The Sign of Directional Flow Asymmetry Depends on the Benign Reference and the Measurement Pipeline: Preregistered Tests and Robustness Audits in Network Intrusion Detection
Year: 2026
Type: preprint
Abstract: The imbalance between the bytes a flow sends and receives is a common intrusion-detection signal, and using it assumes that the direction of the imbalance is a property of the attack, so that it can be learned on one benchmark and carried to another. We test that assumption with preregistered external tests and post-hoc robustness audits, labeling every result by its evidential status. Graph topology did not outlast per-flow asymmetry under information restriction (H2, rejected). The CICIDS-2018 observation that the sign of per-flow asymmetry varied by attack did not replicate on UNSW-NB15 and met the frozen criterion on BoT-IoT (STRONG REPLICATION). The BoT-IoT verdict does not support the reading it invited: resampling host pairs instead of flows widened intervals from 0.014-0.018 to 0.27-0.41, only 3 of 8 evaluable groups kept a confident orientation, and restricting to answered flows made all ten point estimates positive. On CICIDS-2018, HOIC and Hulk reverse when the same traffic is re-exported as NetFlow-v2 with its own benign reference. UNSW-NB15 is the counterexample: its orientation was largely stable. A sign change did not consistently reduce separability. Signed directional orientation is a relational measurement, not an intrinsic attack attribute; a fixed-sign directional feature should be validated against the target pipeline and a named benign reference before its sign is assumed to transport. The evidence is limited to three benchmark datasets and does not include production traffic. Preprint, not peer reviewed. The first file is the paper; the second is its supplementary material (full methods, tables and sensitivity analyses). Code, pre-registration freezes, confirmatory outputs and audits: https://doi.org/10.5281/zenodo.24238405
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7233604817
Cites:
Cited by:
Related to:
