As online social networks grow rapidly, the emergence of a large number of virtual accounts, named social bots, poses great challenges to social security. In response to the bot invasion, the bot detection method has attracted considerable attention. Especially in recent years, with the widespread application of graphs, graph representation learning is widely applied in social bot detection. However, existing detection methods fall short in simultaneously representing with diverse network structures. In this work, we propose a graph-based and structure-aware framework to alleviate this problem. Specifically, we jointly encode user semantics, attributes and neighborhood information. Moreover, we employ a refined graph attention network model for parallel computation on large-scale graphs via subgraph sampling. In particular, we construct local and remote feature extractors, which can achieve multiple network feature extraction. Finally, we adopt a multitask learning approach to construct auxiliary tasks for self-supervised training and conduct bot detection. Extensive experiments show that our model outperforms state-of-the-art methods. Further exploration also demonstrates that our model has a strong generalization ability.