Classification federal learning method and device based on dual decoupling self-balancing strategy

Abstract

The invention provides a classification federated learning method and device based on a dual decoupling self-balancing strategy, and the method comprises the steps: selecting clients participating in federated learning, and transmitting global model parameters and training hyper-parameters to each client; for the classification model of each client, performing representation training based on the global model parameters and the training hyper-parameters to obtain decoupling feature parameters; classifier retraining is carried out based on the decoupling characteristic parameters and the global model parameters, and new classifier parameters are obtained; and receiving and aggregating the decoupling feature parameters and the new classifier parameters sent by each client to obtain a global classification model. The local model with balanced classification is trained through each client, and the balanced global model is obtained through aggregation, so that dependence on any auxiliary data is avoided, and data privacy is protected. Besides, a two-stage training mode is adopted, decoupling of the feature extractor and the classifier is achieved, the re-training process of the classifier is optimized through a decoupling loss function, and the model performance is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a classification federated learning method and device based on a dual decoupling self-balancing strategy. Background Art

[0002] In the financial sector, banks' risk management and credit assessment are crucial to asset quality and stability. Traditional risk classification models often rely on a single bank's internal data. This makes it difficult to identify high-risk customers when data scale and diversity are insufficient, especially for small and medium-sized banks. Financial data often exhibits a long-tail distribution, with the majority of customers having low credit risk, while a small number of high-risk customers are crucial to a bank's asset quality and risk management. Therefore, accurately identifying these high-risk customers becomes a core challenge.

[0003] Federated learning, an emerging distributed machine learning technology, provides a method for banks to jointly train models while protecting data privacy. However, federated learning faces challenges in the financial sector. First, due to the heterogeneity of bank data—that is, differences in the distribution of data from different banks—this heterogeneity slows the convergence of the global model and may even cause oscillation or non-convergence, seriously affecting model performance. Second, federated learning faces the problem of data imbalance. In financial risk control, high-risk customers are often in the minority. This makes it difficult for the global model to fully utilize minority data during training, further reducing the model's ability to identify high-risk customers.

[0004] Therefore, how to alleviate the challenges brought by data imbalance and data heterogeneity under the federated learning framework and improve the model's ability to identify high-risk customers has become a core issue in current financial research. Summary of the Invention

[0005] In view of this, an embodiment of the present invention provides a classification federated learning method and apparatus based on a dual decoupling self-balancing strategy to solve the problems of data heterogeneity and imbalance.

[0006] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0007] A first aspect of the present invention discloses a classification federated learning method based on a dual decoupling self-balancing strategy, the method comprising:

[0008] Selecting clients to participate in federated learning and sending global model parameters and training hyperparameters to each of the clients;

[0009] For each of the client's classification models, characterization training is performed based on the global model parameters and the training hyperparameters to obtain decoupled feature parameters;

[0010] Retraining the classifier based on the decoupling feature parameters and the global model parameters to obtain new classifier parameters;

[0011] The decoupled feature parameters and the new classifier parameters sent by each client are received and aggregated to obtain a global classification model.

[0012] Preferably, the classification model for each client is characterized and trained based on the global model parameters and the training hyperparameters to obtain decoupled feature parameters, including:

[0013] For each classification model of the client, extracting local model copy parameters corresponding to the classification model from the global model parameters;

[0014] Use cross entropy as the loss function for training disentangled features;

[0015] According to the local model copy parameters, the training hyperparameters and the loss function, representation training is performed on the local data set of the classification model to obtain decoupled feature parameters.

[0016] Preferably, the retraining of the classifier based on the decoupling feature parameters and the global model parameters to obtain new classifier parameters includes:

[0017] Adjusting the feature extractor of the classification model of the client according to the decoupled feature parameters to obtain a new classification model;

[0018] Determine a target loss function based on the logits values ​​of the new classification model trained on the local dataset and the logits values ​​of the missing classes trained on the global model;

[0019] Based on the target loss function and the classifier parameters in the global model parameters, the classifier is retrained to obtain new classifier parameters.

