DLA-FCIL: Federated Class-Incremental Learning for Dynamic Data with Forgetting Compensation and Auxiliary Generators

Federated Class-Incremental Learning (FCIL) enables dynamic model updates, but suffers from local and global catastrophic forgetting due to limited client storage and cross-client class non-i.i.d. issues. To address catastrophic forgetting, we propose a multi-scale FCIL scheme, DLA-FCIL, which incorporates a Double-Loss-assisted forgetting compensation mechanism and the Auxiliary generators based on data characteristics. Specifically, to mitigate local catastrophic forgetting, we incorporate an auxiliary generator on local clients for knowledge replay, augmenting the training datasets with generated samples to form hybrid datasets. Then, to fully leverage the hybrid datasets, we design a double-loss forgetting compensation mechanism. This mechanism includes a gradient-weighted compensation loss that normalizes forgetting rates across old class knowledge, and a semantic-transition compensation loss that extracts the semantic relationships between old and new classes, preventing abrupt shifts in semantic consistency during the class transitions. Besides, to effectively alleviate the catastrophic forgetting problem caused by global class imbalance, the trained auxiliary generator is sent to a proxy server with minimal communication cost to build an i.i.d. dataset, enabling the development of an optimal global model. Finally, comparison experiments on CIFAR100, ImageNet-Subset, and Tiny-ImageNet datasets demonstrate that DLA-FCIL consistently outperforms other FCIL baselines by approximately 3–15% in test accuracy.