Federated Class-Incremental Learning: A Survey

Federated learning has recently gained significant attention due to its unique distributed training approach and secure aggregation mechanisms. However, clients in federated learning often have datasets that vary in characteristics, computational resources, and model architectures, leading to data heterogeneity, where the data are non-IID (identically and independently distributed). Additionally, in practical scenarios, local model training frequently encounters new data, leading to catastrophic forgetting of previously learned knowledge. The challenges of data heterogeneity and catastrophic forgetting require effective integration of federated learning with incremental learning to ensure the sustainable development of federated ecosystems. We provides a comprehensive review and analysis of federated class-incremental learning. We take a novel perspective categorizing these methods into two main approaches: rehearsal-based and rehearsal-free, and provide a detailed comparison of the techniques within each category. Finally, we outline future research directions for federated class-incremental learning.