Federal learning privacy protection method and device oriented to device isomerism, and storage medium
By setting training delay and data similarity constraints in federated learning, and using multi-arm robbers and Pareto algorithms for server clustering and model aggregation, the problem of model training delay caused by device heterogeneity is solved, and training efficiency and communication efficiency are improved.
CN119646884AActive Publication Date: 2025-03-18HUNAN TIAN HE GUO YUN TECH CO LTD
12 Cites 4 Cited by
Patent Information
- Application Number
- CN202510163949.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-14
- Publication Date
- 2025-03-18
- Estimated Expiration
- 2045-02-14
AI Technical Summary
Technical Problem
The existing federated learning privacy protection method results in a longer training delay and low training efficiency when the device is heterogeneous.
Method used
By setting training delay constraints and data similarity constraints, the multi-arm robber algorithm is used to predict the computing delay of the server, cluster the server based on the Pareto algorithm, and select cluster head servers for model aggregation and communication, reducing direct communication between the server and the central server.
Benefits of technology
It effectively improves the efficiency of model training in federated learning and the communication efficiency of model parameter exchange, and reduces the communication consumption of federated learning.
✦ Generated by Eureka AI based on patent content.
Abstract
The invention is suitable for the technical field of information security, and provides a federated learning privacy protection method and device oriented to device isomerism and a storage medium, the method comprises the following steps: setting a training time delay constraint and a data similarity constraint; on the basis of a multi-arm theft algorithm, predicting the training time by using the historical training time of each server on the local model to obtain the calculation time delay of each server; based on a Pareto algorithm, clustering each server by using training time delay constraint, data similarity constraint and calculation time delay of each server to obtain at least one cluster, each cluster comprising at least two servers; and determining a cluster head server in each cluster based on the training time delay of each server so as to aggregate local models of the servers in each cluster through each cluster head server, and communicating with the central server, the training time delay comprising a calculation time delay and a communication time delay. According to the technical scheme provided by the invention, the communication consumption of federated learning is reduced, and the model training efficiency in federated learning is improved.
Need to check novelty before this filing date? Find Prior Art