A personalized federated learning method and device for incremental data

Abstract

The application relates to the technical field of federated learning, in particular to a personalized federated learning method and device for incremental data, which can solve the problem of the deficiency of traditional federated learning in processing incremental data and personalization to a certain extent. The method comprises the following steps: introducing a weight decay regularization term in the model training process, applying a regularization constraint to the model parameters to prevent model overfitting and improve the generalization ability of the model; fusing the idea of meta-learning and using an inner-outer loop mechanism to train the model; in the outer loop, the personalized algorithm is iterated, so that the model can adapt to new incremental data more quickly, improve the training efficiency and model performance, improve the real-time performance and adaptability of the model, and through the iteration of the personalized algorithm, the model of each client can be optimized according to the characteristics of the client, thereby improving the adaptability and accuracy of the model on different clients.

Description

Technical Field

[0001] This application relates to the field of federated learning technology, and more specifically, to a personalized federated learning method and apparatus for incremental data. Background Technology

[0002] Federated Learning (FL) is an emerging machine learning paradigm with great potential in data privacy protection and resource utilization. Unlike traditional centralized fraud detection learning methods, federated learning allows multiple clients to train their own models locally without uploading sensitive raw data to a central server. After training the model locally, clients only share the model's updated parameters, such as weights and gradients, rather than the actual data content. This process reduces the risk of data leakage and effectively protects user privacy. Furthermore, since data does not need to be frequently transmitted to a central server, federated learning significantly reduces data transmission latency and bandwidth overhead. This efficient data processing method not only improves the training speed of the model but also enhances the model's generalization ability under different data distributions, thereby improving the model's accuracy and real-time performance.

[0003] In practical applications, the traditional method of federated learning is to aggregate and train based on existing data in stages, which has a lag and makes it difficult to process dynamic incremental data from some clients in a timely manner. Furthermore, the global model lacks personalization capabilities. Summary of the Invention

[0004] To address the problems in practical applications where traditional federated learning methods rely on aggregated training based on existing data in stages, resulting in lag, difficulty in timely processing of dynamic incremental data from some clients, and a lack of personalization capabilities in the global model, this application provides a personalized federated learning method and apparatus for incremental data.

[0005] The embodiments of this application are implemented as follows:

[0006] Firstly, this application provides a personalized federated learning method for incremental data, including:

[0007] The global model is trained based on the inner loop, and the global model is optimized by adding a weight decay regularization term;

[0008] Personalized model optimization training is performed based on incremental training data from the outer loop.

[0009] In one possible implementation, the step of training the global model based on an inner loop and optimizing the global model by adding a weight decay regularization term further includes:

[0010] The server initializes the global model parameter W0 and sends the global model parameter to all clients k.

[0011] Each client k receives the global model parameter W. t Then, use its local stored dataset D k Perform model training;

[0012] Client k uploads the updated local model parameters to the server;

[0013] After receiving the model parameters from all clients, the server aggregates them using methods such as weighted average or simple average to obtain the updated global model parameters.

[0014] The server will assign new global model parameters w t+1 Distribute to each client to begin the next round of training.

[0015] 5. In one possible implementation, after receiving the global model parameters, each client uses its local stored dataset for model training, further comprising:

[0016] Optimize the loss function L, which includes a weight decay regularization term, for each client. k (w), which is defined as the sum of the average loss of the local dataset and the weight decay regularization term of the model parameters;

[0017] The loss function is in the form of:

[0018]

[0019] Wherein, l(w;x) i y i ) is the prediction error, and λ is the weight coefficient of the weight decay regularization term.

[0020] In one possible implementation, during model training using its local stock dataset, each client performs multiple iterations using gradient descent, updating the model parameters w in each iteration.

[0021] The specific update steps are as follows:

[0022]

[0023] Where η is the learning rate, and after multiple iterations, the client k obtains the updated local model parameters.

[0024] In one possible implementation, the training process is repeated until a predetermined number of iterations E is reached or an early stopping condition is met.

