Federated Class-Incremental Learning with New-Class Augmented Self-Distillation

Federated Learning (FL) enables collaborative model training among participants while guaranteeing the privacy of raw data. Mainstream FL methodologies overlook the dynamic nature of real-world data, particularly its tendency to grow in volume and diversify in classes over time. This oversight results in FL methods suffering from catastrophic forgetting, where models inadvertently discard previously learned information upon assimilating new data. In response to this challenge, we propose a novel Federated Class-Incremental Learning (FCIL) method, named FCIL with New-Class Augmented Self-Distillation (FedNASD). FedNASD combines new class scores, which are inferred from current models, with historical models' predictions. Based on the combined past and present knowledge, it incorporates self-distillation over models on clients, aiming to achieve effective knowledge transfer from historical models to current models. Theoretical analysis demonstrates that FedNASD is equivalent to modeling old class scores as conditional probabilities in the absence of new classes. Additionally, it reconciles the predictions of new classes with current models to refine the conditional probabilities of historical scores where new classes do not exist. Empirical experiments demonstrate the superiority of FedNASD over four baseline algorithms in reducing the average forgetting rate and boosting global accuracy.