Full weight aggregation method based on proxy model in heterogeneous federated learning

Abstract

The invention provides a full weight aggregation method based on an agent model in heterogeneous federated learning, and belongs to the field of artificial intelligence and machine learning. The method comprises the steps of system setting and proxy model construction, local model training and model uploading, transfer of weight knowledge to a proxy model, aggregation of loyalty guidance and global model updating and distribution. According to the method, the learning effect of the equipment model and the overall operation efficiency of the system are optimized, the model training efficiency on heterogeneous equipment can be remarkably improved, the energy consumption is reduced, the accuracy and generalization ability of the model are improved, and the method has wide application prospects.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and machine learning, and relates to a full-weight aggregation technology based on a proxy model, and specifically to a full-weight aggregation method based on a proxy model in heterogeneous federated learning. Background Art

[0002] In modern mobile and Internet of Things (IoT) environments, due to the diversity of device types, Federated Learning (FL) technology has become an important method to support collaborative training models between devices. Traditional federated learning methods usually require all mobile devices to have the same model architecture. However, in reality, devices such as smartphones, smart watches, and home automation systems have different processing and storage capabilities, resulting in the need for models to be structurally adjusted to adapt to various devices. For example, high-end devices may use complex deep learning models for image processing or speech recognition, while low-end devices require more simplified models to ensure operational efficiency. This high degree of heterogeneity of devices requires the development of a new federated learning framework that can effectively manage and optimize models of various different architectures. In the existing technology, pruning-based methods and distillation-based methods are mainly used to deal with the heterogeneity of models.

[0003] For example, pruning-based techniques usually involve initializing a standard model and then adapting it to devices of different computing power by pruning non-critical weights. These methods include but are not limited to techniques demonstrated by Hermes and LotteryFL. Although these techniques alleviate the deployment problem of the model to some extent, they usually lead to severe degradation of model performance in highly heterogeneous environments.

[0004] On the other hand, although distillation-based techniques such as FedMD and FedGEMS are able to handle models of different architectures, these methods usually ignore the knowledge of the intermediate layers of the model, resulting in prolonged training convergence time and increased computational burden when the computing power of the device is limited.

[0005] These existing technologies not only have low training efficiency, but also face the problems of reduced accuracy and increased energy consumption when applied on heterogeneous devices. This is mainly because the existing heterogeneous model federated learning technologies can only aggregate part of the model's weights, thus losing the knowledge learned by the model. Therefore, there is an urgent need for a full weight aggregation method that can fully utilize heterogeneous models, which can not only improve the model's training efficiency, but also significantly improve the model's performance and energy efficiency on various devices. Summary of the invention

[0006] In response to the above problems existing in the prior art, the present invention provides a full-weight aggregation method based on an agent model in heterogeneous federated learning, which optimizes the learning effect of the device model and the overall operating efficiency of the system by intelligently aggregating the complete weights of all device models in the system.

[0007] The technical solution of the present invention is as follows:

[0008] A full-weight aggregation method based on an agent model in heterogeneous federated learning includes the following steps:

[0009] S1. System setup and agent model construction

[0010] In the system initialization phase, the corresponding models, training data sets, and test data sets are deployed on different types of devices. A series of proxy models are built in the system server, one for each device type. These proxy models have a unified architecture and can carry key information of different device models while maintaining computational efficiency.

[0011] S2. Local model training and model upload

[0012] Each device uses the training data set locally on the device to independently train its device model according to its function and computing power, and uploads the model weights to the system server after training is completed;

[0013] S3. Transfer of weight knowledge to the proxy model

[0014] After receiving the device model weights uploaded by each device, the system server first pre-aggregates the device models with the same architecture, and then uses meta-learning technology to transfer the knowledge of the pre-aggregated model to the corresponding proxy model to ensure effective integration of knowledge;

[0015] S4. Loyalty-guided aggregation

[0016] After completing the knowledge transfer of the agent models, the system server guides the aggregation process according to the loyalty between each agent model and the pre-aggregated model, thereby aggregating the agent models into a global model;

[0017] S5. Global model update and distribution

[0018] After the aggregation is completed, the system server performs a second knowledge transfer to transfer the knowledge of the global model to the pre-aggregated model and sends the updated pre-aggregated model back to each device;

[0019] The device repeats steps S2 to S5 for multiple rounds of training and aggregation until the device model of each device converges and achieves the expected performance and accuracy.