[0020] Preferably, determining the target loss function according to the new classification model, the logits value trained by the local dataset, and the missing class logits value trained by the global model includes:

[0021] The logits values ​​trained on the local data set are divided into missing class logits values ​​and activated class logits values ​​according to the new classification model;

[0022] Determine the degree of difference between the softmax prediction vector of the missing class logits value and the softmax prediction vector of the missing class logits value trained by the global model as a truncated knowledge distillation loss function;

[0023] Determining a truncated balanced loss function based on the activation class logits value;

[0024] The truncated knowledge distillation loss function and the truncated balance loss function are combined to obtain the target loss function.

[0025] Preferably, the receiving and aggregating the decoupled feature parameters and the new classifier parameters sent by each client to obtain a global classification model includes:

[0026] receiving the decoupled feature parameters and the new classifier parameters sent by each of the clients;

[0027] Performing a weighted average operation on the decoupling feature parameters and the new classifier parameters according to the amount of data uploaded by each client to obtain new global model parameters;

[0028] A global classification model is determined based on the new global model parameters.

[0029] A second aspect of the present invention discloses a classification federated learning device based on a dual decoupling self-balancing strategy, the device comprising:

[0030] A selection unit, configured to select a client to participate in federated learning and send global model parameters and training hyperparameters to each of the clients;

[0031] A representation training unit, configured to perform representation training on the classification model of each client based on the global model parameters and the training hyperparameters to obtain decoupled feature parameters;

[0032] a classifier retraining unit, configured to retrain the classifier based on the decoupling feature parameters and the global model parameters to obtain new classifier parameters;

[0033] An aggregation unit is used to receive the decoupled feature parameters and the new classifier parameters sent by each client and aggregate them to obtain a global classification model.

[0034] Preferably, the representation training unit includes:

[0035] An extraction module, configured to extract, for each classification model of the client, local model copy parameters corresponding to the classification model from the global model parameters;

[0036] Loss function determination module, used to use cross entropy as the loss function for training decoupled features;

[0037] A representation training module is used to perform representation training on the local data set of the classification model according to the local model copy parameters, the training hyperparameters and the loss function to obtain decoupled feature parameters.

[0038] Preferably, the classifier retraining unit includes:

[0039] An adjustment module, configured to adjust a feature extractor of the classification model of the client according to the decoupled feature parameters to obtain a new classification model;

[0040] A determination module, configured to determine a target loss function based on the logits values ​​of the new classification model trained on the local dataset and the logits values ​​of the missing classes trained on the global model;

[0041] The classifier retraining module is used to retrain the classifier based on the target loss function and the classifier parameters in the global model parameters to obtain new classifier parameters.

[0042] Preferably, the determining module includes:

[0043] A division submodule is used to divide the logits values ​​trained by the local data set into missing class logits values ​​and activated class logits values ​​according to the new classification model;

[0044] A first determination submodule is used to determine the difference between the softmax prediction vector of the missing class logits value and the softmax prediction vector of the missing class logits value trained by the global model as a truncated knowledge distillation loss function;

[0045] A second determination submodule is used to determine a truncated balanced loss function according to the activation class logits value;

[0046] The merging submodule is used to merge the truncated knowledge distillation loss function and the truncated balance loss function to obtain a target loss function.

[0047] Preferably, the aggregation unit is specifically used to: receive the decoupling feature parameters and the new classifier parameters sent by each of the clients; perform a weighted averaging operation on the decoupling feature parameters and the new classifier parameters according to the amount of data uploaded by each of the clients to obtain new global model parameters; and determine a global classification model based on the new global model parameters.

[0048] Based on the above-mentioned embodiment of the present invention, a classification federated learning method and device based on a dual decoupling self-balancing strategy is provided. The clients participating in the federated learning are selected, and the global model parameters and training hyperparameters are sent to each client; for the classification model of each client, representation training is performed based on the global model parameters and training hyperparameters to obtain decoupled feature parameters; the classifier is retrained based on the decoupled feature parameters and the global model parameters to obtain new classifier parameters; the decoupled feature parameters and new classifier parameters sent by each client are received and aggregated to obtain a global classification model. By training a local model with balanced classification on each client and aggregating to obtain a balanced global model, dependence on any auxiliary data is avoided, and data privacy is effectively protected. In addition, a two-stage training method is adopted to achieve decoupling of the feature extractor and the classifier, and the retraining process of the classifier is optimized by decoupling the loss function, which significantly improves the model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0050] Figure 1 A schematic diagram of the basic architecture of federated learning provided by an embodiment of the present invention;

