Adaptive Local Steps and Server-Side Optimization for Asynchronous Federated Continual Learning

Asynchronous Federated Continual Learning (AFCL) represents a realistic yet chal- lenging scenario where heterogeneous clients learn distinct tasks over time without central coordination. While prototype-based methods like FedSpace effectively address catastrophic forgetting, they often suffer from high computational costs and performance degradation in non-IID settings. This thesis proposes a framework designed to reconcile efficiency and robustness. First, we introduce FedAlt (Fed- erated Adaptive Local Training), a strategy that accelerates convergence by dynam- ically adjusting local training steps, significantly reducing communication overhead. To recover the accuracy loss induced by accelerated training, we propose FedSpo (Federated Server-Side Prototype Optimization), a privacy-preserving module where the server refines the global model using synthetic data generated from global pro- totypes. We investigate advanced tuning strategies, including Quality Filtering for medium fragmentation and Soft Mixup for extreme fragmentation. Experiments on CIFAR-100 (with 50, 100, and 500 clients) demonstrate that our approach estab- lishes a highly effective trade-off between computational efficiency and classification accuracy compared to state of the art baselines