Text-Enhanced Data-free Approach for Federated Class-Incremental Learning

TL;DR

The paper tackles FCIL under privacy constraints by integrating Data-Free Knowledge Transfer with training through Label Text Embeddings (LTE) as anchors. LANDER enforces LT-centered constraints during client training and uses LTE-driven data generation on the server with a Bounding Loss to preserve natural embedding diversity. Learnable Data Stats further protect privacy while ensuring data-like quality in synthetic samples. Across CIFAR-100, Tiny-ImageNet, and ImageNet, LANDER achieves state-of-the-art performance with reduced forgetting, demonstrating the practical potential of LTE anchors for robust, data-free continual learning in federated settings.

Abstract

Federated Class-Incremental Learning (FCIL) is an underexplored yet pivotal issue, involving the dynamic addition of new classes in the context of federated learning. In this field, Data-Free Knowledge Transfer (DFKT) plays a crucial role in addressing catastrophic forgetting and data privacy problems. However, prior approaches lack the crucial synergy between DFKT and the model training phases, causing DFKT to encounter difficulties in generating high-quality data from a non-anchored latent space of the old task model. In this paper, we introduce LANDER (Label Text Centered Data-Free Knowledge Transfer) to address this issue by utilizing label text embeddings (LTE) produced by pretrained language models. Specifically, during the model training phase, our approach treats LTE as anchor points and constrains the feature embeddings of corresponding training samples around them, enriching the surrounding area with more meaningful information. In the DFKT phase, by using these LTE anchors, LANDER can synthesize more meaningful samples, thereby effectively addressing the forgetting problem. Additionally, instead of tightly constraining embeddings toward the anchor, the Bounding Loss is introduced to encourage sample embeddings to remain flexible within a defined radius. This approach preserves the natural differences in sample embeddings and mitigates the embedding overlap caused by heterogeneous federated settings. Extensive experiments conducted on CIFAR100, Tiny-ImageNet, and ImageNet demonstrate that LANDER significantly outperforms previous methods and achieves state-of-the-art performance in FCIL. The code is available at https://github.com/tmtuan1307/lander.

Figures (8)

• Figure 1: LANDER's motivation centers on using shared LTE as a key constraint for new task features and applying feature distillation from the previous server to anchor and organize the latent space of the client/server model. Subsequently, generating samples around the shared anchor LTE facilitates efficient data-free knowledge transfer from the previous to the current task.
• Figure 2: General Architecture of LANDER: (a) We utilize the previous server model (trained on task $t-1$) to synthesize the data. (b) Subsequently, we use this data to train the $k$-th task. We use the LTE as the anchor to constrain the features in both the client and generator, enhancing performance.
• Figure 3: Server latent space when using only CE loss; CE with MSE Loss; and CE with B Loss to constrain the feature embedding. By using our B Loss, the latent features are organized around but still remain flexible within a defined radius $r$ of the LTE center, mitigating the embedding overlap problem.
• Figure 4: Average accuracy on incremental tasks.
• Figure 5: (a) Accuracies in 5 and 10 tasks of our method with and without LT-centered generation. (b) Incremental Accuracy on CIFAR-100 for different values of $r$ in Bounding Loss. (c) Accuracies for different kinds of data normalization values.