[0051] Figure 2 A flowchart of a classification federated learning method based on a dual decoupling self-balancing strategy provided by an embodiment of the present invention;

[0052] Figure 3 A schematic diagram of a classification federated learning method based on a dual decoupling self-balancing strategy provided by an embodiment of the present invention;

[0053] Figure 4 A structural block diagram of a classification federated learning device based on a dual decoupling self-balancing strategy provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] In this application, the terms "comprises," "comprising," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0056] As can be seen from the background technology, federated learning currently faces the problems of data heterogeneity and imbalance in the financial field, which slows down the convergence of the global model and reduces the ability to identify high-risk customers.

[0057] Therefore, an embodiment of the present invention provides a classification federated learning method and device based on a dual decoupling self-balancing strategy, which selects clients participating in federated learning and sends global model parameters and training hyperparameters to each client; for the classification model of each client, characterization training is performed based on the global model parameters and training hyperparameters to obtain decoupled feature parameters; classifier retraining is performed based on the decoupled feature parameters and global model parameters to obtain new classifier parameters; the decoupled feature parameters and new classifier parameters sent by each client are received and aggregated to obtain a global classification model. By training a local model with balanced classification on each client and aggregating to obtain a balanced global model, dependence on any auxiliary data is avoided and data privacy is effectively protected. In addition, a two-stage training method is adopted to achieve decoupling of the feature extractor and the classifier, and the retraining process of the classifier is optimized by decoupling the loss function, which significantly improves the model performance.

[0058] It should be noted that federated learning is a new distributed machine learning framework. Its basic architecture is as follows: Figure 1 As shown in the figure, it consists of a central server and multiple edge devices (clients such as smartphones and tablets). These clients are connected via a network and collaboratively train the model. During training, each client independently updates the model parameters using its local private data, while the central server aggregates the parameter updates from all clients to optimize the global model. Through multiple iterations, the model gradually converges or reaches the expected performance. The core advantage of federated learning is that it avoids direct sharing of raw data, protecting user privacy, while integrating dispersed data resources to significantly improve the robustness and generalization of the model.

[0059] It is understood that the information (including but not limited to user device information, user personal information) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0060] See also Figure 2 , shows a flowchart of a classification federated learning method based on a dual decoupling self-balancing strategy provided by an embodiment of the present invention.

[0061] It is understandable that in the problem of multi-bank joint training of customer risk classification, suppose there are N banks participating in federated learning. Let X∈R be the instance space, Z∈R be a transition feature space, and Y∈R be the output space.<Y,f( )> The local task T of each bank is defined as a task on domain D consisting of X and prior distribution P(X). n satisfy Model parameters It consists of two parts: one part consists of parameters Mapping of the implemented feature extractor f , and the other part consists of parameters Mapping of the implemented classifier p It can achieve ,in Is a simplex of Y. Given a non-negative convex loss function L: The goal of federated learning is to learn a global model parameter , in order to minimize the risk loss, as shown in formula (1).

[0062] (1)

[0063] In formula (1) Represents the feature extractor parameters; Represents the classifier parameters; L represents the non-negative convex loss function; h represents the mapping process of the classifier; Represents the mapping process of the feature extractor; represents the sample data (i.e. input) of task n; Y n represents the label (i.e., output) of task n.

[0064] Therefore, federated learning is essentially an optimization problem, and the optimization goal is the global model , so that the global model The expected loss value of all local tasks can be minimized. In practice, this optimization goal can be expressed empirically as follows (such as formula (2)):

[0065] (2)

[0066] In formula (2), D n Represents the local dataset of bank n, and the global dataset D conforms to the long-tail distribution, and each independent data D n There is a Non-IID distribution (non-independent distribution) between them.

[0067] The global dataset D satisfies and L represents a non-negative convex loss function; h represents the mapping process of the classifier; Represents the mapping process of the feature extractor; Y n Represents the label of task n.

