Federated Unsupervised Domain Generalization Using Global and Local Alignment of Gradients

• DOI: 10.1609/aaai.v39i19.34197
• Publication Date: 2025-04-11T13:01:21Z
• Authors: Farhad Pourpanah; Mahdiyar Molahasani; Milad Soltany; Michael Greenspan; Ali Etemad

We address the problem of federated domain generalization in an unsupervised setting for first time. theoretically establish a connection between shift and alignment gradients learning show that aligning at both client server levels can facilitate model to new (target) domains. Building on this insight, we propose novel method named FedGaLA, which performs gradient level encourage clients learn domain-invariant features, as well global obtain more generalized aggregated model. To empirically evaluate our method, perform various experiments four commonly used multi-domain datasets, PACS, OfficeHome, DomainNet, TerraInc. The results demonstrate effectiveness outperforms comparable baselines. Ablation sensitivity studies impact different components parameters approach.