[0025] In one possible implementation, the personalized model optimization training based on the outer loop incremental training data further includes:

[0026] After receiving the global model parameters, each client performs personalized model optimization training and updates the parameters of its local model to the global model parameters.

[0027] The client uses a local incremental dataset for training;

[0028] The client needs to choose a smaller learning rate to ensure that the fine-tuning process is stable and does not disrupt the pre-trained weights.

[0029] The client performs fewer training iterations to prevent overfitting;

[0030] Each client fine-tunes the model based on its local incremental dataset, adjusting the model weights using the standard backpropagation algorithm.

[0031] After the model training is completed, the client uses a local validation set to evaluate the performance of the fine-tuned personalized model. Based on the evaluation results, the client can decide whether further iterative adjustments to the model parameters are needed.

[0032] In one possible implementation, the incremental dataset contains the latest user interaction data, which can reflect the current user's latest preferences and behaviors.

[0033] In one possible implementation, during the process of performing local computation within multiple clients to obtain the corresponding model parameters, training using a weight decay regularization method can effectively constrain the complexity of the model, prevent overfitting, and thus improve the model's generalization ability.

[0034] In one possible implementation, during the personalized training process, the final global model is adjusted and optimized according to different data characteristics and business needs to obtain a personalized local model that can meet their specific requirements.

[0035] Secondly, this application provides a personalized federated learning device for incremental data, comprising:

[0036] The inner loop training module is used to train the global model based on the inner loop and optimize the global model by adding a weight decay regularization term;

[0037] The outer loop training module is used for personalized model optimization training based on the incremental training data from the outer loop.

[0038] The technical solution provided in this application can achieve at least the following beneficial effects:

[0039] The personalized federated learning method for incremental data provided in this application improves the generalization of the model, reduces model complexity, and prevents overfitting by adding a weight decay regularization term during the training process. It also proposes the idea of ​​fusion meta-learning, which divides the model into an inner loop and an outer loop. The inner loop performs phased global training to obtain a global model of existing data, while the outer loop performs personalized incremental training on the global model to ensure the accuracy of personalized predictions for incremental data. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart illustrating a personalized federated learning method for incremental data, as shown in an exemplary embodiment of this application.

[0042] Figure 2 This is a schematic diagram illustrating the implementation process of training a federated learning model according to an exemplary embodiment of this application;

[0043] Figure 3 This is a schematic diagram of incremental personalized federated learning training shown in an exemplary embodiment of this application;

[0044] Figure 4 This is a schematic diagram illustrating the structure of a personalized federated learning device for incremental data, as shown in an exemplary embodiment of this application.

[0045] Figure label:

[0046] 1. Inner loop training module; 2. Outer loop training module. Detailed Implementation

[0047] To make the objectives, implementation methods and advantages of this application clearer, the exemplary implementation methods of this application will be clearly and completely described below with reference to the accompanying drawings of the exemplary embodiments of this application. Obviously, the exemplary embodiments described are only some embodiments of this application, and not all embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0048] It should be noted that the brief descriptions of terms in this application are only for the convenience of understanding the embodiments described below, and are not intended to limit the embodiments of this application. Unless otherwise stated, these terms should be understood in their ordinary and common meaning.

[0049] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar or related objects or entities, and do not necessarily imply a specific order or sequence, unless otherwise specified. It should be understood that such terms are interchangeable where appropriate.

[0050] The terms “comprising” and “having”, and any variations thereof, are intended to cover but not exclude inclusion, for example, a product or device that includes a range of components is not necessarily limited to all of the components that are clearly listed, but may include other components that are not clearly listed or that are inherent to such product or device.

[0051] Before explaining the personalized federated learning method for incremental data provided in the embodiments of this application, the application scenarios and implementation environment of the embodiments of this application will be introduced first.

