Recently, attributed community search, a related but different problem to community detection and graph clustering, has been widely studied in the literature. Compared with the community detection that finds all existing static communities from a graph, the attributed community search (ACS) is more challenging since it aims to find dynamic communities with both cohesive structures and homogeneous node attributes given arbitrary queries. To solve the ACS problem, the most popular paradigm is to simplify the problem as two sub-problems, that is, structural matching and attribute filtering and deal with them separately. However, in real-world graphs, the community structure and the node attributes are actually correlated to each other. In this vein, current studies cannot capture these correlations which are vital for the ACS problem. In this paper, we propose Query-Driven Graph Convolutional Networks (QD-GCN), an end-to-end framework that unifies the community structure as well as node attribute to solve the ACS problem. In particular, QD-GCN leverages the Graph Convolutional Networks, which is a powerful tool to encode the graph topology and node attributes concurrently, as the backbones to extract the query-dependent community information from the original graph. By utilizing this query-dependent community information, QD-GCN is able to predict the target community given any queries. Experiments on real-world graphs with ground-truth communities demonstrate that QD-GCN outperforms existing attributed community search algorithms in terms of both efficiency and effectiveness.