A federated learning adaptive pruning regulation method based on model recovery evaluation

By designing hierarchical models and using dynamic pruning techniques, combined with model restorative evaluation and knowledge distillation, the problems of high communication overhead and insufficient model generalization ability in federated learning are solved, thus realizing an efficient, safe, and adaptive federated learning system.

CN122242576APending Publication Date: 2026-06-19CHONGQING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
CHONGQING UNIV OF POSTS & TELECOMM
Filing Date
2026-03-12
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Federated learning suffers from high communication overhead, insufficient model generalization ability, and static pruning strategies that cannot adapt to real-time performance changes on the client side, affecting the aggregation efficiency and robustness of the global model.

Method used

A hierarchical model design is adopted, which combines dynamic pruning and sparsity transport. Based on the selective knowledge distillation mechanism of model restorability assessment, the pruning rate and aggregation weight are dynamically adjusted. Through knowledge distillation, the performance of the damaged model is restored, and adaptive pruning regulation is achieved.

Benefits of technology

It significantly reduces communication costs, improves model robustness and generalization ability, adapts to dynamic changes in different heterogeneous environments, and realizes an efficient, secure, and adaptive federated learning system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122242576A_ABST
    Figure CN122242576A_ABST
Patent Text Reader

Abstract

This invention relates to the field of federated learning technology, specifically to a federated learning adaptive pruning control method based on model restorative evaluation. The method includes: Step 1: The server initializes a global common layer model and distributes it to each client; the client locally initializes a personalized layer model, forming an initial complete model for the client. This federated learning adaptive pruning control method based on model restorative evaluation, through hierarchical model design, divides the model into a shared common layer and a local personalized layer. Only the common layer is transmitted and pruned, significantly reducing the amount of communication data while ensuring that the client can fully adapt to the local data distribution through the local personalized layer, meeting personalized needs. The combination of dynamic pruning and sparse transmission further compresses the transmission volume of the common layer model, significantly reducing communication overhead. It is suitable for resource-constrained edge device scenarios, effectively alleviating the contradiction between model compression and performance preservation, and improving model robustness.
Need to check novelty before this filing date? Find Prior Art