<p>Graph neural networks (GNNs) are deep learning methods for graphs that differ from traditional deep learning methods in that they can be applied to non-Euclidean data, such as user categorization in social networks and molecular structures. In computer networks, network traffic can be expressed very naturally in terms of graph structures, such as the GNN field of intrusion detection. There are many GNN-based intrusion detection methods, but some issues have not been fully explored. (1) More attention has been given to the topological information of the graph, while the characteristics of the network traffic itself have not been fully explored. (2) Most intrusion detection methods based on GNNs consider only the characteristic information of their first-order neighbors while ignoring the characteristic information of high-order neighbors in the graph. Most methods rely on high-quality labeled samples, which entail significant costs. To address these issues, we propose a semi-supervised network intrusion detection system (NIDS) method based on dual-channel high-order graph neural networks (DHGNNs). Our model employs a dual-channel architecture, with one channel dedicated to node attribute learning. This channel employs a deep learning approach to fully exploit the attribute feature information inherent to network traffic. In the other channel, we design a high-order message-passing mechanism to efficiently pass and aggregate the feature information of the high-order neighboring nodes and use an attention mechanism to obtain the node embeddings. Finally, we obtain the final node embedding by using channel attention to fuse the features of the two channels. The experimental results on four datasets show that our method achieves 98.63–99.86% accuracy in anomaly detection with a small amount of labeled data, as well as state-of-the-art multiclassification performance on these four datasets, especially on NF-BoT-IoT-V2, which is much higher than that of state-of-the-art algorithms.</p>