[0068] Therefore, the optimization goal of this optimization problem is to design an effective loss function L. Based on this loss function L, each bank can train a balanced local model. .

[0069] During the training process, the bank's local model updates its own model parameters, and then sends the updated model parameters and local data volume |D n |Upload to the central server together.

[0070] It should be noted that the entire updating process can be expressed as (as shown in formula (3)):

[0071] (3)

[0072] In formula (3), represents the updated local model of the bank; represents the updated classifier parameters; represents the updated feature extractor parameters; Represents local data Features after k-1 global rounds of feature extractor mapping; L represents a non-negative convex loss function; h represents the classifier mapping.

[0073] Finally, the central server will update the updated model parameters uploaded by all banks And the amount of local data |D n | Perform aggregation to obtain the global model .

[0074] The following combination Figure 2 The content in this article explains the classification federated learning method based on the dual decoupling self-balancing strategy proposed in the embodiment of the present invention to solve the above optimization problem. The method includes:

[0075] Step S201: Select clients participating in federated learning and send global model parameters and training hyperparameters to each client.

[0076] It should be noted that in classic federated learning scenarios, due to the large number of participating clients, a random selection of clients is generally considered for each round of training. However, considering the relatively small number of banks in the customer risk classification scenario, each round of training can either randomly select a subset of banks or select all of them.

[0077] In the specific implementation of step S201, n clients are selected to participate in the kth round of federated learning, and the global model parameters of the previous round are and training hyperparameters (such as learning rate, etc.) are sent to each client.

[0078] It is understandable that the global model parameters By feature extractor parameters and classifier parameters composition.

[0079] Step S202: For each client's classification model, representation training is performed based on global model parameters and training hyperparameters to obtain decoupled feature parameters.

[0080] Combine Figure 3 The schematic diagram of the classification federated learning method based on the dual decoupling self-balancing strategy shown in the figure, in the process of specifically implementing step S202, for the classification model of each client, the local model copy parameters corresponding to the classification model are extracted from the global model parameters; secondly, the cross entropy is used as the loss function for training the decoupled features; then, according to the local model copy parameters, training hyperparameters and loss function, the local data set of the classification model is represented and trained to obtain the decoupled feature parameters.

[0081] It is understandable that in the global model parameters In the example, the local model copy parameters corresponding to the classification model of client n can be extracted Then, cross entropy (CE) is used as the loss function to train the decoupled feature parameters to achieve the decoupled feature extractor and classifier.

[0082] Step S203: retraining the classifier based on the decoupled feature parameters and the global model parameters to obtain new classifier parameters.

[0083] In the specific implementation of step S203 , the classifier is retrained based on the classifier parameters and the decoupling feature parameters in the global model parameters to obtain new classifier parameters.

[0084] Specifically, the training process is as follows (process B1 to process B3):

[0085] Process B1: According to the decoupled feature parameters, the feature extractor of the client's classification model is adjusted to obtain a new classification model.

[0086] It can be understood that the purpose of process B1 is to use the trained decoupled feature parameters as fixed feature extractors in the subsequent local training phase of client n, and to retrain the classifier based on them.

[0087] In the specific implementation process B1, the feature extractor of the client's classification model is adjusted according to the trained decoupling feature parameters to obtain a new classification model.

[0088] Process B2: Determine the target loss function based on the new classification model, the logits values ​​trained on the local dataset, and the missing class logits values ​​trained on the global model.

[0089] It can be understood that the local dataset in the client is divided into a missing class dataset and an activated class dataset.

[0090] The specific process of determining the target loss function is as follows (process C1 to process C4):

[0091] Process C1: Based on the new classification model, the logits values ​​trained on the local dataset are divided into missing class logits values ​​and activated class logits values.

[0092] In the specific implementation process C1, the local dataset is trained according to the new classification model, and the logits values ​​output during the training process (that is, the model's prediction scores for each category) are segmented and processed to distinguish between missing class logits values ​​and activated class logits values.

[0093] Combine Figure 3 As shown, and denote the missing class and the activated class on client n, respectively, where .

