FedART: Enhancing Replay in Federated Incremental Learning

Zijiang Tan, Haodi Wang, Libin Jiao, Rongfang Bie

Abstract: Federated Class-Incremental Learning (FCIL) enables distributed models to continuously learn new categories while preserving privacy, which suffers from the problem of catastrophic forgetting. To address this issue, generative replay has emerged as a mainstream solution, yet its performance is hampered by two fundamental bottlenecks: (1) low-fidelity synthesis, where generated visual samples fail to effectively represent historical knowledge, and (2) class imbalance in FCIL, which undermines fair learning across classes. In this paper, we propose a novel generative replay framework called FedART (Federated Adaptive Replay with Text-anchors). To combat low-fidelity synthesis, FedART employs a text-anchored initialization strategy. Instead of optimizing from a random start, this approach provides strong semantic priors to guide the generation process. To tackle class imbalance, we design a dual adaptive aggregation mechanism. This mechanism applies tailored weighting strategies at both the generator and classifier levels, leveraging local training dynamics to ensure both the quality of generative knowledge and the fairness of classifier aggregation. Extensive experiments on CIFAR-100 and Tiny-ImageNet demonstrate that FedART significantly outperforms state-of-the-art methods, achieving an accuracy of up to 43.62% and establishing a new and robust benchmark for enhancing the effectiveness of generative replay in FCIL.

External IDs:dblp:conf/mmasia/TanWJB25