supervised learning is trained based on a labeled data set, and the unsupervised learning parameters β are frozen during the training process to obtain updated supervised learning parameters; the unsupervised learning is trained based on an unlabeled data set; The supervised learning parameters α are frozen during the training process to obtain updated unsupervised learning parameters; After each round of unsupervised learning, the pseudo-labels generated by the model are sampled, and the data corresponding to the sampled pseudo-labels are added to the labeled data set for the next round of supervised learning; The loss function used in the unsupervised learning is as follows: wherein, is a loss function, CE is a cross-entropy loss function, is a label, y is an output, is a local client semi-supervised learning model for an input an output, is an input data after data augmentation, H is a number of high-reliability models, KL is a relative entropy function, is a th high-reliability model for an input an output, is a local client semi-supervised learning model for an input an output; wherein the label is expressed using the following equation: wherein is a one-hot function.

5. The system of claim 4, wherein, The high-reliability model selection module is further configured to: Calculate the classification accuracy score of each of the client semi-supervised learning models on a validation set according to the supervised learning parameters and the unsupervised learning parameters; Calculate the reliability of each of the client semi-supervised learning models according to the classification accuracy score, and the formula used is as follows: = wherein, reliability of a client semi-supervised learning model numbered i, denotes the classification accuracy score of the client semi-supervised learning model numbered i on the validation set, denotes the sum of the classification accuracy scores from all client semi-supervised learning models numbered 1 through i. Select at least one high-reliability model according to the reliability.

6. The system of claim 4, wherein, The parameter aggregation updating module is further configured to: The aggregated supervised learning parameters use the following formula: wherein, is aggregated supervised learning parameters of all client semi-supervised learning models from No. 1 to No. I, is reliability of the client semi-supervised learning model No. i, is supervised learning parameters of the client semi-supervised learning model No. i, and I is the number of client semi-supervised learning models. The formula of the unsupervised learning parameter after the aggregation is as follows: wherein, is the aggregated unsupervised learning parameter of all client semi-supervised learning models from number 1 to number I, is the reliability of the client semi-supervised learning model numbered i, is the unsupervised learning parameter of the client semi-supervised learning model numbered i, and I is the number of client semi-supervised learning models.

Images