MAGNet: Multimodal Attention-based Graph Network for Robust Cross-Dataset Intrusion Detection

Bingjun Luo

Abstract: Intrusion Detection Systems (IDS) are critical for safeguarding networks against malicious attacks. However, existing methods often suffer from performance degradation when applied to datasets from different network environments due to the discrepancy in data distributions. Moreover, they typically neglect the rich multimodal information inherent in network traffic. To address these issues, we propose MAGNet, a Multimodal Attention-based Graph Network that effectively learns from graph-structured network traffic and fuses features from multiple modalities (e.g., packet-level and flow-level data) for robust intrusion detection. MAGNet employs graph neural networks to model relationships between network entities and a multimodal fusion mechanism with attention to focus on the most informative features across modalities. Additionally, we integrate domain adversarial training to enhance cross-dataset generalization. Extensive experiments on three benchmark datasets (CIC-IDS2017, UNSW-NB15, and ToN-IoT) demonstrate that MAGNet outperforms state-of-the-art methods in both intra-dataset and cross-dataset scenarios, particularly improving detection rates for sophisticated attacks like DDoS and port scanning.