VX-One: Typed, Calibrated Decision Heads for Real-Time Perimeter Triage - An empirical study of unknown-attack honesty, cascade gray zones, and calibration under drift
arXiv cs.CL · 22 Oct 2026, 03:39 AEST
VX-One is a small multi-head network designed for real-time perimeter triage, offering typed decision primitives. The study evaluates its performance against various benchmarks, highlighting its strengths and weaknesses in detection and calibration under drift.
Key facts
- Perimeter triage faces a latency/semantics trade-off: signature engines are fast and format-safe but blind to unseen attacks, while generative models are semantic but slow and produce unparseable output.
- We present VX-One: a small multi-head network (100K-10M parameters) whose outputs are these primitives - a calibrated gate, a choice over a closed candidate set that always includes an explicit unknown option, and an ordinal score, trained multi-task, temperature-calibrated per head (~0.7-1.7 ms p50 on CPU).
- Findings: (i) VX-One wins detection (AUROC/FPR@95%TPR) on UNSW-NB15 and under temporal drift on CIC-IDS2017 (0.711 vs 0.637), but XGBoost wins multi-class attribution on all three benchmarks; (ii) leave-one-family-out gate detects never-seen families on 9/9 UNSW folds yet drops below chance on CIC recon family.
- (iv) calibration collapses under temporal drift (ECE 0.04->0.40) for every method while ranking partially survives.
Read aloud
VX-One is a small multi-head network designed for real-time perimeter triage, offering typed decision primitives. The study evaluates its performance against various benchmarks, highlighting its strengths and weaknesses in detection and calibration under drift. Perimeter triage faces a latency/semantics trade-off: signature engines are fast and format-safe but blind to unseen attacks, while generative models are semantic but slow and produce unparseable output. We present VX-One: a small multi-head network 100K-10M parameters whose outputs are these primitives - a calibrated gate, a choice over a closed candidate set that always includes an explicit unknown option, and an ordinal score, trained multi-task, temperature-calibrated per head ~zero point seven-one point seven ms p50 on CPU. Findings: i VX-One wins detection AUROC/FPR@ninety five percentTPR on UNSW-NB15 and under temporal drift on CIC-IDS2017 zero point seven one one vs zero point six three seven, but XGBoost wins multi-class attribution on all three benchmarks; ii leave-one-family-out gate detects never-seen families on nine/nine UNSW folds yet drops below chance on CIC recon family. iv calibration collapses under temporal drift ECE zero point zero four- zero point four zero for every method while ranking partially survives.