Continual Federated Learning with Aggregated Gradients

The ultimate goal of machine learning models is to emulate the human ability to master diverse tasks over time without forgetting previous knowledge. Continual Federated Learning (CFL) improves federated learning systems by allowing them to learn new tasks while retaining old ones, thus enhancing efficiency, privacy, and scalability. Our work introduces a novel aggregation strategy for memory-based CFL and provides a detailed convergence analysis. We address important challenges like global catastrophic forgetting, client drift, and bias, proving that our proposed algorithm converges at a rate of \(\mathcal {O}(1/\sqrt {T})\) and performs better than existing methods in accuracy and mitigating forgetting.