Collaborative Adversarial Learning for Unsupervised Federated Domain Adaptation

Unsupervised Federated Domain Adaptation (UFDA) endeavors to mitigate the domain shift problem within federated learning paradigms to enhance the target model's performance. Nonetheless, it confronts two challenges: (1) The accuracy enhancement for each target domain after domain adaptation remains constrained. (2) The training process exhibits inefficiency. To address the above problems, we propose a novel Collaborative Adversarial Learning for Unsupervised Federated Domain Adaptation called CALUFDA. Firstly, we design a Collaborative Adversarial Learning Framework, integrating domain-level and semantic-level alignments into a deep learning framework. This framework employs adversarial loss to align cross-domain features, simultaneously leveraging an attention mechanism and two task classifiers to facilitate semantic-level alignment. Secondly, we introduce a Knowledge Contribution Mechanism to improve the communication efficiency between nodes, thus promoting global model aggregation. Finally, extensive experiments on benchmark datasets demonstrate the effectiveness of CALUFDA on the UFDA problem, guaranteeing training efficiency while obtaining substantial improvements on various baselines.