[0094] Given an input sample , local model Will generate dimensional logits: . and They represent missing class logits and activation class logits respectively. The former is used to calculate the loss of knowledge distillation, and the latter is used to calculate the prediction loss of local activation class.

[0095] This approach decouples the calculation of the KD loss and the LA loss, allowing for both filling in missing classes and reducing the imbalance of activation classes. The final loss function, L, is the result of combining the KD and LA loss functions. This loss function, L, is used as the loss function for retraining the classifier.

[0096] Process C2: The degree of difference between the softmax prediction vector of the missing class logits value and the softmax prediction vector of the missing class logits value trained by the global model is determined as the truncated knowledge distillation loss function.

[0097] It should be noted that for client n, the truncated knowledge distillation loss function, also known as the TKD loss function It can be defined as the softmax prediction vector of the missing class logits inferred by the local model and the softmax prediction vector of the missing class logits inferred by the global model The KL divergence loss between .

[0098] TKD loss function The specific definition of is shown in formula (4).

[0099] (4)

[0100] In formula (4), represents the missing class on client n; represents the softmax output of the local model, Represents the softmax output of the global model; Is a hyperparameter, equivalent to the temperature parameter.

[0101] Process C3: Determine the truncated balanced loss function based on the activation class logits value.

[0102] It should be noted that the truncated balanced loss function, that is, the TLA loss function is based on the activation class logits This is achieved by modifying the LA loss function and expressing it as .

[0103] TLA loss function The specific definition of is shown in formula (5).

[0104] (5).

[0105] In formula (5), Represents the active class on client n; represents the softmax prediction vector of the activation class logits, A one-hot vector representing the activation class. Representing category prior In actual operation, the proportion of target samples to total samples can be used as an alternative; It is a hyperparameter, equivalent to the temperature parameter in knowledge distillation.

[0106] Process C4: Combine the truncated knowledge distillation loss function and the truncated balance loss function to obtain the target loss function.

[0107] When implementing process C4, the truncated knowledge distillation loss function, TKD loss function , and truncated balanced loss function, TLA loss function Merge and get the target loss function.

[0108] Specifically as shown in formula (6):

[0109] (6)

[0110] In formula (6), Represents the missing class one-hot vector; Represents the softmax prediction vector of the missing class; One-hot vector representing the activation class; Represents the softmax prediction vector of the activation class.

[0111] Process B3: Based on the target loss function and the classifier parameters in the global model parameters, the classifier is retrained to obtain new classifier parameters.

[0112] When implementing process B3, the objective function L is used as the loss function for retraining the classifier, thereby performing the gradient regression algorithm. The detailed steps of the entire process are shown in Table 1.

[0113] Table 1

[0114]

[0115] Step S204: receiving the decoupled feature parameters and new classifier parameters sent by each client and aggregating them to obtain a global classification model.

[0116] In the specific implementation of step S204, the central server receives the decoupled feature parameters and new classifier parameters sent by each client; then performs a weighted average operation on the decoupled feature parameters and the new classifier parameters according to the amount of data uploaded by each client to obtain new global model parameters; finally, based on the new global model parameters, the global classification model is determined. , thus completing the k-th round of training.

[0117] Specifically, the aggregation process is based on the amount of data uploaded by each client. n |Yes Perform weighted averaging, N k represents the client selected for the kth round of training. The entire aggregation process can be defined as formula (7).

[0118] (7)

[0119] In formula (7), represents the global classification model; N k represents the client selected for the kth round of training; n represents the number of clients; |D n | represents the amount of data uploaded by client n; |D| represents the total amount of data; Represents the local model trained by client n.

[0120] In this embodiment of the present invention, a balanced global model is obtained by training a local model with balanced classification. When the local model reaches a balance, it indicates that the method has effectively overcome the challenges posed by data imbalance and heterogeneity. Furthermore, this method effectively protects data privacy by eliminating the need for auxiliary data. Furthermore, the two-stage training decouples the feature extractor and classifier, and the decoupling of the loss function further enables retraining of the classifier, effectively improving model performance.

[0121] Corresponding to the classification federated learning method based on the dual decoupling self-balancing strategy provided by the above embodiment of the present invention, see Figure 4 , shows a structural block diagram of a classification federated learning device based on a dual decoupling self-balancing strategy provided by an embodiment of the present invention.