[0020] Furthermore, the specific steps of the proxy model building process in step S1 include:

[0021] S1-1. The system server selects the device type with the largest model capacity as the infrastructure of the proxy model according to the device type, ensuring that the proxy model can contain the knowledge of all device models;

[0022] S1-2. Use public datasets to pre-train the proxy model to ensure that it has sufficient generalization ability.

[0023] Furthermore, the specific steps of the local model training and uploading process in step S2 include:

[0024] S2-1. Each device performs model training on its local training data set to optimize device model performance;

[0025] S2-2. After training is completed, each device uploads its own model weights to the system server.

[0026] Furthermore, the specific steps of the process of transferring the weight knowledge to the proxy model in step S3 include:

[0027] S3-1. After the system server receives the device model weights uploaded by each device, it pre-aggregates the device models of the same architecture to generate a pre-aggregated model;

[0028] S3-2. Using meta-learning technology, the system server establishes a weight mapping relationship from the pre-aggregation model to the proxy model to achieve accurate knowledge transfer. Meta-learning technology is used to learn the mapping relationship between different layers and neurons to ensure the accuracy and efficiency of knowledge transfer;

[0029] S3-3. By optimizing the mapping relationship and reducing information loss, the proxy model and the pre-aggregation model can maintain the same accuracy.

[0030] Furthermore, the specific steps of the loyalty-guided aggregation process in step S4 include:

[0031] S4-1. Calculate the cosine similarity between the proxy model output and the original model output as the loyalty score;

[0032] S4-2. According to the loyalty score, the weights of different proxy models in the aggregation process are adjusted. Models with high loyalty have a larger weight in the aggregation process, thereby optimizing the performance of the global model.

[0033] Furthermore, the specific steps of the global model updating and distribution process in step S5 include:

[0034] S5-1. The system server uses a public dataset and meta-learning technology to transfer the knowledge of the global model to the pre-aggregation model, ensuring that the pre-aggregation model can reflect the update of global knowledge;

[0035] S5-2. The updated pre-aggregation model is sent back to each participating device. After the device receives the weight, it is used as the updated device model weight and the next round of device model training is carried out;

[0036] S5-3. Update the weights of the proxy model to the weights of the global model.

[0037] The present invention provides a full-weight aggregation method based on an agent model in heterogeneous federated learning. By intelligently aggregating the complete weights of all device models in the system, the learning effect of the device model and the overall operating efficiency of the system are optimized, and the model training efficiency on heterogeneous devices can be significantly improved, energy consumption can be reduced, and the accuracy and generalization ability of the model can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is the execution process of the full-weight aggregation method based on the proxy model in the heterogeneous federated learning of the present invention. DETAILED DESCRIPTION

[0039] In order to better understand the present invention, a specific embodiment is provided below to describe the application of the full-weight aggregation method based on the agent model in a heterogeneous federated learning environment.

[0040] Taking the image classification task as an example, this embodiment describes in detail how to implement the present invention in a heterogeneous environment consisting of a smartphone, a smart watch, and an IoT device. These devices have different processing capabilities and storage capacities and are suitable for heterogeneous models of different sizes and structures.

