Adversarial Graph Neural Diffusion for Zero-Day Network Anomaly Detection: A Controlled Evaluation
Zero-day intrusions cannot be matched against signatures, so a detector must judge whether a host is behaving abnormally rather than whether it matches a known attack.Continuous-depth graph neural networks that treat message passing as a diffusion process have advanced node classification, and this paper asks whether they transfer to host-level anomaly detection.We present AGND, which combines graph neural diffusion with a time-aware attention operator whose diffusivity decays with the time since two hosts last communicated, and with an adversarial generator that synthesises pseudo-anomalous behaviour.On a controlled benchmark of 1440 windows, 100 hosts and 140 attack episodes from six unseen families, evaluated over 10 independently generated networks, AGND reaches 0.939 AUC-ROC and 0.270 AUC-PR under the original protocol and 0.937 and 0.254 under the corrected protocol of attack-free training and label-free selection.Neither proposed component changes this, and removing the diffusion raises AUC-PR to 0.352 and 0.307 respectively.Isolation forests and one-class support vector machines on the same features exceed every graph model on AUC-PR and a two-layer autoencoder matches the best of them, and the ordering survives training on attack-free windows only, label-free model selection and family-held-out selection.A factorial ablation shows that two hand-engineered relational summaries, not learned graph context, carry the relational signal: without them every method loses between 35 and 60 per cent of its AUC-PR, graph models included.Sweeps of depth, source weight and horizon over all networks show that a single Euler step matches the discrete model, that the loss appears at the second step together with a drop in Dirichlet energy, and that changing the energy through the horizon alone leaves detection quality unchanged, so oversmoothing is part of the mechanism rather than the whole of it.A host-window experiment on CIC-IDS2017 traffic reproduces the absence of a graph advantage but not the diffusion penalty.We report a negative result with a bounded explanation.
Authors
- Hayder Kareem Algabri (ORCID: https://orcid.org/0000-0002-5359-9147)
- Zaid Ibrahim Rasool (ORCID: https://orcid.org/0000-0001-6900-3443)
- Wafaa Sallal Abbood
Institutions
- University of Babylon (IQ)
- University of Hillah (IQ)
Publication Details
- Journal
- International journal of intelligent engineering and systems
- Published
- 2026-11-17
- DOI
- https://doi.org/10.22266/ijies2026.1231.07
- Primary Topic
- Network Security and Intrusion Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00