Federated Hypothesis Transfer for Decentralized Multi-Source Domain Adaptation

Unsupervised domain adaptation (UDA) aims to transfer knowledge from a labeled source domain to an unlabeled target domain. Multi-source unsupervised domain adaptation (MUDA) is a challenging paradigm of UDA in which the labeled data are collected from multiple labeled source domains with different distributions. Most existing studies on MUDA assume all source data can be accessed directly for training. However, in reality, data from different domains are usually decentralized and cannot be shared due to data privacy issues or legal constraints. To tackle this practical MUDA setting, we propose a novel decentralized MUDA framework named federated hypothesis transfer (FedHT). FedHT optimizes a local model for each source domain and aggregates the local models iteratively to construct a global model. To perform domain adaptation without accessing source data, FedHT fine-tunes the global model by optimizing the feature extractor with information maximization and self-supervised learning while freezing the classifier (hypothesis). In addition, to further alleviate negative transfer caused by irrelevant source domains, we devise a novel dynamic weighting strategy to generate source weights for model aggregation. The experiments reveal that the proposed FedHT significantly outperforms baseline methods, and also show that FedHT can effectively alleviate the negative transfer.