Existing intrusion detection methods place insufficient emphasis on the traffic interaction patterns contained in edge features and struggle to capture bidirectional communication relationships through nodes, resulting in an incomplete representation of intrusion behavior and compromised detection accuracy. To this end, this study proposes a bidirectional graph convolutional network (BiGCN)-based intrusion detection model that aims to enhance detection performance through edge feature enhancement and a bidirectional fusion mechanism. First, network traffic is segmented by adopting an adaptive time window, and a dynamic graph is constructed with hosts as nodes and communication sessions as directed edges. Second, a decision tree-based edge feature enhancement mechanism is designed, which leverages the Gini index to select highly discriminative features and construct high-order cross features to enrich the semantic information of edge features. Finally, BiGCN is introduced to separately capture forward and backward graph structural information. A dynamic attention fusion mechanism is combined to adaptively compute fusion weights according to node representations and enhanced edge features, enabling deep integration of the node and edge features. Finally, the multi-layer graph convolution is applied to complete the classification prediction. Experimental results on the CIC-IDS2017 and UNSW-NB15 datasets demonstrate that the built model outperforms other comparative methods in terms of precision, recall, and F1-score, thus validating its effectiveness for intrusion detection in network environments.