Disentangling Multi-View Scanning in Mamba for Network Traffic Anomaly Detection

This paper identifies a structural issue in multi-view Mamba models for network traffic anomaly detection: different scanning branches may repeatedly amplify view-invariant information while diluting view-specific cues, causing representation homogenization and multi-view degradation. It introduces DisenMamba, which reformulates multi-view scanning as a two-stage disentangle-then-fuse process. The method explicitly separates invariant and view-specific components before fusion, aiming to preserve complementary contextual information and improve sensitivity to subtle traffic anomalies. The authors report extensive experiments and provide an implementation on GitHub, but the supplied abstract does not include datasets, baselines, or quantitative results.