[0052] Federated Learning (FL) is an emerging machine learning paradigm with great potential in data privacy protection and resource utilization. Unlike traditional centralized fraud detection learning methods, federated learning allows multiple clients to train their own models locally without uploading sensitive raw data to a central server. After training the model locally, clients only share the model's updated parameters, such as weights and gradients, rather than the actual data content. This process reduces the risk of data leakage and effectively protects user privacy. Furthermore, since data does not need to be frequently transmitted to a central server, federated learning significantly reduces data transmission latency and bandwidth overhead. This efficient data processing method not only improves the training speed of the model but also enhances the model's generalization ability under different data distributions, thereby improving the model's accuracy and real-time performance.

[0053] In practical applications, the traditional method of federated learning is to aggregate and train based on existing data in stages, which has a lag and makes it difficult to process dynamic incremental data from some clients in a timely manner. Furthermore, the global model lacks personalization capabilities.

[0054] Based on this, this application provides a personalized federated learning method and apparatus for incremental data, which integrates the idea of ​​meta-learning. It introduces weight decay regularization through inner and outer loops to improve the generalization ability of the model. In the outer loop, personalized algorithms are iterated to ensure that the model can achieve better performance in incremental data training.

[0055] This application considers scenarios involving efficient processing of incremental training data: traditional methods only optimize for training on existing data in federated learning, neglecting the impact of incremental training data on model training in real-world scenarios. This invention considers both existing and incremental training data, enabling dynamic processing of incremental data from the client and overcoming the lag bottleneck of traditional federated learning models.

[0056] This application fully utilizes the global model for personalized training: The algorithm proposed in this invention introduces the concept of meta-learning, optimizing the personalized model through inner and outer loop mechanisms. Iterating the personalized algorithm within the outer loop allows the model to adapt to new incremental training data more quickly, improving training efficiency and model performance.

[0057] Next, the technical solutions of this application and how they solve the aforementioned technical problems will be described in detail through embodiments and in conjunction with the accompanying drawings. The embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0058] Figure 1 This is a flowchart illustrating a personalized federated learning method for incremental data, as shown in an exemplary embodiment of this application.

[0059] In one exemplary embodiment, such as Figure 1 As shown, a personalized federated learning method for incremental data is provided. In this embodiment, the method may include the following steps:

[0060] Step 100: Train the global model based on the inner loop, and optimize the global model by adding a weight decay regularization term.

[0061] Step 200: Perform personalized model optimization training based on the incremental training data from the outer loop.

[0062] Figure 2 This is a schematic diagram illustrating the implementation process of training a federated learning model according to an exemplary embodiment of this application.

[0063] In one possible implementation, such as Figure 2 As shown, the specific implementation process of its learning method is as follows:

[0064] Training process for a global model based on an inner loop:

[0065] Considering the uneven quantity and distribution of data collected by each client, a weight decay regularization term is added to optimize the model in the global model training process based on the inner loop.

[0066] Suppose there are K clients. The server initializes the global model parameter W0 and sends it to all clients.

[0067] Each client k receives the global model parameters W. t Then, use its local stored dataset D k Perform model training.

[0068] During training, each client optimizes the loss function L, which includes a weight decay regularization term. k (w), which is defined as the sum of the average loss of the local dataset and the weight decay regularization term of the model parameters.

[0069] Specifically, the loss function takes the form of:

[0070]

[0071] Wherein, l(w;x) i y i ) is the prediction error, and λ is the weight coefficient of the weight decay regularization term.

[0072] During the local training phase, each client performs multiple iterations using gradient descent (assuming the number of iterations is ), updating the model parameters w in each iteration.

[0073] The specific update steps are as follows:

[0074]

[0075] Where η is the learning rate. After multiple iterations, the client obtains the updated local model parameters.

[0076] Next, the client uploads the updated local model parameters to the server.

[0077] After receiving the model parameters from all clients, the server aggregates them using methods such as weighted average or simple average to obtain the updated global model parameters.

[0078] The server will assign new global model parameters w t+1 Distribute to each client to begin the next round of training.