[0041] The present invention provides a full-weight aggregation method based on a proxy model in heterogeneous federated learning, such as Figure 1 As shown:

[0042] S1. System setup and agent model construction

[0043] In this embodiment, the system is initially set up to include 10 smartphones as high-end devices, 10 smart watches as mid-end devices, and 10 IoT devices as low-end devices. During the system initialization phase, the ResNet-34 model is deployed in the smartphone as the device model, and 10 equally divided Office-HomeArt training data sets and Office-HomeArt test data sets are deployed. The smart watch deploys the MobileNet-V2 model as the device model, and 10 equally divided Office-Home Clipart training data sets and Office-Home Clipart test data sets are deployed. The IoT device deploys the LeNet-5 model as the device model, and 10 equally divided Office-Home Product training data sets and Office-Home Product test data sets are deployed. Three ResNet-34 proxy models are built in the system server, one for each device, so as to carry the key information of different device models while maintaining computational efficiency;

[0044] The specific steps of the proxy model construction process include:

[0045] S1-1. The system server selects the device type with the largest model capacity as the basic architecture of the proxy model based on the device type. In order to ensure that the proxy model can carry the knowledge of all device models, considering the complexity of the model, the ResNet-34 model deployed in the most powerful device (i.e., smartphone) is selected as the benchmark of the proxy model architecture.

[0046] S1-2. Use a public dataset (in this embodiment, the Office-Home Realworld dataset is used) to pre-train the agent model to ensure that it has sufficient generalization ability.

[0047] S2. Local model training and model upload

[0048] Each device uses its local training data set to independently train its device model based on its function and computing power. After training, the model weights are uploaded to the system server.

[0049] The specific steps of the local model training and uploading process include:

[0050] S2-1. Each device performs model training on its local training data set to optimize device model performance.

[0051] S2-2. After training is completed, each device uploads its own device model weight to the system server.

[0052] S3. Transfer of weight knowledge to the proxy model

[0053] After the system server receives the device model weights uploaded by each device, it first pre-aggregates the device models of the same architecture, that is, pre-aggregates the device models of 10 smartphones, pre-aggregates the device models of 10 smart watches, and pre-aggregates the device models of 10 IoT devices. Then, meta-learning technology is used to transfer the knowledge of the pre-aggregated model to the corresponding proxy model to ensure effective integration of knowledge. The specific steps of the weight knowledge transfer process include:

[0054] S3-1. After receiving the device model weights uploaded by each device, the system server pre-aggregates the device models with the same architecture to generate a pre-aggregated model.

[0055] S3-2. Using meta-learning technology, the system server establishes a weight mapping relationship from the pre-aggregation model to the proxy model to achieve accurate knowledge transfer. Meta-learning technology is used to learn the mapping relationship between different layers and neurons to ensure the accuracy and efficiency of knowledge transfer.

[0056] S3-3. By optimizing the mapping relationship and reducing information loss, the proxy model and the pre-aggregation model can maintain the same accuracy.

[0057] S4. Loyalty-guided aggregation

[0058] After completing the knowledge transfer of the proxy models, the system server guides the aggregation process based on the loyalty between each proxy model and the pre-aggregated model (obtained by calculating the cosine similarity between the outputs), thereby aggregating the proxy models into a global model. The specific steps of the loyalty-guided aggregation process include:

[0059] S4-1. Calculate the cosine similarity between the proxy model output and the pre-aggregated model output as the loyalty score.

[0060] S4-2. According to the loyalty score, adjust the weights of different proxy models in the aggregation process. Proxy models with high loyalty scores have greater weights in the aggregation process, thereby optimizing the performance of the global model.

[0061] S5. Global model update and distribution

[0062] After the aggregation is completed, the system server performs the second knowledge transfer, transfers the knowledge of the global model to the pre-aggregated model, and sends the updated pre-aggregated model back to each device. The specific steps of the global model update and distribution process include:

[0063] S5-1. The server uses a public dataset and meta-learning technology to transfer the knowledge of the global model to the pre-aggregate model, ensuring that the pre-aggregate model can reflect the updates of the global model knowledge.

[0064] S5-2. The updated pre-aggregated model is sent back to each participating device. After receiving the weight, the device uses it as the updated device model weight and performs the next round of device model training.

[0065] S5-3. Update the weights of the proxy model to the weights of the global model.

[0066] The device repeats steps S2 to S5 for multiple rounds of training and aggregation until the accuracy of the device model of each device in the test data set deployed locally on the device tends to be stable within 10 rounds without increasing. The training of each device model is considered to have converged and the training is terminated.

