FedProK: Trustworthy Federated Class-Incremental Learning via Prototypical Feature Knowledge Transfer

TL;DR

This paper tackles Federated Class-Incremental Learning (FCIL) under a trustworthy AI lens, addressing catastrophic forgetting and data heterogeneity while preserving privacy and efficiency. It introduces FedProK, which combines a client-side feature translation for temporal knowledge transfer and a server-side prototypical knowledge fusion for spatial knowledge transfer, all without sharing raw data. The framework is evaluated in both synchronous and asynchronous FCIL settings on CIFAR-10/100, showing superior final accuracy and improved trustworthiness across continual utility, efficiency, and privacy metrics. The results demonstrate that selective, prototypical knowledge transfer can robustly balance stability and plasticity in dynamic, distributed learning environments, enabling more practical deployments of FCIL. These findings suggest a scalable approach to trustworthy FCIL that mitigates forgetting and data heterogeneity without compromising data privacy.

Abstract

Federated Class-Incremental Learning (FCIL) focuses on continually transferring the previous knowledge to learn new classes in dynamic Federated Learning (FL). However, existing methods do not consider the trustworthiness of FCIL, i.e., improving continual utility, privacy, and efficiency simultaneously, which is greatly influenced by catastrophic forgetting and data heterogeneity among clients. To address this issue, we propose FedProK (Federated Prototypical Feature Knowledge Transfer), leveraging prototypical feature as a novel representation of knowledge to perform spatial-temporal knowledge transfer. Specifically, FedProK consists of two components: (1) feature translation procedure on the client side by temporal knowledge transfer from the learned classes and (2) prototypical knowledge fusion on the server side by spatial knowledge transfer among clients. Extensive experiments conducted in both synchronous and asynchronous settings demonstrate that our FedProK outperforms the other state-of-the-art methods in three perspectives of trustworthiness, validating its effectiveness in selectively transferring spatial-temporal knowledge.

Figures (6)

• Figure 1: Illustration of the catastrophic forgetting and data heterogeneity in FCIL. The bar chart shows the test accuracy when vanilla FedAvg is employed in this scenario. $a$: The heterogeneous local data leads to poor performance of the global model. $b$: After the completion of training on Task 2, the model suffers catastrophic forgetting of Task 1.
• Figure 2: Illustration of the knowledge transfer via FedProK. The clients conduct feature translation locally to mitigate forgetting of the previous classes of the entire federated system. The server conducts prototypical knowledge fusion to alleviate the adverse impact of data heterogeneity among clients. $\mathcal{F}$ and $\mathcal{L}$ denote the feature extractor and the classification layer.
• Figure 3: Experimental results on CIFAR-10 (left) and CIFAR-100 (right) with 5 incremental tasks, where $\alpha$ denotes the Dirichlet concentration parameter.
• Figure 4: Experimental results on CIFAR-100 with 10 incremental tasks, where $\alpha$ denotes the Dirichlet concentration parameter.
• Figure 5: Illustration of the reconstructed images of malicious attackers, where the first line is the ground truth, the second line is the reconstructed images of FediCaRL, the third line is the reconstructed images of GLFC and the bottom line is the reconstructed images of our method.