[0079] This training process is repeated until the predetermined number of iterations T is reached or the early stopping condition is met. Through this inner and outer loop training process, federated learning not only effectively utilizes distributed data, but also prevents overfitting through weight decay regularization, thereby improving the model's generalization ability and stability.

[0080] Furthermore, since each client only transmits model update parameters and not the original data, this method significantly reduces network bandwidth usage and data transmission latency while protecting data privacy.

[0081] Ultimately, after multiple iterations, the server is able to obtain an accurate global model.

[0082] Personalized optimization process based on incremental training data from the outer loop:

[0083] In a distributed learning environment, since incremental data is not collected simultaneously from all clients, we need to perform personalized model optimization training for those clients that have incremental data. The entire process can be divided into two stages: an inner loop and an outer loop. In the inner loop, the server trains a global model based on the currently available global data. In the outer loop, the server sends this global model to each client that has incremental data, and these clients further customize the model based on this global data.

[0084] First, assume there are N clients with incremental data. The server trains the global model parameters W from the currently available global data through an inner loop. t Specifically, the server aggregates all available global data and obtains the current global model parameters W through multiple rounds of training. t Then, the server distributes these model parameters to all clients that have incremental data.

[0085] Each client receives the global model parameter W. t Then, personalized model optimization training will be performed according to the following steps. The client first updates the parameters of the local model to the global model parameters W. t This ensures that all client models begin personalized training from the same initial state. Specifically, this can be achieved by setting the model weights.

[0086] Next, the client uses its local incremental dataset D n Training is performed on the incremental dataset D. n Including the latest user interaction data, reflecting current user preferences and behaviors, the client needs to choose a smaller learning rate for model fine-tuning to ensure stability and avoid disrupting pre-trained weights. In terms of training settings, the client performs fewer training iterations to prevent overfitting. This meticulous adjustment allows the model to better adapt to local incremental data. Each client uses its local incremental dataset D... n Fine-tune the model by adjusting the model weights using the standard backpropagation algorithm.

[0087] After model training is complete, the client uses a local validation set to evaluate the performance of the fine-tuned personalized model. Based on the evaluation results, the client can decide whether further iterative adjustments to the model parameters are needed.

[0088] By combining internal and external loops, the system can effectively utilize incremental data for personalized model optimization training, improving the model's personalization capabilities and overall performance. This approach not only continuously optimizes model performance but also adapts to constantly changing user needs and behaviors, ensuring the model remains in optimal condition.

[0089] Figure 3 This is a schematic diagram illustrating incremental personalized federated learning training, as shown in an exemplary embodiment of this application.

[0090] In one possible implementation, such as Figure 3 As shown, the implementation method of the learning method is as follows:

[0091] Suppose there are three insurance data centers, which will be operated using the incremental federated learning method described in detail in this paper. Specifically, these are data center 1, data center 2, and data center 3.

[0092] First, local computation is performed within their respective data centers to calculate the corresponding model parameters. During this process, they use a weight decay regularization training method, which effectively constrains the complexity of the model, prevents overfitting, and thus improves the model's generalization ability.

[0093] Once local computations are complete, these data centers send the encrypted results to a dedicated aggregation server. The aggregation server then uses federated learning dynamic weight aggregation to process these results from different data centers, generating a global model which is then sent back to all participating parties.

[0094] After receiving the global model, each participating party encrypts the received gradients to ensure data security and privacy. Next, they use the decrypted gradient results to update the model parameters within their respective data centers. After this series of operations, all parties obtain the final global model.

[0095] Once all parties obtain the final global model and incremental data, they will use this incremental data to perform personalized training on the already constructed global model. During this personalized training process, the model will be adjusted and optimized according to different data characteristics and business needs, ultimately resulting in a personalized local model that meets each party's specific requirements. This personalized local model can better adapt to different scenarios and data environments in practical applications, providing more accurate and efficient services.

