It is well known that domain-adaptive object detection can make use of knowledge transfer from the source domain with abundant labeled data to the target domain with scarce or unavailable labeled data for boosting model capability of detecting out-of-distribution objects. Although the previous method leverages semi-supervised algorithms to fully utilize unlabeled data, it has not fully exploited the domain information of an image, and suffers degraded domain-transfer performance. To alleviate this limitation, we have proposed an Attention-based domain-adaptive YOLO framework, termed ATDA-YOLO, for domain-adaptive object detection. On the one hand, we have designed an attention-based domain adaptation network (Attn-DAN) and incorporated it into our ATDA-YOLO framework. Designed for domain adaptation of feature maps at different levels, this module enables our method to identify domain-invariant features for improving model adaptability with reduced domain difference, and extract semantic information across various scales from shallow to deep layers. On the other hand, our ATDA-YOLO explicitly encodes the domain information by introducing a domain classification loss, which enhances the identification of domain information via supervised learning. Extensive domain-transfer experiments for cross-domain detection task suggest that our ATDA-YOLO achieves significant improvements over the state-of-the-art, demonstrating the promise and potential of our proposed method in boosting domain adaptation performance.