Federated learning classification model training method based on adaptive model perturbation

Through the combination of adaptive allocation of privacy budgets and differential privacy technology, the degree of perturbation of the local model is dynamically adjusted, and the problem of insufficient privacy protection capabilities and classification accuracy in federated learning is solved, achieving more efficient privacy protection and classification effects.

CN117056785BActive Publication Date: 2025-09-05XIDIAN UNIV
3 Cites 0 Cited by

Patent Information

Application Number
CN202311113980.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-08-31
Publication Date
2025-09-05
Estimated Expiration
2043-08-31

AI Technical Summary

Technical Problem

Among the existing federated learning methods, the privacy protection capability is weak and the classification accuracy is low, especially the problem of imbalance in the allocation of privacy budgets caused by local data heterogeneity, which affects the model's privacy protection capability and classification effect.

Method used

The federated learning method of adaptive model perturbation is adopted, and the privacy budget is adaptively allocated to each client through a reinforcement learning algorithm. The adaptive perturbation is added on the local model weight parameters in combination with differential privacy technology, and the degree of perturbation is dynamically adjusted according to the contribution of the local model to the global model.

Benefits of technology

The privacy protection capabilities and model classification effects of federated learning are improved, and the privacy protection capabilities or classification effects are reduced due to excessive or small privacy budgets are avoided, thus achieving more efficient privacy protection and classification accuracy.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

This invention proposes a federated learning classification model training method based on adaptive model perturbation. The implementation steps are as follows: constructing a federated learning system; the client obtains a training sample set; the server obtains a global test sample set and initializes the global classification model and reinforcement learning model; the server adaptively allocates a privacy budget to the client; the client iteratively trains the local classification model; the client adaptively perturbs the local classification model; the server aggregates the weight parameters of the local classification model; and the server obtains the training results of the federated learning classification model. The server adopts a reinforcement learning algorithm to adaptively allocate a privacy budget to the local classification model based on stored local classification model performance, privacy budget, and reward information. This method controls the noise added to the local classification model, avoids the impact of excessive or insufficient perturbation on the model, and thus improves the privacy protection capability and performance of federated learning.
Need to check novelty before this filing date? Find Prior Art