Benchmark Results · v3.1.2

External performance
specification

Locked results from the external model card. Benign calibration only. No attack labels for detection. Methodology is not published on this site.

Numbers below match Model Card v3.1.2 (external). Four independent measurement perspectives calibrate from benign only. The operator controls one parameter, α (false positive budget). Full operating points and limitations are on the model card.


CICIDS-2017 (multi-scale)

5-fold CV · benign calibration only · independently reproduced AUC to 4 decimals

DayFlowsAttack %AUCF1 @ α=0.05F1 @ α=0.005
Friday500,00046.9%0.99900.9720.992
Wednesday496,64135.7%0.99860.9570.995
Tuesday322,0782.2%0.99850.4700.898
Thursday362,07620.4%0.99840.9100.969

Tuesday shows α at work: low attack prevalence with a 5% FPR budget hurts F1 while AUC stays high. Tightening α recovers F1 without changing the instrument.


MAWIFlow (real backbone)

YearMethodAUCF1 @ α=0.05Peak F1
2016Flow-only0.9540.6080.634 (α=0.07)
2021Multi-scale0.9240.5210.547 (α=0.07)
2021Flow-only0.8600.5170.525 (α=0.07)
2011Multi-scale0.892n/an/a

Multi-scale on 2016 is not stable and is not reported. Do not tighten α on MAWI: at α=0.005, 2016 F1 falls to 0.150 and 2021 to 0.118.


CSE-CIC-IDS-2018 (10 daily captures)

500K random sample per day · attack-weighted F1 across days: 0.876

DayAUCF1Attack types
Fri 16-021.0000.948DoS Hulk
Tue 20-021.0000.773DDoS-LOIC
Wed 21-021.0000.929DDoS-HOIC
Wed 14-020.9930.781FTP/SSH Brute Force
Thu 15-020.9930.348DoS GoldenEye/Slowloris
Fri 02-030.9940.592Botnet Ares
Wed 28-020.9950.318Infiltration
Thu 01-030.9910.287Infiltration/NMAP
Fri 23-020.9970.003Web Attack (20 flows in sample)
Thu 22-020.9980.002Web Attack (18 flows in sample)

Other domains (flow-level)

DatasetAUCF1FPR
NSL-KDD Train0.9930.9532.5%
UNSW Train0.9930.9612.3%
UNSW Test0.9800.9242.7%
NSL-KDD Test0.9690.8782.6%
TON-IoT0.9480.6463.6%
DoH0.9380.6815.0%
IoT-DIAD0.8640.5043.0%

BCCC-IoT withheld pending end-to-end reproduction of a previously published figure.


Operational profile

Labels for detectionNone (benign calibration only)
Cold-start calibration~60 seconds
Total footprint~50 MB
Measurement perspectives4
Operator parametersOne (α)
HardwareCPU only, no GPU
Air-gappedYes
Encoding throughput15,000+ to 70,000+ flows/s
Detection throughput1,300 to 2,100 flows/s
Network requiredNone

Disclosure

This page publishes performance only. Clustering recipes, composition formulas, threshold math, named scale internals, and the internal model card are not on the public website, not in downloadable public cards, and not in browser-reachable demo assets.