Heterogeneity federated learning-oriented local model adaptive aggregation method and system

By using a central server to determine the model in federated learning and determining the enantioscopic information and client to determine the local model, and perform weight combination and adaptive initialization processing, the problem of poor federated learning performance in heterogeneous environments is solved, and the model accuracy and reliability are improved.

CN119940569APending Publication Date: 2025-05-06BEIHANG UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411847840.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-12-16
Publication Date
2025-05-06

AI Technical Summary

Technical Problem

The existing federated learning methods do not perform well in heterogeneous environments and are difficult to meet the heterogeneous differentiation needs of different clients, resulting in a decrease in overall model accuracy and operating reliability.

Method used

The central server determines the model distribution enantiosity information based on the global model training feature information, and the client determines the final local model based on the local training log and uploads it to aggregate it into a global model. Then, select the global model and the local model to combine according to the corresponding weights and perform adaptive initialization processing to match the client's needs.

Benefits of technology

It improves the overall accuracy and reliability of the model, adapts to the heterogeneous needs of different clients, and enhances the performance and matching applicability of federated learning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119940569A_ABST
    Figure CN119940569A_ABST
Patent Text Reader

Abstract

The invention provides a local model adaptive aggregation method and system oriented to heterogeneity federated learning, and the method comprises the steps: determining model distribution mapping information between a global model completing each round of training and all clients connected with a central server based on global model training feature information in a training process of the global model by the central server; a reliable basis is provided for distributing proper global models to different clients; the local training result feature information is obtained on the basis of a log obtained by carrying out local training on the distributed global model by the client, so that the local model obtained by final training of the client is determined, and heterogeneity training on the global model is realized; the local models of all the clients are uploaded and aggregated into a global model, then the global model and the corresponding local models are selected to be combined according to corresponding weights to obtain an aggregation model, self-adaptive initialization processing is carried out, and initialized local models matched with the clients are obtained; the trained model meets the requirements of different clients in heterogeneity federation learning, the overall precision and reliability of the model are improved, and heterogeneity requirements of different clients are met.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Cited By

  • Model training method, system, equipment and medium

    CN120781110A