[0122] The device includes: a selection unit 401 , a representation training unit 402 , a classifier retraining unit 403 and an aggregation unit 404 .

[0123] The selection unit 401 is used to select clients participating in federated learning and send global model parameters and training hyperparameters to each client.

[0124] The representation training unit 402 is configured to perform representation training on the classification model of each client based on the global model parameters and the training hyperparameters to obtain decoupled feature parameters.

[0125] The classifier retraining unit 403 is configured to retrain the classifier based on the decoupled feature parameters and the global model parameters to obtain new classifier parameters.

[0126] The aggregation unit 404 is configured to receive the decoupled feature parameters and new classifier parameters sent by each client and aggregate them to obtain a global classification model.

[0127] Aggregation unit 404 is specifically used to: receive the decoupled feature parameters and new classifier parameters sent by each client; perform weighted averaging operation on the decoupled feature parameters and the new classifier parameters according to the amount of data uploaded by each client to obtain new global model parameters; and determine the global classification model based on the new global model parameters.

[0128] In this embodiment of the present invention, a balanced global model is obtained by training a local model with balanced classification. When the local model reaches a balance, it indicates that the method has effectively overcome the challenges posed by data imbalance and heterogeneity. Furthermore, this method effectively protects data privacy by eliminating the need for auxiliary data. Furthermore, the two-stage training decouples the feature extractor and classifier, and the decoupling of the loss function further enables retraining of the classifier, effectively improving model performance.

[0129] Combine Figure 4 The content shown, the representation training unit 402, includes: an extraction module, a loss function determination module and a representation training module.

[0130] The extraction module is used to extract the local model copy parameters corresponding to the classification model of each client from the global model parameters.

[0131] The loss function determination module is used to use cross entropy as the loss function for training decoupled features.

[0132] The representation training module is used to perform representation training on the local dataset of the classification model based on the local model copy parameters, training hyperparameters and loss function to obtain decoupled feature parameters.

[0133] Combine Figure 4 The content shown, the classifier retraining unit 403, includes: an adjustment module, a determination module and a classifier retraining module.

[0134] The adjustment module is used to adjust the feature extractor of the client's classification model according to the decoupled feature parameters to obtain a new classification model.

[0135] The determination module is used to determine the target loss function based on the logits value of the new classification model trained on the local dataset and the missing class logits value trained on the global model.

[0136] The classifier retraining module is used to retrain the classifier based on the target loss function and the classifier parameters in the global model parameters to obtain new classifier parameters.

[0137] Combine Figure 4 The content shown, the determination module, includes: a division submodule, a first determination submodule, a second determination submodule and a merging submodule.

[0138] The division submodule is used to divide the logits values ​​trained on the local dataset into missing class logits values ​​and activated class logits values ​​according to the new classification model.

[0139] The first determination submodule is used to determine the degree of difference between the softmax prediction vector of the missing class logits value and the softmax prediction vector of the missing class logits value trained by the global model as a truncated knowledge distillation loss function.

[0140] The second determination submodule is used to determine the truncated balanced loss function according to the activation class logits value.

[0141] The merging submodule is used to merge the truncated knowledge distillation loss function and the truncated balance loss function to obtain the target loss function.

[0142] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0143] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0144] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the present invention. Therefore, the present invention is not limited to the embodiments shown herein, but is intended to conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A classification federated learning method based on a dual decoupling self-balancing strategy, characterized in that: The method comprises: Selecting clients to participate in federated learning and sending global model parameters and training hyperparameters to each of the clients; For each of the client's classification models, characterization training is performed based on the global model parameters and the training hyperparameters to obtain decoupled feature parameters; Retraining the classifier based on the decoupling feature parameters and the global model parameters to obtain new classifier parameters; The decoupled feature parameters and the new classifier parameters sent by each client are received and aggregated to obtain a global classification model.

