Rehearsal-free Federated Domain-incremental Learning

TL;DR

The paper tackles catastrophic forgetting in federated domain-incremental learning by proposing RefFiL, a rehearsal-free framework that leverages global prompt sharing. It introduces a client-wise domain adaptive prompt generator, a global prompt sharing and clustering scheme, and a domain-specific prompt contrastive loss with temperature decay, all aimed at learning domain-invariant representations without storing old data. Experiments across four datasets demonstrate state-of-the-art performance and robustness to domain order changes, highlighting RefFiL's practicality for privacy-sensitive and resource-constrained FL settings. The approach provides a scalable, memory-efficient alternative to rehearsal-based methods with strong cross-domain transfer capabilities.

Abstract

We introduce a rehearsal-free federated domain incremental learning framework, RefFiL, based on a global prompt-sharing paradigm to alleviate catastrophic forgetting challenges in federated domain-incremental learning, where unseen domains are continually learned. Typical methods for mitigating forgetting, such as the use of additional datasets and the retention of private data from earlier tasks, are not viable in federated learning (FL) due to devices' limited resources. Our method, RefFiL, addresses this by learning domain-invariant knowledge and incorporating various domain-specific prompts from the domains represented by different FL participants. A key feature of RefFiL is the generation of local fine-grained prompts by our domain adaptive prompt generator, which effectively learns from local domain knowledge while maintaining distinctive boundaries on a global scale. We also introduce a domain-specific prompt contrastive learning loss that differentiates between locally generated prompts and those from other domains, enhancing RefFiL's precision and effectiveness. Compared to existing methods, RefFiL significantly alleviates catastrophic forgetting without requiring extra memory space, making it ideal for privacy-sensitive and resource-constrained devices.

Figures (2)

• Figure 1: Key steps of RefFiL framework. Left panel (a) shows the common setting in existing FCL works, characterized by a cliff-style data transition. Right panel (a) shows our approach, where a subset of participants gradually transitions to new tasks. Panel (b) provides an overview example of a key step in our methodology: the 1st participant processes new domain data using global prompts from the 2nd to $m$-th participants and local prompts, enhancing robustness by aligning the model's predictions across diverse domain prompts as inputs.
• Figure 2: Overview of the RefFiL Framework. Each participant first encodes local prompts using the tokenized feature map and task ID embedding. These local prompts are then concatenated with the feature map to compute the loss $\mathcal{L}_{CE}$. Simultaneously, the feature map is combined with global prompts to calculate the loss $\mathcal{L}_{GPL}$, and the loss $\mathcal{L}_{DPCL}$ is determined between global and local prompts. Subsequently, all local prompts, along with the updated local models, are transmitted to the central server. The server then clusters the prompts domain-wise and aggregates the local models for distribution in the next training round.