[0067] Through the above specific implementation, the present invention can significantly improve the model training efficiency on heterogeneous devices, reduce energy consumption, and improve the accuracy and generalization ability of the model. Each functional module of the present invention can be implemented by software, hardware or a combination thereof to adapt to different heterogeneous federated learning application scenarios.

[0068] Finally, it should be noted that the purpose of publishing the embodiments is to help further understand the present invention, but those skilled in the art can understand that various substitutions and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments, and the scope of protection claimed by the present invention shall be subject to the scope defined in the claims.

Claims

1. A full-weight aggregation method based on an agent model in heterogeneous federated learning, characterized in that: The following steps are involved: S1. System setup and agent model construction In the system initialization phase, the corresponding models, training data sets, and test data sets are deployed on different types of devices. A series of proxy models are built in the system server, one for each device type. These proxy models have a unified architecture and can carry key information of different device models while maintaining computational efficiency. S2. Local model training and model upload Each device uses the training data set locally on the device to independently train its device model according to its function and computing power, and uploads the model weights to the system server after training is completed; S3. Transfer of weight knowledge to the proxy model After receiving the device model weights uploaded by each device, the system server first pre-aggregates the device models with the same architecture, and then uses meta-learning technology to transfer the knowledge of the pre-aggregated model to the corresponding proxy model to ensure effective integration of knowledge; S4. Loyalty-guided aggregation After completing the knowledge transfer of the agent models, the system server guides the aggregation process according to the loyalty between each agent model and the pre-aggregated model, thereby aggregating the agent models into a global model; S5. Global model update and distribution After the aggregation is completed, the system server performs a second knowledge transfer to transfer the knowledge of the global model to the pre-aggregated model and sends the updated pre-aggregated model back to each device; The device repeats steps S2 to S5 for multiple rounds of training and aggregation until the device model of each device converges and achieves the expected performance and accuracy.

2. The method according to claim 1, characterized in that The specific steps of the agent model building process in step S1 include: S1-1. The system server selects the device type with the largest model capacity as the infrastructure of the proxy model according to the device type, ensuring that the proxy model can contain the knowledge of all device models; S1-2. Use public datasets to pre-train the proxy model to ensure that it has sufficient generalization ability.

3. The method according to claim 1, characterized in that The specific steps of the local model training and uploading process in step S2 include: S2-1. Each device performs model training on its local training data set to optimize device model performance; S2-2. After training is completed, each device uploads its own model weights to the system server.

4. The method according to claim 1, characterized in that The specific steps of the process of transferring the weight knowledge to the proxy model in step S3 include: S3-1. After the system server receives the device model weights uploaded by each device, it pre-aggregates the device models of the same architecture to generate a pre-aggregated model; S3-2. Using meta-learning technology, the system server establishes a weight mapping relationship from the pre-aggregation model to the proxy model to achieve accurate knowledge transfer. Meta-learning technology is used to learn the mapping relationship between different layers and neurons to ensure the accuracy and efficiency of knowledge transfer; S3-3. By optimizing the mapping relationship and reducing information loss, the proxy model and the pre-aggregation model can maintain the same accuracy.

5. The method according to claim 1, characterized in that The specific steps of the loyalty-guided aggregation process in step S4 include: S4-1. Calculate the cosine similarity between the proxy model output and the original model output as the loyalty score; S4-2. According to the loyalty score, the weights of different proxy models in the aggregation process are adjusted. Models with high loyalty have a larger weight in the aggregation process, thereby optimizing the performance of the global model.

6. The method according to claim 1, characterized in that The specific steps of the global model update and distribution process in step S5 include: S5-1. The system server uses a public dataset and meta-learning technology to transfer the knowledge of the global model to the pre-aggregation model, ensuring that the pre-aggregation model can reflect the update of global knowledge; S5-2. The updated pre-aggregation model is sent back to each participating device. After the device receives the weight, it is used as the updated device model weight and the next round of device model training is carried out; S5-3. Update the weights of the proxy model to the weights of the global model.

Images