2. The method according to claim 1, characterized in that The classification model for each client is characterized and trained based on the global model parameters and the training hyperparameters to obtain decoupled feature parameters, including: For each classification model of the client, extracting local model copy parameters corresponding to the classification model from the global model parameters; Use cross entropy as the loss function for training disentangled features; According to the local model copy parameters, the training hyperparameters and the loss function, representation training is performed on the local data set of the classification model to obtain decoupled feature parameters.

3. The method according to claim 1, characterized in that The retraining of the classifier based on the decoupled feature parameters and the global model parameters to obtain new classifier parameters includes: Adjusting the feature extractor of the classification model of the client according to the decoupled feature parameters to obtain a new classification model; Determine a target loss function based on the logits values ​​of the new classification model trained on the local dataset and the logits values ​​of the missing classes trained on the global model; Based on the target loss function and the classifier parameters in the global model parameters, the classifier is retrained to obtain new classifier parameters.

4. The method according to claim 3, characterized in that Determining a target loss function based on the new classification model, the logits value trained by the local dataset, and the missing class logits value trained by the global model includes: The logits values ​​trained on the local data set are divided into missing class logits values ​​and activated class logits values ​​according to the new classification model; Determine the degree of difference between the softmax prediction vector of the missing class logits value and the softmax prediction vector of the missing class logits value trained by the global model as a truncated knowledge distillation loss function; Determining a truncated balanced loss function based on the activation class logits value; The truncated knowledge distillation loss function and the truncated balance loss function are combined to obtain the target loss function.

5. The method according to claim 1, wherein The receiving and aggregating the decoupled feature parameters and the new classifier parameters sent by each client to obtain a global classification model includes: receiving the decoupled feature parameters and the new classifier parameters sent by each of the clients; Performing a weighted average operation on the decoupling feature parameters and the new classifier parameters according to the amount of data uploaded by each client to obtain new global model parameters; A global classification model is determined based on the new global model parameters.

6. A classification federated learning device based on a dual decoupling self-balancing strategy, characterized in that: The device comprises: A selection unit, configured to select a client to participate in federated learning and send global model parameters and training hyperparameters to each of the clients; A representation training unit, configured to perform representation training on the classification model of each client based on the global model parameters and the training hyperparameters to obtain decoupled feature parameters; a classifier retraining unit, configured to retrain the classifier based on the decoupling feature parameters and the global model parameters to obtain new classifier parameters; An aggregation unit is used to receive the decoupled feature parameters and the new classifier parameters sent by each client and aggregate them to obtain a global classification model.

7. The device according to claim 6, characterized in that The representation training unit includes: An extraction module, configured to extract, for each classification model of the client, local model copy parameters corresponding to the classification model from the global model parameters; Loss function determination module, used to use cross entropy as the loss function for training decoupled features; A representation training module is used to perform representation training on the local data set of the classification model according to the local model copy parameters, the training hyperparameters and the loss function to obtain decoupled feature parameters.

8. The device according to claim 6, characterized in that The classifier retraining unit comprises: An adjustment module, configured to adjust a feature extractor of the classification model of the client according to the decoupled feature parameters to obtain a new classification model; A determination module, configured to determine a target loss function based on the logits values ​​of the new classification model trained on the local dataset and the logits values ​​of the missing classes trained on the global model; The classifier retraining module is used to retrain the classifier based on the target loss function and the classifier parameters in the global model parameters to obtain new classifier parameters.

9. The device according to claim 8, characterized in that The determining module includes: A division submodule is used to divide the logits values ​​trained by the local data set into missing class logits values ​​and activated class logits values ​​according to the new classification model; A first determination submodule is used to determine the difference between the softmax prediction vector of the missing class logits value and the softmax prediction vector of the missing class logits value trained by the global model as a truncated knowledge distillation loss function; A second determination submodule is used to determine a truncated balanced loss function according to the activation class logits value; The merging submodule is used to merge the truncated knowledge distillation loss function and the truncated balance loss function to obtain a target loss function.

10. The device according to claim 6, characterized in that The aggregation unit is specifically used to: receive the decoupling feature parameters and the new classifier parameters sent by each of the clients; perform a weighted average operation on the decoupling feature parameters and the new classifier parameters according to the amount of data uploaded by each of the clients to obtain new global model parameters; and determine a global classification model based on the new global model parameters.

Images