FMDADA: Federated multi-discriminative adversarial domain adaptation

Federated domain adaptation system aims to address the problem of domain shift in a federated learning (FL) framework, where knowledge learned from distributed source domains can be readily transferred to the target domain. However, federated domain adaptation suffers from two challenges: (1) Inefficient assignment of source domain weights. (2) The joint distributions of feature and category across domains are poorly aligned. To solve the above problems, we propose a novel unsupervised federated domain adaptation (UFDA) approach called Federated Multi-Discriminative Adversarial Domain Adaptation (FMDADA). Firstly, we propose a FL aggregation scheme (F-DIS), which assigns weights to distributed source domains with different contribution rates based on a measure of cross-domain discrepancy. Secondly, we facilitate the joint distribution alignment of feature and category by designing multiple tightly coupled joint classifiers, which facilitates the positive transfer of source domain knowledge. Finally, extensive experimental results on three datasets demonstrate the effectiveness of FMDADA for UFDA problem. Compared to the currently advanced comparison approaches, the accuracy of FMDADA is significantly improved, reaching 54.7% and achieving an improvement of 5.9% on the large-scale dataset DomainNet.