This paper introduces HePCo, a novel approach for Continual Federated Learning (CFL) that addresses the problem of catastrophic forgetting and data heterogeneity across clients without requiring access to any stored data. HePCo leverages a prompting-based approach and proposes a lightweight generation and distillation scheme to consolidate client models at the server. The method is formulated for image classification and tested on CIFAR-100, ImageNet-R, and DomainNet datasets. HePCo outperforms both existing methods and the authors’ own baselines by as much as 7% while significantly reducing communication and client-level computation costs.
Publication date: June 16, 2023
Project Page: N/A
Paper: https://arxiv.org/pdf/2306.09970.pdf