[0096] Working principle:

[0097] Some embodiments of this application introduce weight decay regularization terms during model training. By imposing regularization constraints on model parameters, overfitting is prevented and the generalization ability of the model is improved. By incorporating the idea of ​​meta-learning, the model is trained using an inner and outer loop mechanism. The outer loop iterates personalized algorithms, enabling the model to adapt to new incremental data more quickly and improving training efficiency and model performance.

[0098] Regularization is introduced in both the inner and outer loops to ensure model stability and generalization ability throughout the training process. The inner loop primarily updates the global model, while the outer loop optimizes individual models.

[0099] Furthermore, some embodiments of this application design a mechanism for dynamically processing incremental client data, which can update and process new client data in real time, breaking through the lag bottleneck of traditional federated learning models, improving the model's real-time performance and adaptability, and enabling each client's model to be optimized according to its own characteristics through personalized algorithm iteration, thereby improving the model's adaptability and accuracy on different clients and solving the problem of the global model lacking personalization capabilities.

[0100] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially as indicated, these steps are not necessarily executed in the indicated order. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps.

[0101] Corresponding to the aforementioned embodiments of the personalized federated learning method for incremental data, and employing the same technical concept, this application also provides embodiments of a personalized federated learning apparatus for incremental data.

[0102] Figure 4 This is a schematic diagram illustrating the structure of a personalized federated learning device for incremental data, as shown in an exemplary embodiment of this application.

[0103] In one exemplary embodiment, such as Figure 4 As shown, this personalized federated learning device for incremental data includes:

[0104] Inner loop training module 1 is used to train the global model based on the inner loop and optimize the global model by adding a weight decay regularization term;

[0105] The outer loop training module 2 is used for personalized model optimization training based on the incremental training data from the outer loop.

[0106] Specific limitations regarding the personalized federated learning apparatus for incremental data can be found in the limitations of the personalized federated learning method for incremental data described above, and will not be repeated here. Each module in the aforementioned personalized federated learning apparatus for incremental data can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0107] As can be seen, some embodiments of this application take into account the uneven data distribution among clients during federated learning training, and propose to add a weight decay regularization term during training to improve the generalization of the model, reduce model complexity, and prevent overfitting.

[0108] Furthermore, to address the issue that incremental training data cannot be added to the global model training in a timely manner for some clients, the idea of ​​fusion meta-learning is proposed. The model is divided into an inner loop and an outer loop. The inner loop performs phased global training to obtain a global model of existing data, while the outer loop performs personalized incremental training on the global model to ensure the accuracy of personalized predictions for incremental data.

[0109] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for personalized federated learning oriented to incremental data, characterized in that, Comprise: Training the global model based on inner loop, and optimizing the global model by adding weight decay regularization term, comprising: initializing global model parameters by a server and sending the global model parameters to all clients ; Each client Upon receiving the current global model parameters It then uses its local inventory dataset to perform model training; Client uploading the updated local model parameters to the server; After the server receives the model parameters of all clients, the updated global model parameters are obtained by weighted average or simple average method ; The server will change the global model parameters. Distribute the data to each client to begin the next round of training; Incremental training data based on outer loop for personalized model optimization training, comprising: Each client receives the global model parameters, and performs personalized model optimization training to update the local model parameters to the global model parameters; The client uses the local incremental data set for training; The client needs to select a smaller learning rate to ensure that the fine-tuning process is stable and will not damage the pre-training weights; The client performs a smaller number of training to prevent overfitting; Each client adjusts the model weights using the standard backpropagation algorithm based on the local incremental data set; After the model training is completed, the client evaluates the performance of the fine-tuned personalized model using the local validation set, and according to the evaluation results, the client can decide whether to further iterate and adjust the model parameters; Apply the incremental data-oriented personalized federated learning method to the insurance data center, comprising: The insurance data center performs local calculation in each data center to calculate the corresponding model parameters; After local calculation is completed, the insurance data center sends the calculation results processed by encryption technology to a special aggregation server, and the aggregation server processes the results from different data centers using the federated learning dynamic weight aggregation method to generate a global model, and sends the global model back to each participant; Each participant encrypts the received gradient to ensure data security and privacy; Update the model parameters in each data center using the decrypted gradient results, and each party gets the final global model; When each party gets the final global model and obtains incremental data, the incremental data is trained on the basis of the already constructed global model, and in the process of personalized training, the model is adjusted and optimized according to different data characteristics and business requirements, and finally a personalized local model that meets the specific requirements of each party is obtained.

