Anomaly detection on attributed networks is intended to find instances that dramatically different from other instances in terms of attributes or structure. However, most existing methods ignore the adverse effects of abnormal nodes on normal nodes and are unable to capture high-dimensional cross-modal information from the structure and attributes. To tackle this drawback, we propose an SVDD-based anomaly detection framework, named STEAM, which detects abnormal nodes by fusing attributes and structure. First, STEAM uses node structure and its attributes information to learn two independent representations of nodes, respectively. Secondly, to alleviate the interference of abnormal nodes, we introduce Support Vector Data Description (SVDD), which helps us to learn the hypersphere from the structure representations of normal nodes. Thirdly, the two representations of the nodes are fused to reconstruct the attributed networks. Finally, we use a combination of the distance of every node to the hypersphere center and the reconstruction error of the attributed network to determine whether the node is anomalous or not. The performance of STEAM is assessed by 6 public datasets. Numerous experimental results have illustrated that STEAM is effective in detecting anomalous nodes in attributed networks.