With the advancement of network information technology and smart device technology, cyberspace is gradually evolved into Human-Cyber-Physical Networks (HCPNs). At the same time, the security problems caused by malicious nodes are becoming more and more serious. It is urgent to propose an efficient approach for malicious node detection. In this paper, we apply graph attention network (GAT) to detect malicious nodes layer by layer in HCPN. In addition, we investigate the influence of graph structure features on the detection performance in terms of accuracy, precision, recall, F1-score by comparing with graph convolutional network-based approach. Experimental results show that our approach has better performance as well as stronger generalizability than graph convolutional network-based approach in general.