Abstract
We introduce a rehearsal-free federated domain incremental learningframework, RefFiL, based on a global prompt-sharing paradigm to alleviatecatastrophic forgetting challenges in federated domain-incremental learning,where unseen domains are continually learned. Typical methods for mitigatingforgetting, such as the use of additional datasets and the retention of privatedata from earlier tasks, are not viable in federated learning (FL) due todevices' limited resources. Our method, RefFiL, addresses this by learningdomain-invariant knowledge and incorporating various domain-specific promptsfrom the domains represented by different FL participants. A key feature ofRefFiL is the generation of local fine-grained prompts by our domain adaptiveprompt generator, which effectively learns from local domain knowledge whilemaintaining distinctive boundaries on a global scale. We also introduce adomain-specific prompt contrastive learning loss that differentiates betweenlocally generated prompts and those from other domains, enhancing RefFiL'sprecision and effectiveness. Compared to existing methods, RefFiL significantlyalleviates catastrophic forgetting without requiring extra memory space, makingit ideal for privacy-sensitive and resource-constrained devices.