Unsupervised domain adaptation serves as an effective approach for transferring information from labeled source domains to unlabeled target domains. While many existing methods focus on aligning global distributions to achieve domain-invariant feature representations, they often overlook fine-grained, domain-specific features. To address these limitations, a Domain Attention and Confidence-Aware Network (DACAN) is proposed. DACAN incorporates a domain attention module that dynamically assigns attention weights to features from different domains, enhancing the model’s focus on task-relevant features while minimizing dependence on domain-specific attributes. Additionally, by integrating a confidence-aware optimization strategy, DACAN effectively captures finer-grained features across domains. Experimental results demonstrate that DACAN significantly improves the model’s generalization capability.