Federated Class-Incremental Learning With Dynamic Feature Extractor Fusion

Federated class-incremental learning (FCIL) allows multiple clients in a distributed environment to learn models collaboratively from evolving data streams, where new classes arrive continually at each client. Some existing works in FCIL combine traditional federated learning methods with class-incremental methods. However, the global model affected by data heterogeneity can aggravate local forgetting through the direct combination of traditional methods. To tackle this issue, we propose FCIDF, a novel Federated Class-Incremental learning approach based on Dynamic feature extractor Fusion. FCIDF learns personalized and incremental models for each client by introducing personalized fusion rates to integrate global knowledge into local features. Leveraging meta-learning during each incremental round, FCIDF ensures involvement of both old and new task knowledge in personalized training. Besides, we further propose a new Storing strategy based on Accumulated Global Feature Means (AGFMS), which helps the model review unbiased old knowledge and compensates for local forgetting. Experiment results show that FCIDF outperforms the baseline methods in both accuracy and forgetting on most settings, and AGFMS improves the performance of FCIDF on most evaluated scales.