In multi-source domain adaptation, the main challenge is effectively integrating information from various source domains and adapting it to the target domain. Existing methods either align feature distributions of each source domain with the target domain separately and fuse at the classifier level, or jointly align feature distributions of all domains. The former approach fragments shared information, while the latter sacrifices discriminative properties. To address this, we propose Collaborative Domain Alignment (CoDA). CoDA utilizes an integrated feature encoder with domain attention masks to capture diverse shared information within a unified framework, thereby preserving both robustness and discriminability. Specifically, each source domain is elastically aligned with the target domain using a source-specific domain attention mask on the shared feature representation. Activated masks highlight features shared between individual source domains and the target domain, while overlapping masks highlight features shared by multiple source domains and the target domain. To optimize CoDA, we devise a domain-collaborative training strategy that includes domain-specific training loss, domain-consistency training loss, and pseudo-labeling loss. Extensive experiments on diverse datasets confirm the effectiveness and superiority of our approach.