Identification of abnormal and malicious traffic in the Internet-of-Things (IoT) network is critical for IoT security. However, it is worth noting that the majority of recent efforts demand a large amount of tagged traffic to train a machine-learning model. In this paper, we develop MetaIoT, an intelligent approach for identifying malicious traffic. MetaIoT is more accurate and more difficult for attackers to circumvent by taking into account both the local attributes of each traffic source and their global relationships. In MetaIoT, we begin by considering the heterogeneous and dynamic nature of traffic. Then, we introduce a heterogeneous graph (HG) to model the relationships between traffic and employ a relation-based heterogeneous graph attention network to learn node (i.e., traffic) representations over the built HG. Alternatively, MetaIoT addresses the issue of needing enough data for model training through the meta-learning technique. After conducting a comprehensive comparison with the baseline through experiments, our model demonstrated superior performance in few-shot learning scenarios, obtaining an accuracy score of 91.65% and an F1 score of 90.33%. When compared with current state-of-the-art IoT traffic detection models, our model showed the best results.