Answers are grounded in this paper and cite their sources.
Conventional facial expression recognition (FER) methods often train the model on large-scale centralized expression data. However, expression data are often distributed across heterogeneous devices and incrementally collected in practice. In this paper, we study a little-studied yet important task named federated class-incremental FER (FCI-FER), which aims to incrementally identify unseen compound expressions with the model initially trained on easily accessible basic expressions under the federated learning settings. To address the unique problems of catastrophic forgetting and heterogeneous disturbances across clients in FCI-FER, we propose a disentanglement-guided multi-teacher relational distillation (DMT-RD) method which follows a novel "distillation and disentanglement" paradigm. For local training, we design a multi-teacher hypergraph relational distillation (MTD) module to capture semantic relations between expressions and learn expression features. Meanwhile, we develop a disturbance encoding (DE) module to learn client-specific disturbance information. By minimizing the mutual information (MI) loss between the outputs of the MTD and DE modules, we can achieve feature disentanglement and reduce the negative influence of specific disturbances on local training. For global training, we introduce a multi-teacher learning strategy by clustering historical clients to comprehensively capture client diversity across communication rounds. Our method can effectively preserve old knowledge while learning new compound expressions under federated settings. Extensive experiments demonstrate the effectiveness of our method against several state-of-the-art federated class-incremental learning methods on benchmark datasets.
No takes yet. Share an insight, caveat, or question.
Jiang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: