The present network infrastructure is safeguarded against cyber threats using Network Intrusion Detection Systems (NIDS). Many existing methods, including basic deep learning approaches on graph data, struggle to capture the spatiotemporal relationships between network nodes. They often don’t consider the data in a continuous time format. To address these issues, we propose NID-TGN, an encoder-decoder model for intrusion detection in IoT dynamic networks. The encoder enhances the Temporal Graph Network (TGN) framework by incorporating a learnable aggregation mechanism that better processes continuous time dynamic graph data. The decoder combines feature selection techniques with a random forest classifier, using only the node embeddings generated by the encoder to predict cyber attacks with an accuracy of 97%.