2. The incremental data-oriented personalized federated learning method of claim 1, wherein, Each client receives the global model parameters, and uses its local stock data set to train the model, comprising: Optimizing a loss function including a weight decay regularizer by each client defined as the sum of the average loss over the local dataset and a weight decay regularizer on the model parameters; The loss function is in the form of: ； wherein, is the prediction error, is a weight coefficient of the weight decay regular term; for updating the model parameters.

3. The method of federated learning for incremental data personalization of claim 2, wherein, In the process of model training using its local inventory dataset, each client performs multiple iterations by gradient descent method, updating model parameters in each iteration ; The specific update steps are: ； wherein, is the learning rate, after multiple iterations, the client obtains the updated local model parameters .

4. The incremental data-oriented personalized federated learning method of claim 3, wherein, The training process is repeated continuously until a predetermined number of iterations is reached or a premature stop condition is met.

5. The incremental data-oriented personalized federated learning method of claim 4, wherein, The incremental data set contains the latest user interaction data, which can reflect the latest preferences and behaviors of the current user.

6. The incremental data-oriented personalized federated learning method of claim 1, wherein, During the local calculation in multiple clients to obtain corresponding model parameters, the weight decay regularization training method is used for training, which can effectively constrain the complexity of the model and prevent overfitting, thereby improving the generalization ability of the model.

7. The incremental data-oriented personalized federated learning method of claim 1, wherein, In the process of personalized training, the final global model will be adjusted and optimized according to different data characteristics and business requirements to obtain a personalized local model that meets the specific requirements of each party.

8. An apparatus for personalized federated learning oriented to incremental data, comprising: Comprise: The inner loop training module is used to train the global model based on the inner loop and optimize the global model by adding a weight decay regularization term, including: initializing the global model parameters through the server. And send the global model parameters to all clients. Each client Upon receiving the current global model parameters Then, use its local stored dataset. Model training; Client The updated local model parameters are uploaded to the server; after receiving the model parameters from all clients, the server aggregates them using a weighted average or simple average method to obtain the updated global model parameters. The server will assign new global model parameters. Distribute the data to each client to begin the next round of training; The outer loop training module is used for personalized model optimization training based on the outer loop incremental training data, and includes: each client updates the parameters of the local model to the global model parameters after receiving the global model parameters; the client trains using the local incremental data set; the client needs to select a smaller learning rate to ensure that the fine-tuning process is stable and does not damage the pre-training weights, and the client performs a smaller number of training to prevent overfitting; each client fine-tunes the model according to the local incremental data set, adjusts the model weights using the standard back propagation algorithm; after the model training is completed, the client evaluates the performance of the fine-tuned personalized model using the local validation set, and according to the evaluation result, the client can decide whether further cycle iteration is needed to adjust the model parameters; the incremental data-oriented personalized federated learning method is applied to the insurance data center, including: the insurance data center performs local calculation in the respective data center to calculate the corresponding model parameters; after the local calculation is completed, the insurance data center sends the calculation results processed by the encryption technology to a special aggregation server, the aggregation server processes the results from different data centers using the federated learning dynamic weight aggregation method to generate a global model, and sends the global model back to each participant; each participant encrypts the received gradient after receiving the global model to ensure the security and privacy of the data; the gradient results after decryption are used to update the model parameters in the respective data center, and each party obtains the final global model; when each party obtains the final global model and obtains the incremental data, the incremental data is trained on the basis of the already constructed global model, and in the process of personalized training, the model is adjusted and optimized according to different data characteristics and business requirements, and finally a personalized local model that meets the specific requirements of each party is obtained.

Images