A Data-Free Approach to Mitigate Catastrophic Forgetting in Federated Class Incremental Learning for Vision Tasks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Deep learning models face catastrophic forgetting when learning new data.
• Federated learning (FL) exacerbates this issue due to distributed and changing data.
• Existing solutions for centralized settings are not directly applicable to FL's privacy and resource constraints.

Purpose of the Study:

• To present a novel framework for federated class incremental learning (FCIL).
• To mitigate catastrophic forgetting in FL without compromising data privacy or user flexibility.
• To introduce a new dataset, SuperImageNet, for evaluating FCIL.

Main Methods:

• A generative model synthesizes past data distributions to combat forgetting.
• The generative model is trained server-side using data-free methods, preserving client privacy.
• The framework supports dynamic client participation (join/leave) without requiring clients to store old data or models.

Main Results:

• The proposed framework significantly reduces catastrophic forgetting in federated class incremental learning.
• Experimental results demonstrate substantial improvements over existing baselines on multiple datasets.
• The SuperImageNet dataset provides a tailored benchmark for FCIL.

Conclusions:

• The developed framework effectively addresses catastrophic forgetting in federated class incremental learning.
• The data-free, server-side training of the generative model ensures privacy and efficiency.
• The framework offers a flexible and robust solution for continual learning in federated environments.