Federated forgetting learning method and apparatus

By identifying key layers in federated learning and embedding lightweight adaptation modules for feature space optimization, the high cost and irreversibility issues of federated forgetting methods are resolved, achieving low-cost, reversible forgetting and recovery, which is applicable to various federated learning scenarios.

CN122242649APending Publication Date: 2026-06-19INSTITUTE OF INFORMATION ENGINEERING CHINESE ACADEMY OF SCIENCES
0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
INSTITUTE OF INFORMATION ENGINEERING CHINESE ACADEMY OF SCIENCES
Filing Date
2026-03-05
Publication Date
2026-06-19

AI Technical Summary

⚠Technical Problem

Existing federated forgetting methods are computationally and time-intensive in federated scenarios, easily leading to catastrophic forgetting, high communication overhead, and irreversible operations.

⚗Method used

By determining key layers based on the parameters of the pre-trained global model, freezing non-key layers, implanting a lightweight adaptation module, using the local dataset for feature space-oriented optimization and updating, and restoring the model state when needed, a sparse representation adapter and nested loop training mechanism are adopted, and a federated contrastive representation alignment strategy is combined to construct the total loss function for optimization.

🎯Benefits of technology

It achieves precise forgetting, reduces computation and communication costs, ensures stable model performance on retained data after forgetting, and supports fast and lossless recovery, making it suitable for various federated learning scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
Patent Text Reader

Abstract

This invention provides a federated forgetting learning method and apparatus, comprising: determining a key layer based on the parameters of a pre-trained global model and their sensitivity to target forgotten data; freezing the parameters of the global model excluding the key layer as frozen parameters; embedding a lightweight adaptation module in the key layer; distributing the model containing the frozen parameters and the lightweight adaptation module to each participating client; controlling each participating client to perform feature space-oriented optimization and update of the lightweight adaptation module using the locally retained dataset and the forgotten dataset; aggregating the updated lightweight adaptation module parameters from each participating client to generate a forgotten global model; and, in response to a knowledge recovery request, restoring the forgotten global model to its pre-trained global model state by removing the lightweight adaptation module from the forgotten global model, effectively solving problems such as catastrophic forgetting, high communication overhead, and irreversible operation.
Need to check novelty before this filing date? Find Prior Art